CN116628507A - Data processing method, device, equipment and readable storage medium - Google Patents
Data processing method, device, equipment and readable storage medium Download PDFInfo
- Publication number
- CN116628507A CN116628507A CN202310893001.2A CN202310893001A CN116628507A CN 116628507 A CN116628507 A CN 116628507A CN 202310893001 A CN202310893001 A CN 202310893001A CN 116628507 A CN116628507 A CN 116628507A
- Authority
- CN
- China
- Prior art keywords
- data
- feature
- class
- training
- features
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000003672 processing method Methods 0.000 title claims abstract description 47
- 238000012549 training Methods 0.000 claims abstract description 347
- 238000000605 extraction Methods 0.000 claims abstract description 127
- 238000000034 method Methods 0.000 claims abstract description 87
- 238000012545 processing Methods 0.000 claims description 104
- 238000004891 communication Methods 0.000 claims description 9
- 230000010354 integration Effects 0.000 claims description 6
- 230000006870 function Effects 0.000 description 40
- 238000004590 computer program Methods 0.000 description 10
- 238000010586 diagram Methods 0.000 description 8
- 102100040160 Rabankyrin-5 Human genes 0.000 description 6
- 101710086049 Rabankyrin-5 Proteins 0.000 description 6
- 238000010606 normalization Methods 0.000 description 6
- 238000012360 testing method Methods 0.000 description 6
- 230000009286 beneficial effect Effects 0.000 description 5
- 238000004458 analytical method Methods 0.000 description 4
- 238000013459 approach Methods 0.000 description 4
- 238000013473 artificial intelligence Methods 0.000 description 3
- 238000006243 chemical reaction Methods 0.000 description 3
- 230000004438 eyesight Effects 0.000 description 3
- 230000003044 adaptive effect Effects 0.000 description 2
- 125000003275 alpha amino acid group Chemical group 0.000 description 2
- 238000013528 artificial neural network Methods 0.000 description 2
- 150000001875 compounds Chemical class 0.000 description 2
- 230000001186 cumulative effect Effects 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 230000001965 increasing effect Effects 0.000 description 2
- 230000010365 information processing Effects 0.000 description 2
- 230000003993 interaction Effects 0.000 description 2
- 230000000452 restraining effect Effects 0.000 description 2
- 230000004304 visual acuity Effects 0.000 description 2
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 238000012935 Averaging Methods 0.000 description 1
- 208000037170 Delayed Emergence from Anesthesia Diseases 0.000 description 1
- 230000002159 abnormal effect Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000013527 convolutional neural network Methods 0.000 description 1
- 238000007405 data analysis Methods 0.000 description 1
- 238000013480 data collection Methods 0.000 description 1
- 230000002708 enhancing effect Effects 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 230000002349 favourable effect Effects 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 238000010295 mobile communication Methods 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 230000002093 peripheral effect Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/0464—Convolutional networks [CNN, ConvNet]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
-
- 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
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Abstract
The application provides a data processing method, a device, equipment and a readable storage medium, wherein the method comprises the following steps: determining data characteristics of each training data in the training data set; determining update data comprising similar data feature update data and heterogeneous data feature update data for a first class data feature in the cache data feature set, wherein the first class data feature is any one class data feature in the cache data feature set; updating the first class data features according to the update data to obtain updated first class data features; and training the feature extraction model according to the updated class data features corresponding to each data class and the data features of each training data to obtain a trained feature extraction model. The method provided by the application can restrict the category data characteristics from two dimensions of the same class and different classes at the same time, improves the accuracy of the category data characteristics and improves the training efficiency of the model.
Description
Technical Field
The present application relates to the field of artificial intelligence, and in particular, to a data processing method, apparatus, device, and readable storage medium.
Background
The caching feature refers to caching part or all of model output features in the model training process and is used for calculating a loss function in the model training. The cache feature can reserve features of a plurality of samples or categories, so that the loss function determined based on the cache feature can contain more distance relations among the samples, and accuracy of model training can be improved. At present, the caching feature method is widely applied to training tasks such as unsupervised training, fine granularity classification training and the like.
In the practical application process, in order to ensure the effectiveness of the cache features, the cache features are prevented from being invalid due to model updating, and the cache features need to be updated along with training. In the existing data processing method, the method for updating the cache features generally performs updating processing according to new data features and similar cache features. The method can restrict the cache features to approach to the similar features, but the method also easily causes inaccurate loss functions determined according to the cache features, so that the training time of the model is long and the training efficiency is low. Therefore, finding a method for improving the training efficiency of the model is a problem to be solved.
Disclosure of Invention
The embodiment of the application provides a data processing method, a device, equipment and a readable storage medium, which can restrict category data characteristics from two dimensions of similar characteristics and heterogeneous characteristics simultaneously, and improve the accuracy of the category data characteristics, thereby improving the training efficiency of a model.
In one aspect, an embodiment of the present application provides a data processing method, where the method includes:
acquiring a training data set, and calling a feature extraction model to perform feature extraction processing on each training data included in the training data set to obtain data features of each training data;
for a first class data feature in a cache data feature set, determining update data of the first class data feature according to the first class data feature and data features of the respective training data, wherein the update data comprises class data feature corresponding to at least one data class, the first class data feature is class data feature corresponding to a first data class in the cache data feature set, the class data feature update data is obtained according to a class difference feature between a data feature of training data belonging to the first data class in the training data set and the first class data feature, and the class data feature update data is obtained according to a class difference feature between a second class data feature and the first class data feature, and the second class data feature comprises class data features except the first class data feature in the cache data feature set, and the first data class is any one of the at least one data class;
Updating the first class data feature according to the similar data feature updating data and the heterogeneous data feature updating data to obtain an updated first class data feature, wherein the difference feature between the updated first class data feature and the data feature of the training data belonging to the first data class is smaller than the same class difference feature, and the difference feature between the updated first class data feature and the second class data feature is larger than the heterogeneous difference feature;
and training the feature extraction model according to the updated class data features corresponding to each data class in the at least one data class and the data features of each training data to obtain a trained feature extraction model.
In one aspect, an embodiment of the present application provides a data processing apparatus, including:
the device comprises an acquisition unit, a feature extraction unit and a feature extraction unit, wherein the acquisition unit is used for acquiring a training data set, and invoking a feature extraction model to perform feature extraction processing on each training data included in the training data set to obtain the data features of each training data;
the processing unit is used for determining updating data of the first class data feature according to the first class data feature and the data feature of each training data in a cache data feature set, wherein the updating data comprises class data feature corresponding to at least one data class in the cache data feature set, the first class data feature is class data feature corresponding to the first data class in the cache data feature set, the class data feature updating data is obtained according to a class difference feature between the data feature of the training data belonging to the first data class in the training data set and the first class data feature, and the class data feature updating data is obtained according to a class difference feature between a second class data feature and the first class data feature, and the second class data feature comprises class data features except the first class data feature in the cache data feature set;
The processing unit is further configured to update the first type data feature according to the similar type data feature update data and the heterogeneous type data feature update data to obtain an updated first type data feature, a difference feature between the updated first type data feature and the data feature of the training data belonging to the first type data is smaller than the same type difference feature, and a difference feature between the updated first type data feature and the second type data feature is larger than the heterogeneous type difference feature;
and the training unit is used for training the feature extraction model according to the updated class data features corresponding to each data class in the at least one data class and the data features of each training data to obtain a trained feature extraction model.
In one aspect, an embodiment of the present application provides a computer device, including: the data processing method comprises a processor, a communication interface and a memory, wherein the processor, the communication interface and the memory are mutually connected, executable program codes are stored in the memory, and the processor is used for calling the executable program codes to realize the data processing method provided by the embodiment of the application.
Correspondingly, the embodiment of the application also provides a computer readable storage medium, wherein the computer readable storage medium stores instructions which, when run on a computer, cause the computer to realize the data processing method provided by the embodiment of the application.
Accordingly, embodiments of the present application also provide a computer program product comprising a computer program or computer instructions stored in a computer readable storage medium. The processor of the computer device reads the computer program or the computer instructions from the computer readable storage medium, and the processor executes the computer program or the computer instructions, so that the computer device realizes the data processing method provided by the embodiment of the application.
In the application, a training data set is obtained, and a feature extraction model is called to perform feature extraction processing on each training data included in the training data set, so as to obtain the data features of each training data; for a first class data feature in the cache data feature set, determining update data of the first class data feature according to the first class data feature and data features of each training data, wherein the update data comprises class data feature corresponding to at least one class data feature in the cache data feature set, the first class data feature is a class data feature corresponding to the first data class in the cache data feature set, the class data feature update data is obtained according to a class difference feature between a data feature of training data belonging to the first data class in the training data set and the first class data feature, the class data feature update data is obtained according to a class difference feature between a second class data feature and the first class data feature, the second class data feature comprises a class data feature except the first class data feature in the cache data feature set, and the first data class is any one of the at least one data class; and carrying out updating processing on the first class data characteristics according to the updating data to obtain updated first class data characteristics, wherein the difference characteristics between the updated first class data characteristics and the data characteristics of the training data belonging to the first data class are smaller than the same class difference characteristics, the difference characteristics between the updated first class data characteristics and the second class data characteristics are larger than the different class difference characteristics, and carrying out training processing on the characteristic extraction model according to the updated class data characteristics corresponding to each data class in at least one data class and the data characteristics of each training data to obtain a trained characteristic extraction model. The data processing method provided by the embodiment of the application can restrict the category data characteristics from the similar characteristics and the heterogeneous characteristics at the same time, so that the category data characteristics are close to the similar characteristics and far from the heterogeneous characteristics in the updating process, and the accuracy of the category data characteristics is improved; the loss function required by model training can be determined according to the data characteristics of the class, so that the loss function can contain the distance relation among the data characteristics of various different classes, and the accuracy of the loss function is effectively improved; the feature extraction model can be trained by using the loss function with higher accuracy, so that the training times of model training are effectively reduced, and the model training efficiency is improved.
Drawings
In order to more clearly illustrate the embodiments of the application or the technical solutions in the prior art, the drawings that are required in the embodiments or the description of the prior art will be briefly described, it being obvious that the drawings in the following description are only some embodiments of the application, and that other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.
FIG. 1 is a schematic diagram of a system architecture of a data processing system according to an embodiment of the present application;
FIG. 2 is a schematic flow chart of a data processing method according to an embodiment of the present application;
FIG. 3 illustrates a feature update method provided by an embodiment of the present application;
FIG. 4 is a flowchart of another data processing method according to an embodiment of the present application;
fig. 5 illustrates a vector update method according to an embodiment of the present application;
FIG. 6 is a schematic diagram of a model training method according to an embodiment of the present application;
FIG. 7 is a block diagram of a data processing apparatus according to an embodiment of the present application;
fig. 8 is a block diagram of a computer device according to an embodiment of the present application.
Detailed Description
The following description of the embodiments of the present application will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present application, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the application without making any inventive effort, are intended to be within the scope of the application.
It should be noted that the descriptions of "first," "second," and the like in the embodiments of the present application are for descriptive purposes only and are not to be construed as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a technical feature defining "first", "second" may include at least one such feature, either explicitly or implicitly.
The caching feature refers to caching part or all of model output features in the model training process, and is used for calculating a loss function in the model training. The existing cache feature updating method generally adopts a momentum updating method to restrict the cache features to approach to the similar features, so that the distance between the cache features and the similar features is reduced during updating, and the method easily causes inaccurate loss functions determined according to the cache features, so that the model training efficiency is low.
Based on the above, the embodiment of the application provides a data processing method, which can acquire a training data set, and call a feature extraction model to perform feature extraction processing on each training data included in the training data set to obtain the data features of each training data; for a first class data feature in the cache data feature set, determining update data of the first class data feature according to the first class data feature and data features of each training data, wherein the update data comprises class data feature corresponding to at least one class data feature in the cache data feature set, the first class data feature is a class data feature corresponding to the first data class in the cache data feature set, the class data feature update data is obtained according to a class difference feature between a data feature of training data belonging to the first data class in the training data set and the first class data feature, the class data feature update data is obtained according to a class difference feature between a second class data feature and the first class data feature, the second class data feature comprises a class data feature except the first class data feature in the cache data feature set, and the first data class is any one of the at least one data class; updating the first type data feature according to the updated data to obtain an updated first type data feature, wherein the difference feature between the updated first type data feature and the data feature of the training data belonging to the first data type is smaller than the same type difference feature, and the difference feature between the updated first type data feature and the second type data feature is larger than the different type difference feature; and training the feature extraction model according to the updated class data features corresponding to each data class in at least one data class and the data features of each training data to obtain a trained feature extraction model. According to the data processing method provided by the embodiment of the application, the class data features (namely the cache features) can be restrained from the two dimensions of the similar features and the different features at the same time, so that the class data features are close to the similar features and far away from the different features, the accuracy of the class data features can be effectively improved, the accuracy of a loss function determined according to the class data features is effectively improved, and the training efficiency of the model is further improved.
The data processing method provided by the embodiment of the application can be applied to the fields of Computer Vision (CV) and the like in artificial intelligence. The computer vision means that a camera and a computer are used for replacing human eyes to perform machine vision such as identification and measurement on targets, and further graphic processing is performed, so that the computer is processed into images which are more suitable for human eyes to observe or transmit to an instrument to detect. The data processing method provided by the embodiment of the application can be used in the training process of the target image retrieval model in the field of computer vision, and the feature extraction model trained by the method provided by the embodiment of the application can also be used as a part of the target image retrieval model. For example: the method provided by the application can be used for determining the update data, updating the category data features by using the update data, and training the feature extraction model by using the updated category data features, so that the training efficiency and accuracy of the model can be effectively improved. When the data is an image, a target image retrieval model can be determined according to the feature extraction model, the target image retrieval model can be applied to a sports match analysis scene, and intelligent analysis can be performed on the competition situation according to real-time image data of athletes; the target image retrieval model can also be applied to film and television analysis scenes, and intelligent analysis can be carried out on film and television dramas according to actor image data in different scenes. The data processing method provided by the embodiment of the application can be used in the training process of the image-text understanding model in the field of computer vision, and the feature extraction model trained by the method provided by the embodiment of the application can also be a part of the image-text understanding model, so that the accuracy of the image-text understanding model is effectively improved, and the image-text understanding model can be applied to scenes such as network content auditing, abnormal text or picture identification and the like.
The architecture of a data processing system provided by embodiments of the present application will be described below with reference to the accompanying drawings.
With reference to fig. 1, a schematic system architecture of a data processing system according to an exemplary embodiment of the present application is shown, where the data processing system includes: terminal equipment 101, server 102 and database 103, server 102 can interact with terminal equipment 101, database 103. Wherein:
the terminal device 101 may interact with the server 102, and the terminal device 101 may be a handheld device (such as a smart phone, a tablet computer) with a camera function, a computing device (such as a personal computer (Personal Computer, PC)), a vehicle-mounted terminal, an intelligent voice interaction device, a wearable device or other intelligent apparatus, and the like, but is not limited thereto.
The server 102 may perform model training on the feature extraction model, or may perform data analysis processing using the trained feature extraction model. The server 102 may be an independent physical server, a server cluster or a distributed system formed by a plurality of physical servers, or may be a cloud server providing cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, a content delivery network (Content Delivery Network, CDN), basic cloud computing services such as big data and an artificial intelligence platform.
The database 103 may be used to store training data, and may also be used to store the results of processing the data by the trained feature extraction model. The database 103 may be a database provided in the server 102, that is, may be a database built in or self-contained in the server; the cloud database may be a peripheral database connected to the server 102 (i.e., a cloud database (i.e., a database deployed in the cloud), and may be specifically deployed based on any one of a private cloud, a public cloud, a hybrid cloud, an edge cloud, etc., so that the cloud database has different focused functions. For example: the base cloud hardware is the personal equipment of the user, more emphasis is placed on serving a small part of users, and the database deployed in the public cloud is deployed based on a cloud platform provided by a third party, so that data stored in the database can be shared, any user data can be stored in the database, and any user can also use the data in the database.
The working principle of the data processing system as shown in fig. 1 will be explained in detail below:
the server 102 may acquire a training data set from the database 103, and invoke a feature extraction model to perform feature extraction processing on each training data included in the training data set, so as to obtain data features of each training data; for a first class data feature in the cache data feature set, determining update data of the first class data feature according to the first class data feature and data features of each training data, wherein the update data comprises class data feature update data corresponding to at least one class data feature in the cache data feature set, the first class data feature is a class data feature corresponding to the first data class in the cache data feature set, the class data feature update data is obtained according to a class difference feature between a data feature of training data belonging to the first data class in the training data feature set and the first class data feature, the class data feature update data is obtained according to a class difference feature between a second class data feature and the first class data feature, the second class data feature comprises a class data feature except the first class data feature in the cache data feature set, and the first data class is any one of the at least one data class; updating the first class data feature according to the similar data feature updating data and the heterogeneous data feature updating data to obtain an updated first class data feature, wherein the difference feature between the updated first class data feature and the data feature of the training data belonging to the first data class is smaller than the same class difference feature, and the difference feature between the updated first class data feature and the second class data feature is larger than the heterogeneous difference feature; and training the feature extraction model according to the updated class data features corresponding to each data class in at least one data class and the data features of each training data to obtain a trained feature extraction model. After model training is completed, the terminal equipment 101 can interact with the object to obtain data to be processed, the data to be processed is sent to the server 102, and after the server 102 receives the data to be processed, the data to be processed can be processed by utilizing the feature extraction model after model training is completed to obtain a data processing result; the server 102 transmits the data processing result to the terminal device 101. The data processing method provided by the embodiment of the application can restrict the class data characteristics from the two dimensions of the similar characteristics and the heterogeneous characteristics at the same time, and improve the accuracy of the class data characteristics, thereby improving the accuracy of the loss function, improving the training efficiency of the model and improving the accuracy of the model.
It will be appreciated that the architecture diagram of the data processing system described in the embodiments of the present application is for more clearly describing the data processing method of the embodiments of the present application, and is not limited to the data processing method provided in the embodiments of the present application. For example, the data processing method provided by the embodiment of the present application may be performed by a server 102, but may also be performed by a server or a server cluster other than the server 102 and capable of communicating with the terminal device 101 and/or the server 102. Those of ordinary skill in the art will recognize that the number of terminal devices and servers in fig. 1 is merely illustrative. Any number of terminal devices and servers may be configured according to service implementation needs. Moreover, with the evolution of the system architecture and the appearance of new service scenarios, the data processing method provided by the embodiment of the application is also applicable to similar technical problems.
In the present application, the related biometric identification technology, when the above embodiments of the present application are applied to specific products or technologies, the relevant data collection, use and processing processes should comply with the relevant legal regulations, the information processing rules should be notified and the individual consent of the target object should be solicited before the biometric information is collected, and the biometric information is processed in strict compliance with the legal regulations and the personal information processing rules, and technical measures are taken to ensure the security of the relevant data.
Referring to fig. 2, fig. 2 is a flowchart illustrating a data processing method according to an embodiment of the application. The data processing method may be implemented by the server 102, or may be implemented by other server devices. The flow of the data processing method provided in the embodiment of the application includes, but is not limited to:
s201, acquiring a training data set, and calling a feature extraction model to perform feature extraction processing on each training data included in the training data set to obtain data features of each training data.
In the embodiment of the application, the training data set can comprise a plurality of training data, and the training data can be image, text, audio and other types of data. The training data set may further include data category information of each training data, for example: the training data is data of an image type, and it is assumed that a training image P is included in the training data set, and data category information of the training image P may be "river". Also for example: the training data is text type data, and the training data set is assumed to include training text T, and the data category information of the training text can be "noun". And calling a feature extraction model to perform feature extraction processing on each training data included in the training data set, so that the data features of each training data can be obtained.
S202, aiming at first class data features in a cache data feature set, determining update data of the first class data features according to the first class data features and the data features of each training data, wherein the update data comprises similar data feature update data and heterogeneous data feature update data.
In the embodiment of the present application, the cached data feature set may include a class data feature corresponding to at least one data class, and the first class data feature may be a class data feature belonging to a first data class in the cached data feature set, and the first data class may be any one of at least one data class in the cached data feature set. For example: the cache data feature set includes category data features corresponding to three data categories, and the first category data feature may be any one of the three category data features. For a first category data feature in the cached data feature set, determining update data of the first category data feature according to the first category data feature and the data feature of each training data, wherein the update data comprises similar data feature update data and heterogeneous data feature update data. The similar data feature update data may be obtained according to a similar difference feature between a data feature of training data belonging to the first data category in the training data set and the first data feature, and the dissimilar data feature update data may be obtained according to a dissimilar difference feature between a second category data feature and the first data feature, the second category data feature including a category data feature other than the first category data feature in the cache data feature set. For example: the cache data feature set comprises class data features corresponding to three data classes, wherein the three data classes are class A, class B and class C respectively; assuming that the first data category is a category A, determining similar data feature update data according to the same-category difference features between the data features of the training data corresponding to the category A in the training data set and the first-category data features; and determining heterogeneous data characteristic updating data according to the heterogeneous difference characteristics between the class data characteristics corresponding to the class B and the class data characteristics corresponding to the class C in the cache data characteristic set and the first class data characteristics.
In an embodiment, the feature extraction model may be trained for multiple times, and then, during the 1 st training, the feature extraction model may be invoked to perform feature extraction processing on each sample data included in the sample data set, so as to obtain data features of each sample data, and perform integration processing on the data features of each sample data, so as to obtain a cached data feature set; each training data included in the training data set is included in the sample data set; and in the N-th training, determining a cache data characteristic set according to updated class data characteristics corresponding to each data class in the N-1 th training, wherein N is a positive integer greater than 1. The sample data set may include a plurality of sample data, and data category information of each sample data; and the training data set is included in the sample data set. For example: the sample data set may contain 10000 sample data, and then 100 data may be obtained from the sample data set to form a training data set. In the training of the 1 st time, a feature extraction model can be called to perform feature extraction processing on each sample data included in the sample data to obtain data features of each sample data, and the data features of each sample data are integrated to obtain a cache data feature set; and in the N-th training, the updated class data characteristics corresponding to each data class in the N-1 th training can be determined to be a cache data characteristic set, and N is a positive integer greater than 1. The method provided by the embodiment of the application can ensure that the category data characteristics which are updated in the previous training are contained in the cache data characteristic set during each training, so that the category data characteristics are updated along with the training, and the effectiveness of the category data characteristics is effectively ensured; meanwhile, in the first training, the cache data feature set is determined according to the data features of the sample data, so that the category data features can be fused with the data features of the training data and the data features of the sample data, the accuracy of a loss function determined according to the category data features is further improved, and the training efficiency of the model is improved.
In an embodiment, the implementation manner of integrating the data features of each sample data to obtain the cached data feature set may be: acquiring data characteristics of at least one sample data corresponding to any one data category of at least one data category; carrying out mean value processing on the data characteristics of at least one sample data corresponding to any data category to obtain mean value data characteristics corresponding to any data category; normalizing the mean value data characteristic corresponding to any data category to obtain the category data characteristic corresponding to any data category; and determining a cache data characteristic set according to the class data characteristics corresponding to each data class. For any one of a plurality of data categories, acquiring data characteristics of at least one sample data corresponding to the data category; carrying out mean processing on the data characteristics of the acquired at least one sample data to obtain mean data characteristics corresponding to the data category; and carrying out normalization processing on the mean value data characteristics corresponding to the data category to obtain category data characteristics corresponding to the data category. For example: for the class A data category in at least one data category, the data characteristics of 10 sample data corresponding to the class A data category can be obtained, namely, the data category information of the 10 sample data is class A; averaging the data characteristics of the 10 sample data to obtain average data characteristics corresponding to class A data types; and carrying out normalization processing on the mean value data characteristics corresponding to the class A data category to obtain category data characteristics corresponding to the class A data category. And determining a cache data characteristic set according to the class data characteristics corresponding to each data class. The method provided by the embodiment of the application determines the cache data characteristic set according to the sample data, is favorable for subsequent updating operation of the category data characteristic, and further improves the accuracy of the loss function.
And S203, updating the first class data features according to the similar data feature updating data and the heterogeneous data feature updating data to obtain updated first class data features.
In the embodiment of the application, the first-class data feature can be updated according to the similar data feature updating data and the heterogeneous data feature updating data to obtain the updated first-class data feature, the difference feature between the updated first-class data feature and the data feature of the training data belonging to the first data class is smaller than the same-class difference feature, and the difference feature between the updated first-class data feature and the second-class data feature is larger than the heterogeneous difference feature. The similar data feature updating data can be used for indicating the updating direction of the first similar data feature on the similar feature, the dissimilar data feature updating data can be used for indicating the updating direction of the first similar data feature on the dissimilar feature, and the similar data feature updating data and the dissimilar data feature updating data can simultaneously restrict the class data feature in two dimensions according to the similar data feature updating data and the dissimilar data feature updating data, so that the first similar data feature has larger difference with the feature information of other data classes while fully expressing the feature information of the first data class, and the first similar data feature has stronger distinguishing capability.
Referring to fig. 3, a feature updating method provided by an embodiment of the present application is shown. As shown in fig. 3, when training is performed for the 1 st time, a feature extraction model is called to perform feature extraction processing on each sample data in a sample data set, so as to obtain data features of each sample data, and a cache data feature set containing a plurality of category data features is determined according to the data features of the sample data; and inputting the training data set into a feature extraction model to perform feature extraction processing to obtain data features of each training data, and determining update data according to the data features of each training data and the cached data feature set, wherein the update data comprises similar data feature update data and heterogeneous data feature update data. And updating each category data feature by using the updating data to obtain updated category data features, and adjusting model parameters of the feature extraction model according to the updated category data features. And in the N-th training (N is a positive integer greater than 1), acquiring updated class data characteristics in the N-1 th training, inputting a training data set into a characteristic extraction model to perform characteristic extraction processing to obtain data characteristics of each training data, determining updating data according to the data characteristics of each training data and the updated class data characteristics in the N-1 th training, performing updating processing on the updated class data characteristics in the N-1 th training by using the updating data to obtain updated class data characteristics in the N-th training, and performing model parameter adjustment on the characteristic extraction model by using the updated class data characteristics in the N-th training. The method provided by the embodiment of the application can effectively update the category data characteristics and ensure the validity of the category data characteristics; meanwhile, in the updating process, the category data features can be restrained from two dimensions at the same time, so that the category data features are more accurate; since the class data features are determined according to the sample data in the first training, the class data features also comprise the data features of the sample data, so that the class data features are more accurate and more representative.
S204, training the feature extraction model according to the updated class data features corresponding to each data class in the at least one data class and the data features of the training data to obtain a trained feature extraction model.
In the embodiment of the application, the loss function can be determined according to the updated class data characteristics corresponding to each data class in at least one data class and the data characteristics of each training data, and the characteristic extraction model is trained according to the loss function to obtain the trained characteristic extraction model. And performing model training treatment on the feature extraction model for multiple times, and completing model training when the training times reach preset times or the feature extraction model converges.
It should be noted that, the data processing method provided by the embodiment of the application not only can be applied to the training process of the feature extraction model for extracting the image features, but also can be applied to the training process of the feature extraction model for extracting various data features such as texts, audios and the like. The feature extraction model trained by the method provided by the embodiment of the application can be used as a part of models such as image retrieval, image-text understanding, image classification, semantic understanding and the like.
Based on the embodiment, the application has the following beneficial effects: the data processing method provided by the embodiment of the application can determine the update data comprising the similar data feature update data and the heterogeneous data feature update data, and update the category data features (namely the cache features) according to the update data, so that the category data features are restrained from approaching to and being far away from the heterogeneous features towards the similar features, the effectiveness of the category data features is ensured, the resolving power of the category data features to different data categories is effectively improved, and more accurate supervision information is provided for model training; the loss function can be determined according to the updated category data characteristics, so that the accuracy of the loss function is higher, the model training efficiency is improved, and the prediction accuracy of the feature extraction model after training is improved.
Referring to fig. 4, fig. 4 is a flowchart illustrating another data processing method according to an embodiment of the present application. The data processing method may be implemented by the server 102, or may be implemented by other server devices. The flow of the data processing method provided in the embodiment of the application includes, but is not limited to:
s401, acquiring a training data set, and calling a feature extraction model to perform feature extraction processing on each training data included in the training data set to obtain data features of each training data.
In the embodiment of the application, the training data set may include a plurality of training data and data category information of each training data. The data type of the training data may be image, text, audio, video, etc. And calling a feature extraction model to perform feature extraction processing on each training data included in the training data set, so that the data features of each training data can be obtained.
In one embodiment, a feature extraction model (feature extraction model is denoted as f #)X), wherein x represents input data, +.>Model parameters representing the feature extraction model) performs feature extraction processing on each training data, and performs normalization processing on the extracted features to obtain data features of each training data. The data characteristic of each training data can be expressed as {>，/>I= … N }, where N represents the number of training data included in the training data set, +.>Data characteristic representing the ith training data, < +.>Representing the data category of the ith training data.
In an embodiment, the Network structure of the feature extraction model may be a Residual Network (res net) structure. The residual network has the characteristics of easy optimization and can improve the accuracy by increasing the equivalent depth. The residual blocks inside the residual network use jump connection, so that the gradient vanishing problem caused by adding depth in the deep neural network is relieved.
In an embodiment, after the data enhancement processing is performed on the training data, the feature extraction model is called to perform feature extraction processing on the training data after the data enhancement processing. For example: when the data type of the training data is an image, image enhancement processing (such as image cutting processing, image inversion processing, image division processing and the like) can be performed on the training images in the training data set to obtain a plurality of enhanced training images, and feature extraction models are called to perform feature extraction processing on each enhanced training image to obtain image features of each enhanced training image, and the image features of each enhanced training image are utilized to perform subsequent operations.
It should be noted that, the embodiment of the present application is not limited to the type of the data features extracted by the feature extraction model. The extracted data features may be image features (when the data type of the training data is an image) or other types of features, such as: scale-invariant feature transform (SIFT) features, direction gradient histogram (Histogram of Oriented Gradient, HOG) features, contrast language Image Pre-Training (CLIP) features, deep convolutional neural network (e.g., alexent neural network) features, and the like.
S402, acquiring at least one training data belonging to the first data category in the training data set, wherein the cache data feature set comprises category data features corresponding to at least one data category, and the first data category is any one of the at least one data category.
In the embodiment of the application, the cache data feature set comprises category data features corresponding to at least one data category, and the first data category is any one of the at least one data category. And determining at least one training data corresponding to the first data category from the training data set according to the first data category. For example: and if the first data category is the X category, determining that at least one data category information is the training data of the X category from the training data set according to the first data category.
In an embodiment, the implementation manner of determining the cached data feature set may be: in the training of the 1 st time, a feature extraction model can be called to perform feature extraction processing on each sample data included in a sample data set to obtain data features of each sample data, and the data features of each sample data are integrated to obtain a cache data feature set; each training data included in the training data set is included in the sample data set; and in the N-th training, determining a cache data characteristic set according to updated class data characteristics corresponding to each data class in the N-1 th training, wherein N is a positive integer greater than 1. The sample data set may include a plurality of sample data and data category information of the respective sample data. During training for the 1 st time, a cache data characteristic set can be determined according to the data characteristics of each sample data; and in the nth training, determining a cache data characteristic set in the nth training according to the class data characteristics updated in the (N-1) th training.
In an embodiment, the implementation manner of integrating the data features of each sample data to obtain the cached data feature set may be: acquiring data characteristics of at least one sample data corresponding to any one data category of at least one data category; carrying out mean value processing on the data characteristics of at least one sample data corresponding to any data category to obtain mean value data characteristics corresponding to any data category; normalizing the mean value data characteristic corresponding to any data category to obtain the category data characteristic corresponding to any data category; and determining a cache data characteristic set according to the class data characteristics corresponding to each data class. For any one data category, the average value of the data characteristics of a plurality of sample data corresponding to the data category can be calculated, and the obtained average value is normalized to obtain the category data characteristics corresponding to the data category. In some casesEach category data feature in the cached data feature set may be represented as: {I= … K }, where K represents the number of data categories in the cached data feature set, +.>And representing class data characteristics corresponding to the ith data class.
In an embodiment, the implementation manner of integrating the data features of each sample data to obtain the cached data feature set may be: determining target data characteristics corresponding to any data category from the data characteristics of at least one sample data corresponding to any data category; carrying out normalization processing on the target data characteristics corresponding to any data category to obtain category data characteristics corresponding to any data category; and determining a cache data characteristic set according to the class data characteristics corresponding to each data class. The target data characteristics corresponding to the data category can be determined from the data characteristics of one or more sample data corresponding to the data category according to a preset rule, and the category data characteristics corresponding to the data category can be determined according to the target data characteristics corresponding to the data category. The preset rules can be modified according to different actual application requirements.
S403, determining a first distance between the data feature of each training data in the at least one training data and a first class data feature, obtaining at least one first distance, determining a same class difference feature according to the at least one first distance, and determining the same class difference feature as similar data feature update data of the first class data feature, wherein the first class data feature is a class data feature corresponding to a first data class in the cache data feature set.
In the embodiment of the application, the first class data feature is a class data feature corresponding to a first data class in the cache data feature set, and the at least one training data is training data corresponding to the first data class in the training data set, that is, the data class corresponding to the first class data feature is the same as the data class corresponding to each training data in the at least one training data, and is a similar feature. A first distance between the data features of each of the at least one training data and the first class data features may be calculated (e.g., the first distance may be a euclidean distance between the data features of the training data and the first class data features), resulting in at least one first distance. And determining the same-category difference feature according to at least one first distance, and determining the same-category difference feature as similar data feature update data of the first-category data feature.
In an embodiment, when the at least one first distance includes only one first distance, calculating a difference feature between a data feature of training data corresponding to the first distance and a first class data feature, to obtain a same class difference feature; when the at least one first distance includes two or more first distances, an implementation manner of determining the same-category difference feature according to the at least one first distance may be: comparing at least one first distance to obtain a maximum first distance; and calculating the difference characteristic between the data characteristic of the training data corresponding to the maximum first distance and the first category data characteristic to obtain the category difference characteristic. Calculating the difference characteristic between the data characteristic of the training data corresponding to the maximum first distance and the first class data characteristic to obtain the implementation mode of the same class difference characteristic, wherein the implementation mode can be shown in the following formula (1):
In the formula (1), the components are as follows,representing a first category data feature->Data characteristics of training data corresponding to the first data category are represented, and D1 represents the same-category difference characteristics. When the training data set contains training data corresponding to a plurality of first data categories, the training data set can be trained according to the plurality of training dataThe data features of the training data and the first class data features determine a plurality of first distances reflecting distances between the first class data features and the data features of the training data of the same class. In order to enable the first class data feature to be closer to the similar data feature, difference features between the data feature of training data corresponding to the largest first distance in the first distances and the first class data feature can be calculated, the similar difference features are obtained, the similar difference features are determined to be similar data feature updating data of the first class data feature, the first class data feature is enabled to be closer to the similar data feature updating data in the following updating process, and accordingly the first class data feature is enabled to be closer to the similar data feature and can represent the data feature of the first data class. The method provided by the embodiment of the application can restrict the approach of the category data features to the similar features through the similar data feature update data, so that the category data features are more representative and more accurate.
In an embodiment, a difference feature between a data feature of each training data in the at least one training data and the first category data feature may be determined, resulting in at least one same category difference feature; and carrying out summation processing on at least one same-category difference feature to obtain similar data feature updating data of the first-category data feature. And accumulating the data characteristics of each training data of the first data category and the difference characteristics between the first data characteristics to obtain similar data characteristic updating data of the first data characteristics.
S404, determining at least one different type difference feature according to each second type data feature and the first type data feature, and determining different type data feature update data of the first type data feature according to the at least one different type difference feature, wherein the second type data feature comprises type data features except the first type data feature in the cache data feature set.
In the embodiment of the application, the cache data feature set comprises category data features corresponding to a plurality of data categories; the first class data feature is a class data feature corresponding to a first data class in the cache data feature set, and the second class data feature comprises class data features except the first class data feature in the cache data feature set, namely, the first class data feature is different from the data class corresponding to the second class data feature, and the first class data feature and the second class data feature are different from each other. At least one heterogeneous data characteristic of the first class data characteristic may be determined from the respective second class data characteristic and the first class data characteristic, and heterogeneous data characteristic update data of the first class data characteristic may be determined from the determined at least one heterogeneous differential characteristic. The method provided by the embodiment of the application can update the data constraint category data characteristics far away from the heterogeneous data characteristics by utilizing the heterogeneous data characteristics, so that the category data characteristics can be more obviously distinguished from the heterogeneous data characteristics, the distinction degree is further provided, and the accuracy of determining the loss function according to the category data characteristics is improved.
In an embodiment, the implementation manner of determining at least one heterogeneous difference feature according to each of the second class data features and the first class data features may be: and determining difference features between the second class data features and the first class data features to obtain at least one different class difference feature. At this time, an implementation manner of determining the heterogeneous difference feature may be as shown in the following formula (2):
in the formula (2), the amino acid sequence of the compound,representing a first category data feature->Representing a second category of data features->Representing a heterogeneous difference between the first class data feature and the second class data featureFeatures.
In an embodiment, the method shown in the above formula (2) is adopted to determine at least one heterogeneous difference feature, and after determining heterogeneous data feature update data according to the at least one heterogeneous difference feature, the similar data feature update data and the heterogeneous data feature update data may be weighted respectively to obtain weighted similar data feature update data and weighted heterogeneous data feature update data; performing inverse processing on the weighted heterogeneous data characteristic updating data to obtain inverse weighted heterogeneous data characteristic updating data; and updating the first class data characteristic by using the weighted similar data characteristic updating data and the inverted weighted heterogeneous data characteristic updating data, so that the updated first class data characteristic can be obtained.
In an embodiment, the implementation manner of determining at least one heterogeneous difference feature according to each second-class data feature and the first-class data feature may be: performing inverse processing on each second-class data feature to obtain opposite data features of each second-class data feature; and calculating difference features between opposite data features of each second-class data feature and the first-class data feature to obtain at least one different-class difference feature. And performing inverse processing on the second-class data features (namely, taking the inverse number of each element in the second-class data features) to obtain the inverse data features of the second-class data features. After feature normalization, the modulus of the opposite data feature and the original data feature are both one, and the opposite data feature is the feature farthest from the original data feature (i.e., the distance between the two is 2). An implementation manner of determining the heterogeneous difference feature according to the opposite data feature of the second-class data feature and the first-class data feature may be as shown in the following formula (3):
in the formula 3, the components are mixed,representing the second kindInverse data characteristic of other data characteristic, +.>Representing a first category data feature, representing a heterogeneous difference feature between an opposite data feature of a second category data feature and the first category data feature. Referring to fig. 5, a vector update method according to an embodiment of the present application is shown. As shown in fig. 5, it is assumed that in the two-dimensional coordinate system, a feature vector corresponding to a certain second-class data feature is m1, and a feature vector corresponding to an opposite data feature of the second-class data feature is-m 1. In fig. 5, the feature vector m2 is distant from the feature vector m1 in the manner shown in the above formula (2) (i.e., the first-type data feature is distant from the second-type data feature), and the feature vector m3 is close to the feature vector-m 1 in the manner shown in the above formula (3) (i.e., the first-type data feature is close to the opposite data feature of the second-type data feature). As can be seen from fig. 5, to achieve the purpose that the first class data feature is far away from the second class data feature, the efficiency of making the first class data feature approach to the opposite data feature of the second class data feature is higher, that is, the efficiency of adopting the feature update method corresponding to the feature vector m3 is higher. When the heterogeneous difference feature is determined in the manner shown in the formula (3), the heterogeneous difference feature actually represents a difference feature between the opposite data features of the first class data feature and the second class data feature. In order to achieve the purpose that the first class data feature is far away from the second class data feature, namely, far away from the second class data feature, in the updating process of the first class data feature, the first class data feature can be made to be close to the opposite data feature of the second class data feature, namely, in the updating process, the difference feature between the opposite data feature of the first class data feature and the opposite data feature of the second class data feature is gradually reduced, so that the difference feature between the first class data feature and the second class data feature is gradually increased, and the effect that the first class data feature is far away from the second class data feature is achieved. The method provided by the embodiment of the application can determine the different class difference characteristics between the opposite data characteristics of the second class data characteristics and the first class data characteristics And determining the heterogeneous data characteristic updating data of the first class data characteristic according to the heterogeneous data characteristic difference, so that the first class data characteristic is far away from the heterogeneous data characteristic by the opposite data characteristic close to the second class data characteristic, thereby enhancing the distinguishing degree between the first class data characteristic and the heterogeneous data characteristic, and meanwhile, the method of far away from the heterogeneous data characteristic by the opposite data characteristic close to the second class data characteristic can effectively improve the updating efficiency of the first class data characteristic and further improve the training efficiency of the model.
In an embodiment, the implementation manner of determining the heterogeneous data feature update data of the first type of data feature according to the at least one heterogeneous difference feature may be: and carrying out summation processing on at least one heterogeneous difference feature to obtain heterogeneous data feature updating data of the first class data feature. For example: the cache data feature set comprises category data features corresponding to K data categories; the first category of data is characterized byThe difference of different categories is characterized by->Wherein->Representing a second category of data features; the implementation of the heterogeneous data feature update data to determine the first type of data feature may be as shown in the following equation (4):
Equation (4) indicates that the heterogeneous data feature update data D2 is equal to the sum of heterogeneous difference features between the first-type data feature and the respective second-type data features. The heterogeneous data feature update data may bring the first type of data feature close to the opposite data feature of the respective second type of data feature, equivalent to bringing the first type of data feature away from the respective second type of data feature, i.e. away from all the heterogeneous data features. Equation conversion of the above formula (4) can give formula (5):
in the formula (5), C represents an average value of each class of data features in the cached data feature set, and the expression of C may be represented by the following formula (6):
in the formula (6), K represents the number of category data features included in the cached data feature set,representing the ith category data feature. By the method provided by the embodiment of the application, the heterogeneous data characteristic updating data can be determined according to the heterogeneous difference characteristics between the first type data characteristic and each second type data characteristic, so that the first type data characteristic updated according to the heterogeneous data characteristic updating data can be far away from all the heterogeneous data characteristics, the first type data characteristic is more differentiated, and the loss function determined according to the first type data characteristic can be more accurate and has better supervision.
And S405, updating the first class data features according to the similar data feature updating data and the heterogeneous data feature updating data to obtain updated first class data features.
In the embodiment of the application, the update data of two dimensions, namely the similar data feature update data and the heterogeneous data feature update data, can be overlapped to update the first class data feature, so that the updated first class data feature is obtained, the difference feature between the updated first class data feature and the data feature of the training data belonging to the first data class is smaller than the same class difference feature, and the difference feature between the updated first class data feature and each second class data feature is larger than the heterogeneous difference feature between the first class data feature and the corresponding second class data feature. For example: the first data category is a category A, the second data category is a category B and a category C, and the difference characteristic a2 between the updated first data category data characteristic and the data characteristic of the training data belonging to the first data category is smaller than the same category difference characteristic a1 between the first data category data characteristic and the data characteristic of the training data belonging to the first data category; the difference characteristic B2 between the updated first class data characteristic and the second class data characteristic of the class B is larger than the difference characteristic B1 between the first class data characteristic and the second class data characteristic of the class B; the difference feature C2 between the updated first class data feature and the second class data feature of class C is larger than the difference feature C1 between the first class data feature and the second class data feature of class C. The method provided by the embodiment of the application can ensure that the updated first class data characteristics are far away from the heterogeneous data characteristics while being close to the similar data characteristics, thereby improving the accuracy and the distinguishing degree of the first class data characteristics.
In an embodiment, the updating the first class data feature according to the similar data feature updating data and the heterogeneous data feature updating data to obtain an updated implementation manner of the first class data feature may be: acquiring first weight data and second weight data; weighting the similar data characteristic updating data according to the first weight data to obtain weighted similar data characteristic updating data, and weighting the heterogeneous data characteristic updating data according to the second weight data to obtain weighted heterogeneous data characteristic updating data; and carrying out summation processing on the weighted similar data characteristic updating data, the weighted heterogeneous data characteristic updating data and the first class data characteristic to obtain the updated first class data characteristic. In the embodiment of the present application, at least one heterogeneous difference feature is determined by using opposite data features of the second type data feature (i.e., the method shown in the above formula (3)), and after heterogeneous data feature update data is determined according to the at least one heterogeneous difference feature, the similar data feature update data and the heterogeneous data feature update data may be weighted respectively, and then the weighted similar data feature update data and the weighted heterogeneous data feature update data are used to update the first type data feature, where a specific implementation manner may be as shown in the following formula (7):
In the formula (7), the amino acid sequence of the compound,represents first weight data, D1 represents similar data feature update data,/and/or the like>Representing second weight data, D2 representing heterogeneous data feature update data. In some cases, the first weight data +.>The value of (2) may be 0.7, the second weight data +.>The value of (2) may be 0.3. The method provided by the embodiment of the application can integrate the similar data feature updating data and the heterogeneous data feature updating data to update the category data features, thereby ensuring the effectiveness of the category data features, restricting the updating of the category data features from two dimensions and being beneficial to improving the accuracy of determining the loss function according to the category data features. Substituting the above formula (1) and formula (5) into the above formula (7) and performing an equation conversion process can obtain the following formula (8):
in the formula (8), D represents the sum of the weighted similar data feature update data and the weighted heterogeneous data feature update data. D may indicate an update direction of the first class data feature, and summing the D with the first class data feature may obtain an updated first class cache feature, as shown in the following formula (9):
equation conversion processing is performed by substituting the above equation (8) into equation (9), and equation (10) can be obtained. Wherein the method comprises the steps of Representing the first class cache feature after addition of the update direction. After determining the first class cache feature added with the updated direction, normalization processing can be performed on the first class cache feature added with the updated direction to obtain updated first class data features. The method provided by the embodiment of the application can superimpose the update data of two dimensions and update the category data characteristics by utilizing the update data, so that the update of the category data characteristics simultaneously considers the characteristics of the data close to the same category and the characteristics of the data far away from the different category, thereby ensuring that the category data characteristics have stronger distinguishing capability for the characteristics of different categories while expressing the category information more fully.
S406, training the feature extraction model according to the updated class data features corresponding to each data class in the at least one data class and the data features of the training data to obtain a trained feature extraction model.
In the embodiment of the application, a loss function is determined according to the updated class data characteristics corresponding to each data class in at least one data class and the data characteristics of each training data, and the characteristic extraction model is trained according to the loss function, so that the trained characteristic extraction model is obtained.
In an embodiment, the implementation manner of determining the loss function according to the updated class data feature corresponding to each data class in at least one data class and the data feature of each training data may be as shown in the following formulas (11), (12), (13) and (14):
wherein:
in the formula (11), L%) Model parameters representing a model for feature extraction +.>Exp represents a power function of the natural exponent e. Please refer to fig. 6, which is a schematic diagram of a model training method according to an embodiment of the present application. In fig. 6, for the nth training (N is a positive integer greater than 1), each training data in the training data set is input into the feature extraction model to be processed, so as to obtain the data feature of each training data; as shown in the above steps S402 to S404, update data is determined according to the data characteristics of each training data and each category data characteristic updated in the N-1 th training, and the update data specifically includes similar data characteristic update data and heterogeneous data characteristic update data; after determining the update data, performing feature update processing as shown in the above step S405 on each category data feature updated during the nth training according to the update data, to obtain each category data feature updated during the nth training; finally, according to the data characteristics of each training data and the updated category number of each training data in the Nth training The loss function is determined according to the characteristics (the method of determining the loss function may be as shown in the above formulas (11), (12), (13) and (14)), and the model parameter adjustment is performed on the feature extraction model according to the loss function. The method provided by the embodiment of the application can update the category data characteristics from two dimensions at the same time, so that the category data characteristics are more accurate, and the loss function determined according to the category data characteristics is also more accurate, thereby properly reducing the training times of the model and improving the training efficiency of the model; meanwhile, the accuracy of the model can be effectively improved by training the model through the accurate loss function.
In one embodiment, after determining the loss function, model parameters of the feature extraction model may be optimized using an adaptive moment estimation (Adaptive Moment Estimation, adam) optimizer, the learning rate of which may be set to 0.0035.
In an embodiment, the above steps may be repeated multiple times to perform multiple model parameter adjustment on the feature extraction model, and when the number of model training times reaches a preset number of times, model training is completed. For example: the sample data set comprises 10000 sample data, the training data set comprises 100 training data determined from the sample data set, model parameter adjustment calculation is performed on the feature extraction model according to the data characteristics of each training data in one training data set to perform model training once, and when the training data set traverses all sample data in the sample data set and completes the training of the feature extraction model (namely, 100 training data sets are determined according to the sample data included in the sample data set and the feature extraction model is trained by using the 100 training data sets), the model can be marked as one training generation (epoch); the preset number of times may be for model training for 20 training generations. After the model training for the preset number of times is completed, the feature extraction model at this time may be determined as a trained feature extraction model.
In an embodiment, after determining the trained feature extraction model, the accuracy of the feature extraction model may be detected using test data in the test data set. For the accuracy of the feature extraction model after training, rank1 and Rank5 in the average accuracy and accumulated matching features can be used as evaluation standards of the model accuracy. The average accuracy (mean Average Precision, mAP) refers to the average of the accuracy of the search results determined by the feature extraction model for the test data. The cumulative matching feature (Cumulative Matching Characteristics, CMC) may be used to reflect the accuracy of the search result of the test data, rank1 indicates that after searching according to a certain data search rule, the ratio of the number of test data of the correct data tag to the total number of test sample data can be determined for the first time, and Rank5 indicates that there are five opportunities (selecting the five items with the greatest matching degree) to determine whether there is a correct search result. The effect data of the various kinds of data feature updating methods and the data processing methods provided by the present application are shown in table 1 below:
TABLE 1
As shown in table 1, when the update method of the category data features is "constraint only category data features are close to the same category", the mAP of the feature extraction model is 78.0%, rank1 is 90.5%, rank5 is 96.3%; the updating method of the category data features is that when the category data features are only constrained to be far away from each other in a first mode (the first mode refers to a mode shown in the formula (2)) the mAP of the feature extraction model is 80.5%, the Rank1 is 91.5% and the Rank5 is 96.6%; the updating method of the category data features is that when the category data features are only constrained to be far away from each other in a second mode (the second mode refers to a mode shown in the formula (3)) the mAP of the feature extraction model is 81.1%, rank1 is 92.4% and Rank5 is 97.0%; the updating method of the category data features is that when the category data features are simultaneously constrained to be close to the similar features and far away from the heterogeneous features (namely, the data processing method provided by the application), mAP of the feature extraction model is 83.3%, rank1 is 93.7%, and Rank5 is 97.1%. As shown in table 1, compared with the method of mutually separating constraint category data features in the first mode, the method of mutually separating constraint category data features in the second mode is adopted, so that the updating efficiency of category data features is improved, and the average accuracy (mAP) of a feature extraction model is improved by 0.6%; compared with the method for restraining the class data features to be close to the similar features only, the method for restraining the class data features to be close to the similar features and away from the heterogeneous features simultaneously is adopted, so that the resolving power of the class cache features to different data classes is effectively improved, more accurate supervision information can be provided for the model, the training efficiency of the model is improved, the average accuracy (mAP) of the feature extraction model is improved by 5.3%, and Rank1 is improved by 3.2%.
Based on the embodiment, the application has the following beneficial effects: the data processing method provided by the embodiment of the application can determine the similar data feature updating data and the heterogeneous data feature updating data of the category data features, and update the category data features according to the updating data of the two dimensions, so that the category data features can be restrained from being close to the similar data features and being far away from the heterogeneous data features, the category data features corresponding to different data categories have a far-near distance relationship, the distinction degree of the category data features is enhanced, and more accurate supervision information can be provided for model training; the loss function can be determined according to the updated category data characteristics, so that the accuracy of the loss function is improved, and the training efficiency and the prediction accuracy of the model are improved; when determining the category data characteristics, the method provided by the embodiment of the application is determined according to the data characteristics of the sample data, so that the category data characteristics not only comprise the data characteristics of the training data, but also comprise a large number of data characteristics of the sample data in the model training process, and the category data characteristics can be more accurate; when the update data of the heterogeneous data features is determined, the data processing method provided by the embodiment of the application adopts the method of the opposite data features close to the category data features to determine the distance between the features, so that the update efficiency of the category data features can be effectively improved, the distinguishing force of the category data features on different category features can be more effectively improved, and the prediction accuracy of the feature extraction model is improved.
Referring to fig. 7, fig. 7 is a block diagram illustrating a data processing apparatus according to an embodiment of the present application. The device comprises:
an obtaining unit 701, configured to obtain a training data set, and invoke a feature extraction model to perform feature extraction processing on each training data included in the training data set, so as to obtain data features of each training data;
a processing unit 702, configured to determine, for a first class data feature in a cached data feature set, update data of the first class data feature according to the first class data feature and data features of the respective training data, where the update data includes class data feature corresponding to at least one data class in the cached data feature set, the first class data feature is a class data feature corresponding to a first data class in the cached data feature set, the class data feature update data is obtained according to a class difference feature between a data feature of the training data belonging to the first data class in the training data set and the first class data feature, and the class data feature update data is obtained according to a class difference feature between a second class data feature and the first class data feature, where the second class data feature includes class data features other than the first class data feature in the cached data feature set, and the first data class is any one of the at least one data class;
The processing unit 702 is further configured to update the first type of data feature according to the similar type of data feature update data and the heterogeneous type of data feature update data, so as to obtain an updated first type of data feature, wherein a difference feature between the updated first type of data feature and the data feature of the training data belonging to the first type of data is smaller than the same type of difference feature, and a difference feature between the updated first type of data feature and the second type of data feature is larger than the heterogeneous type of difference feature;
the training unit 703 is configured to perform training processing on the feature extraction model according to the updated class data feature corresponding to each data class in the at least one data class and the data feature of each training data, so as to obtain a trained feature extraction model.
In an embodiment, the processing unit 702 is specifically configured to, when determining the update data of the first type data feature according to the first type data feature and the data feature of the respective training data: acquiring at least one training data belonging to the first data category in the training data set; determining a first distance between the data feature of each training data in the at least one training data and the first class data feature to obtain at least one first distance; determining a same-category difference feature according to the at least one first distance, and determining the same-category difference feature as similar data feature update data of the first-category data feature; at least one heterogeneous data feature update data of the first class data feature is determined from each of the second class data features and the first class data features.
In an embodiment, the processing unit 702 is specifically configured to, when determining the peer class difference feature according to the at least one first distance: comparing the at least one first distance to obtain a maximum first distance; and calculating the difference characteristic between the data characteristic of the training data corresponding to the maximum first distance and the first class data characteristic to obtain the same class difference characteristic.
In an embodiment, the processing unit 702 is specifically configured to, when determining at least one heterogeneous difference feature according to each of the second class data features and the first class data features: performing inverse processing on each second-class data feature to obtain opposite data features of each second-class data feature; and calculating difference features between opposite data features of the second-class data features and the first-class data features to obtain at least one different-class difference feature.
In an embodiment, the processing unit 702 is specifically configured to, when determining the heterogeneous data feature update data of the first type of data feature according to the at least one heterogeneous difference feature: and carrying out summation processing on the at least one heterogeneous difference feature to obtain heterogeneous data feature updating data of the first type of data feature.
In an embodiment, the processing unit 702 is further configured to: when training is performed for the 1 st time, calling the feature extraction model to perform feature extraction processing on each sample data included in a sample data set to obtain data features of each sample data, and performing integration processing on the data features of each sample data to obtain the cache data feature set; each training data included in the training data set is included in the sample data set; and in the N-th training, determining the cache data characteristic set according to the updated class data characteristics corresponding to each data class in the N-1 th training, wherein N is a positive integer greater than 1.
In an embodiment, the processing unit 702 is specifically configured to, when performing an integration process on the data features of the respective sample data to obtain the cached data feature set: acquiring data characteristics of at least one sample data corresponding to any one data category of the at least one data category; carrying out mean value processing on the data characteristics of at least one sample data corresponding to any data category to obtain mean value data characteristics corresponding to any data category; normalizing the mean value data characteristic corresponding to any data category to obtain the category data characteristic corresponding to any data category; and determining the cache data characteristic set according to the class data characteristics corresponding to the data classes.
In an embodiment, the processing unit 702 is specifically configured to, when performing update processing on the first class data feature according to the similar data feature update data and the heterogeneous data feature update data to obtain an updated first class data feature: acquiring first weight data and second weight data; weighting the similar data characteristic updating data according to the first weight data to obtain weighted similar data characteristic updating data, and weighting the heterogeneous data characteristic updating data according to the second weight data to obtain weighted heterogeneous data characteristic updating data; and carrying out summation processing on the weighted similar data feature updating data, the weighted heterogeneous data feature updating data and the first class data feature to obtain an updated first class data feature.
It may be understood that the functions of each functional unit of the data processing apparatus according to the embodiments of the present application may be specifically implemented according to the data processing method in the embodiments of the method, and the specific implementation process may refer to the relevant description in the embodiments of the data processing method, which is not repeated herein.
Based on the embodiment, the application has the following beneficial effects: the data processing method provided by the embodiment of the application can determine the similar data feature updating data and the heterogeneous data feature updating data of the category data features, and update the category data features according to the updating data of the two dimensions, so that the category data features can be restrained from being close to the similar data features and being far away from the heterogeneous data features, the category data features corresponding to different data categories have a far-near distance relationship, the distinction degree of the category data features is enhanced, and more accurate supervision information can be provided for model training; the loss function can be determined according to the updated category data characteristics, so that the accuracy of the loss function is improved, and the training efficiency and the prediction accuracy of the model are improved; when determining the category data characteristics, the method provided by the embodiment of the application is determined according to the data characteristics of the sample data, so that the category data characteristics not only comprise the data characteristics of the training data, but also comprise a large number of data characteristics of the sample data in the model training process, and the category data characteristics can be more accurate; when the update data of the heterogeneous data features is determined, the data processing method provided by the embodiment of the application adopts the method of the opposite data features close to the category data features to determine the distance between the features, so that the update efficiency of the category data features can be effectively improved, the distinguishing force of the category data features on different category features can be more effectively improved, and the prediction accuracy of the feature extraction model is improved.
Referring to fig. 8, fig. 8 is a block diagram of a computer device according to an embodiment of the present application. The computer device described in the embodiment of the application comprises: a processor 801, a communication interface 802, and a memory 803. The processor 801, the communication interface 802, and the memory 803 may be connected by a bus or other means, for example, in the embodiment of the present application.
Among them, the processor 801 (or CPU (Central Processing Unit, central processing unit)) is a computing core and a control core of a computer device, which can parse various instructions in the computer device and process various data of the computer device, for example: the CPU can be used for analyzing a startup and shutdown instruction sent by a user to the computer equipment and controlling the computer equipment to perform startup and shutdown operation; and the following steps: the CPU may transmit various types of interaction data between internal structures of the computer device, and so on. The communication interface 802 may optionally include a standard wired interface, a wireless interface (e.g., wi-Fi, mobile communication interface, etc.), controlled by the processor 801 for transceiving data. The Memory 803 (Memory) is a Memory device in the computer device for storing programs and data. It will be appreciated that the memory 803 herein may include both built-in memory of the computer device and extended memory supported by the computer device. Memory 803 provides storage space that stores the operating system of the computer device, which may include, but is not limited to: android systems, iOS systems, windows Phone systems, etc., the application is not limited in this regard.
In an embodiment of the present application, the processor 801 performs the following operations by executing executable program code in the memory 803:
acquiring a training data set, and calling a feature extraction model to perform feature extraction processing on each training data included in the training data set to obtain data features of each training data;
for a first class data feature in a cache data feature set, determining update data of the first class data feature according to the first class data feature and data features of the respective training data, wherein the update data comprises class data feature corresponding to at least one data class, the first class data feature is class data feature corresponding to a first data class in the cache data feature set, the class data feature update data is obtained according to a class difference feature between a data feature of training data belonging to the first data class in the training data set and the first class data feature, and the class data feature update data is obtained according to a class difference feature between a second class data feature and the first class data feature, and the second class data feature comprises class data features except the first class data feature in the cache data feature set, and the first data class is any one of the at least one data class;
Updating the first class data feature according to the similar data feature updating data and the heterogeneous data feature updating data to obtain an updated first class data feature, wherein the difference feature between the updated first class data feature and the data feature of the training data belonging to the first data class is smaller than the same class difference feature, and the difference feature between the updated first class data feature and the second class data feature is larger than the heterogeneous difference feature;
and training the feature extraction model according to the updated class data features corresponding to each data class in the at least one data class and the data features of each training data to obtain a trained feature extraction model.
In an embodiment, the processor 801 is specifically configured to, when determining the update data of the first type data feature according to the first type data feature and the data feature of the respective training data: acquiring at least one training data belonging to the first data category in the training data set; determining a first distance between the data feature of each training data in the at least one training data and the first class data feature to obtain at least one first distance; determining a same-category difference feature according to the at least one first distance, and determining the same-category difference feature as similar data feature update data of the first-category data feature; at least one heterogeneous data feature update data of the first class data feature is determined from each of the second class data features and the first class data features.
In one embodiment, the processor 801 is specifically configured to, when determining the peer class difference feature according to the at least one first distance: comparing the at least one first distance to obtain a maximum first distance; and calculating the difference characteristic between the data characteristic of the training data corresponding to the maximum first distance and the first class data characteristic to obtain the same class difference characteristic.
In one embodiment, the processor 801 is specifically configured to, when determining at least one heterogeneous differential feature according to each of the second class data feature and the first class data feature: performing inverse processing on each second-class data feature to obtain opposite data features of each second-class data feature; and calculating difference features between opposite data features of the second-class data features and the first-class data features to obtain at least one different-class difference feature.
In an embodiment, the processor 801 is specifically configured to, when determining the heterogeneous data feature update data of the first type of data feature according to the at least one heterogeneous difference feature: and carrying out summation processing on the at least one heterogeneous difference feature to obtain heterogeneous data feature updating data of the first type of data feature.
In an embodiment, the processor 801 is further configured to: when training is performed for the 1 st time, calling the feature extraction model to perform feature extraction processing on each sample data included in a sample data set to obtain data features of each sample data, and performing integration processing on the data features of each sample data to obtain the cache data feature set; each training data included in the training data set is included in the sample data set; and in the N-th training, determining the cache data characteristic set according to the updated class data characteristics corresponding to each data class in the N-1 th training, wherein N is a positive integer greater than 1.
In an embodiment, when the processor 801 performs an integration process on the data features of the respective sample data to obtain the cached data feature set, the processor is specifically configured to: acquiring data characteristics of at least one sample data corresponding to any one data category of the at least one data category; carrying out mean value processing on the data characteristics of at least one sample data corresponding to any data category to obtain mean value data characteristics corresponding to any data category; normalizing the mean value data characteristic corresponding to any data category to obtain the category data characteristic corresponding to any data category; and determining the cache data characteristic set according to the class data characteristics corresponding to the data classes.
In an embodiment, when the processor 801 performs update processing on the first class data feature according to the similar data feature update data and the heterogeneous data feature update data to obtain an updated first class data feature, the processor is specifically configured to: acquiring first weight data and second weight data; weighting the similar data characteristic updating data according to the first weight data to obtain weighted similar data characteristic updating data, and weighting the heterogeneous data characteristic updating data according to the second weight data to obtain weighted heterogeneous data characteristic updating data; and carrying out summation processing on the weighted similar data feature updating data, the weighted heterogeneous data feature updating data and the first class data feature to obtain an updated first class data feature.
In a specific implementation, the processor 801, the communication interface 802, and the memory 803 described in the embodiments of the present application may execute an implementation manner of a computer device described in a data processing method provided in the embodiments of the present application, or may execute an implementation manner described in a data processing apparatus provided in the embodiments of the present application, which is not described herein again.
Based on the embodiment, the application has the following beneficial effects: the data processing method provided by the embodiment of the application can determine the similar data feature updating data and the heterogeneous data feature updating data of the category data features, and update the category data features according to the updating data of the two dimensions, so that the category data features can be restrained from being close to the similar data features and being far away from the heterogeneous data features, the category data features corresponding to different data categories have a far-near distance relationship, the distinction degree of the category data features is enhanced, and more accurate supervision information can be provided for model training; the loss function can be determined according to the updated category data characteristics, so that the accuracy of the loss function is improved, and the training efficiency and the prediction accuracy of the model are improved; when determining the category data characteristics, the method provided by the embodiment of the application is determined according to the data characteristics of the sample data, so that the category data characteristics not only comprise the data characteristics of the training data, but also comprise a large number of data characteristics of the sample data in the model training process, and the category data characteristics can be more accurate; when the update data of the heterogeneous data features is determined, the data processing method provided by the embodiment of the application adopts the method of the opposite data features close to the category data features to determine the distance between the features, so that the update efficiency of the category data features can be effectively improved, the distinguishing force of the category data features on different category features can be more effectively improved, and the prediction accuracy of the feature extraction model is improved.
The embodiment of the application also provides a computer readable storage medium, wherein a computer program is stored in the computer readable storage medium, and when the computer program runs on a computer, the computer is caused to execute the data processing method according to the embodiment of the application. The specific implementation manner may refer to the foregoing description, and will not be repeated here.
Embodiments of the present application also provide a computer program product comprising a computer program or computer instructions stored in a computer readable storage medium. A processor of a computer device reads the computer program or computer instructions from the computer readable storage medium, and the processor executes the computer program or computer instructions to cause the computer device to perform a data processing method according to an embodiment of the present application. The specific implementation manner may refer to the foregoing description, and will not be repeated here.
It should be noted that, for simplicity of description, the foregoing method embodiments are all expressed as a series of action combinations, but it should be understood by those skilled in the art that the present application is not limited by the order of action described, as some steps may be performed in other order or simultaneously according to the present application. Further, those skilled in the art will also appreciate that the embodiments described in the specification are all preferred embodiments, and that the acts and modules referred to are not necessarily required for the present application.
Those of ordinary skill in the art will appreciate that all or part of the steps in the various methods of the above embodiments may be implemented by a program to instruct related hardware, the program may be stored in a computer readable storage medium, and the storage medium may include: flash disk, read-Only Memory (ROM), random-access Memory (Random Access Memory, RAM), magnetic or optical disk, and the like.
The above disclosure is illustrative only of some embodiments of the application and is not intended to limit the scope of the application, which is defined by the claims and their equivalents.
Claims (11)
1. A method of data processing, the method comprising:
acquiring a training data set, and calling a feature extraction model to perform feature extraction processing on each training data included in the training data set to obtain data features of each training data;
for a first class data feature in a cache data feature set, determining update data of the first class data feature according to the first class data feature and data features of the respective training data, wherein the update data comprises class data feature corresponding to at least one data class, the first class data feature is class data feature corresponding to a first data class in the cache data feature set, the class data feature update data is obtained according to a class difference feature between a data feature of training data belonging to the first data class in the training data set and the first class data feature, and the class data feature update data is obtained according to a class difference feature between a second class data feature and the first class data feature, and the second class data feature comprises class data features except the first class data feature in the cache data feature set, and the first data class is any one of the at least one data class;
Updating the first class data feature according to the similar data feature updating data and the heterogeneous data feature updating data to obtain an updated first class data feature, wherein the difference feature between the updated first class data feature and the data feature of the training data belonging to the first data class is smaller than the same class difference feature, and the difference feature between the updated first class data feature and the second class data feature is larger than the heterogeneous difference feature;
and training the feature extraction model according to the updated class data features corresponding to each data class in the at least one data class and the data features of each training data to obtain a trained feature extraction model.
2. The method of claim 1, wherein said determining updated data for the first category data feature based on the first category data feature and the data features of the respective training data comprises:
acquiring at least one training data belonging to the first data category in the training data set;
determining a first distance between the data feature of each training data in the at least one training data and the first class data feature to obtain at least one first distance;
Determining a same-category difference feature according to the at least one first distance, and determining the same-category difference feature as similar data feature update data of the first-category data feature;
at least one heterogeneous data feature update data of the first class data feature is determined from each of the second class data features and the first class data features.
3. The method of claim 2, wherein said determining the peer class difference feature from the at least one first distance comprises:
comparing the at least one first distance to obtain a maximum first distance;
and calculating the difference characteristic between the data characteristic of the training data corresponding to the maximum first distance and the first class data characteristic to obtain the same class difference characteristic.
4. The method of claim 2, wherein said determining at least one heterogeneous differential feature from each of said second class data features and said first class data features comprises:
performing inverse processing on each second-class data feature to obtain opposite data features of each second-class data feature;
And calculating difference features between opposite data features of the second-class data features and the first-class data features to obtain at least one different-class difference feature.
5. The method of claim 2, wherein said determining disparate data feature update data for the first category data feature based on the at least one disparate difference feature comprises:
and carrying out summation processing on the at least one heterogeneous difference feature to obtain heterogeneous data feature updating data of the first type of data feature.
6. The method according to claim 1, wherein the method further comprises:
when training is performed for the 1 st time, calling the feature extraction model to perform feature extraction processing on each sample data included in a sample data set to obtain data features of each sample data, and performing integration processing on the data features of each sample data to obtain the cache data feature set; each training data included in the training data set is included in the sample data set;
and in the N-th training, determining the cache data characteristic set according to the updated class data characteristics corresponding to each data class in the N-1 th training, wherein N is a positive integer greater than 1.
7. The method according to claim 6, wherein the integrating the data features of the respective sample data to obtain the cached data feature set includes:
acquiring data characteristics of at least one sample data corresponding to any one data category of the at least one data category;
carrying out mean value processing on the data characteristics of at least one sample data corresponding to any data category to obtain mean value data characteristics corresponding to any data category;
normalizing the mean value data characteristic corresponding to any data category to obtain the category data characteristic corresponding to any data category;
and determining the cache data characteristic set according to the class data characteristics corresponding to the data classes.
8. The method according to any one of claims 1-7, wherein the updating the first category data feature according to the homogeneous data feature updating data and the heterogeneous data feature updating data to obtain an updated first category data feature includes:
acquiring first weight data and second weight data;
weighting the similar data characteristic updating data according to the first weight data to obtain weighted similar data characteristic updating data, and weighting the heterogeneous data characteristic updating data according to the second weight data to obtain weighted heterogeneous data characteristic updating data;
And carrying out summation processing on the weighted similar data feature updating data, the weighted heterogeneous data feature updating data and the first class data feature to obtain an updated first class data feature.
9. A data processing apparatus, the apparatus comprising:
the device comprises an acquisition unit, a feature extraction unit and a feature extraction unit, wherein the acquisition unit is used for acquiring a training data set, and invoking a feature extraction model to perform feature extraction processing on each training data included in the training data set to obtain the data features of each training data;
the processing unit is used for determining updating data of the first class data feature according to the first class data feature and the data feature of each training data in a cache data feature set, wherein the updating data comprises class data feature corresponding to at least one data class in the cache data feature set, the first class data feature is class data feature corresponding to the first data class in the cache data feature set, the class data feature updating data is obtained according to a class difference feature between the data feature of the training data belonging to the first data class in the training data set and the first class data feature, and the class data feature updating data is obtained according to a class difference feature between a second class data feature and the first class data feature, and the second class data feature comprises class data features except the first class data feature in the cache data feature set;
The processing unit is further configured to update the first type data feature according to the similar type data feature update data and the heterogeneous type data feature update data to obtain an updated first type data feature, a difference feature between the updated first type data feature and the data feature of the training data belonging to the first type data is smaller than the same type difference feature, and a difference feature between the updated first type data feature and the second type data feature is larger than the heterogeneous type difference feature;
and the training unit is used for training the feature extraction model according to the updated class data features corresponding to each data class in the at least one data class and the data features of each training data to obtain a trained feature extraction model.
10. A computer device, comprising: the data processing method according to any one of claims 1-8, comprising a processor, a communication interface and a memory, said processor, said communication interface and said memory being interconnected, wherein said memory stores executable program code, said processor being adapted to invoke said executable program code.
11. A computer readable storage medium, characterized in that the computer readable storage medium has stored therein computer instructions, which when run on a computer, cause the computer to implement the data processing method according to any of claims 1-8.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310893001.2A CN116628507B (en) | 2023-07-20 | 2023-07-20 | Data processing method, device, equipment and readable storage medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310893001.2A CN116628507B (en) | 2023-07-20 | 2023-07-20 | Data processing method, device, equipment and readable storage medium |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN116628507A true CN116628507A (en) | 2023-08-22 |
| CN116628507B CN116628507B (en) | 2023-10-27 |
Family
ID=87597596
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310893001.2A Active CN116628507B (en) | 2023-07-20 | 2023-07-20 | Data processing method, device, equipment and readable storage medium |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116628507B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116910100A (en) * | 2023-09-08 | 2023-10-20 | 湖南立人科技有限公司 | Cache data processing method for low-code platform |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110288085A (en) * | 2019-06-20 | 2019-09-27 | 厦门市美亚柏科信息股份有限公司 | A kind of data processing method, device, system and storage medium |
| CN111242199A (en) * | 2020-01-07 | 2020-06-05 | 中国科学院苏州纳米技术与纳米仿生研究所 | Training method and classification method of image classification model |
| CN112085041A (en) * | 2019-06-12 | 2020-12-15 | 北京地平线机器人技术研发有限公司 | Training method and training device for neural network and electronic equipment |
| CN114241273A (en) * | 2021-12-01 | 2022-03-25 | 电子科技大学 | Multi-modal image processing method and system based on Transformer network and hypersphere space learning |
| WO2022161380A1 (en) * | 2021-01-30 | 2022-08-04 | 华为技术有限公司 | Model training method and apparatus, and image retrieval method and apparatus |
| CN115221389A (en) * | 2022-08-03 | 2022-10-21 | 北京芯联心科技发展有限公司 | Training method, device and equipment of cross-modal retrieval model and storage medium |
-
2023
- 2023-07-20 CN CN202310893001.2A patent/CN116628507B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112085041A (en) * | 2019-06-12 | 2020-12-15 | 北京地平线机器人技术研发有限公司 | Training method and training device for neural network and electronic equipment |
| CN110288085A (en) * | 2019-06-20 | 2019-09-27 | 厦门市美亚柏科信息股份有限公司 | A kind of data processing method, device, system and storage medium |
| CN111242199A (en) * | 2020-01-07 | 2020-06-05 | 中国科学院苏州纳米技术与纳米仿生研究所 | Training method and classification method of image classification model |
| WO2022161380A1 (en) * | 2021-01-30 | 2022-08-04 | 华为技术有限公司 | Model training method and apparatus, and image retrieval method and apparatus |
| CN114241273A (en) * | 2021-12-01 | 2022-03-25 | 电子科技大学 | Multi-modal image processing method and system based on Transformer network and hypersphere space learning |
| CN115221389A (en) * | 2022-08-03 | 2022-10-21 | 北京芯联心科技发展有限公司 | Training method, device and equipment of cross-modal retrieval model and storage medium |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116910100A (en) * | 2023-09-08 | 2023-10-20 | 湖南立人科技有限公司 | Cache data processing method for low-code platform |
| CN116910100B (en) * | 2023-09-08 | 2023-11-28 | 湖南立人科技有限公司 | Cache data processing method for low-code platform |
Also Published As
| Publication number | Publication date |
|---|---|
| CN116628507B (en) | 2023-10-27 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US10936919B2 (en) | Method and apparatus for detecting human face | |
| CN111602147B (en) | Machine learning model based on non-local neural network | |
| US10824874B2 (en) | Method and apparatus for processing video | |
| US10599709B2 (en) | Object recognition device, object recognition method, and program for recognizing an object in an image based on tag information | |
| US20190057164A1 (en) | Search method and apparatus based on artificial intelligence | |
| CN107590255B (en) | Information pushing method and device | |
| CN110020009B (en) | Online question and answer method, device and system | |
| CN108197592B (en) | Information acquisition method and device | |
| CN116628507B (en) | Data processing method, device, equipment and readable storage medium | |
| CN112188306B (en) | Label generation method, device, equipment and storage medium | |
| CN110781413A (en) | Interest point determining method and device, storage medium and electronic equipment | |
| CN110765286A (en) | Cross-media retrieval method and device, computer equipment and storage medium | |
| CA2865062A1 (en) | Method and apparatus for enhancing context intelligence in random index based system | |
| US20210089825A1 (en) | Systems and methods for cleaning data | |
| CN114612743A (en) | Deep learning model training method, target object identification method and device | |
| CN110807472A (en) | Image recognition method and device, electronic equipment and storage medium | |
| CN113157956B (en) | Picture searching method, system, mobile terminal and storage medium | |
| CN117409419A (en) | Image detection method, device and storage medium | |
| CN112434533B (en) | Entity disambiguation method, entity disambiguation device, electronic device, and computer-readable storage medium | |
| CN109657710B (en) | Data screening method and device, server and storage medium | |
| CN116994021A (en) | Image detection method, device, computer readable medium and electronic equipment | |
| CN114610938A (en) | Remote sensing image retrieval method and device, electronic equipment and computer readable medium | |
| CN113742525A (en) | Self-supervision video hash learning method, system, electronic equipment and storage medium | |
| CN112070022A (en) | Face image recognition method and device, electronic equipment and computer readable medium | |
| CN111599363A (en) | Voice recognition method and device |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |