CN111860568B - Method and device for balanced distribution of data samples and storage medium - Google Patents

Method and device for balanced distribution of data samples and storage medium Download PDF

Info

Publication number
CN111860568B
CN111860568B CN202010403741.XA CN202010403741A CN111860568B CN 111860568 B CN111860568 B CN 111860568B CN 202010403741 A CN202010403741 A CN 202010403741A CN 111860568 B CN111860568 B CN 111860568B
Authority
CN
China
Prior art keywords
sample
weight
data
determining
samples
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN202010403741.XA
Other languages
Chinese (zh)
Other versions
CN111860568A (en
Inventor
王艳
张修宝
沈海峰
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Beijing Didi Infinity Technology and Development Co Ltd
Original Assignee
Beijing Didi Infinity Technology and Development Co Ltd
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Beijing Didi Infinity Technology and Development Co Ltd filed Critical Beijing Didi Infinity Technology and Development Co Ltd
Priority to CN202010403741.XA priority Critical patent/CN111860568B/en
Publication of CN111860568A publication Critical patent/CN111860568A/en
Application granted granted Critical
Publication of CN111860568B publication Critical patent/CN111860568B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • 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
    • G06N3/045—Combinations of networks
    • 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
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/50—Context or environment of the image
    • G06V20/59—Context or environment of the image inside of a vehicle, e.g. relating to seat occupancy, driver state or inner lighting conditions
    • G06V20/597—Recognising the driver's state or behaviour, e.g. attention or drowsiness
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/16—Human faces, e.g. facial parts, sketches or expressions
    • G06V40/161—Detection; Localisation; Normalisation
    • G06V40/166—Detection; Localisation; Normalisation using acquisition arrangements
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/16—Human faces, e.g. facial parts, sketches or expressions
    • G06V40/168—Feature extraction; Face representation
    • G06V40/171—Local features and components; Facial parts ; Occluding parts, e.g. glasses; Geometrical relationships

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Health & Medical Sciences (AREA)
  • General Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Oral & Maxillofacial Surgery (AREA)
  • Data Mining & Analysis (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Multimedia (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computing Systems (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Computational Linguistics (AREA)
  • Biomedical Technology (AREA)
  • Human Computer Interaction (AREA)
  • Molecular Biology (AREA)
  • Biophysics (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Image Analysis (AREA)
  • Image Processing (AREA)

Abstract

The method not only adds different class weights to samples of different classes, but also considers the difference of the data samples in the same class, and applies the sample weights to the samples at a sample level, so that the weights of the samples in the same class are different, the effect of the classes with small sample number can be increased, the difference between the classes with small sample number and the classes with large sample number can be reduced, and the unbalance of the model caused by unbalanced data distribution can be effectively relieved integrally. In addition, the sample weight of each sample is dynamically set according to the output result of the sample in the network training process, and the method is not limited to setting the same weight for each sample, so that the contribution of the sample to the network is better adjusted, and the influence of unbalanced sample distribution on the model performance is effectively relieved.

Description

Method and device for balanced distribution of data samples and storage medium
Technical Field
The present application relates to artificial intelligence technologies, and in particular, to a method and an apparatus for uniform distribution of data samples, and a storage medium.
Background
With the rapid development of techniques in the field of Artificial Intelligence (AI) from neural networks to deep learning, people have been able to use these AI techniques to implement a perception function for the surrounding environment. For example, in automatic driving, a neural network may be used to perform target detection on an image captured by a camera mounted on a vehicle, thereby identifying an attribute thereof. For example, the evaluation of the meter for the driver's appearance requires detection of images collected by a camera on the vehicle, such as whether the driver's hair style is normal, whether glasses and sunglasses are worn, whether a mask is worn, whether a face has obvious scars, and the like, so as to identify attributes of the driver, classify the driver, and the like, so as to determine the state of the driver based on the attributes later, and further avoid driving the vehicle when the driver is in an abnormal state.
In order to obtain a high-accuracy recognition result, the neural network used needs to be trained first. For example, if a neural network is needed to identify an image to obtain a target object and attributes thereof, the neural network needs to be trained by using a data set (or called a training set) first, so as to identify and classify the target attributes by using the trained neural network.
However, the data samples in the data set are not distributed equally, for example, when the evaluation is performed on the driver appearance instrument, the singular hair style data in the hair style data set is very small, and the data of the normal hair style is very large, so that the trained neural network cannot perform the target attribute identification, classification and the like well.
Disclosure of Invention
In order to solve the problems in the prior art, the present application provides a method and an apparatus for uniform distribution of data samples, and a storage medium.
In a first aspect, an embodiment of the present application provides a method for balanced distribution of data samples, where the data samples are used for training a neural network, and the data samples are image samples, and the method includes:
acquiring an output result of a data sample in a target class in the neural network training process;
determining a sample weight of the data sample according to the output result;
determining a target weight of the data sample according to the sample weight and the class weight of the target class;
and performing balanced distribution processing on the data samples based on the target weight.
In one possible implementation, the determining the sample weight of the data sample according to the output result includes:
determining the sample weight by a first weight function and the output result;
wherein the first weighting function is determined according to the output result, a preset result corresponding to the data sample, and one or more of a training target and a training phase of the neural network.
In a possible implementation manner, the method further includes:
judging whether the number of the data samples is lower than a preset number lower limit or not;
if the number is lower than the preset number lower limit, acquiring a preset result corresponding to the data sample;
the determining a sample weight of the data sample according to the output result comprises:
and determining the sample weight according to the preset result and the output result.
In a possible implementation manner, the method further includes:
judging whether the number of the data samples is higher than a preset number upper limit or not;
if the quantity is higher than the preset quantity upper limit, determining the influence parameters of the data sample on the neural network training;
the determining a sample weight of the data sample according to the output result comprises:
and determining the sample weight according to the influence parameters and the output result.
In a possible implementation manner, the determining the sample weight according to the preset result and the output result includes:
calculating the difference between the output result and the preset result;
determining the sample weight based on the difference.
In one possible implementation, the determining the sample weight based on the difference includes:
determining the sample weight by a second weight function and the difference;
wherein the second weighting function is determined based on the output result, the preset result, and one or more of a training goal and a training phase of the neural network.
In a possible implementation manner, the determining the influence parameters of the data samples on the neural network training includes:
obtaining a difference between the output result and a preset result corresponding to the data sample;
and determining the influence parameters according to the difference and the current training state of the neural network.
In a possible implementation manner, after the determining the sample weight according to the influence parameter and the output result, the method further includes:
judging whether the sample weight is lower than a preset weight threshold value or not;
and if the sample weight is lower than the preset weight threshold value, setting the corresponding data sample as a data sample which does not participate in the neural network training any more.
In one possible implementation, the determining the target weight of the data sample according to the sample weight and the class weight of the target class includes:
calculating a product of the sample weight and a class weight of the target class;
taking the product as the target weight.
In one possible implementation, the image samples are face image samples.
In a second aspect, an embodiment of the present application provides a training method for a neural network, including:
training the neural network using the data samples processed by the method of the first aspect.
In a third aspect, an embodiment of the present application provides an apparatus for balanced distribution of data samples, where the data samples are used for training a neural network, and the data samples are image samples, and the apparatus includes:
the first acquisition module is used for acquiring the output result of the data sample in the target class in the neural network training process;
a first determining module, configured to determine a sample weight of the data sample according to the output result;
a second determining module, configured to determine a target weight of the data sample according to the sample weight and the class weight of the target class;
and the processing module is used for carrying out balanced distribution processing on the data samples based on the target weight.
In a possible implementation manner, the first determining module is specifically configured to:
determining the sample weight by a first weight function and the output result;
wherein the first weighting function is determined according to the output result, a preset result corresponding to the data sample, and one or more of a training target and a training phase of the neural network.
In a possible implementation manner, the apparatus further includes:
the first judgment module is used for judging whether the number of the data samples is lower than a preset number lower limit or not;
the second obtaining module is used for obtaining a preset result corresponding to the data sample if the number is lower than the preset number lower limit;
the first determining module is specifically configured to:
and determining the sample weight according to the preset result and the output result.
In a possible implementation manner, the apparatus further includes:
the second judgment module is used for judging whether the number of the data samples is higher than a preset number upper limit;
a third determining module, configured to determine an influence parameter of the data sample on the neural network training if the quantity is higher than the preset quantity upper limit;
the first determining module is specifically configured to:
and determining the sample weight according to the influence parameters and the output result.
In a possible implementation manner, the determining, by the first determining module, the sample weight according to the preset result and the output result includes:
calculating the difference between the output result and the preset result;
determining the sample weight based on the difference.
In one possible implementation, the determining the sample weight by the first determining module based on the difference includes:
determining the sample weight by a second weight function and the difference;
wherein the second weighting function is determined based on the output result, the preset result, and one or more of a training goal and a training phase of the neural network.
In a possible implementation manner, the third determining module is specifically configured to:
if the quantity is higher than the preset quantity upper limit, obtaining the difference between the output result and a preset result corresponding to the data sample;
and determining the influence parameters according to the difference and the current training state of the neural network.
In a possible implementation manner, after the first determining module determines the sample weight according to the influence parameter and the output result, the method further includes:
the third judging module is used for judging whether the sample weight is lower than a preset weight threshold value;
and the setting module is used for setting the corresponding data sample as the data sample which does not participate in the neural network training any more if the sample weight is lower than the preset weight threshold.
In a possible implementation manner, the second determining module is specifically configured to:
calculating a product of the sample weight and a class weight of the target class;
taking the product as the target weight.
In one possible implementation, the image samples are face image samples.
In a fourth aspect, an embodiment of the present application provides a training apparatus for a neural network, including:
a training module, configured to train the neural network by using the data sample processed by the apparatus of the first aspect.
In a fifth aspect, an embodiment of the present application provides a server, including:
a processor;
a memory; and
a computer program;
wherein the computer program is stored in the memory and configured to be executed by the processor, the computer program comprising instructions for performing the method of the first aspect.
In a sixth aspect, an embodiment of the present application provides a computer-readable storage medium, where a computer program is stored, and the computer program causes a server to execute the method in the first aspect.
According to the method, different class weights are added to different classes of samples, the difference of the data samples in the same class is considered, the sample weights are applied to the samples on the sample level, the weights of the samples in the same class are different, the effect of the classes with small sample number can be increased, the difference between the classes with small sample number and the classes with large sample number can be reduced, and the unbalance of the model caused by unbalanced data distribution is effectively relieved integrally. In addition, the sample weight of each sample is dynamically set according to the output result of the sample in the network training process, and the method is not limited to setting the same weight for each sample, so that the contribution of the sample to the network is better adjusted, and the influence of unbalanced sample distribution on the model performance is effectively relieved.
Drawings
In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings needed to be used in the description of the embodiments or the prior art will be briefly introduced below, and it is obvious that the drawings in the following description are only some embodiments of the present application, and it is obvious for those skilled in the art that other drawings can be obtained according to the drawings without inventive exercise.
Fig. 1 is a schematic diagram of a system architecture for balanced distribution of data samples according to an embodiment of the present disclosure;
fig. 2 is a schematic flowchart of a method for equalizing distribution of data samples according to an embodiment of the present disclosure;
fig. 3 is a schematic flowchart of another method for equalizing distribution of data samples according to an embodiment of the present disclosure;
fig. 4 is a schematic flowchart of another method for equalizing distribution of data samples according to an embodiment of the present application;
fig. 5 is a schematic flowchart of another method for equalizing distribution of data samples according to an embodiment of the present application;
fig. 6 is a schematic structural diagram of an apparatus for equalizing distribution of data samples according to an embodiment of the present disclosure;
fig. 7 is a schematic structural diagram of another apparatus for equalizing distribution of data samples according to an embodiment of the present disclosure;
fig. 8 is a schematic structural diagram of another apparatus for equalizing distribution of data samples according to an embodiment of the present application;
FIG. 9A is a schematic diagram of a possible architecture of a server of the present application;
fig. 9B shows another possible structure diagram of the server of the present application.
Detailed Description
The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application, and it is obvious that the described embodiments are only a part of the embodiments of the present application, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present application.
The terms "first," "second," "third," and "fourth," if any, in the description and claims of this application and the above-described figures are used for distinguishing between similar elements and not necessarily for describing a particular sequential or chronological order. It is to be understood that the data so used is interchangeable under appropriate circumstances such that the embodiments of the application described herein are capable of operation in sequences other than those illustrated or described herein. Furthermore, the terms "comprises," "comprising," and "having," and any variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, system, article, or apparatus that comprises a list of steps or elements is not necessarily limited to those steps or elements expressly listed, but may include other steps or elements not expressly listed or inherent to such process, method, article, or apparatus.
The balanced distribution of the data samples in the embodiment of the application refers to adding different class weights to different classes of samples, and considering that the data samples in the same class are different, the sample weights are applied to the samples on a sample level, so that the weights of the samples in the same class are different, on one hand, the function of the class with small sample number is increased, on the other hand, the difference between the class with small sample number and the class with large sample number is reduced, and the unbalance of the model caused by unbalanced data distribution is integrally and effectively relieved.
The method and the device for the balanced distribution of the data samples provided by the embodiment of the application can be applied to neural network training, wherein the neural system can be used for target attribute recognition, classification and the like, for example, whether the hairstyle of a driver is normal, whether glasses and sunglasses are worn, whether a mask is worn, whether the face has obvious scars and the like are detected based on the neural network, and the embodiment of the application is not particularly limited.
Optionally, the method and apparatus for balanced distribution of data samples provided in the embodiment of the present application may be applied to the application scenario shown in fig. 1. Fig. 1 only describes, by way of example, one possible application scenario of the method for evenly distributing data samples provided in the embodiment of the present application, and the application scenario of the method for evenly distributing data samples provided in the embodiment of the present application is not limited to the application scenario illustrated in fig. 1.
FIG. 1 is a schematic diagram of a system architecture for equalizing distribution of data samples. In fig. 1, an example of detecting whether the hairstyle of the driver is normal is taken, wherein the sample category is hair. The above-described architecture includes the analysis device 11 and a plurality of cameras, here exemplified by the first camera 12, the second camera 13 and the third camera 14.
It is understood that the illustrated structure of the embodiments of the present application does not constitute a specific limitation on the architecture of the balanced distribution of data samples. In other possible embodiments of the present application, the foregoing architecture may include more or less components than those shown in the drawings, or combine some components, or split some components, or arrange different components, which may be determined according to practical application scenarios, and is not limited herein. The components shown in fig. 1 may be implemented in hardware, software, or a combination of software and hardware.
In a specific implementation process, in the embodiment of the present application, the first camera 12, the second camera 13, and the third camera 14 may be respectively mounted on different vehicles, for example, the first camera 12 is mounted on the vehicle 1, the second camera 13 is mounted on the vehicle 2, the third camera 14 is mounted on the vehicle 3, the first camera 12 collects an image of a driver in the vehicle 1, the second camera 13 collects an image of the driver in the vehicle 2, and the third camera 14 collects an image of the driver in the vehicle 3. In the application scenario, before the driver enters the vehicle and starts driving, the camera mounted on the vehicle can be used for collecting the image of the driver. After the acquisition of the images, the first camera 12, the second camera 13 and the third camera 14 transmit the acquired images to the analysis device 11. The analyzing device 11 uses the received image as a data sample for training a neural network. The analyzer 11 considers the difference between the data samples, and applies sample weights to the samples at the sample level to make the weights of the samples in the same class different, thereby increasing the effect of the class with a small number of samples, and reducing the difference between the class with a small number of samples and the class with a large number of samples, so as to effectively alleviate the imbalance of the model caused by the imbalance of data distribution as a whole. Therefore, the neural network can be trained by using the data sample subjected to the balanced distribution processing, so that the trained neural network can better detect whether the hair style of the driver is normal.
In addition, the network architecture and the service scenario described in the embodiment of the present application are for more clearly illustrating the technical solution of the embodiment of the present application, and do not constitute a limitation to the technical solution provided in the embodiment of the present application, and it can be known by a person skilled in the art that along with the evolution of the network architecture and the appearance of a new service scenario, the technical solution provided in the embodiment of the present application is also applicable to similar technical problems.
The following describes a method for equalizing distribution of data samples according to an embodiment of the present application in detail with reference to the accompanying drawings. The subject of execution of the method may be the analysis device 11 in fig. 1. The workflow of the analysis device 11 mainly comprises an acquisition phase and a determination phase. In the acquisition stage, the analysis device 11 acquires the output result of the data sample in the neural network training process. In the determining stage, the analyzing device 11 determines the sample weight of the data sample according to the output result, and further performs the balanced distribution processing on the data sample, thereby effectively relieving the imbalance of the model caused by the unbalanced data distribution as a whole.
The technical solutions of the present application are described below with several embodiments as examples, and the same or similar concepts or processes may not be described in detail in some embodiments.
Fig. 2 is a flowchart of a method for equalizing distribution of data samples according to an embodiment of the present disclosure, where the method for equalizing distribution of data samples is applicable to processing of equalizing distribution of data samples, and the method may be executed by any apparatus that executes the method for equalizing distribution of data samples, and the apparatus may be implemented by software and/or hardware. As shown in fig. 2, on the basis of the application scenario shown in fig. 1, the method for equalizing distribution of data samples provided in the embodiment of the present application includes the following steps:
s201: and acquiring the output result of the data sample in the target class in the neural network training process.
Here, the above data samples are used to train a neural network, which can be used for object attribute recognition, classification, and the like, for example, to detect whether the driver's hairstyle is normal, whether glasses and sunglasses are worn, whether a mask is worn, whether a face has a significant scar, and the like, based on the neural network. The data samples are image samples, such as driver images collected by a camera mounted on the vehicle.
Illustratively, the image samples are human face image samples.
The target category may be set according to actual conditions, and is not particularly limited in the embodiment of the present application. Taking the above-mentioned detection of whether the driver's hair style is normal or not as an example, the attribute is hair, the category may be long hair, short hair, no hair, etc., each category includes a plurality of data samples, for example, the long hair category includes a plurality of long hair data samples. The target category may be the long hair, short hair, no hair, etc. described above.
In the embodiment of the present application, taking the execution subject as the analysis device 11 as an example, the analysis device respectively obtains the output results of all the data samples in a certain category in the neural network training process, and taking the long hair category as an example, if the category includes 10 long hair data samples, the analysis device respectively obtains the output results of the 10 long hair data samples in the neural network training process, and may further establish a corresponding relationship between the data samples and the output results of the samples, so as to perform corresponding processing based on the corresponding relationship in the following, and is suitable for application.
S202: and determining the sample weight of the data sample according to the output result.
According to the embodiment of the application, the weight of the sample is determined according to the output result of the sample in the training process, so that the output of the sample can change along with the training of the network, the weight of the sample can also change dynamically, the weight of each sample is determined according to the actual situation of the network, and the weight is based on the dynamic change of the sample, so that the contribution of the sample to the network is adjusted better, and the influence of sample distribution imbalance on model performance is relieved effectively.
For example, determining the sample weight of the data sample according to the output result may include: and determining the sample weight of the data sample according to the preset result corresponding to the data sample and the output result. Here, the preset result corresponding to the data sample may be a known real output result corresponding to the data sample, and the preset result may be predefined or configured, which is not particularly limited in the embodiment of the present application.
S203: and determining the target weight of the data sample according to the sample weight and the class weight of the target class.
The category weight of the target category may be obtained by:
and obtaining the category weight of the target category according to the pre-stored corresponding relation between the category and the category weight. Here, the correspondence relationship may be set according to actual conditions, and is not particularly limited in the embodiment of the present application.
In one possible implementation manner, the determining the target weight of the data sample according to the sample weight and the class weight of the target class includes:
calculating a product of the sample weight and the class weight of the target class;
this product is taken as the above target weight.
S204: and performing balanced distribution processing on the data samples based on the target weight.
In the embodiment of the application, different class weights are added to samples of different classes, the difference between data samples in the same class is considered, and the sample weights are applied to the samples on a sample level, so that the weights of the samples in the same class are different, the effect of the class with a small number of samples can be increased on the one hand, the difference between the class with a small number of samples and the class with a large number of samples can be reduced on the other hand, and the unbalance of the model caused by unbalanced data distribution is effectively relieved on the whole.
In a possible implementation manner, after performing the equalized distribution processing on the data samples, the method further includes:
and sending the data samples to the network, sending the weights of the data samples to the network together, and directly multiplying the weights of the data samples by the weights when calculating the loss to obtain the final loss.
And determining whether to continue to perform the balanced distribution processing on the data samples according to the loss. If so, the operation of the balanced distribution processing is continuously executed on the data samples, otherwise, the corresponding processing is stopped, and therefore the problem of unbalanced distribution of the data samples is effectively solved.
According to the embodiment of the application, different class weights are added to samples of different classes, the difference existing between data samples in the same class is considered, the sample weights are applied to the samples on a sample level, the weights of the samples in the same class are different, the effect of the class with small sample number can be increased on one hand, the difference between the class with small sample number and the class with large sample number is reduced on the other hand, and the unbalance of a model caused by unbalanced data distribution is relieved integrally and effectively. In addition, the sample weight of each sample is dynamically set according to the output result of the sample in the network training process, and the method is not limited to setting the same weight for each sample, so that the contribution of the sample to the network is better adjusted, and the influence of unbalanced sample distribution on the model performance is effectively relieved.
In addition, when determining the sample weight of the data sample, the embodiment of the application not only considers the output result of the data sample in the neural network training process, but also utilizes the weight function. Fig. 3 is a flowchart illustrating another method for equalizing distribution of data samples according to an embodiment of the present disclosure. As shown in fig. 3, the method includes:
s301: and acquiring an output result of the data sample in the target class in the neural network training process.
Step S301 is the same as the implementation of step S201, and is not described herein again.
S302: and determining the sample weight of the data sample through a first weight function and the output result.
The first weighting function is determined according to the output result, a preset result corresponding to the data sample, and one or more of a training target and a training stage of the neural network.
Here, the preset result corresponding to the data sample may be a known real output result corresponding to the data sample, and the preset result may be predefined or configured.
The training target of the neural network may be determined according to actual situations, for example, the form of the output result is a numerical value or a probability, and the like, which is not particularly limited in the embodiment of the present application.
The neural network may be divided into different stages in the training process, such as a first stage, a second stage, and the like, and the training stage of the neural network may be understood as the different stages divided in the training process of the neural network.
The first weighting function may be determined according to one or more of the output result, the preset result corresponding to the data sample, and the training target and the training phase of the neural network, and the first weighting function may also be determined according to other factors, which is not limited in this embodiment of the present application.
The first weight function may be used to derive sample weights of various data samples, for example, the output results of the data samples in the neural network training process may be used as input parameters to generate sample weights of the data samples, which are, for example: p is a radical ofi＝g1(f(xi) Wherein p isiSample weight, g, representing data sample i1() Representing a first weight function, f (x)i) Representing the output of data sample i.
In addition, the first weighting function may further include other input parameters, such as the preset result corresponding to the data samples, which is not particularly limited in the embodiments of the present application.
When the sample weight is determined, firstly, an output result of a data sample in a neural network training process is obtained, and then, the sample weight of the data sample is quickly and accurately generated through a weight function and the output result, so that the method is suitable for application.
S303: and determining the target weight of the data sample according to the sample weight and the class weight of the target class.
S304: and performing balanced distribution processing on the data samples based on the target weight.
The steps S303 to S304 are the same as the steps S203 to S304, and are not described herein again.
According to the embodiment of the application, the sample weight of the data sample is generated quickly and accurately through the weight function and the output result, the application is suitable, different class weights are added to samples of different classes, the difference between the data samples in the same class is also considered, the sample weight is applied to the sample at the sample level, the weights of the samples in the same class are different, the effect of the class with small sample number can be increased on one hand, the difference between the class with small sample number and the class with large sample number is reduced on the other hand, and the unbalance of the model caused by unbalanced data distribution is integrally and effectively relieved. In addition, the sample weight of each sample is dynamically set according to the output result of the sample in the network training process, and the method is not limited to setting the same weight for each sample, so that the contribution of the sample to the network is better adjusted, and the influence of unbalanced sample distribution on the model performance is effectively relieved.
In addition, in the embodiment of the application, in the process of carrying out balanced distribution on the data samples, the number of the data samples is also considered, and different treatments are given to the small number of the samples. Fig. 4 is a flowchart illustrating another method for equalizing distribution of data samples according to an embodiment of the present disclosure. As shown in fig. 4, the method includes:
s401: and acquiring an output result of the data sample in the target class in the neural network training process.
Step S401 is the same as the implementation of step S201, and is not described herein again.
S402: and judging whether the number of the data samples is lower than a preset number lower limit or not.
Here, the lower limit of the preset number may be set according to practical situations, and the embodiment of the present application does not particularly limit this.
S403: and if the number is lower than a preset number lower limit, acquiring a preset result corresponding to the data sample.
If the number of the data samples is lower than the lower limit of the preset number, which indicates that the number of the samples is small, in the embodiment of the present application, different processing is performed on such samples, for example, the preset result corresponding to the data samples is obtained, and further, the sample weight of the data samples is determined according to the preset result and the output result.
S404: and determining the sample weight of the data sample according to the preset result and the output result.
In a possible implementation manner, the determining the sample weight according to the preset result and the output result includes:
calculating the difference between the output result and the preset result;
based on the difference, the sample weight is determined.
Here, in determining the sample weight, the output result and the preset result are taken into consideration, so that the determined sample weight is more suitable for the actual situation.
In addition, in addition to the output result and the preset result, other factors may be considered when determining the sample weight, and this embodiment of the present application is not particularly limited thereto.
Illustratively, the determining the sample weight based on the difference includes:
the sample weight is determined by a second weight function and the difference.
Wherein the second weighting function is determined according to the output result, the preset result, and one or more of a training target and a training phase of the neural network.
Here, the second weighting function may be determined by other factors besides the above, and this is not particularly limited by the embodiment of the present application.
The second weighting function may be used to derive sample weights for various data samples, such as the difference between the output result and the predetermined result, which may be used as an input parameter to generate the sample weights for the data samples, for example, the square of the difference between the output result and the predetermined result, such as: p is a radical ofi＝g2((f(xi)-yi)2) Wherein p isiSample weight, g, representing data sample i2() Representing a second weight function, f (x)i) Output result, y, representing data sample iiAnd representing a preset result corresponding to the data sample i.
In addition, the second weighting function may further include other input parameters, which may be set according to practical situations, and this is not particularly limited in this embodiment of the application.
In the embodiment of the application, when determining the sample weight, not only the output result of the data sample in the neural network training process is considered, but also the preset result corresponding to the data sample is considered, further, the difference between the output result and the preset result is calculated, and the sample weight of the data sample is generated through the weight function and the difference, so that the effect of each sample on the network is fully excavated under the condition of small number of samples.
S405: and determining the target weight of the data sample according to the sample weight and the class weight of the target class.
S406: and performing balanced distribution processing on the data samples based on the target weight.
The steps S405 to S406 are the same as the implementation of the steps S203 to S304, and are not described herein again.
According to the embodiment of the application, under the condition that the number of samples is small, the effect of each sample on a network is fully excavated, and then, the subsequent data samples can be better subjected to balanced distribution processing, in addition, different class weights are added to the samples of different classes, the difference between the data samples in the same class is also considered, the sample weights are applied to the samples at the sample level, the weights of the samples in the same class are different, the effect of the classes with the small number of samples can be increased on the one hand, the difference between the classes with the small number of samples and the classes with the large number of samples is reduced on the one hand, and the unbalance of the model caused by unbalanced data distribution is integrally and effectively relieved. In addition, the sample weight of each sample is dynamically set according to the output result of the sample in the network training process, and the method is not limited to setting the same weight for each sample, so that the contribution of the sample to the network is better adjusted, and the influence of unbalanced sample distribution on the model performance is effectively relieved.
In addition, in the embodiment of the present application, in the process of performing the uniform distribution on the data samples, different processing is performed on a data sample having a large number of samples in addition to the data sample having a small number of samples. Fig. 5 is a flowchart illustrating another method for equalizing distribution of data samples according to an embodiment of the present application. As shown in fig. 5, the method includes:
s501: and acquiring an output result of the data sample in the target class in the neural network training process.
Step S501 is the same as the implementation of step S201, and is not described herein again.
S502: and judging whether the number of the data samples is higher than a preset upper limit of the number.
Here, the preset upper limit of the number may be set according to practical situations, and the embodiment of the present application does not particularly limit this.
S503: and if the quantity is higher than the preset quantity upper limit, determining the influence parameters of the data samples on the neural network training.
In a possible implementation manner, the determining the parameters of influence of the data samples on the neural network training includes:
obtaining the difference between the output result and a preset result corresponding to the data sample;
and determining the influence parameters according to the difference and the current training state of the neural network.
Here, the preset result corresponding to the data sample may be a known real output result corresponding to the data sample, and the preset result may be predefined or configured, which is not particularly limited in the embodiment of the present application. The current training state of the neural network may be determined according to actual conditions, which is not particularly limited in the embodiments of the present application.
In the embodiment of the application, when the number of samples is large, the samples are different, some samples are difficult, and some samples are very simple. In order to prevent the difficult samples from being masked by the simple samples and not playing the true role, for the data, the influence parameters of the samples on the network are considered to be set with different weights while the output result is considered.
S504: and determining the sample weight according to the influence parameters and the output result.
The sample weight may be determined, for example, by a third weight function, as well as the influence quantity and the output result.
Wherein the third weighting function is determined according to the output result, the preset result, and one or more of a training target and a training phase of the neural network.
In addition, the third weighting function may be determined by other factors besides the above, and the embodiment of the present application is not particularly limited thereto.
In a possible implementation manner, after the determining the sample weight according to the influence parameter and the output result, the method further includes:
judging whether the sample weight is lower than a preset weight threshold value or not;
and if the sample weight is lower than a preset weight threshold value, setting the corresponding data sample as a data sample which does not participate in the neural network training any more.
The preset weight threshold may be set according to an actual situation, and is not particularly limited in the embodiment of the present application.
Here, since the number of samples is large, in order to make difficult samples play a role in a large amount along with the training process of the network, when the weight of a certain sample is low, the sample does not participate in the training of the network any more. The number of samples involved in the training is reduced, and the truly useful samples can also function better, and the difference in number between classes with small number of samples can also be reduced.
S505: and determining the target weight of the data sample according to the sample weight and the class weight of the target class.
S506: and performing balanced distribution processing on the data samples based on the target weight.
The implementation of steps S505 to S506 is the same as that of steps S203 to S304, and is not described herein again.
According to the embodiment of the application, under the condition that the number of samples is large, different weights are set for the samples according to the influence parameters of the samples on the network while the output results are considered, further, the data samples can be better subjected to balanced distribution processing subsequently, in addition, the difference existing between the data samples in the same category is considered while different category weights are added to the samples of different categories, the sample weights are applied to the samples at the sample level, the weights of the samples in the same category are different, the effect of the category with the small number of samples can be increased on the one hand, the difference between the category with the small number of samples and the category with the large number of samples is reduced, and therefore the unbalance of the model caused by unbalanced data distribution is integrally and effectively relieved. In addition, the sample weight of each sample is dynamically set according to the output result of the sample in the network training process, and the method is not limited to setting the same weight for each sample, so that the contribution of the sample to the network is better adjusted, and the influence of unbalanced sample distribution on the model performance is effectively relieved.
The embodiment of the present application further provides a training method for a neural network, including:
the neural network is trained using the data samples processed as described above.
According to the embodiment of the application, the data samples processed by the method are used for carrying out the neural network, so that the trained neural network can well carry out target attribute recognition, classification and the like, and the actual application requirements are met.
Fig. 6 is a schematic structural diagram of an apparatus for equalizing distribution of data samples according to an embodiment of the present application, which corresponds to the method for equalizing distribution of data samples according to the foregoing embodiment. For convenience of explanation, only portions related to the embodiments of the present application are shown. Fig. 6 is a schematic structural diagram of an apparatus for evenly distributing data samples according to an embodiment of the present application, where the apparatus 60 for evenly distributing data samples includes: a first obtaining module 601, a first determining module 602, a second determining module 603, and a processing module 604. Here, the data samples are used for training a neural network, and the data samples are image samples. The data sample distribution device may be the analysis device itself, or a chip or an integrated circuit that implements the functions of the analysis device. It should be noted here that the division of the first obtaining module, the first determining module, the second determining module, and the processing module is only a division of logic functions, and the two may be integrated or independent physically.
The first obtaining module 601 is configured to obtain an output result of the data sample in the target class in the neural network training process.
A first determining module 602, configured to determine a sample weight of the data sample according to the output result.
A second determining module 603, configured to determine a target weight of the data sample according to the sample weight and the class weight of the target class.
A processing module 604, configured to perform an equal distribution process on the data samples based on the target weight.
In one possible design, the first determining module 602 is specifically configured to:
determining the sample weight by a first weight function and the output result;
wherein the first weighting function is determined according to the output result, a preset result corresponding to the data sample, and one or more of a training target and a training phase of the neural network.
In a possible design, the second determining module 603 is specifically configured to:
calculating a product of the sample weight and a class weight of the target class;
taking the product as the target weight.
In one possible design, the image samples are human face image samples.
The apparatus provided in the embodiment of the present application may be configured to implement the technical solution of the method embodiment, and the implementation principle and the technical effect are similar, which are not described herein again in the embodiment of the present application.
Fig. 7 is a schematic structural diagram of another apparatus for equalizing distribution of data samples according to an embodiment of the present disclosure. As shown in fig. 7, in addition to fig. 6, the apparatus 60 for evenly distributing data samples further includes: a first determining module 605 and a second obtaining module 606.
The first determining module 605 is configured to determine whether the number of the data samples is lower than a preset lower number limit.
A second obtaining module 606, configured to obtain a preset result corresponding to the data sample if the number is lower than the preset number lower limit.
The first determining module 602 is specifically configured to:
and determining the sample weight according to the preset result and the output result.
In one possible design, the determining module 602 determines the sample weight according to the preset result and the output result, including:
calculating the difference between the output result and the preset result;
determining the sample weight based on the difference.
In one possible design, the first determining module 602 determines the sample weights based on the differences, including:
determining the sample weight by a second weight function and the difference;
wherein the second weighting function is determined based on the output result, the preset result, and one or more of a training goal and a training phase of the neural network.
The apparatus provided in the embodiment of the present application may be configured to implement the technical solution of the method embodiment, and the implementation principle and the technical effect are similar, which are not described herein again in the embodiment of the present application.
Fig. 8 is a schematic structural diagram of another apparatus for equalizing distribution of data samples according to an embodiment of the present application. As shown in fig. 8, in addition to fig. 6, the apparatus 60 for evenly distributing data samples further includes: a second determination module 607 and a third determination module 608.
The second determining module 607 is configured to determine whether the number of the data samples is higher than a preset upper limit;
a third determining module 608, configured to determine an influence parameter of the data sample on the neural network training if the quantity is higher than the preset quantity upper limit;
the first determining module 602 is specifically configured to:
and determining the sample weight according to the influence parameters and the output result.
In one possible design, the third determining module 608 is specifically configured to:
if the quantity is higher than the preset quantity upper limit, obtaining the difference between the output result and a preset result corresponding to the data sample;
and determining the influence parameters according to the difference and the current training state of the neural network.
In one possible design, after the first determining module 602 determines the sample weight according to the influence parameter and the output result, the method further includes:
the third judging module is used for judging whether the sample weight is lower than a preset weight threshold value;
and the setting module is used for setting the corresponding data sample as the data sample which does not participate in the neural network training any more if the sample weight is lower than the preset weight threshold.
The apparatus provided in the embodiment of the present application may be configured to implement the technical solution of the method embodiment, and the implementation principle and the technical effect are similar, which are not described herein again in the embodiment of the present application.
The embodiment of the present application further provides a training apparatus for a neural network, including:
and the training module is used for training the neural network by utilizing the data samples processed by the data sample balanced distribution device.
Fig. 9A shows a schematic diagram of a possible structure of the server of the present application. The server 100 includes: a processing unit 102 and a communication unit 103. The processing unit 102 is used for controlling and managing the actions of the server 100, for example, the processing unit 102 is used for supporting the server 100 to execute the above-mentioned method steps and/or other processes for the technology described herein. The communication unit 103 is used to support communication between the server 100 and other network entities, for example, terminal devices. The server 100 may further comprise a storage unit 101 for storing computer program codes and data of the server 100.
The processing unit 102 may be a processor or a controller, such as a CPU, a general purpose processor, a Digital Signal Processor (DSP), an Application-Specific Integrated Circuit (ASIC), a Field Programmable Gate Array (FPGA) or other Programmable logic device, a transistor logic device, a hardware component, or any combination thereof. Which may implement or perform the various illustrative logical blocks, modules, and circuits described in connection with the disclosure. The processor may also be a combination of computing functions, e.g., comprising one or more microprocessors, DSPs, and microprocessors, among others. The communication unit 103 may be a communication interface, a transceiver, a transceiving circuit, etc., wherein the communication interface is a generic term and may comprise one or more interfaces. The storage unit 101 may be a memory.
When the processing unit 102 is a processor, the communication unit 103 is a communication interface, and the storage unit 101 is a memory, the server according to the present application may be the server shown in fig. 9B.
Referring to fig. 9B, the server 110 includes: a processor 112, a communication interface 113, and a memory 111. Optionally, server 110 may also include bus 114. Wherein, the communication interface 113, the processor 112 and the memory 111 may be connected to each other by a bus 114; the bus 114 may be a Peripheral Component Interconnect (PCI) bus, an Extended Industry Standard Architecture (EISA) bus, or the like. The bus 114 may be divided into an address bus, a data bus, a control bus, and the like. For ease of illustration, only one thick line is shown in FIG. 9B, but this is not intended to represent only one bus or type of bus.
In addition, a computer program is stored in the memory 111 and configured to be executed by the processor 112, the computer program comprising instructions for performing the method as described in the embodiments shown above.
The embodiment of the present application further provides a computer-readable storage medium, where the computer-readable storage medium stores a computer program, and the computer program enables a server to execute the method for evenly distributing data samples provided in the foregoing illustrated embodiment. The readable storage medium may be implemented by any type of volatile or non-volatile memory device or combination thereof, such as Static Random Access Memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic or optical disk.
Those of ordinary skill in the art will understand that: all or a portion of the steps of implementing the above-described method embodiments may be performed by hardware associated with program instructions. The program may be stored in a computer-readable storage medium. When executed, the program performs steps comprising the method embodiments described above; and the aforementioned storage medium includes: various media that can store program codes, such as ROM, RAM, magnetic or optical disks.
Finally, it should be noted that: the above embodiments are only used for illustrating the technical solutions of the present application, and not for limiting the same; although the present application has been described in detail with reference to the foregoing embodiments, it should be understood by those of ordinary skill in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some or all of the technical features may be equivalently replaced; and the modifications or the substitutions do not make the essence of the corresponding technical solutions depart from the scope of the technical solutions of the embodiments of the present application.

Claims (13)

1. The method for the balanced distribution of the data samples is characterized in that the data samples are used for training a neural network, the data samples are image samples, the image samples are driver image samples collected by a camera mounted on a vehicle, the purpose of the neural network comprises the steps of detecting whether the hairstyle of a driver is normal, whether glasses are worn, whether a mask is worn and whether the face has obvious scars, and the method comprises the following steps:
obtaining an output result of the data sample in the target class in the neural network training process;
determining a sample weight of the data sample according to the output result;
determining a target weight of the data sample according to the sample weight and the class weight of the target class;
based on the target weight, carrying out balanced distribution processing on the data sample;
wherein the class weight of the target class is obtained by:
obtaining the category weight of the target category according to the corresponding relation between the pre-stored category and the category weight;
the method further comprises the following steps:
judging whether the number of the data samples is higher than a preset number upper limit or not;
if the quantity is higher than the preset quantity upper limit, determining the influence parameters of the data sample on the neural network training;
the determining a sample weight of the data sample according to the output result comprises:
and determining the sample weight according to the influence parameters and the output result.
2. The method of claim 1, wherein determining the sample weight of the data sample according to the output comprises:
determining the sample weight by a first weight function and the output result;
wherein the first weighting function is determined according to the output result, a preset result corresponding to the data sample, and one or more of a training target and a training phase of the neural network.
3. The method of claim 1 or 2, further comprising:
judging whether the number of the data samples is lower than a preset number lower limit or not;
if the number is lower than the preset number lower limit, acquiring a preset result corresponding to the data sample;
the determining a sample weight of the data sample according to the output result comprises:
and determining the sample weight according to the preset result and the output result.
4. The method of claim 3, wherein determining the sample weight according to the preset result and the output result comprises:
calculating the difference between the output result and the preset result;
determining the sample weight based on the difference.
5. The method of claim 4, wherein the determining the sample weight based on the difference comprises:
determining the sample weight by a second weight function and the difference;
wherein the second weighting function is determined based on the output result, the preset result, and one or more of a training goal and a training phase of the neural network.
6. The method of claim 1, wherein determining the parameters of influence of the data samples on the neural network training comprises:
obtaining a difference between the output result and a preset result corresponding to the data sample;
and determining the influence parameters according to the difference and the current training state of the neural network.
7. The method of claim 1, further comprising, after said determining said sample weights based on said influencing quantities and said output,:
judging whether the sample weight is lower than a preset weight threshold value or not;
and if the sample weight is lower than the preset weight threshold value, setting the corresponding data sample as a data sample which does not participate in the neural network training any more.
8. The method of claim 1 or 2, wherein determining the target weight for the data sample based on the sample weight and the class weight for the target class comprises:
calculating a product of the sample weight and a class weight of the target class;
taking the product as the target weight.
9. The method of claim 1, wherein the image sample is a human face image sample.
10. A method of training a neural network, comprising:
training the neural network with the data samples processed by the method for the balanced distribution of the data samples according to any one of claims 1 to 9.
11. The device for the balanced distribution of the data samples is characterized in that the data samples are used for training a neural network, the data samples are image samples, the image samples are driver image samples collected by a camera mounted on a vehicle, the purpose of the neural network comprises the steps of detecting whether the hair style of a driver is normal, whether glasses are worn, whether a mask is worn and whether the face has obvious scars, and the device comprises:
the first acquisition module is used for acquiring the output result of the data sample in the target class in the neural network training process;
a first determining module, configured to determine a sample weight of the data sample according to the output result;
a second determining module, configured to determine a target weight of the data sample according to the sample weight and the class weight of the target class;
the processing module is used for carrying out balanced distribution processing on the data samples based on the target weight;
wherein the second determination module obtains the class weight of the target class by:
obtaining the category weight of the target category according to the corresponding relation between the pre-stored category and the category weight;
the device, still include:
the second judgment module is used for judging whether the number of the data samples is higher than a preset number upper limit;
a third determining module, configured to determine an influence parameter of the data sample on the neural network training if the quantity is higher than the preset quantity upper limit;
the first determining module is specifically configured to:
and determining the sample weight according to the influence parameters and the output result.
12. A server, comprising:
a processor;
a memory; and
a computer program;
wherein the computer program is stored in the memory and configured to be executed by the processor, the computer program comprising instructions for performing the method of any of claims 1-9.
13. A computer-readable storage medium, characterized in that it stores a computer program that causes a server to execute the method of any one of claims 1-9.
CN202010403741.XA 2020-05-13 2020-05-13 Method and device for balanced distribution of data samples and storage medium Active CN111860568B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010403741.XA CN111860568B (en) 2020-05-13 2020-05-13 Method and device for balanced distribution of data samples and storage medium

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010403741.XA CN111860568B (en) 2020-05-13 2020-05-13 Method and device for balanced distribution of data samples and storage medium

Publications (2)

Publication Number Publication Date
CN111860568A CN111860568A (en) 2020-10-30
CN111860568B true CN111860568B (en) 2022-02-08

Family

ID=72985162

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010403741.XA Active CN111860568B (en) 2020-05-13 2020-05-13 Method and device for balanced distribution of data samples and storage medium

Country Status (1)

Country Link
CN (1) CN111860568B (en)

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113052165A (en) * 2021-01-28 2021-06-29 北京迈格威科技有限公司 Target detection method and device, electronic equipment and storage medium
CN113255670A (en) * 2021-06-28 2021-08-13 湖南星汉数智科技有限公司 Unbalanced small sample target detection method and device and computer equipment
CN116821647B (en) * 2023-08-25 2023-12-05 中国电子科技集团公司第十五研究所 Optimization method, device and equipment for data annotation based on sample deviation evaluation

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP3355270A1 (en) * 2017-01-27 2018-08-01 AGFA Healthcare Multi-class image segmentation method
CN109272115A (en) * 2018-09-05 2019-01-25 宽凳（北京）科技有限公司 A kind of neural network training method and device, equipment, medium
CN109948478A (en) * 2019-03-06 2019-06-28 中国科学院自动化研究所 The face identification method of extensive lack of balance data neural network based, system
CN110443280A (en) * 2019-07-05 2019-11-12 北京达佳互联信息技术有限公司 Training method, device and the storage medium of image detection model

Family Cites Families (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20160092789A1 (en) * 2014-09-29 2016-03-31 International Business Machines Corporation Category Oversampling for Imbalanced Machine Learning
CN104573013A (en) * 2015-01-09 2015-04-29 上海大学 Category weight combined integrated learning classifying method
US11449756B2 (en) * 2018-09-24 2022-09-20 Samsung Electronics Co., Ltd. Method to balance sparsity for efficient inference of deep neural networks
CN109472345A (en) * 2018-09-28 2019-03-15 深圳百诺名医汇网络技术有限公司 A kind of weight update method, device, computer equipment and storage medium
CN111091191B (en) * 2018-10-23 2023-07-21 中国移动通信有限公司研究院 Data processing method and device and data deep learning system
CN110287932B (en) * 2019-07-02 2021-04-13 中国科学院空天信息创新研究院 Road blocking information extraction method based on deep learning image semantic segmentation
CN110705489B (en) * 2019-10-09 2022-07-26 北京迈格威科技有限公司 Training method and device for target recognition network, computer equipment and storage medium
CN111028224B (en) * 2019-12-12 2020-12-01 广西医准智能科技有限公司 Data labeling method, model training device, image processing method, image processing device and storage medium
CN111091164A (en) * 2020-03-25 2020-05-01 魔视智能科技(上海)有限公司 Semantic segmentation model training method and system based on oem

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP3355270A1 (en) * 2017-01-27 2018-08-01 AGFA Healthcare Multi-class image segmentation method
CN109272115A (en) * 2018-09-05 2019-01-25 宽凳（北京）科技有限公司 A kind of neural network training method and device, equipment, medium
CN109948478A (en) * 2019-03-06 2019-06-28 中国科学院自动化研究所 The face identification method of extensive lack of balance data neural network based, system
CN110443280A (en) * 2019-07-05 2019-11-12 北京达佳互联信息技术有限公司 Training method, device and the storage medium of image detection model

Also Published As

Publication number Publication date
CN111860568A (en) 2020-10-30

Similar Documents

Publication Publication Date Title
CN111860568B (en) Method and device for balanced distribution of data samples and storage medium
WO2019214309A1 (en) Model test method and device
CN109272016B (en) Target detection method, device, terminal equipment and computer readable storage medium
WO2018112783A1 (en) Image recognition method and device
CN106789844B (en) Malicious user identification method and device
CN112200392B (en) Service prediction method and device
CN110837852A (en) Fault diagnosis method and device for rolling mill gearbox and terminal equipment
CN110741387A (en) Face recognition method and device, storage medium and electronic equipment
CN111881706B (en) Living body detection, image classification and model training method, device, equipment and medium
CN110111311B (en) Image quality evaluation method and device
CN114612011A (en) Risk prevention and control decision method and device
CN109190757B (en) Task processing method, device, equipment and computer readable storage medium
CN110348215A (en) Exception object recognition methods, device, electronic equipment and medium
CN113989519A (en) Long-tail target detection method and system
JP2019158684A (en) Inspection system, identification system, and discriminator evaluation device
CN110704614B (en) Information processing method and device for predicting user group type in application
CN115617998A (en) Text classification method and device based on intelligent marketing scene
CN112434717B (en) Model training method and device
CN109436980B (en) Method and system for detecting state of elevator component
CN115248779A (en) Code coverage testing method and device, computer equipment and storage medium
CN110569475A (en) Evaluation method, device, equipment and storage medium for netizen influence
WO2022051974A1 (en) Code inspection method and apparatus applied to embedded platform, device and computer-readable storage medium
CN115168016B (en) Task scheduling method and related device, chip, device and medium
US20230298329A1 (en) Information processing apparatus, information processing method, and storage medium
CN113257329A (en) Memory fault diagnosis method based on machine learning

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