CN114332550A - Model training method, system, storage medium and terminal equipment - Google Patents
Model training method, system, storage medium and terminal equipment Download PDFInfo
- Publication number
- CN114332550A CN114332550A CN202110952438.XA CN202110952438A CN114332550A CN 114332550 A CN114332550 A CN 114332550A CN 202110952438 A CN202110952438 A CN 202110952438A CN 114332550 A CN114332550 A CN 114332550A
- Authority
- CN
- China
- Prior art keywords
- sample
- training
- subset
- type
- object recognition
- 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.)
- Pending
Links
Images
Abstract
The embodiment of the invention discloses a model training method, a model training system, a storage medium and terminal equipment, which are applied to the technical field of information processing. The model training system respectively determines the sample weight value of each training sample in each type of sample subset of the training sample set, respectively selects at least one training sample from each sample subset according to the sample weight value, and combines the at least one training sample respectively selected from the multiple sample subsets into a training subset of a current batch to train the object recognition model. Therefore, the training sample set can be divided into a plurality of types of sample subsets, the probability of the training samples in each sample subset being selected is measured through the sample weight values, the training subsets of the current batch are formed, the probability of the training samples being selected can be balanced through adjusting the sample weight values, and the accuracy of the object recognition model obtained through training for recognizing various types of training samples is improved.
Description
Technical Field
The invention relates to the technical field of information processing, in particular to a model training method, a model training system, a storage medium and terminal equipment.
Background
With the development of machine learning models, computer vision technology is used to perform specific tasks in various scenes, such as unmanned driving, intelligent retail, visual content understanding and the like. At present, the industry is trending to combine deep learning, on one hand, the recognition effect can be improved, and on the other hand, the labor cost can be reduced, wherein the basis of computer vision related tasks based on deep learning is a machine learning model, and a good machine learning model can be directly migrated to other related tasks, so that the upgrading of a series of tasks is realized. Generally, an excellent machine learning model does not depend on an excellent basic network architecture, and a training method and a data sampling method capable of stably improving model performance are also one of the current research hotspots. And the excellent data sampling method and the model training strategy can be migrated to a plurality of downstream tasks, such as image classification, video content understanding and the like.
In particular, for news reading applications, a large number of news contents of various categories are generated by a content producer every day, and thus a problem of long-tailed distribution exists, namely, the number of contents in some common categories is far larger than that in rare categories, so that the number is unbalanced, for example, in a news application, the content related to the television series flowers in the video content occupies 7% of the total number, and the scientifically-relevant content occupies only 1% of the total data. Due to the phenomenon, the performance of the machine learning model based on deep learning can be obviously improved along with the increase of the data volume, so if the data volume of some classes is large, the machine learning model can have better effect on the classes in the training process, but has poorer effect on other classes.
Disclosure of Invention
The embodiment of the invention provides a model training method, a model training system, a storage medium and terminal equipment, which can be used for realizing more accurate training of an object recognition model.
An embodiment of the present invention provides a model training method, including:
obtaining a training sample set, wherein the training sample set comprises a plurality of types of sample subsets;
respectively determining the sample weight value of each training sample in each type of sample subset;
respectively selecting at least one training sample from the sample subsets of the corresponding types according to the sample weight values of the training samples in the sample subsets of each type, and forming the at least one training sample respectively selected from the sample subsets of the multiple types into a training subset of the current batch;
and training an object recognition model according to the training subset of the current batch, wherein the object recognition model is used for recognizing the type of the target object.
Another aspect of an embodiment of the present invention provides a model training system, including:
the device comprises a sample acquisition unit, a data processing unit and a data processing unit, wherein the sample acquisition unit is used for acquiring a training sample set which comprises a plurality of types of sample subsets;
the sample weight unit is used for respectively determining the sample weight value of each training sample in each type of sample subset;
the sample selection unit is used for respectively selecting at least one training sample from the sample subsets of the corresponding types according to the sample weight values of the training samples in the sample subsets of each type, and forming the at least one training sample respectively selected from the sample subsets of the multiple types into a training subset of the current batch;
and the training unit is used for training an object recognition model according to the training subset of the current batch, and the object recognition model is used for recognizing the type of the target object.
In another aspect, an embodiment of the present invention further provides a computer-readable storage medium, which stores a plurality of computer programs, where the computer programs are adapted to be loaded by a processor and execute the model training method according to an embodiment of the present invention.
In another aspect, an embodiment of the present invention further provides a terminal device, including a processor and a memory;
the memory is used for storing a plurality of computer programs, and the computer programs are used for being loaded by the processor and executing the model training method according to the embodiment of the invention; the processor is configured to implement each of the plurality of computer programs.
As can be seen, in the method of this embodiment, the model training system respectively determines the sample weight values of the training samples in each type of sample subset of the training sample set, respectively selects at least one training sample from the sample subsets according to the sample weight values, and combines the at least one training sample selected from the sample subsets into the training subset of the current batch to train the object recognition model. Therefore, the training sample set can be divided into a plurality of types of sample subsets, the probability of the training samples in each sample subset being selected is measured through the sample weight values, the training subsets of the current batch are formed, the probability of the training samples being selected can be balanced through adjusting the sample weight values, and the accuracy of the object recognition model obtained through training for recognizing various types of training samples is improved.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and for those skilled in the art, other drawings can be obtained according to these drawings without creative efforts.
FIG. 1 is a schematic diagram of a model training method according to an embodiment of the present invention;
FIG. 2 is a flow chart of a model training method provided by one embodiment of the present invention;
FIG. 3 is a flow diagram of a method of training an object recognition model in accordance with an embodiment of the present invention;
FIG. 4 is a flow chart of a model training method provided by an embodiment of the present invention;
FIG. 5 is a diagram illustrating the selection of training samples from a subset of samples of a type in an embodiment of the present invention;
FIG. 6 is a diagram of training subsets that are grouped into batches in accordance with an exemplary embodiment of the present invention;
FIG. 7 is a schematic diagram of a distributed system to which a model training method is applied in another embodiment of the present invention;
FIG. 8 is a block diagram illustrating an exemplary block structure according to another embodiment of the present invention;
FIG. 9 is a schematic diagram of a logical structure of a model training system according to an embodiment of the present invention;
fig. 10 is a schematic logical structure diagram of a terminal device according to an embodiment of the present invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, 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 invention.
The terms "first," "second," "third," "fourth," and the like in the description and in the claims, as well as in the drawings, if any, 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 invention described herein are, for example, capable of operation in sequences other than those illustrated or otherwise 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 embodiment of the invention provides a model training method, which is mainly used for training an object recognition model according to a training sample set after certain processing is performed on the training sample set, and specifically, as shown in fig. 1, a model training system can train the object recognition model according to the following steps:
acquiring a training sample set, wherein the training sample set comprises sample subsets of a plurality of types (n types are taken as examples in the figure); respectively determining the sample weight value of each training sample in each type of sample subset; respectively selecting at least one training sample from the sample subsets of the corresponding types according to the sample weight values of the training samples in the sample subsets of each type, and forming the at least one training sample respectively selected from the sample subsets of the multiple types into a training subset of the current batch; and training an object recognition model according to the training subset of the current batch, wherein the object recognition model is used for recognizing the type of the target object.
Therefore, the training sample set can be divided into a plurality of types of sample subsets, the probability of the training samples in each sample subset being selected is measured through the sample weight values, the training subsets of the current batch are formed, the probability of the training samples being selected can be balanced through adjusting the sample weight values, and the accuracy of the object recognition model obtained through training for recognizing various types of training samples is improved.
In practical application, the method of the embodiment can be applied to not only video/image-text classification and identification products, but also training of a series of models such as video/text label identification, high-quality content identification and low-quality content identification.
The object recognition model is a machine learning model based on artificial intelligence, and may specifically be a Multi-task convolutional neural Network (MTCNN), such as an Output Network (ONet) of the MTCNN. Among them, Artificial Intelligence (AI) is a theory, method, technique and application system that simulates, extends and expands human Intelligence using a digital computer or a machine controlled by a digital computer, senses the environment, acquires knowledge and uses the knowledge to obtain the best result. In other words, artificial intelligence is a comprehensive technique of computer science that attempts to understand the essence of intelligence and produce a new intelligent machine that can react in a manner similar to human intelligence. Artificial intelligence is the research of the design principle and the realization method of various intelligent machines, so that the machines have the functions of perception, reasoning and decision making.
The artificial intelligence technology is a comprehensive subject and relates to the field of extensive technology, namely the technology of a hardware level and the technology of a software level. The artificial intelligence infrastructure generally includes technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technologies, operation/interaction systems, mechatronics, and the like. The artificial intelligence software technology mainly comprises a computer vision technology, a voice processing technology, a natural language processing technology, machine learning/deep learning and the like.
Machine Learning (ML) is a multi-domain cross subject, and relates to multiple subjects such as probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and the like. The special research on how a computer simulates or realizes the learning behavior of human beings so as to acquire new knowledge or skills and reorganize the existing knowledge structure to continuously improve the performance of the computer. Machine learning is the core of artificial intelligence, is the fundamental approach for computers to have intelligence, and is applied to all fields of artificial intelligence. Machine learning and deep learning generally include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and teaching learning.
An embodiment of the present invention provides a model training method, which is a method executed by a model training system, and a flowchart is shown in fig. 2, and includes:
It can be understood that, when any machine learning model is trained, parameter values in the machine learning model of any structure need to be learned according to a large number of training samples, the training samples form a training sample set, each training sample includes a sample object and label information of the sample object, where which kind of information the label information of the sample object specifically is needs to be determined according to the function of the machine learning model to be trained, for example, the machine learning model is mainly used for identifying a certain type of object in an image, and the label information of the sample object is mainly position information of the corresponding type of object in the sample object.
In this embodiment, the machine learning model to be trained is mainly an object recognition model, and is used to recognize attribute information of a target object, such as whether any image is a face image or not, or recognize user emotion represented by any piece of voice data, and the like, and the fact is to recognize which type the target object belongs to, where the target object may be multimedia information such as an image, a video, or a text. Accordingly, the training samples in the training sample set also need to include a plurality of types of sample subsets, specifically which types of sample subsets to include, need to be consistent with the types that can be recognized by the object recognition model, for example, if the object recognition model can recognize n types of target objects, the training sample set needs to include corresponding n types of sample subsets.
In order to improve the accuracy of the trained object recognition model for recognizing the target object as any type of object, it is necessary to ensure that the number of training samples in each type of sample subset is sufficient, and the difference between the numbers of training samples in any two types of sample subsets is not large, i.e. balance needs to be achieved between the sample subsets of each type. However, in practical applications, some types of data in the existing data set that can be used as training samples are relatively small, for example, in the data set of news videos, the number of news videos related to drama flowers is relatively large, and the number of science-related news videos is relatively small, so that the obtained sample subsets of each type are unbalanced, and in order to reduce the influence of the sample subset unbalance on the training object recognition model, in the embodiment of the present invention, after the training sample set is obtained, the following steps 102 to 104 need to be performed.
In particular, the model training system may consider only all training samples in a sample subset of one type, and not training samples in subsets of other types, when determining the sample weight values of the training samples in the sample subset.
In a specific case, the model training system may determine a difficult sample and a simple sample in all training samples of a type of sample subset, determine that a sample weight value of the difficult sample is greater than a threshold, and determine that a sample weight value of the simple sample is less than another threshold, so that a probability that the difficult sample is selected to form the training subset is improved, and accuracy of the object recognition model in learning the difficult sample is improved. Here one threshold may be greater than another threshold.
The hard samples are training samples which are easily predicted by the object recognition model to be of wrong types, or training samples with low prediction confidence, and the simple samples are training samples which are easily learned by the object recognition model. Particularly, when difficult samples are determined, the difficult samples can be screened out in a mode of calculating cosine clustering or Euclidean clustering equidistance among training samples.
And 103, respectively selecting at least one training sample from the sample subsets of the corresponding types according to the sample weight values of the training samples in the sample subsets of each type, and combining the at least one training sample respectively selected from the sample subsets of the multiple types into a training subset of the current batch.
Specifically, the model training system needs to separately select at least one training sample from each sample subset and combine the training samples into a batch of training subsets. The number of training samples selected from each sample subset may be the same or different, for example, 3 training samples are selected from one sample subset, 4 training samples are selected from another sample subset, and the like.
When at least one training sample is selected from a sample subset, at least one training sample with the largest sample weight value may be selected from the sample subset, or at least one training sample with a sample weight value larger than a preset value may be selected. At least one training sample in any subset of samples may be selected specifically by:
(1) at least one training sample is selected based on a method of binary search.
Specifically, the sample weight values of all training samples in a type of sample subset are used as vectors, and at least one training sample with the largest sample weight value is obtained through binary search.
The basic idea of binary search is as follows: if the elements in the table are arranged in ascending order, comparing the key word recorded in the middle position of the table with the search key word, and if the two are equal, the search is successful; otherwise, the table is divided into a front sub-table and a rear sub-table by using the middle position record, if the key word of the middle position record is larger than the search key word, the front sub-table is further searched, and if not, the rear sub-table is further searched. The above process is repeated until a record is found that satisfies the condition, such that the lookup is successful, or until a sub-table does not exist, at which point the lookup is unsuccessful. When the method is adopted to select the training sample, the searching speed is high, and the average performance is good.
(2) At least one training sample is selected based on the method of the wheel rotation.
Specifically, the sample weight value of each training sample in a type of sample subset may be corresponding to one unit component of the preset wheel, and then the preset wheel is rotated, and the training sample corresponding to the sample weight value pointed by the pointer in the preset wheel is selected.
The preset wheel disc mainly comprises a plurality of unit components and pointers, and the pointers can circularly point to each unit component in sequence in the process of rotating the wheel disc, wherein the unit components can be in any shapes, such as fan shapes or square shapes.
When the sample weight value of each training sample in a sample subset of one type is corresponding to one unit component of a preset wheel disc, the relation between the number a of the training samples in the sample subset and the number b of the unit components of the wheel disc needs to be considered, and if a is the same as b, the sample weight values can be respectively corresponding to the unit components one by one; if a is smaller than b, a unit components can be selected from b unit components, and the sample weight values are respectively in one-to-one correspondence with the selected unit components.
If a is larger than b, x × b sample weight values are respectively in one-to-one correspondence with b unit components, then a-x × b unit components are selected from the b unit components, and a-x × b sample weight values are respectively in one-to-one correspondence with the selected a-x × b unit components. Wherein x is a natural number greater than or equal to 1.
(3) At least one training sample is selected based on a method of presetting the number of times of being sampled.
Specifically, according to the sample weight value of each training sample in a sample subset, distributing corresponding sampled times for each training sample; selecting a training sample with the largest sample weight value from a sample subset according to the sampled times of each training sample, and updating the sampled times of the selected training sample; in this way, the steps of selecting the training samples and updating the sampled times are executed in a loop, so that at least one training sample corresponding to one sample subset can be obtained.
When distributing corresponding sampled times for each training sample, the sampled times corresponding to the training samples with higher sample weight values are more, and the sampled times corresponding to the training samples with lower sample weight values are less. When the number of times of sampling of the selected training sample is updated, the current number of times of sampling is mainly reduced by 1, and when the number of times of sampling of a certain training sample is 0, the training sample is not selected.
The number of cycles for performing the steps of selecting training samples and updating the sampled number of times in a loop may be determined by the number of cycles preset in the model training system in advance, or may be determined by the number of training samples in the training subset of the current batch, for example, the number of training samples in the training subset of the current batch is a1, and there are a2 types of sample subsets, then a1/a2 training samples may be selected from each sample subset, and for a sample subset, a3 training samples may be selected in each loop process, so that the number of cycles for performing the steps of selecting training samples and updating the sampled number of times in a loop is (a1/a2)/a 3.
And 104, training an object recognition model according to the training subset of the current batch, wherein the object recognition model is used for recognizing the type of the target object.
Specifically, each of the training samples includes a sample object and its label information, and the sample weight value of one training sample is the sample weight value of the sample object, as shown in fig. 3, the model training system may train the object recognition model according to the training subset of the current batch according to the following steps:
in step 201, an object recognition initial model is determined.
It can be understood that, when determining the initial model of object recognition, the model training system mainly determines the initial values of parameters in the multilayer structure and each layer mechanism included in the initial model of object recognition.
Wherein the object recognition initial model may include: the device comprises a feature extraction module and an identification module, wherein the feature extraction module is used for extracting feature information of a sample object in a training sample, the identification module is used for determining the type of the sample object according to the feature information extracted by the feature extraction module, and specifically, the identification module can output probability information that the sample object belongs to a certain type.
The parameters of the object identification initial model refer to fixed parameters used in the calculation process of each layer structure in the object identification initial model, and the parameters do not need to be assigned at any time, such as parameters of parameter scale, network layer number, user vector length and the like.
And step 202, respectively identifying the type of each sample object in the training subset of the current batch through the object identification initial model.
Specifically, the feature extraction module in the object identification initial model extracts feature information of each sample object, and then the identification module determines the type of each sample object according to the feature information extracted by the feature extraction module.
Specifically, the model training system device calculates a loss function associated with the object recognition initial model according to the type of each sample object obtained by the object recognition initial model in step 202, and the label information and the sample weight value of the corresponding sample object in the training subset of the current batch, where the loss function specifically includes: the type of each sample object identified by the object identification initial model, the error between the type of each sample object and the actual type of the corresponding sample object (obtained according to the labeling information of the sample object), and the product of the error and the sample weight value of the corresponding sample object, such as a cross entropy loss function; and then adjusting the parameter values in the object recognition initial model according to the calculated loss function.
It should be noted that the process of training the object recognition model is to reduce the error as much as possible, and the training process is to continuously optimize the parameter values of the parameters in the object recognition initial model determined in the step 201 by a series of mathematical optimization means such as back propagation derivation and gradient descent, and to minimize the calculated value of the loss function. Specifically, when the calculated loss function has a large function value, for example, a function value larger than a preset value, it is necessary to change a parameter value, for example, to reduce a weight value of a neuron connection, so that the calculated loss function has a small function value according to the adjusted parameter value.
In a specific embodiment, the model training system may first calculate the loss function, and then consider the type weight value of each sample object. Specifically, the method comprises the following steps:
the model training system respectively determines a type weight value of each type of sample subset, wherein the type weight value is a weight value corresponding to any type of sample subset, and all training samples in one sample subset correspond to one type weight value. In this way, when the model training system calculates the loss function, the loss function related to the object recognition initial model may be calculated specifically according to the type of each sample object obtained by the object recognition initial model, and the label information of the corresponding sample object in the training subset of the current batch, and the sample weight value and the type weight value thereof. For example, the loss function may include: the type of each sample object identified by the object identification initial model, the error between the type of each sample object and the actual type of the corresponding sample object (obtained according to the labeling information of the sample object), and the product of the sample weight value and the type weight value of the corresponding sample object.
When the type weight value of each type of sample subset is determined separately, this may be achieved by, but is not limited to, the following several ways:
(1) the corresponding class weight value is determined by the number of training samples included in the sample subset of a class.
Specifically, the number of training samples included in a type of sample subset may be directly inverted to obtain an inverse value, and the inverse value is used as a type weight value of the type of sample subset, so that a sample subset with a small number of training samples may have a higher type weight value, and a sample subset with a large number of training samples may have a lower type weight value. Therefore, the importance of the sample subsets with less quantity is improved, and the long tail distribution phenomenon of the training samples is relieved.
(2) The type weight values are determined by a smoothing policy.
Specifically, a ratio of the reciprocal of the number of training samples included in the sample subset of one type to the sum of the reciprocals of the number of training samples in each sample subset is calculated, and a value calculated as a function of the ratio is used as a type weight value of the sample subset of one type. Wherein the number Count of training samples included in the sample subset of one typeiIs ofiCan be expressed by the following formula 1, and the weight value weight of the type of the sample subset of one typeiCan be expressed by the following equation 2:
where M is the number of sample subsets, j represents the sample subset for each type, and i represents the sample subset for one of the types.
The type weight values determined by such a smoothing strategy are such that the type weight values of a subset of samples with a high number of training samples are not too low.
It should be noted that the above steps 202 to 203 are performed by once adjusting the parameter values in the object recognition initial model according to the type of each sample object determined by the object recognition initial model, and in practical applications, the above steps 202 to 203 need to be continuously executed in a loop until the adjustment of the parameter values meets a certain stop condition.
Therefore, after the steps 201 to 203 of the above embodiment are executed, the model training system needs to determine whether the current adjustment on the parameter value meets the preset stop condition, and when the current adjustment on the parameter value meets the preset stop condition, the parameter value obtained by the adjustment in the step 203 is used as the parameter value of the object recognition model obtained by the final training, and the process is ended; if not, identifying the initial model for the object after adjusting the parameter value, and returning to execute the above steps 202 to 203. Wherein the preset stop condition includes but is not limited to any one of the following conditions: the difference value between the current adjusted parameter value and the last adjusted parameter value is smaller than a threshold value, namely the adjusted parameter value reaches convergence; and the adjustment times of the parameter values are equal to the preset times, and the like.
It should be further noted that, because the number of all training samples in the training sample set obtained in the step 101 is large, in the actual training process, the training samples need to be divided into a plurality of batches (batch) of training subsets, there may be overlapped training samples between any two batches of training subsets, then, for each batch of training subsets, the training of the object recognition model may be performed according to the methods of the steps 202 to 203, respectively, and when the training of the object recognition model is performed for all batches of training subsets, a round of training of the object recognition model is completed. Further, the model training system can perform multiple rounds of training on the object recognition model, so that the trained object recognition model is more accurate. When the object recognition model is trained on a batch of training subsets, the above steps 202 and 203 may be executed repeatedly.
In this process, the sample weight values of the sample objects may be continuously updated, and the training samples correspond to the sample objects one to one, that is, the sample weight values of the sample objects are the sample weight values of the training samples, so that the sample weight values of the training samples are continuously updated, which may specifically include, but not limited to, the following two ways:
(1) after training of the object recognition model is performed on the training subsets of the current batch, the sample weight values of the corresponding sample objects can be updated according to the types of the sample objects obtained by the object recognition initial model and the labeling information of the corresponding sample objects in the training subsets of the current batch, so that the updated sample weight values of the sample objects in the training subsets of the current batch are obtained; then, for the updated sample weight values, the steps 103 and 104 are executed again, that is, the steps of forming the training subsets and the training object recognition models of another current batch are executed, so that the steps 103 and 104 and the step of updating the sample weight values can be executed in a loop until the training subsets of all batches are processed.
When the sample weight value of a sample object is updated, the difference between the type of the sample object obtained by the object identification initial model and the labeling information of the corresponding sample object in the training subset of the current batch is updated, if the difference is greater than a threshold, the sample weight value of the sample object can be increased, and if the difference is less than another threshold, the sample weight value of the sample object can be decreased. Therefore, when the type prediction of a certain sample object by the object training initial model is inaccurate, the sample weight value of the sample object needs to be increased, so that the consideration proportion of the sample object is increased when the parameter value in the object training initial model is adjusted, and the training of the object training model is more accurate.
(2) After one round of training of the object recognition model is performed on all batches of training subsets corresponding to the training sample set, updating the sample weight values of all training samples in the training sample set to obtain updated sample weight values of all training samples; then, the next round of training for the object recognition model is started according to the updated sample weight values, that is, the steps of forming the training subset and the training object recognition model of the current batch in the above steps 103 and 104 are executed again.
In this case, the updating of the sample weight value of a training sample is specifically to update the sample weight value of a sample object in the training sample, and specifically, a difference between a type of the sample object obtained by the object recognition initial model and labeling information of a corresponding sample object in a batch of training subsets may be first counted in a round of training and obtained when the object recognition model is respectively trained based on the sample object in at least one batch of training subsets, and then an average value of the difference obtained for one batch of training subsets is calculated, if the average value is greater than a threshold, the sample weight value of the sample object may be increased, and if the average value is less than another threshold, the sample weight value of the sample object may be decreased.
As can be seen, in the method of this embodiment, the model training system respectively determines the sample weight values of the training samples in each type of sample subset of the training sample set, respectively selects at least one training sample from the sample subsets according to the sample weight values, and combines the at least one training sample selected from the sample subsets into the training subset of the current batch to train the object recognition model. Therefore, the training sample set can be divided into a plurality of types of sample subsets, the probability of the training samples in each sample subset being selected is measured through the sample weight values, the training subsets of the current batch are formed, the probability of the training samples being selected can be balanced through adjusting the sample weight values, and the accuracy of the object recognition model obtained through training for recognizing various types of training samples is improved.
The model training method of the present invention is described below with a specific application example, as shown in fig. 4, the model training method of the present embodiment may include the following steps:
Any type of sample subset may be denoted as D { (x)i,yi) J ∈ (1, 2.. times, N), where N is the number of training samples in the sample subset, where x isiAs a sample object, yiLabel information for the sample object.
At step 302, type weight values for sample subsets of respective types are determined.
Specifically, the number of training samples included in each type of sample subset is counted first, and a type weight value of the corresponding type of sample subset is determined according to the counted number, where a method for specifically determining the type weight value is described in the above embodiment and is not described herein again.
It should be noted that, when the object recognition model predicts different types of training samples, different expressions are available, some training samples can quickly learn corresponding characteristics, and the training samples are simple samples easy to learn for the object recognition model; when some training samples are predicted by the object recognition model, the training samples are predicted to be of an incorrect type, or the prediction confidence is low, the training samples are under-fitted to the object recognition model, so that important learning needs to be performed in a subsequent learning stage, and the training samples are considered to be difficult samples.
Based on the situation, when determining the sample weight value of each training sample in a sample subset, the model training system may determine the difficult samples in the training samples included in the sample subset, and assign higher sample weight values to the difficult samples, thereby increasing the probability that the difficult samples are selected in the training subset, and also ensuring the probability that the simple samples are selected, that is, increasing the accuracy of the object recognition model in recognizing the difficult samples and the simple samples.
For example, as shown in FIG. 5, all training sample items included from a sample subset of a type0,iI 1, 2.. k.. n.one training sample item with the largest sample weight value is selected from n0,kPut into a batch of training subsets.
Specifically, in the training process of the object recognition model, due to the limitation of hardware conditions, all training samples in the entire training sample set need to be divided into multiple batches of training subsets, and in order to ensure that the object recognition model has the same importance degree on all types of training samples, when a batch of training subsets can be formed, the training subsets of each type can be numbered i ∈ (0,1, 2.. and M), and at least one training sample selected from each training subset in the step 304 is sequentially filled in the training subsets of a batch according to the number, so that each type of training sample can be guaranteed to have the same probability to be learned by the object recognition model.
For example, as shown in fig. 6, the numbers of the sample subsets of the respective types are 0,1,2 …, and 9, and the numbers are sequentially countedAt least one training sample obtained from each sample subset needs to be filled into each batch of training subsets, so that the training sample item in any batch of training subsetsi,jRepresenting training samples selected from one sample subset i and filled into the training subset of batch j.
It should be noted that, in this embodiment, the model training system may obtain one training subset corresponding to the training sample set as the training subset of the current batch at a time, perform the following steps of training the object recognition model, and after adjusting the sample weight value, obtain another training subset corresponding to the training sample set as the training subset of the current batch to train the object recognition model, so that the operation is performed in a loop, so that all the training subsets corresponding to the training sample set are processed.
Or, in other embodiments, the model training system may obtain training subsets of all batches corresponding to the training sample set at one time, sequentially select one training subset as the training subset of the current batch, respectively perform the following training steps of the object recognition model, and then adjust the sample weight values of all the training samples after completing one round of training.
It should be noted that, in this embodiment, each time the object recognition model is trained for a batch of training subsets, the sample weight values of the sample objects in the batch of training subsets are adjusted, and in another specific embodiment, after one round of training of the object recognition model is completed, that is, after the object recognition model is trained for all batches of training subsets corresponding to the training set, the sample weight values of all training samples are adjusted.
Therefore, the method in the embodiment is mainly used for training the object recognition model based on a dynamic sampling method of a training sample, can be directly applied to training of various machine learning models, and can relieve the problem of long tail distribution existing in the existing data set; meanwhile, in the iterative process of each round of training, the sample weight value of each training sample is dynamically updated according to the newly trained object recognition model to realize dynamic sampling of the training samples, and the training samples with poor prediction effect of the object recognition model can be subjected to targeted training, so that the object recognition model can be used for key learning of some samples such as difficult samples, the performance of the object recognition model can be remarkably improved, and the general performance is improved by about 1-2%.
The model training method in the present invention is described below with another specific application example, the model training system in the embodiment of the present invention is mainly a distributed system 100, and the distributed system may include a client 300 and a plurality of nodes 200 (any form of computing devices in an access network, such as servers and user terminals), and the client 300 and the nodes 200 are connected by way of network communication.
Taking a distributed system as an example of a blockchain system, referring To fig. 7, which is an optional structural schematic diagram of the distributed system 100 applied To the blockchain system provided in the embodiment of the present invention, the system is formed by a plurality of nodes 200 (computing devices in any form in an access network, such as servers and user terminals) and clients 300, a Peer-To-Peer (P2P, Peer To Peer) network is formed between the nodes, and the P2P Protocol is an application layer Protocol operating on a Transmission Control Protocol (TCP). In a distributed system, any machine, such as a server or a terminal, can join to become a node, and the node comprises a hardware layer, a middle layer, an operating system layer and an application layer.
Referring to the functions of each node in the blockchain system shown in fig. 7, the functions involved include:
1) routing, a basic function that a node has, is used to support communication between nodes.
Besides the routing function, the node may also have the following functions:
2) the application is used for being deployed in a block chain, realizing specific services according to actual service requirements, recording data related to the realization function to form recording data, carrying a digital signature in the recording data to represent a source of task data, and sending the recording data to other nodes in the block chain system, so that the other nodes add the recording data to a temporary block when the source and integrity of the recording data are verified successfully.
For example, the services implemented by the application include: code for implementing a model training function, the model training function comprising:
obtaining a training sample set, wherein the training sample set comprises a plurality of types of sample subsets; respectively determining the sample weight value of each training sample in each type of sample subset; respectively selecting at least one training sample from the sample subsets of the corresponding types according to the sample weight values of the training samples in the sample subsets of each type, and forming the at least one training sample respectively selected from the sample subsets of the multiple types into a training subset of the current batch; and training an object recognition model according to the training subset of the current batch, wherein the object recognition model is used for recognizing the type of the target object.
3) And the Block chain comprises a series of blocks (blocks) which are mutually connected according to the generated chronological order, new blocks cannot be removed once being added into the Block chain, and recorded data submitted by nodes in the Block chain system are recorded in the blocks.
Referring to fig. 8, an optional schematic diagram of a Block Structure (Block Structure) provided in the embodiment of the present invention is shown, where each Block includes a hash value of a transaction record stored in the Block (hash value of the Block) and a hash value of a previous Block, and the blocks are connected by the hash values to form a Block chain. The block may include information such as a time stamp at the time of block generation. A block chain (Blockchain), which is essentially a decentralized database, is a string of data blocks associated by using cryptography, and each data block contains related information for verifying the validity (anti-counterfeiting) of the information and generating a next block.
An embodiment of the present invention further provides a model training system, a schematic structural diagram of which is shown in fig. 9, and the model training system specifically includes:
the apparatus includes a sample acquiring unit 10 configured to acquire a training sample set, where the training sample set includes a plurality of types of sample subsets.
A sample weighting unit 11, configured to determine a sample weighting value of each training sample in each type of sample subset acquired by the sample acquiring unit 10, respectively.
The sample weighting unit 11 is specifically configured to determine hard samples and simple samples in all training samples of a type of sample subset; determining that a sample weight value of the difficult sample is greater than a threshold, determining that a sample weight value of the simple sample is less than another threshold.
The sample weighting unit 11 is specifically configured to correspond the sample weighting values of the training samples in a sample subset to a unit component of the preset wheel disc, and rotate the preset wheel disc; selecting a training sample corresponding to the sample weight value pointed by the pointer in the preset wheel disc; or, taking the sample weight values of all training samples in a sample subset as vectors, and obtaining at least one training sample with the maximum sample weight value by a binary search method.
Or, the sample weighting unit 11 is specifically configured to allocate, according to a sample weight value of each training sample in one sample subset, a corresponding number of times of sampling to each training sample; selecting the training sample with the largest sample weight value from the sample subset according to the sampling times of the training samples; updating the sampled times of the selected training samples; and circularly executing the steps of selecting the training samples and updating the sampled times to obtain at least one training sample corresponding to the sample subset.
A sample selecting unit 12, configured to select at least one training sample from the sample subsets of the corresponding types according to the sample weight values of the training samples in the sample subsets of each type determined by the sample weight unit 11, and combine the at least one training sample selected from the sample subsets of the multiple types into a training subset of the current batch.
A training unit 13, configured to train an object recognition model according to the training subset of the current batch obtained by the sample selection unit 12, where the object recognition model is used to recognize the type of the target object.
The training unit 13 is specifically configured to determine an object recognition initial model; respectively identifying the type of each sample object in the training subset of the current batch through the object identification initial model; and adjusting the object recognition initial model according to the type of each sample object obtained by the object recognition initial model, and the labeling information and the sample weight value of the corresponding sample object in the training subset of the current batch to obtain a final object recognition model.
When the training unit 13 adjusts the object recognition initial model according to the type of each sample object obtained by the object recognition initial model, and the label information and sample weight value of the corresponding sample object in the training subset of the current batch, the training unit is specifically configured to calculate a loss function related to the object recognition initial model according to the type of each sample object obtained by the object recognition initial model, and the label information and sample weight value of the corresponding sample object in the training subset of the current batch; and adjusting parameter values in the object identification initial model according to the loss function.
Further, the model training system of the present embodiment further includes: a type weighting unit 14 and a weight adjusting unit 15, wherein:
a type weight unit 14 for determining a type weight value of each type of sample subset, respectively; the training unit 13 is specifically configured to calculate a loss function related to the object identification initial model according to the type of each sample object obtained by the object identification initial model, the label information of the corresponding sample object in the training subset of the current batch, the sample weight value of the corresponding sample object, and the type weight value determined by the type weight unit 14.
The type weighting unit 14 is specifically configured to reciprocal the number of training samples included in a type of sample subset to obtain a reciprocal value, and use the reciprocal value as a type weighting value of the type of sample subset; or, calculating a ratio of the reciprocal of the number of training samples included in the sample subset of one type to the sum of the reciprocals of the number of training samples in the sample subset of each type, and calculating a value as a function of the ratio as a type weight value of the sample subset of one type.
The weight adjusting unit 15 is configured to update the sample weight values of the corresponding sample objects according to the types of the sample objects obtained by the object identification initial model and the label information of the corresponding sample objects in the training subset of the current batch, so as to obtain updated sample weight values of the sample objects; and for the updated sample weight values, informing the sample selection unit 12 to execute the training subsets forming the current batch and informing the training unit 13 to execute the steps of training the object recognition model.
Or, the weight adjusting unit 15 is configured to update the sample weight values of all training samples in the training sample set after performing a round of training of the object recognition model on all batches of training subsets corresponding to the training sample set, so as to obtain updated sample weight values of all training samples; and for the updated sample weight values, informing the sample selection unit 12 to execute the training subsets forming the current batch and informing the training unit 13 to execute the steps of training the object recognition model.
As can be seen, in the model training system of this embodiment, the sample weighting unit 11 determines the sample weight value of each training sample in each type of sample subset of the training sample set, the sample selecting unit 12 selects at least one training sample from each sample subset according to the sample weight value, and the training unit 13 combines the at least one training sample selected from the plurality of sample subsets into the training subset of the current batch to train the object recognition model. Therefore, the training sample set can be divided into a plurality of types of sample subsets, the probability of the training samples in each sample subset being selected is measured through the sample weight values, the training subsets of the current batch are formed, the probability of the training samples being selected can be balanced through adjusting the sample weight values, and the accuracy of the object recognition model obtained through training for recognizing various types of training samples is improved.
The present invention further provides a terminal device, a schematic structural diagram of which is shown in fig. 10, where the terminal device may generate a relatively large difference due to different configurations or performances, and may include one or more Central Processing Units (CPUs) 20 (e.g., one or more processors) and a memory 21, and one or more storage media 22 (e.g., one or more mass storage devices) storing the application programs 221 or the data 222. Wherein the memory 21 and the storage medium 22 may be a transient storage or a persistent storage. The program stored in the storage medium 22 may include one or more modules (not shown), each of which may include a series of instruction operations for the terminal device. Still further, the central processor 20 may be arranged to communicate with the storage medium 22, and to execute a series of instruction operations in the storage medium 22 on the terminal device.
Specifically, the application program 221 stored in the storage medium 22 includes an application program for model training, and the program may include the sample obtaining unit 10, the sample weighting unit 11, the sample selecting unit 12, the training unit 13, the type weighting unit 14, and the weight adjusting unit 15 in the model training system, which will not be described herein again. Further, the central processor 20 may be configured to communicate with the storage medium 22, and execute a series of operations corresponding to the application program of the model training stored in the storage medium 22 on the terminal device.
The terminal equipment may also include one or more power supplies 23, one or more wired or wireless network interfaces 24, one or more input-output interfaces 25, and/or one or more operating systems 223, such as Windows Server, Mac OS XTM, UnixTM, LinuxTM, FreeBSDTM, and the like.
The steps performed by the model training system in the above method embodiment may be based on the structure of the terminal device shown in fig. 10.
In another aspect, an embodiment of the present invention further provides a computer-readable storage medium, which stores a plurality of computer programs, the computer programs being adapted to be loaded by a processor and to perform a model training method as performed by the above model training system.
In another aspect, an embodiment of the present invention further provides a terminal device, including a processor and a memory;
the memory is used for storing a plurality of computer programs which are used for being loaded by the processor and executing the model training method executed by the model training system; the processor is configured to implement each of the plurality of computer programs.
Further, according to an aspect of the application, a computer program product or a computer program is provided, comprising computer instructions, which are stored in a computer readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions to cause the computer device to perform the model training method provided in the various alternative implementations described above.
Those skilled in the art will appreciate that all or part of the steps in the methods of the above embodiments may be implemented by associated hardware instructed by a program, which may be stored in a computer-readable storage medium, and the storage medium may include: read Only Memory (ROM), Random Access Memory (RAM), magnetic or optical disks, and the like.
The model training method, system, storage medium and terminal device provided by the embodiments of the present invention are described in detail above, and a specific example is applied in the present disclosure to explain the principle and the implementation of the present invention, and the description of the above embodiments is only used to help understanding the method and the core idea of the present invention; meanwhile, for a person skilled in the art, according to the idea of the present invention, there may be variations in the specific embodiments and the application scope, and in summary, the content of the present specification should not be construed as a limitation to the present invention.
Claims (13)
1. A method of model training, comprising:
obtaining a training sample set, wherein the training sample set comprises a plurality of types of sample subsets;
respectively determining the sample weight value of each training sample in each type of sample subset;
respectively selecting at least one training sample from the sample subsets of the corresponding types according to the sample weight values of the training samples in the sample subsets of each type, and forming the at least one training sample respectively selected from the sample subsets of the multiple types into a training subset of the current batch;
and training an object recognition model according to the training subset of the current batch, wherein the object recognition model is used for recognizing the type of the target object.
2. The method of claim 1, wherein the determining the sample weight values of the training samples in each type of sample subset respectively comprises:
determining hard samples and simple samples in all training samples of a type of sample subset;
determining that a sample weight value of the difficult sample is greater than a threshold, determining that a sample weight value of the simple sample is less than another threshold.
3. The method according to claim 1, wherein the selecting at least one training sample from the sample subsets of the corresponding types according to the sample weight values of the training samples in the sample subsets of each type respectively comprises:
respectively corresponding the sample weight value of each training sample in a sample subset to a unit component of a preset wheel disc, and rotating the preset wheel disc; selecting a training sample corresponding to the sample weight value pointed by the pointer in the preset wheel disc;
or, taking the sample weight values of all training samples in a sample subset as vectors, and obtaining at least one training sample with the maximum sample weight value by a binary search method.
4. The method according to claim 1, wherein the selecting at least one training sample from the sample subsets of the corresponding types according to the sample weight values of the training samples in the sample subsets of each type respectively comprises:
distributing corresponding sampled times for each training sample according to the sample weight value of each training sample in a sample subset;
selecting the training sample with the largest sample weight value from the sample subset according to the sampling times of the training samples;
updating the sampled times of the selected training samples;
and circularly executing the steps of selecting the training samples and updating the sampled times to obtain at least one training sample corresponding to the sample subset.
5. The method according to any one of claims 1 to 4, wherein any training sample includes a sample object and its label information, and the training of the object recognition model according to the training subset of the current batch specifically includes:
determining an object recognition initial model;
respectively identifying the type of each sample object in the training subset of the current batch through the object identification initial model;
and adjusting the object recognition initial model according to the type of each sample object obtained by the object recognition initial model, and the labeling information and the sample weight value of the corresponding sample object in the training subset of the current batch to obtain a final object recognition model.
6. The method of claim 5, wherein the adjusting the object recognition initial model according to the type of each sample object obtained by the object recognition initial model, and the label information and the sample weight value of the corresponding sample object in the training subset of the current batch comprises:
calculating a loss function related to the object recognition initial model according to the type of each sample object obtained by the object recognition initial model, and the labeling information and the sample weight value of the corresponding sample object in the training subset of the current batch;
and adjusting parameter values in the object identification initial model according to the loss function.
7. The method of claim 6, wherein prior to training the object recognition model based on the training subset of the current batch, further comprising:
respectively determining a type weight value of each type of sample subset;
calculating a loss function related to the object recognition initial model according to the types of the sample objects obtained according to the object recognition initial model, and the labeling information and the sample weight values of the corresponding sample objects in the training subset of the current batch, specifically including:
and calculating a loss function related to the object recognition initial model according to the type of each sample object obtained by the object recognition initial model, the labeling information of the corresponding sample object in the training subset of the current batch, the sample weight value and the type weight value of the corresponding sample object.
8. The method of claim 7, wherein the determining the type weight value for each type of sample subset comprises:
counting the number of training samples included in a type of sample subset to obtain a reciprocal value, and taking the reciprocal value as a type weight value of the type of sample subset;
or, calculating a ratio of the reciprocal of the number of training samples included in the sample subset of one type to the sum of the reciprocals of the number of training samples in the sample subset of each type, and calculating a value as a function of the ratio as a type weight value of the sample subset of one type.
9. The method of claim 7, wherein the method further comprises:
updating the sample weight values of the corresponding sample objects according to the types of the sample objects obtained by the object recognition initial model and the labeling information of the corresponding sample objects in the training subset of the current batch to obtain the updated sample weight values of the sample objects;
and returning to execute the steps of forming the training subsets and the training object recognition models of the current batch according to the updated sample weight values.
10. The method of claim 7, wherein the method further comprises:
after one round of training of the object recognition model is performed on all batches of training subsets corresponding to the training sample set, updating the sample weight values of all training samples in the training sample set to obtain updated sample weight values of all training samples;
and returning to execute the steps of forming the training subsets and the training object recognition models of the current batch according to the updated sample weight values.
11. A model training system, comprising:
the device comprises a sample acquisition unit, a data processing unit and a data processing unit, wherein the sample acquisition unit is used for acquiring a training sample set which comprises a plurality of types of sample subsets;
the sample weight unit is used for respectively determining the sample weight value of each training sample in each type of sample subset;
the sample selection unit is used for respectively selecting at least one training sample from the sample subsets of the corresponding types according to the sample weight values of the training samples in the sample subsets of each type, and forming the at least one training sample respectively selected from the sample subsets of the multiple types into a training subset of the current batch;
and the training unit is used for training an object recognition model according to the training subset of the current batch, and the object recognition model is used for recognizing the type of the target object.
12. A computer-readable storage medium, characterized in that the computer-readable storage medium stores a plurality of computer programs adapted to be loaded by a processor and to perform the model training method according to any one of claims 1 to 10.
13. A terminal device comprising a processor and a memory;
the memory for storing a plurality of computer programs for loading by the processor and executing the model training method of any one of claims 1 to 10; the processor is configured to implement each of the plurality of computer programs.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110952438.XA CN114332550A (en) | 2021-08-19 | 2021-08-19 | Model training method, system, storage medium and terminal equipment |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110952438.XA CN114332550A (en) | 2021-08-19 | 2021-08-19 | Model training method, system, storage medium and terminal equipment |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114332550A true CN114332550A (en) | 2022-04-12 |
Family
ID=81044274
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110952438.XA Pending CN114332550A (en) | 2021-08-19 | 2021-08-19 | Model training method, system, storage medium and terminal equipment |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114332550A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114612750A (en) * | 2022-05-09 | 2022-06-10 | 杭州海康威视数字技术股份有限公司 | Target identification method and device for adaptive learning rate collaborative optimization and electronic equipment |
| CN114785526A (en) * | 2022-06-16 | 2022-07-22 | 德德市界(深圳)科技有限公司 | Multi-user multi-batch weight distribution calculation and storage processing system based on block chain |
-
2021
- 2021-08-19 CN CN202110952438.XA patent/CN114332550A/en active Pending
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114612750A (en) * | 2022-05-09 | 2022-06-10 | 杭州海康威视数字技术股份有限公司 | Target identification method and device for adaptive learning rate collaborative optimization and electronic equipment |
| CN114785526A (en) * | 2022-06-16 | 2022-07-22 | 德德市界(深圳)科技有限公司 | Multi-user multi-batch weight distribution calculation and storage processing system based on block chain |
| CN114785526B (en) * | 2022-06-16 | 2022-09-02 | 德德市界(深圳)科技有限公司 | Multi-user multi-batch weight distribution calculation and storage processing system based on block chain |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US11349680B2 (en) | Method and apparatus for pushing information based on artificial intelligence | |
| CN108694225A (en) | A kind of image search method, the generation method of feature vector, device and electronic equipment | |
| CN111241345A (en) | Video retrieval method and device, electronic equipment and storage medium | |
| CN113850281B (en) | MEANSHIFT optimization-based data processing method and device | |
| CN113064879A (en) | Database parameter adjusting method and device and computer readable storage medium | |
| CN110909182A (en) | Multimedia resource searching method and device, computer equipment and storage medium | |
| CN113868366B (en) | Streaming data-oriented online cross-modal retrieval method and system | |
| CN113158554B (en) | Model optimization method and device, computer equipment and storage medium | |
| CN114332550A (en) | Model training method, system, storage medium and terminal equipment | |
| CN111382555A (en) | Data processing method, medium, device and computing equipment | |
| CN111382190A (en) | Object recommendation method and device based on intelligence and storage medium | |
| CN113298197B (en) | Data clustering method, device, equipment and readable storage medium | |
| CN113641797A (en) | Data processing method, device, equipment, storage medium and computer program product | |
| CN115062709A (en) | Model optimization method, device, equipment, storage medium and program product | |
| CN112200296A (en) | Network model quantification method and device, storage medium and electronic equipment | |
| CN112800253B (en) | Data clustering method, related device and storage medium | |
| CN111222399A (en) | Method and device for identifying object identification information in image and storage medium | |
| CN114490923A (en) | Training method, device and equipment for similar text matching model and storage medium | |
| CN111767953B (en) | Method and apparatus for training an article coding model | |
| CN113128526A (en) | Image recognition method and device, electronic equipment and computer-readable storage medium | |
| CN114449342B (en) | Video recommendation method, device, computer readable storage medium and computer equipment | |
| CN115600017A (en) | Feature coding model training method and device and media object recommendation method and device | |
| CN111459990B (en) | Object processing method, system, computer readable storage medium and computer device | |
| CN111935259B (en) | Method and device for determining target account set, storage medium and electronic equipment | |
| CN114564516A (en) | Business object classification method, device, equipment and storage medium |
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 |