CN112016633A

CN112016633A - Model training method and device, electronic equipment and storage medium

Info

Publication number: CN112016633A
Application number: CN202011027593.2A
Authority: CN
Inventors: 崔程; 杨敏; 魏凯
Original assignee: Beijing Baidu Netcom Science and Technology Co Ltd
Current assignee: Beijing Baidu Netcom Science and Technology Co Ltd
Priority date: 2020-09-25
Filing date: 2020-09-25
Publication date: 2020-12-01
Anticipated expiration: 2040-09-25
Also published as: CN112016633B

Abstract

The application discloses a model training method, a device, electronic equipment and a storage medium, which relate to the field of artificial intelligence, in particular to a computer vision and deep learning technology, can be used for image recognition, and comprise the following steps: carrying out feature extraction training on the target data set to obtain a feature extraction model; the data in the target data set is label data; performing feature extraction on the label-free data set through the feature extraction model to obtain label-free data features; training a pre-training model according to the label-free data characteristics; data cutting is carried out on each non-label data in the non-label data set, and a cut non-label data set is obtained; performing optimization training on the pre-training model according to the cut label-free data set to obtain a target pre-training model; and training the target pre-training model according to the target data set to obtain a target training model. The method and the device for recognizing the model can improve the training efficiency and recognition accuracy of the model in the specific data set.

Description

Model training method and device, electronic equipment and storage medium

Technical Field

The application relates to the technical field of data processing, in particular to a deep learning technology in the technical field of artificial intelligence.

Background

Machine Learning (ML) is the core of artificial intelligence, and is the fundamental approach to making computers intelligent, and its application is spread over various fields of artificial intelligence. The machine learning algorithm is an algorithm for automatically analyzing and obtaining rules from data and predicting unknown data by using the rules. Typically, deep learning is one of the fields in machine learning, and the process of deep learning can be summarized as a process of determining a model-training a model-using a model. Wherein the process of training the model plays a decisive role in the accuracy of the model.

Disclosure of Invention

The embodiment of the application provides a model training method and device, electronic equipment and a storage medium, so as to improve the training efficiency and the recognition accuracy of a model in a specific data set.

In a first aspect, an embodiment of the present application provides a model training method, including:

carrying out feature extraction training on the target data set to obtain a feature extraction model; the data in the target data set is label data;

performing feature extraction on the label-free data set through the feature extraction model to obtain label-free data features;

training a pre-training model according to the label-free data characteristics;

data cutting is carried out on each non-label data in the non-label data set, and a cut non-label data set is obtained;

performing optimization training on the pre-training model according to the cut label-free data set to obtain a target pre-training model;

and training the target pre-training model according to the target data set to obtain a target training model.

In a second aspect, an embodiment of the present application provides a model training apparatus, including:

the characteristic extraction model acquisition module is used for carrying out characteristic extraction training on the target data set to obtain a characteristic extraction model; the data in the target data set is label data;

the non-tag data feature acquisition module is used for extracting features of the non-tag data set through the feature extraction model to obtain non-tag data features;

the pre-training model training module is used for training a pre-training model according to the label-free data characteristics;

the data cutting module is used for cutting data of each non-label data in the non-label data set to obtain a cut non-label data set;

the target pre-training model acquisition module is used for carrying out optimization training on the pre-training model according to the cut label-free data set to obtain a target pre-training model;

and the target training model acquisition module is used for training the target pre-training model according to the target data set to obtain a target training model.

In a third aspect, an embodiment of the present application provides an electronic device, including:

at least one processor; and

a memory communicatively coupled to the at least one processor; wherein,

the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the model training method provided in the embodiments of the first aspect.

In a fourth aspect, the present application further provides a non-transitory computer-readable storage medium storing computer instructions for causing a computer to execute the model training method provided in the first aspect.

The method and the device for the label-free data feature training obtain the feature extraction model by performing feature extraction training on the target data set comprising the label data, perform feature extraction on the label-free data set through the feature extraction model to obtain the label-free data feature, and then train the pre-training model according to the label-free data feature. And after the pre-training model is obtained, data cutting is carried out on each non-label data in the non-label data set, and the pre-training model is optimally trained according to the obtained cut non-label data set to obtain a target pre-training model. And finally, training the target pre-training model according to the target data set to obtain the target training model, solving the problem of low training timeliness caused by high data labeling cost in the existing model training method, and improving the recognition precision of the model while improving the training efficiency.

It should be understood that the statements in this section do not necessarily identify key or critical features of the embodiments of the present disclosure, nor do they limit the scope of the present disclosure. Other features of the present disclosure will become apparent from the following description.

Drawings

The drawings are included to provide a better understanding of the present solution and are not intended to limit the present application. Wherein:

FIG. 1 is a flow chart of a model training method provided by an embodiment of the present application;

FIG. 2 is a flow chart of a model training method provided by an embodiment of the present application;

FIG. 3 is a block diagram of a model training apparatus according to an embodiment of the present disclosure;

fig. 4 is a schematic structural diagram of an electronic device for implementing the model training method according to the embodiment of the present application.

Detailed Description

The following description of the exemplary embodiments of the present application, taken in conjunction with the accompanying drawings, includes various details of the embodiments of the application for the understanding of the same, which are to be considered exemplary only. Accordingly, those of ordinary skill in the art will recognize that various changes and modifications of the embodiments described herein can be made without departing from the scope and spirit of the present application. Also, descriptions of well-known functions and constructions are omitted in the following description for clarity and conciseness.

In the deep learning technology in the field of machine learning, the accuracy of a model has a significant influence on the application effect of the model. Currently, sample data for training a model includes two types, labeled data and unlabeled data. The model accuracy can be improved well by increasing the data volume of the tag data, however, the data labeling cost is higher due to the mode, and a large amount of tag-free data cannot be well utilized.

The current methods for improving the accuracy of the model by using label-free data are generally divided into two methods, one is an unsupervised learning method, namely, the characteristics of a large amount of sample data are learned by using a similar contrast loss function, and the learned model generally has better data characteristics and can be used for performing transfer learning on a specific data set. The second method is a semi-supervised learning method, namely, the model is trained by the non-label data and the label data together, the characteristics of the non-label data are gradually mined by the characteristics of the supervised learning, and the characteristics learned by the model are further expanded.

In the method for improving the model precision by using the non-tag data, the amount of data required by the non-supervised learning method during learning of the non-tag data is extremely huge and can even reach hundred million level. The huge data volume causes serious bottleneck to the training, the training duration is seriously prolonged, the convergence is difficult in a certain time, and the training timeliness is low. The semi-supervised learning method has strict requirements on training sample data, and requires that the unlabeled data belong to the label range of the target data set as much as possible. The target data set is also the data set of the model which needs to be automatically identified in the process. In actual training, most of the non-label data are difficult to be ensured to belong to the label range of the label data, and before formal training, the non-label data are often sorted and selected at a certain time cost, so that the training timeliness is indirectly reduced.

In an example, fig. 1 is a flowchart of a model training method provided in an embodiment of the present application, which may be applicable to a case where a model is trained quickly to improve model recognition accuracy, and the method may be performed by a model training apparatus, which may be implemented by software and/or hardware, and may be generally integrated in an electronic device. The electronic device may be a computer device or the like. Accordingly, as shown in fig. 1, the method comprises the following operations:

s110, performing feature extraction training on the target data set to obtain a feature extraction model; the data in the target data set is tag data.

The target data set is also the data set of the model which needs to be automatically identified in the process. The data in the target data set may be tag data, and the type of the tag data may be any type, such as image data, audio data, or character data, and the embodiment of the present application does not limit the specific data type of the tag data. The feature extraction model may be a model obtained by performing feature extraction training on a target data set, and is mainly used for repeatedly extracting data features from the target data set.

In this embodiment of the application, before training a target training model for identifying a target data set, feature extraction training may be performed on the target data set to obtain a feature extraction model. Optionally, the feature extraction model may be any type of CNN (Convolutional Neural Networks), and the loss function of the feature extraction model may be arcregion loss or triplet loss, as long as the feature extraction can be performed on the target data set, and the specific model type of the feature extraction model and the corresponding loss function type are not limited in the embodiment of the present application.

And S120, performing feature extraction on the label-free data set through the feature extraction model to obtain the label-free data features.

Wherein the non-tag data set may be a data set composed of non-tag data of the same type as the tag data in the target data set. Illustratively, when the data type of the tag data is image data, the non-tag data in the non-tag data set is also image data. When the data type of the tag data is character data, the non-tag data in the non-tag data set is also character data. However, it should be noted that the tag type of the non-tag data does not need to be within the tag range of the target data set, and the non-tag data does not need to be supported by a huge amount of data. The unlabeled data set may be obtained in various manners, such as downloading related data through the internet or collecting data through other manners, which is not limited in this embodiment of the application.

Correspondingly, after the feature extraction model is obtained, feature extraction can be performed on each non-tag data in the non-tag data set by using the feature extraction model to obtain the non-tag data features. Because the label type of the non-label data is not required to belong to the label range of the target data set, the non-label data set does not need to be sorted and selected, the screening time of the training data is avoided, and the training efficiency can be effectively improved.

And S130, training a pre-training model according to the label-free data characteristics.

The pre-training model can be understood as an intermediate training model before the target pre-training model, and can also be understood as a preliminary training model corresponding to the target pre-training model. The pre-training model and the pre-training model have the same partial weight value.

After the non-label data features of the non-label data set are extracted by using the feature extraction model, feature training can be performed by using the non-label data features to obtain a pre-training model. The pre-training model has an initial recognition function, but the recognition accuracy needs to be further improved. That is, the embodiment of the application can perform model training by using the unlabeled data as sample data, thereby effectively reducing the data labeling cost.

S140, data cutting is carried out on each non-label data in the non-label data set, and a cut non-label data set is obtained.

The clipping non-label data set may be a data set composed of clipped non-label data and the clipped non-label data.

S150, carrying out optimization training on the pre-training model according to the cut label-free data set to obtain a target pre-training model.

The target pre-training model can be a pre-training model which is successfully trained finally, the recognition accuracy of the target pre-training model is higher than that of the pre-training model, and the target pre-training model can be used for training the label data to obtain the final target training model.

It can be understood that although the pre-trained model has been preliminarily provided with the recognition function, the recognition accuracy still does not meet the requirement. In order to further improve the recognition accuracy of the pre-training model, after the pre-training model is obtained, data clipping may be performed on each non-label data in the non-label data set, so as to obtain a clipped non-label data set. Optionally, when data is clipped, an area of interest without label data may be clipped, that is, the area of interest may most represent the data characteristics. That is, each clipped unlabeled data in the clipped unlabeled dataset may reflect the most core data characteristics, which may be of higher purity. For example, an area including a target object (e.g., a person, an animal, or a vehicle) may be cut out from each unlabeled image data, and the cut-out image data may include characteristic information such as a position or a size of the target object. Correspondingly, after the cutting non-label data set is obtained, the cutting non-label data set can be further utilized to carry out optimization training on the pre-training model, and therefore the target pre-training model with higher recognition accuracy is obtained.

And S160, training the target pre-training model according to the target data set to obtain a target training model.

Correspondingly, after the target pre-training model is obtained, the target pre-training model can be subjected to a final learning and training process according to the target data set, so that the target training model is obtained. The target training model can effectively identify each label data in the target data set, and the identification precision is higher.

Therefore, the model training method provided by the embodiment of the application does not need a data labeling process, and can greatly improve the precision of the model in a specific data set while effectively reducing the data labeling cost. In addition, the model training method provided by the embodiment of the application does not need huge label-free data support, and can greatly save the data volume of label-free data, so that the model iteration speed is higher. Meanwhile, the non-tag data only needs to keep the data type the same as that of the tag data, the tag type of the non-tag data does not need to be limited to belong to the tag range of the target data set, and the time for sorting and selecting the non-tag data can be greatly saved.

In an example, fig. 2 is a flowchart of a model training method provided in an embodiment of the present application, and the embodiment of the present application performs optimization and improvement on the basis of the technical solutions of the above embodiments, and provides a plurality of specific implementation manners for performing feature extraction training on a target data set, training a pre-training model according to the feature of the unlabeled data, performing data clipping on each unlabeled data in the unlabeled data set, and training the target pre-training model according to the target data set.

A model training method as shown in fig. 2, comprising:

s210, metric learning is carried out on the target data set through a preset feature extraction model, and a feature extraction model is obtained.

The preset feature extraction model may be an untrained original model with original weights preserved, such as an original CNN model. Optionally, the loss function of the preset feature extraction model may adopt an arccargin loss function.

In the embodiment of the application, when the feature extraction training is performed on the target data set, metric learning can be performed on the target data set through a preset feature extraction model. The metric learning is also called similarity learning. The preset feature extraction model adopts an arcmargin loss function, so that the feature extraction model can be effectively trained, and the confidence coefficient of corresponding cluster classes can be given to the extracted unlabeled data of each cluster class when the unlabeled data features of the unlabeled data set are subsequently extracted.

In an optional embodiment of the present application, the label data in the target data set is label image data, and the target training model is used for performing image recognition on the label image.

The tag image data is also the image data with a tag.

In the embodiment of the application, the target training model can be trained on the target data set formed by the label image data, so that the image recognition technology can be realized by using the target training model. Optionally, the target training model may be widely applied to various computer vision tasks, such as animal and plant classification, dish identification, landmark identification, and the like, and the application field of image identification of the target training model is not limited in the embodiments of the present application.

S220, performing feature extraction on the label-free data set through the feature extraction model to obtain the label-free data features.

And S230, training a pre-training model according to the label-free data characteristics.

Accordingly, S230 may specifically include the following operations:

and S231, clustering the label-free data features through the feature extraction model.

Optionally, the pre-training model may be trained by clustering unlabeled data. Optionally, clustering may be performed on the unlabeled data through a feature extraction model. The feature extraction model may extract a corresponding feature vector for each unlabeled data, and calculate a corresponding confidence for each obtained cluster class.

S232, classifying each non-label data in the non-label data set according to the clustering result of the non-label data characteristics.

And the difference value between the clustering category number included by the clustering result and the category number of the label data in the target data set meets the characteristic clustering condition.

The cluster type is a plurality of data types obtained by clustering the unlabeled data features. The feature clustering condition may be set according to an actual requirement, for example, the feature clustering condition is far greater than a certain set threshold, and the specific condition content of the feature clustering condition is not limited in the embodiment of the present application.

Correspondingly, after the characteristic extraction model is used for clustering the characteristic of the label-free data to obtain a plurality of categories, the label-free data can be automatically labeled through the obtained clustering categories. Here, the method of automatically labeling each non-labeled data by the cluster type does not belong to the method of labeling data by a label, but belongs to a data classification method. That is, each unlabeled data is respectively assigned to its corresponding cluster category set. The clustering method may be any one of the clustering methods in the field of machine learning. In order to better utilize the label-free data characteristics of the label-free data, the difference value between the number of cluster categories included in the clustering result and the number of categories of the label data in the target data set can be limited to meet the characteristic clustering condition. For example, the clustering result may be defined to include a number of cluster categories that is much larger than a number of categories of tag data in the target dataset. For example, assuming that the number of categories of tag data in the target data set is 200 categories, the number of categories of clusters may be 1 ten thousand categories.

S233, updating the unlabeled data set according to the classification processing result of the unlabeled data set to obtain an updated unlabeled data set, and training the pre-training model by using the updated unlabeled data set.

Wherein updating the unlabeled dataset may be the unlabeled dataset that includes the cluster category.

Correspondingly, after each unlabeled data is respectively attributed to the corresponding clustering class set, each clustering class set comprising a certain amount of unlabeled data can be combined into an updated unlabeled data set. It will be appreciated that the non-labeled data included in the updated non-labeled data set is not substantially changed, but a plurality of subsets are further divided into a complete set, each set belonging to a cluster category and including a certain amount of non-labeled data. That is, the data characteristics of each data in the updated label-free data set are more concentrated, and the training effect of the pre-training model is more ideal.

According to the technical scheme, after the label-free data set is clustered, the label-free data in the label-free data set are classified according to the processing result to form a new label-free data set, so that the pre-training model is trained by the new label-free data set with more concentrated label-free data characteristics, and the training efficiency and the recognition accuracy of the pre-training model can be improved.

In an optional embodiment of the present application, the training the pre-training model according to the classification processing result of the unlabeled dataset may include: determining a current classification processing result of the updated unlabeled dataset; performing reliability sequencing on each current clustering category according to the confidence of each current clustering category in the current classification processing result; obtaining label-free data of a preset proportion as a target classification result according to the reliability sorting result, and training the pre-training model according to the target classification result; and returning to execute the operation of performing feature extraction training on the target data set until the reliability sequencing result is kept unchanged. Wherein the current clustering result may be a current clustering result of updating the unlabeled dataset.

Wherein the current classification processing result may be a classification processing result of the currently updated unlabeled dataset. The current cluster category may be a cluster category included in the update unlabeled dataset. The current reliability ranking may be a ranking from large to small according to the value of the confidence. The reliability ranking result may be a ranking result of each non-tag data obtained by the reliability ranking manner. The set proportion can be set according to actual requirements, such as 50% or 80%, and the embodiments of the present application do not limit the specific values of the set proportion. The target classification result may be a classification result whose reliability satisfies a requirement.

In order to further improve the training efficiency and the recognition accuracy of the pre-training model, after the updated unlabeled data set is obtained, the current cluster categories can be subjected to reliability sequencing according to the confidence degrees of the current cluster categories in the current classification processing result of the updated unlabeled data set, that is, the current cluster categories are sequenced according to the sequence of the numerical values of the confidence degrees of the current cluster categories from large to small. The higher the confidence value corresponding to the cluster category with the higher order, the better the reliability. At this time, the unlabeled data of the previously set proportion in the reliability ranking result can be acquired as the target classification result, and the pre-training model is trained according to the acquired target classification result. The confidence degree value of the target classification result is larger, the reliability is more ideal, and the training effect is more ideal. After the training of the current updated unlabeled data set is completed, the operation of performing feature extraction training on the target data set can be returned, that is, iterative training is repeated until the reliability ranking result is kept unchanged, that is, until most of the unlabeled data in the unlabeled data set participates in the feature extraction process of the unlabeled data features. In the iterative training process, namely, in each subsequent iterative training process, the feature extraction training can be performed on the target data set by using the feature extraction model obtained in the previous training, the feature extraction can be performed on the updated unlabeled data set in the previous training by using the new feature extraction model, and the updated unlabeled data set can be trained by using the new pre-training model. In such an iterative training mode, the features for updating the label-free data set are stronger and more detailed. Therefore, the number of cluster categories obtained by each clustering process for updating the unlabeled data set is greater than the number of cluster categories in the previous time, and more detailed features of the unlabeled data can be distinguished. Therefore, the final pre-training model can be acquired more quickly and effectively by means of iterative training.

S240, data cutting is carried out on each non-label data in the non-label data set, and a cut non-label data set is obtained.

Accordingly, S240 may specifically include the following operations:

s241, forecasting the forecasting heat area matched with each non-label data in the non-label data set according to the pre-training model.

Wherein the predicted heat area may be a feature concentration area of the unlabeled data.

Although the pre-trained model trained from each original label-free data in the label-free data set has a certain recognition function, the recognition accuracy is not ideal. In order to further improve the recognition accuracy of the pre-training model, the obtained pre-training model can be used for predicting the prediction heat area matched with each non-label data in the non-label data set.

And S242, performing data clipping on the region of interest of each non-label data according to each predicted heat region.

Correspondingly, after the prediction heat degree area matched with each non-label data is obtained, the data of each non-label data can be cut by using the prediction heat degree area matched with each non-label data, so that the feature concentration area of each non-label data is reserved.

In the scheme, the pre-training model has certain identification precision, and the prediction heat degree region matched with each unlabeled data is predicted by utilizing the pre-training model, so that the region in the feature set is simply determined, and the data feature is not required to be specifically identified, so that the accuracy of data cutting is high, each cut unlabeled data in the correspondingly obtained cut unlabeled data set can reflect the most core data feature, and the data feature purity is high.

And S250, carrying out optimization training on the pre-training model according to the cut label-free data set to obtain a target pre-training model.

And S260, training the target pre-training model according to the target data set to obtain a target training model.

Accordingly, S260 may specifically include the following operations:

s261, determining a target learning rate for the target pre-training model.

The target learning rate may be a learning rate with a small value, such as 0.1 or 0.01, and the specific value of the target learning rate is not limited in the embodiments of the present application.

As the target pre-training model has higher recognition precision, excessive learning is not needed. Therefore, optionally, when the target pre-training model is trained according to the target data set, a target learning rate with a smaller value may be determined for the target pre-training model, so as to prevent the weight change caused by excessive learning of the target pre-training model.

And S262, performing parameter fixing processing on the network parameters of the preset network layer of the target pre-training model.

And S263, training the target pre-training model after parameter fixing treatment according to the target data set.

The preset network layers may be a preset number of network layers of the target pre-training model.

In order to keep the target pre-training model to quickly and effectively extract the basic features of the tag data in the target data set, parameter fixing processing can be performed on the network parameters of the preset network layer of the target pre-training model. For example, assuming that the target pre-training models have a total of 100 network layers, the network layers of the top 20 or 30 layers may be fixed. After the parameter fixing processing, the target pre-training model after the parameter fixing processing can be used for carrying out a final learning training process on the original target data set so as to obtain a target training model with higher identification precision. The parameter fixing processing mode can ensure that the basic characteristics of the label data are not disturbed, thereby further improving the training efficiency and the recognition accuracy of the target training model.

In general, in order to effectively improve the recognition accuracy of a model, it is often necessary to replace a relatively small-scale model with a relatively large-scale model or to expand the amount of data used for model training. However, relatively large-scale models are slow to predict and difficult to use in industrial deployment. And the expansion of the data volume needs a large amount of data marking work, and is time-consuming and labor-consuming. The model training method provided by the embodiment of the application can improve the recognition precision of the model in the target data set without marking data. Compared with the existing unsupervised learning mode, the method can greatly save the data volume of the label-free data, so that the model training iteration speed is higher. Compared with the existing semi-supervised learning mode, whether the label type of the non-label data belongs to the label range of the target data set or not does not need to be concerned, and the time for sorting and selecting the data can be greatly saved.

In an example, fig. 3 is a structural diagram of a model training apparatus provided in an embodiment of the present application, where the embodiment of the present application is applicable to a case where a model is trained quickly to improve model recognition accuracy, and the apparatus is implemented by software and/or hardware and is specifically configured in an electronic device. The electronic device may be a computer device or the like.

A model training apparatus 300 as shown in fig. 3, comprising: a feature extraction model acquisition module 310, a label-free data feature acquisition module 320, a pre-training model training module 330, a data clipping module 340, a target pre-training model acquisition module 350, and a target training model acquisition module 360. Wherein,

the feature extraction model acquisition module 310 is configured to perform feature extraction training on a target data set to obtain a feature extraction model; the data in the target data set is label data;

the non-tag data feature obtaining module 320 is configured to perform feature extraction on the non-tag data set through the feature extraction model to obtain non-tag data features;

a pre-training model training module 330, configured to train a pre-training model according to the label-free data characteristics;

a data clipping module 340, configured to perform data clipping on each non-tag data in the non-tag data set, so as to obtain a clipped non-tag data set;

a target pre-training model obtaining module 350, configured to perform optimization training on the pre-training model according to the clipped unlabeled dataset to obtain a target pre-training model;

and a target training model obtaining module 360, configured to train the target pre-training model according to the target data set, so as to obtain a target training model.

Optionally, the feature extraction model obtaining module 310 is specifically configured to: performing metric learning on the target data set through a preset feature extraction model; and the loss function of the preset feature extraction model adopts an arccargin loss function.

Optionally, the pre-training model training module 330 is specifically configured to: clustering the label-free data features through the feature extraction model; classifying each non-label data in the non-label data set according to the clustering result of the non-label data characteristics; the difference value between the clustering category number included in the clustering result and the category number of the label data in the target data set meets a characteristic clustering condition; and updating the label-free data set according to the classification processing result of the label-free data set to obtain an updated label-free data set, and training the pre-training model by using the updated label-free data set.

Optionally, the pre-training model training module 330 is specifically configured to: determining a current classification processing result of the updated unlabeled dataset; performing reliability sequencing on each current clustering category according to the confidence of each current clustering category in the current classification processing result; obtaining label-free data of a preset proportion as a target classification result according to the reliability sorting result, and training the pre-training model according to the target classification result; and returning to execute the operation of performing feature extraction training on the target data set until the reliability sequencing result is kept unchanged.

Optionally, the data clipping module 340 is specifically configured to: predicting a prediction heat area matched with each unlabeled data in the unlabeled data set according to the pre-training model; and performing data clipping on the interested region of the unlabeled data according to each predicted heat region.

Optionally, the target training model obtaining module 360 is specifically configured to: determining a target learning rate for the target pre-training model; carrying out parameter fixing processing on network parameters of a preset network layer of the target pre-training model; and training the target pre-training model after the parameter fixing treatment according to the target data set.

Optionally, the label data in the target data set is label image data, and the target training model is used for performing image recognition on the label image.

The model training device can execute the model training method provided by any embodiment of the application, and has corresponding functional modules and beneficial effects of the execution method. For details of the technique not described in detail in this embodiment, reference may be made to the model training method provided in any embodiment of the present application.

Since the above-described model training device is a device capable of executing the model training method in the embodiment of the present application, based on the model training method described in the embodiment of the present application, a person skilled in the art can understand the specific implementation of the model training device of the present embodiment and various variations thereof, and therefore, how to implement the model training method in the embodiment of the present application by the model training device is not described in detail herein. The scope of the present application is intended to be covered by the claims so long as those skilled in the art can implement the model training method in the embodiments of the present application.

In one example, the present application also provides an electronic device and a readable storage medium.

Fig. 4 is a schematic structural diagram of an electronic device for implementing the model training method according to the embodiment of the present application. Fig. 4 is a block diagram of an electronic device according to an embodiment of the present disclosure. Electronic devices are intended to represent various forms of digital computers, such as laptops, desktops, workstations, personal digital assistants, servers, blade servers, mainframes, and other appropriate computers. The electronic device may also represent various forms of mobile devices, such as personal digital processing, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions, are meant to be examples only, and are not meant to limit implementations of the present application that are described and/or claimed herein.

As shown in fig. 4, the electronic apparatus includes: one or more processors 401, memory 402, and interfaces for connecting the various components, including high-speed interfaces and low-speed interfaces. The various components are interconnected using different buses and may be mounted on a common motherboard or in other manners as desired. The processor may process instructions for execution within the electronic device, including instructions stored in or on the memory to display graphical information of a GUI on an external input/output apparatus (such as a display device coupled to the interface). In other embodiments, multiple processors and/or multiple buses may be used, along with multiple memories and multiple memories, as desired. Also, multiple electronic devices may be connected, with each device providing portions of the necessary operations (e.g., as a server array, a group of blade servers, or a multi-processor system). In fig. 4, one processor 401 is taken as an example.

Memory 402 is a non-transitory computer readable storage medium as provided herein. Wherein the memory stores instructions executable by at least one processor to cause the at least one processor to perform the model training methods provided herein. The non-transitory computer readable storage medium of the present application stores computer instructions for causing a computer to perform the model training method provided herein.

The memory 402, which is a non-transitory computer readable storage medium, may be used to store non-transitory software programs, non-transitory computer executable programs, and modules, such as program instructions/modules corresponding to the model training method in the embodiments of the present application (e.g., the feature extraction model acquisition module 310, the unlabeled data feature acquisition module 320, the pre-training model training module 330, the data cropping module 340, the target pre-training model acquisition module 350, and the target training model acquisition module 360 shown in fig. 3). The processor 401 executes various functional applications of the server and data processing by executing non-transitory software programs, instructions, and modules stored in the memory 402, that is, implements the model training method in the above method embodiment.

The memory 402 may include a storage program area and a storage data area, wherein the storage program area may store an operating system, an application program required for at least one function; the storage data area may store data created by use of an electronic device implementing the model training method, and the like. Further, the memory 402 may include high speed random access memory, and may also include non-transitory memory, such as at least one magnetic disk storage device, flash memory device, or other non-transitory solid state storage device. In some embodiments, memory 402 optionally includes memory located remotely from processor 401, and such remote memory may be connected over a network to an electronic device implementing the model training method. Examples of such networks include, but are not limited to, the internet, intranets, local area networks, mobile communication networks, and combinations thereof.

The electronic device implementing the model training method may further include: an input device 403 and an output device 404. The processor 401, the memory 402, the input device 403 and the output device 404 may be connected by a bus or other means, and fig. 4 illustrates an example of a connection by a bus.

The input device 403 may receive input numeric or character information and generate key signal inputs related to user settings and function control of an electronic apparatus implementing the model training method, such as a touch screen, a keypad, a mouse, a track pad, a touch pad, a pointing stick, one or more mouse buttons, a track ball, a joystick, or other input devices. The output devices 404 may include a display device, auxiliary lighting devices (e.g., LEDs), and haptic feedback devices (e.g., vibrating motors), among others. The display device may include, but is not limited to, a Liquid Crystal Display (LCD), a Light Emitting Diode (LED) display, and a plasma display. In some implementations, the display device can be a touch screen.

Various implementations of the systems and techniques described here can be realized in digital electronic circuitry, integrated circuitry, application specific ASICs (application specific integrated circuits), computer hardware, firmware, software, and/or combinations thereof. These various embodiments may include: implemented in one or more computer programs that are executable and/or interpretable on a programmable system including at least one programmable processor, which may be special or general purpose, receiving data and instructions from, and transmitting data and instructions to, a storage system, at least one input device, and at least one output device.

These computer programs (also known as programs, software applications, or code) include machine instructions for a programmable processor, and may be implemented using high-level procedural and/or object-oriented programming languages, and/or assembly/machine languages. As used herein, the terms "machine-readable medium" and "computer-readable medium" refer to any computer program product, apparatus, and/or device (e.g., magnetic discs, optical disks, memory, Programmable Logic Devices (PLDs)) used to provide machine instructions and/or data to a programmable processor, including a machine-readable medium that receives machine instructions as a machine-readable signal. The term "machine-readable signal" refers to any signal used to provide machine instructions and/or data to a programmable processor.

To provide for interaction with a user, the systems and techniques described here can be implemented on a computer having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to a user; and a keyboard and a pointing device (e.g., a mouse or a trackball) by which a user can provide input to the computer. Other kinds of devices may also be used to provide for interaction with a user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user may be received in any form, including acoustic, speech, or tactile input.

The systems and techniques described here can be implemented in a computing system that includes a back-end component (e.g., as a data server), or that includes a middleware component (e.g., an application server), or that includes a front-end component (e.g., a user computer having a graphical user interface or a web browser through which a user can interact with an implementation of the systems and techniques described here), or any combination of such back-end, middleware, or front-end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: local Area Networks (LANs), Wide Area Networks (WANs), and the Internet.

The computer system may include clients and servers. The client may be a smart phone, a notebook computer, a desktop computer, a tablet computer, a smart speaker, etc., but is not limited thereto. The server may be an independent physical server, a server cluster or a distributed system formed by a plurality of physical servers, or a cloud server providing basic cloud computing services such as cloud computing, cloud service, a cloud database, cloud storage and the like. A client and server are generally remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other.

It should be understood that various forms of the flows shown above may be used, with steps reordered, added, or deleted. For example, the steps described in the present application may be executed in parallel, sequentially, or in different orders, and the present invention is not limited thereto as long as the desired results of the technical solutions disclosed in the present application can be achieved.

The above-described embodiments should not be construed as limiting the scope of the present application. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions may be made in accordance with design requirements and other factors. Any modification, equivalent replacement, and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A model training method, comprising:

training a pre-training model according to the label-free data characteristics;

2. The method of claim 1, wherein the feature extraction training of the target dataset comprises:

performing metric learning on the target data set through a preset feature extraction model;

and the loss function of the preset feature extraction model adopts an arccargin loss function.

3. The method of claim 1, wherein the training a pre-trained model from the unlabeled data features comprises:

clustering the label-free data features through the feature extraction model;

classifying each non-label data in the non-label data set according to the clustering result of the non-label data characteristics; the difference value between the clustering category number included in the clustering result and the category number of the label data in the target data set meets a characteristic clustering condition;

and updating the label-free data set according to the classification processing result of the label-free data set to obtain an updated label-free data set, and training the pre-training model by using the updated label-free data set.

4. The method of claim 3, wherein the training the pre-trained model according to the classification process results of the unlabeled dataset comprises:

determining a current classification processing result of the updated unlabeled dataset;

performing reliability sequencing on each current clustering category according to the confidence of each current clustering category in the current classification processing result;

obtaining label-free data of a preset proportion as a target classification result according to the reliability sorting result, and training the pre-training model according to the target classification result;

and returning to execute the operation of performing feature extraction training on the target data set until the reliability sequencing result is kept unchanged.

5. The method of claim 1, wherein the data pruning each unlabeled data in the unlabeled data set comprises:

predicting a prediction heat area matched with each unlabeled data in the unlabeled data set according to the pre-training model;

and performing data clipping on the interested region of the unlabeled data according to each predicted heat region.

6. The method of claim 1, wherein the training the target pre-training model from the target dataset comprises:

determining a target learning rate for the target pre-training model;

carrying out parameter fixing processing on network parameters of a preset network layer of the target pre-training model;

and training the target pre-training model after the parameter fixing treatment according to the target data set.

7. The method of any of claims 1-6, wherein the label data in the target dataset is label image data, and the target training model is used for image recognition of the label image.

8. A model training apparatus comprising:

9. The apparatus of claim 8, wherein the feature extraction model acquisition module is specifically configured to:

10. The apparatus of claim 8, wherein the pre-training model training module is specifically configured to:

clustering the label-free data features through the feature extraction model;

11. The apparatus of claim 10, wherein the pre-training model training module is specifically configured to:

12. The apparatus of claim 8, wherein the data cropping module is specifically configured to:

13. The apparatus of claim 8, wherein the target training model acquisition module is specifically configured to:

determining a target learning rate for the target pre-training model;

14. The apparatus according to any one of claims 8-13, wherein the label data in the target data set is label image data, and the target training model is used for image recognition of the label image.

15. An electronic device, comprising:

at least one processor; and

a memory communicatively coupled to the at least one processor; wherein,

the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the model training method of any one of claims 1-7.

16. A non-transitory computer readable storage medium having stored thereon computer instructions for causing the computer to perform the model training method of any one of claims 1-7.