CN111291868A - Network model training method, device, equipment and computer readable storage medium - Google Patents

Abstract

The invention discloses a network model training method, comprising the following steps: acquiring training data and verification data corresponding to preset messages; generating multiple candidate sub-networks through a deep neural sub-network generator, and based on the training data and verification data , determine the training error corresponding to each candidate sub-network; based on each of the training errors, determine the optimal sub-network in each candidate sub-network, and determine the target prediction model based on the optimal sub-network and the main network. The invention also discloses a network model training device, equipment and computer-readable storage medium. The present invention can realize the automatic training of network models through machine learning. For different preset messages, there is no need to manually train and optimize different network models, save a lot of human resources, reduce the workload of model training and optimization, and improve the The training efficiency of the network model.

Description

Network model training method, apparatus, device, and computer-readable storage medium

technical field

The present invention relates to the technical field of artificial intelligence, and in particular, to a network model training method, apparatus, device, and computer-readable storage medium.

Background technique

With the rapid development of artificial intelligence technology, neural network models (which can be referred to as network models) have broad application prospects in the fields of system identification, pattern recognition, and intelligent control. Usually, a preset network model can be trained based on the training sample set to obtain a trained network model, and then the samples to be tested can be input into the trained network model to obtain the actual output result of the network model. The actual output result is The prediction result of the prediction of the sample to be tested.

For example, in programmatic advertising, the characteristics of each user can be input into the network model, and a decision on whether to place an advertisement can be made quickly according to the actual output of the network model.

However, the currently adopted network model requires professional algorithm engineers to conduct model training and optimization based on the characteristics of different advertising services. Different advertising services cannot use the same network model, and algorithm engineers need to conduct different network models for different advertising services. The training and optimization of the model requires a lot of time for algorithm engineers, and the model structure used in each training needs to be manually specified by the algorithm personnel, resulting in the training and optimization of different network models are not intelligent enough.

The above content is only used to assist the understanding of the technical solutions of the present invention, and does not mean that the above content is the prior art.

SUMMARY OF THE INVENTION

The main purpose of the present invention is to provide a network model training method, device, equipment and computer-readable storage medium, aiming at solving the technical problem of unintelligent training and optimization of different network models.

In order to achieve the above object, the present invention provides a network model training method, the network model training method includes the following steps:

Obtain the training data and verification data corresponding to the message;

Generate a plurality of candidate sub-networks by the deep neural sub-network generator, and determine the training error corresponding to each candidate sub-network based on the training data and the verification data;

Based on each of the training errors, an optimal sub-network is determined in each candidate sub-network, and a target prediction model is determined based on the optimal sub-network and the main network.

Further, the step of determining the training error corresponding to each candidate sub-network based on the training data and the verification data includes:

Obtain target candidate sub-networks in each candidate sub-network in turn;

Embed the target candidate sub-network into the main network to obtain an updated main network, and input the training data and verification data into the updated main network to obtain the training error corresponding to the target candidate sub-network.

Further, the step of inputting the training data and verification data into the updated main network to obtain the training error corresponding to the target candidate sub-network includes:

Input the training data and the verification data into the updated main network, and update the training times;

When the number of training times reaches the preset number of training times, the training error corresponding to the updated main network after training is used as the training error corresponding to the target candidate sub-network;

When the number of trainings does not reach the preset number of trainings, the updated main network after training is used as the updated main network, and the training data and verification data are input into the updated main network, and the training times are updated. A step of.

Further, the step of determining the optimal sub-network in each candidate sub-network based on each of the training errors includes:

Obtain the penalty coefficient corresponding to the number of network layers of each candidate sub-network;

Based on each penalty coefficient and training error, the corresponding weight scores of each candidate sub-network are determined respectively;

The candidate sub-network with the smallest weight score among the candidate sub-networks is used as the optimal sub-network.

Further, the network model training method also includes:

When multiple candidate sub-networks are generated by the deep neural sub-network generator, update the number of searches;

The step of determining the target prediction model based on the optimal sub-network and the main network includes:

Determine whether the training error corresponding to the optimal sub-network is smaller than the training error corresponding to the main network;

If the training error corresponding to the optimal sub-network is smaller than the training error corresponding to the main network, determine whether the number of searches reaches a preset number of searches;

When the number of searches reaches the preset number of searches, the optimal sub-network is embedded in the main network to obtain the target prediction model.

Further, after the step of determining whether the training error corresponding to the optimal sub-network is smaller than the training error corresponding to the main network, the method further includes:

If the training error corresponding to the optimal sub-network is greater than or equal to the training error corresponding to the main network, return to the step of generating multiple candidate sub-networks through the deep neural sub-network generator.

Further, after the step of determining whether the number of searches reaches a preset number of searches, the method further includes:

When the number of searches does not reach the preset number of searches, the optimal sub-network is embedded in the main network;

The main network after embedding the optimal sub-network is used as the main network, and the step of generating a plurality of candidate sub-networks by the deep neural sub-network generator is performed back.

Further, after the step of determining the target prediction model based on the optimal sub-network and the main network, the network model training method further includes:

When the duration after the target prediction model is determined reaches the preset duration, the steps of obtaining training data and verification data corresponding to the preset message are returned to.

Further, the training data is user feature data, the verification data is exposure click data, and after the step of determining the target prediction model based on the optimal sub-network and the main network, the network model training further includes:

Obtaining the feature data of the user to be predicted, and inputting the user data to be predicted into the target prediction model to obtain the click probability corresponding to the preset message;

When the click probability is greater than a preset threshold, it is determined that the user corresponding to the to-be-predicted user characteristic data is the to-be-delivered user.

In addition, in order to achieve the above object, the present invention also provides a network model training device, the network model training device includes:

an acquisition module for acquiring training data and verification data corresponding to the preset message;

A generation module, used for generating multiple candidate sub-networks through the deep neural sub-network generator, and determining the training error corresponding to each candidate sub-network based on the training data and the verification data;

A determination module, configured to determine an optimal sub-network in each candidate sub-network based on each of the training errors, and then determine a target prediction model based on the optimal sub-network and the main network.

In addition, in order to achieve the above object, the present invention also provides a network model training device, the network model training device includes: a memory, a processor, and a network model training device stored on the memory and running on the processor A program, when the network model training program is executed by the processor, implements the steps of the aforementioned network model training method.

In addition, in order to achieve the above object, the present invention also provides a computer-readable storage medium, where a network model training program is stored on the computer-readable storage medium, and when the network model training program is executed by a processor, the aforementioned network model is implemented The steps of the training method.

The present invention obtains training data and verification data corresponding to preset messages; then generates a plurality of candidate sub-networks through a deep neural sub-network generator, and determines the training error corresponding to each candidate sub-network based on the training data and verification data; Then, based on each of the training errors, the optimal sub-network is determined in each candidate sub-network, and the target prediction model is determined based on the optimal sub-network and the main network, which can realize automatic training of the network model through machine learning, and realize different networks. The intelligent training and optimization of the model does not require manual training and optimization of different network models for different advertising services, which saves a lot of human resources, reduces the workload of model training and tuning, and improves the training efficiency of network models.

Description of drawings

1 is a schematic structural diagram of a network model training device in a hardware operating environment involved in an embodiment of the present invention;

2 is a schematic flowchart of a first embodiment of a network model training method according to the present invention;

FIG. 3 is a schematic diagram of functional modules of an embodiment of a network model training apparatus according to the present invention.

The realization, functional characteristics and advantages of the present invention will be further described with reference to the accompanying drawings in conjunction with the embodiments.

Detailed ways

It should be understood that the specific embodiments described herein are only used to explain the present invention, but not to limit the present invention.

As shown in FIG. 1 , FIG. 1 is a schematic structural diagram of a network model training device in a hardware operating environment according to an embodiment of the present invention.

The terminal in the embodiment of the present invention may be a PC, a smart phone, a tablet computer, an e-book reader, an MP3 (Moving Picture Experts Group Audio Layer III, moving picture expert compression standard audio layer 3) player, an MP4 (Moving Picture Experts Group Audio Layer IV, moving image expert compression standard audio layer 4) Players, portable computers and other portable terminal devices with display functions.

As shown in FIG. 1 , the network model training device may include: a processor 1001 , such as a CPU, a network interface 1004 , a user interface 1003 , a memory 1005 , and a communication bus 1002 . Among them, the communication bus 1002 is used to realize the connection and communication between these components. The user interface 1003 may include a display screen (Display), an input unit such as a keyboard (Keyboard), and the optional user interface 1003 may also include a standard wired interface and a wireless interface. Optionally, the network interface 1004 may include a standard wired interface and a wireless interface (eg, a WI-FI interface). The memory 1005 may be high-speed RAM memory, or may be non-volatile memory, such as disk memory. Optionally, the memory 1005 may also be a storage device independent of the aforementioned processor 1001 .

Optionally, the network model training device may further include a camera, an RF (Radio Frequency, radio frequency) circuit, a sensor, an audio circuit, a WiFi module, and the like. Among them, sensors such as light sensors, motion sensors and other sensors. Specifically, the light sensor may include an ambient light sensor and a proximity sensor; of course, the network model training device may also be configured with other sensors such as a gyroscope, a barometer, a hygrometer, a thermometer, and an infrared sensor, which will not be repeated here.

Those skilled in the art can understand that the structure of the network model training device shown in FIG. 1 does not constitute a limitation on the network model training device, and may include more or less components than those shown in the figure, or combine some components, or different component layout.

As shown in FIG. 1 , the memory 1005 as a computer storage medium may include an operating system, a network communication module, a user interface module and a network model training program.

In the network model training device shown in FIG. 1, the network interface 1004 is mainly used to connect the background server, and perform data communication with the background server; the user interface 1003 is mainly used to connect the client (client), and perform data communication with the client; And the processor 1001 can be used to call the network model training program stored in the memory 1005 .

In this embodiment, the network model training device includes: a memory 1005, a processor 1001, and a network model training program stored on the memory 1005 and running on the processor 1001, wherein the processor 1001 calls the memory 1005 When the network model training program stored in the network model is performed, the steps of the network model training method in the following embodiments are executed.

The present invention also provides a network model training method. Referring to FIG. 2 , FIG. 2 is a schematic flowchart of the first embodiment of the network model training method of the present invention.

In this embodiment, the network model training method includes the following steps:

Step S100, acquiring training data and verification data corresponding to the preset message;

In this embodiment, the preset message may be preset advertisements, preset products, questionnaires, etc., the verification data is click data, exposure data, viewing data, etc. of the preset message, and the training data is click data, exposure data, or viewing data Corresponding user feature information. When performing model training on the preset message, first obtain training data and verification data corresponding to the preset message, and use the obtained training data and verification data as training data, that is, sample data for subsequent model training. The user characteristic data may include the gender, age, occupation, and hobby of each user who clicks the preset message (eg, several network links that the user clicks frequently, or the network link that the user clicks the most frequently).

For example, taking a preset advertisement as an example, the preset advertisement can be any advertisement placed online (on the Internet, etc.), and after the preset advertisement is placed for a period of time, that is, when there is exposure click data corresponding to the preset advertisement , model training can be performed on the preset advertisement to obtain the target prediction model of the preset advertisement. For example, when the duration after the preset advertisement is placed reaches the preset time length, model training can be performed on the preset advertisement, or , after the preset advertisement is placed, when it is detected that the exposure click data corresponding to the preset advertisement is greater than the preset exposure click value, model training is performed on the preset advertisement.

Step S200, generating a plurality of candidate sub-networks through a deep neural sub-network generator, and determining the training error corresponding to each candidate sub-network based on the training data and the verification data;

In this embodiment, after the training data and verification data corresponding to the preset message are obtained, multiple candidate subnetworks are generated by the deep neural subnetworks generator, wherein the number of layers of each subnetworks in the multiple candidate subnetworks is not calculated. Restriction, that is, the number of network layers in each candidate sub-network is random, and the number of candidate sub-networks generated each time by the deep neural sub-network generator is also random. Of course, in other embodiments, the candidate sub-network can also be limited. The number of sub-networks, for example, a preset number of candidate sub-networks are generated by the deep neural sub-network generator each time.

Then, based on the training data and the verification data, the training error corresponding to each candidate sub-network is determined. Specifically, each candidate sub-network is embedded in the main network in turn, and the training data and the verification data are input into the main network embedded in the candidate sub-network, and the model is carried out. Training to get the training error corresponding to each candidate sub-network.

It should be noted that the deep neural sub-network generator can be specified by the algorithm developer, or, multiple deep neural sub-network generators are preset, and one of the multiple deep neural sub-network generators is randomly selected to generate multiple candidate subnet.

Step S300, based on each of the training errors, determine an optimal sub-network in each candidate sub-network, and determine a target prediction model based on the optimal sub-network and the main network.

In this embodiment, when the training error is obtained, the optimal sub-network is determined in each candidate sub-network based on the training error. For example, the candidate sub-network with the smallest training error is taken as the optimal sub-network, or the training error is compared with the optimal sub-network. The number of layers of each candidate sub-network is weighted, and the candidate sub-network with the smallest weight score after the weight calculation is used as the optimal sub-network.

Then, the target prediction model is determined based on the optimal sub-network and the main network. If only one search/training is required, the neural network structure obtained by embedding the optimal sub-network into the main network is the target prediction model; if multiple searches are required /training, then judge whether the number of searches/training has reached the specified number of times, if so, the neural network structure obtained by embedding the optimal sub-network into the main network is the target prediction model, if not, embedding the optimal sub-network into the main network , and use the main network embedded in the optimal sub-network as the new main network, and return to step S200, so that the automatic training of the network model can be realized through machine learning. For different preset messages, there is no need to manually perform different network models. It only takes a small amount of time for algorithm engineers to obtain the machine learning model for online preset message delivery, which saves a lot of human resources and improves the training efficiency of the network model.

Further, in an embodiment, after step S300, it further includes:

When the duration after the target prediction model is determined reaches the preset duration, the steps of obtaining training data and verification data corresponding to the preset message are returned to.

In this embodiment, when it is determined that the duration of the target prediction model reaches the preset duration, the process returns to step S100, so as to realize the regular update of the target prediction model, and then realize the cyclic update of the target prediction model based on the automatic machine learning technology. Model optimization is automatically carried out with changes in user preferences (user characteristic data) to ensure the delivery effect of preset messages.

It should be noted that, in other embodiments, the verification data corresponding to the preset message can be stored as the first verification data data, and the click data, exposure data or viewing data corresponding to the preset message can be counted in real time to obtain the current data. the second verification data, and then calculate the difference between the second verification data and the first verification data, when the difference is greater than the preset difference, return to step S100 to update according to the number of clicks (verification data) of the user Target prediction model to ensure the delivery effect of preset messages.

The network model training method proposed in this embodiment obtains training data and verification data corresponding to preset messages; then generates a plurality of candidate sub-networks through a deep neural sub-network generator, and determines each sub-network based on the training data and verification data The training error corresponding to the candidate sub-network; then, based on each of the training errors, the optimal sub-network is determined in each candidate sub-network, and the target prediction model is determined based on the optimal sub-network and the main network, and the network can be realized through machine learning. The automatic training of the model realizes the intelligent training and optimization of different network models. For different message services, there is no need to manually train and optimize different network models, which saves a lot of human resources and reduces the workload of model training and tuning. The training efficiency of the network model is improved.

Based on the first embodiment, a second embodiment of the network model training method of the present invention is proposed. In this embodiment, step S200 includes:

Step S210, obtaining target candidate sub-networks in each candidate sub-network in turn;

Step S220: Embed the target candidate sub-network into the main network to obtain an updated main network, and input the training data and verification data into the updated main network to obtain the training error corresponding to the target candidate sub-network.

In this embodiment, after a plurality of candidate sub-networks are obtained, target candidate sub-networks are obtained from each candidate sub-network in turn, and each obtained target candidate sub-network is embedded in the main network to obtain an updated main network, and all the obtained target candidate sub-networks are embedded in the main network. The training data and verification data are input into the updated main network, and the training error corresponding to the target candidate sub-network is obtained by model training, and then the training error corresponding to each candidate sub-network is accurately obtained.

It should be noted that when the training data and verification data are input into the updated main network, and model training is performed, multiple model trainings can be performed. Specifically, the training data and verification data are input into the updated main network. , and update the training times; if the training times are less than the preset training times, input the training data and the verification data into the updated main network again, and the updated main network is the embedded target that has passed the training data and the verification data. For the main network of the candidate sub-network, if the number of training times reaches the preset number of training times, the currently updated loss function value of the main network is used as the training error, where the preset number of training times can be set reasonably.

In this embodiment, after training through the updated main network, the click probability of each user corresponding to the training data for clicking/viewing the preset message can be obtained, and the user whose click probability is less than the preset click probability is regarded as not clicking/viewing the preset message. Users who view the preset message will divide the users who do not click the preset message by the number of verification data to obtain the training error, where the number of verification data may be the number of users of the training data.

In the network model training method proposed in this embodiment, target candidate sub-networks are obtained from each candidate sub-network in turn; then the target candidate sub-network is embedded in the main network to obtain an updated main network, and the training data and verification data are The updated main network is input to obtain the training error corresponding to the target candidate sub-network, and the training error corresponding to each candidate sub-network can be accurately obtained, thereby improving the accuracy of the target prediction model in predicting the preset message, so as to ensure the prediction of the target candidate sub-network. Setting the delivery effect of the message further improves the training efficiency of the network model.

Based on the second embodiment, a third embodiment of the network model training method of the present invention is proposed. In this embodiment, step S220 includes:

Step S221, input the training data and verification data into the updated main network, and update the training times;

Step S222, when the training times reaches the preset training times, take the training error corresponding to the updated main network after training as the training error corresponding to the target candidate sub-network;

Step S223, when the number of times of training does not reach the preset number of times of training, take the updated main network after training as the updated main network, and return the trained data and verification data into the updated main network, and update The number of training steps.

In this embodiment, in order to improve the efficiency of model training, the updated main network needs to be trained multiple times, and then the training data and the verification data are input into the updated main network, and the training of the updated main network is performed. , the number of update training is the number of training of the main network after the current update.

Then, it is judged whether the training times reaches the preset training times, and if the training times reaches the preset training times, the training error corresponding to the updated main network after training is taken as the training error corresponding to the target candidate sub-network, and reset The number of training times, otherwise the updated main network after training is used as the updated main network, and the training data and verification data are input into the updated main network, and the number of training steps is updated, so that each target candidate The updated main network corresponding to the sub-network is trained for preset training times to improve the accuracy of the training error.

The preset training times can be set reasonably.

In the network model training method proposed in this embodiment, the training data and the verification data are input into the updated main network, and the training times are updated; then when the training times reach the preset training times, the updated the main network, the corresponding training error is taken as the training error corresponding to the target candidate sub-network; then when the number of training times does not reach the preset number of training times, the updated main network after training is used as the updated main network, And return to the steps of inputting the training data and verification data into the updated main network, and updating the training times, so as to improve the training of the updated main network with the preset training times, thereby improving the training error corresponding to each candidate sub-network. The accuracy of the target prediction model further improves the accuracy of predicting the preset message and improves the training efficiency of the network model.

Based on the first embodiment, a fourth embodiment of the network model training method of the present invention is proposed. In this embodiment, step S300 includes:

Step S310, obtaining the penalty coefficient corresponding to the number of network layers of each candidate sub-network;

Step S320, based on each penalty coefficient and training error, respectively determine the weight score corresponding to each candidate sub-network;

Step S330, taking the candidate sub-network with the smallest weight score among the candidate sub-networks as the optimal sub-network.

In this embodiment, the preset penalty coefficient corresponding to the number of network layers, or the mapping relationship between the number of network layers and the preset penalty coefficient can be preset. The number of layers, and the mapping relationship, determine the penalty coefficient corresponding to the number of network layers of each candidate sub-network.

Then, based on each penalty coefficient and training error, the corresponding weight scores of each candidate sub-network are determined respectively. Specifically, the weight of the penalty coefficient and the weight of the training error are obtained respectively, based on the penalty coefficient and the training error, the weight of the penalty coefficient and the training error. The weight of the error determines the weight score corresponding to each candidate sub-network, for example, weight score=weight of penalty coefficient*penalty coefficient+weight of training error*training error.

The network model training method proposed in this embodiment obtains the penalty coefficient corresponding to the number of network layers of each candidate sub-network; then, based on each penalty coefficient and the training error, respectively determines the weight score corresponding to each candidate sub-network; The candidate sub-network with the smallest weight score in the sub-network is used as the optimal sub-network. By comprehensively considering the penalty coefficient and the training error, the determined candidate sub-network is a candidate sub-network with a simple network and a small training error. The target prediction model reduces the complexity of the network model while predicting the accuracy of the preset message.

Based on the first embodiment, a fifth embodiment of the network model training method of the present invention is proposed. In this embodiment, the network model training method further includes:

Step S100, when multiple candidate sub-networks are generated by the deep neural sub-network generator, update the number of searches;

Step S300 includes:

Step S340, determining whether the training error corresponding to the optimal sub-network is smaller than the training error corresponding to the main network;

Step S350, if the training error corresponding to the optimal sub-network is smaller than the training error corresponding to the main network, determine whether the number of searches reaches a preset number of searches;

Step S360, when the number of searches reaches the preset number of searches, the optimal sub-network is embedded in the main network to obtain the target prediction model.

In this embodiment, in order to improve the prediction accuracy of the target prediction model, multiple optimal sub-networks need to be superimposed. Therefore, when multiple candidate sub-networks are generated by the deep neural sub-network generator, the number of searches is updated. Search count +1 to get the updated search count.

Then, it is determined whether the training error corresponding to the optimal sub-network is smaller than the training error corresponding to the main network, wherein, when any sub-network is embedded in the main network, the training error of the main network is greater than that of any candidate sub-network For the corresponding training error, when the sub-network has been embedded in the main network, the training error of the main network can be obtained according to the previous training data.

If the training error corresponding to the optimal sub-network is smaller than the training error corresponding to the main network, it is determined whether the number of searches reaches the preset number of searches to determine whether the search/training of the preset number of searches has been completed currently. When the preset number of searches is reached, the optimal sub-network is embedded in the main network to obtain the target prediction model, and the number of searches is reset at the same time.

Further, in an embodiment, after step S340, it further includes:

If the training error corresponding to the optimal sub-network is greater than or equal to the training error corresponding to the main network, return to the step of generating multiple candidate sub-networks through the deep neural sub-network generator.

In this embodiment, if the training error corresponding to the optimal sub-network is greater than or equal to the training error corresponding to the main network, the update of the current main network is abandoned, and the execution of generating multiple candidates through the deep neural sub-network generator is returned. The step of sub-network, wherein, when multiple candidate sub-networks are generated this time, the number of searches may or may not be updated; or, if the number of searches reaches the preset number of searches at this time, multiple candidate sub-networks are generated this time. In the network, the number of searches is not updated, and if the number of searches does not reach the preset number of searches, the number of searches is updated.

Further, in another embodiment, after step S350, it further includes:

Step S360, when the number of searches does not reach the preset number of searches, embed the optimal sub-network into the main network;

Step S370, take the main network after embedding the optimal sub-network as the main network, and return to the step of generating multiple candidate sub-networks through the deep neural sub-network generator.

In this embodiment, the number of searches has not reached the preset number of searches, and the model training has not reached the specified number of times. Therefore, the optimal sub-network is embedded in the main network, and the main network after embedding the optimal sub-network is The network acts as the main network, and returns to perform the step of generating multiple candidate sub-networks through the deep neural sub-network generator to realize cyclic training.

The network model training method proposed in this embodiment determines whether the training error corresponding to the optimal sub-network is smaller than the training error corresponding to the main network; then if the training error corresponding to the optimal sub-network is smaller than the main network corresponding training error, then determine whether the number of searches reaches the preset number of searches; then when the number of searches reaches the preset number of searches, the optimal sub-network is embedded in the main network to obtain the For the target prediction model, the target prediction model is obtained through multiple searches, which further improves the accuracy of the target prediction model in predicting the preset message and improves the training efficiency of the model.

Based on the above embodiments, a sixth embodiment of the network model training method of the present invention is proposed. In this embodiment, the training data is user characteristic data, and the verification data is exposure click data. After step S300, the network model Training methods also include:

Step S400, obtaining the feature data of the user to be predicted, and inputting the user data to be predicted into the target prediction model to obtain the click probability corresponding to the preset message;

Step S500, when the click probability is greater than a preset threshold, determine that the user corresponding to the feature data of the user to be predicted is the user to be placed.

In this embodiment, after the target prediction model is obtained, the characteristic data of the user to be predicted can be obtained at any time, and the user data to be predicted can be input into the target prediction model, and the click probability corresponding to the preset message can be obtained through model training, that is, the click probability corresponding to the preset message is obtained. Predict the click probability that the user characteristic data clicks the preset message.

Then, it is judged whether the click probability is greater than a preset threshold, and if it is greater than the value, it is determined that the user corresponding to the feature data of the user to be predicted is the user to be delivered, and the preset message can be delivered to the user terminal corresponding to the feature data of the user to be predicted. Or, push the user terminal information of the user to be delivered to the delivery server of the preset message, so that the delivery server will deliver the preset message to the user terminal corresponding to the feature data of the user to be predicted, so as to achieve accurate delivery of the preset message, Ensure that the delivered preset messages can be clicked and viewed by users with a high probability, thereby ensuring the delivery effect of the preset messages.

For example, when a user watches a certain video, the user's feature data can be obtained as the user's feature data to be predicted, and model training is performed to obtain the probability that the user clicks on the preset message. If the click probability is greater than the preset threshold, then in When the user finishes watching the video, the preset message is pushed to the user's terminal.

The network model training method proposed in this embodiment obtains the feature data of the user to be predicted, and inputs the user data to be predicted into the target prediction model for model training, so as to obtain the click probability corresponding to the preset message; When the click probability is greater than the preset threshold, it is determined that the user corresponding to the feature data of the user to be predicted is the user to be delivered, and the delivery of the preset message is performed by accurately predicting the click probability of the user, thereby improving the delivery effect of the preset message.

An embodiment of the present invention further provides a network model training device. Referring to FIG. 3 , the network model training device includes:

an acquisition module 100, configured to acquire training data and verification data corresponding to the preset message;

The generating module 200 is used for generating a plurality of candidate sub-networks through a deep neural sub-network generator, and based on the training data and the verification data, determining the training error corresponding to each candidate sub-network;

The determining module 300 is configured to determine an optimal sub-network in each candidate sub-network based on each of the training errors, and then determine a target prediction model based on the optimal sub-network and the main network.

Optionally, the generating module 200 is also used for:

Obtain target candidate sub-networks in each candidate sub-network in turn;

Embed the target candidate sub-network into the main network to obtain an updated main network, and input the training data and verification data into the updated main network to obtain the training error corresponding to the target candidate sub-network.

Optionally, the generating module 200 is also used for:

Input the training data and the verification data into the updated main network, and update the training times;

When the number of training times reaches the preset number of training times, the training error corresponding to the updated main network after training is used as the training error corresponding to the target candidate sub-network;

When the number of trainings does not reach the preset number of trainings, take the updated main network after training as the updated main network, and return the training data and verification data into the updated main network, and update the training times A step of.

Optionally, the determining module 300 is further configured to:

Obtain the penalty coefficient corresponding to the number of network layers of each candidate sub-network;

Based on each penalty coefficient and training error, the corresponding weight scores of each candidate sub-network are determined respectively;

The candidate sub-network with the smallest weight score among the candidate sub-networks is used as the optimal sub-network.

Optionally, the network model training device further includes:

When multiple candidate sub-networks are generated by the deep neural sub-network generator, update the number of searches;

The determination module 300 is also used to:

Determine whether the training error corresponding to the optimal sub-network is smaller than the training error corresponding to the main network;

If the training error corresponding to the optimal sub-network is smaller than the training error corresponding to the main network, determine whether the number of searches reaches a preset number of searches;

When the number of searches reaches the preset number of searches, the optimal sub-network is embedded in the main network to obtain the target prediction model.

Optionally, the determining module 300 is further configured to:

If the training error corresponding to the optimal sub-network is greater than or equal to the training error corresponding to the main network, return to the step of generating multiple candidate sub-networks through the deep neural sub-network generator.

Optionally, the determining module 300 is further configured to:

When the number of searches does not reach the preset number of searches, the optimal sub-network is embedded in the main network;

The main network after embedding the optimal sub-network is used as the main network, and the step of generating a plurality of candidate sub-networks by the deep neural sub-network generator is performed back.

Optionally, the network model training device further includes:

When the duration after the target prediction model is determined reaches the preset duration, the steps of acquiring training data and verification data corresponding to the preset message are returned to.

Optionally, the network model training device further includes:

Obtaining the feature data of the user to be predicted, and inputting the user data to be predicted into the target prediction model to obtain the click probability corresponding to the preset message;

When the click probability is greater than a preset threshold, it is determined that the user corresponding to the to-be-predicted user characteristic data is the to-be-delivered user.

For the methods executed by the above program modules, reference may be made to the various embodiments of the network model training method of the present invention, which will not be repeated here.

In addition, an embodiment of the present invention also provides a computer-readable storage medium, where a network model training program is stored on the computer-readable storage medium, and when the network model training program is executed by a processor, the above-mentioned network model training is implemented steps of the method.

For the method implemented when the network model training program running on the processor is executed, reference may be made to the various embodiments of the network model training method of the present invention, which will not be repeated here.

It should be noted that, herein, the terms "comprising", "comprising" or any other variation thereof are intended to encompass non-exclusive inclusion, such that a process, method, article or system comprising a series of elements includes not only those elements, It also includes other elements not expressly listed or inherent to such a process, method, article or system. Without further limitation, an element qualified by the phrase "comprising a..." does not preclude the presence of additional identical elements in the process, method, article or system that includes the element.

The above-mentioned serial numbers of the embodiments of the present invention are only for description, and do not represent the advantages or disadvantages of the embodiments.

From the description of the above embodiments, those skilled in the art can clearly understand that the method of the above embodiment can be implemented by means of software plus a necessary general hardware platform, and of course can also be implemented by hardware, but in many cases the former is better implementation. Based on such understanding, the technical solutions of the present invention can be embodied in the form of software products in essence or the parts that make contributions to the prior art, and the computer software products are stored in a storage medium (such as ROM/RAM) as described above. , magnetic disk, optical disk), including several instructions to make a terminal device (which may be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) execute the methods described in the various embodiments of the present invention.

The above are only preferred embodiments of the present invention, and are not intended to limit the scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the description and drawings of the present invention, or directly or indirectly applied in other related technical fields , are similarly included in the scope of patent protection of the present invention.

Claims (12)

1. a network model training method, is characterized in that, described network model training method comprises the following steps: Obtain the training data and verification data corresponding to the preset message; Generate a plurality of candidate sub-networks by the deep neural sub-network generator, and determine the training error corresponding to each candidate sub-network based on the training data and the verification data; Based on each of the training errors, an optimal sub-network is determined in each candidate sub-network, and a target prediction model is determined based on the optimal sub-network and the main network. 2. The network model training method according to claim 1, wherein the step of determining the training error corresponding to each candidate sub-network based on the training data and the verification data comprises: Obtain target candidate sub-networks in each candidate sub-network in turn; Embed the target candidate sub-network into the main network to obtain an updated main network, and input the training data and verification data into the updated main network to obtain the training error corresponding to the target candidate sub-network. 3. The network model training method according to claim 2, wherein the step of inputting the training data and the verification data into the updated main network to obtain the training error corresponding to the target candidate sub-network comprises the following steps: : Input the training data and the verification data into the updated main network, and update the training times; When the number of training times reaches the preset number of training times, the training error corresponding to the updated main network after training is used as the training error corresponding to the target candidate sub-network; When the number of trainings does not reach the preset number of trainings, the updated main network after training is used as the updated main network, and the training data and verification data are input into the updated main network, and the training times are updated. A step of. 4. The network model training method according to claim 1, wherein the step of determining an optimal sub-network in each candidate sub-network based on each of the training errors comprises: Obtain the penalty coefficient corresponding to the number of network layers of each candidate sub-network; Based on each penalty coefficient and training error, the corresponding weight scores of each candidate sub-network are determined respectively; The candidate sub-network with the smallest weight score among the candidate sub-networks is used as the optimal sub-network. 5. The network model training method according to claim 1, wherein the network model training method further comprises: When multiple candidate sub-networks are generated by the deep neural sub-network generator, update the number of searches; The step of determining the target prediction model based on the optimal sub-network and the main network includes: Determine whether the training error corresponding to the optimal sub-network is smaller than the training error corresponding to the main network; If the training error corresponding to the optimal sub-network is smaller than the training error corresponding to the main network, determine whether the number of searches reaches a preset number of searches; When the number of searches reaches the preset number of searches, the optimal sub-network is embedded in the main network to obtain the target prediction model. 6. The network model training method according to claim 5, wherein after the step of determining whether the training error corresponding to the optimal sub-network is less than the training error corresponding to the main network, further comprising: If the training error corresponding to the optimal sub-network is greater than or equal to the training error corresponding to the main network, return to the step of generating multiple candidate sub-networks through the deep neural sub-network generator. 7. The network model training method according to claim 5, wherein after the step of determining whether the number of searches reaches a preset number of searches, further comprising: When the number of searches does not reach the preset number of searches, the optimal sub-network is embedded in the main network; The main network after embedding the optimal sub-network is used as the main network, and the step of generating a plurality of candidate sub-networks by the deep neural sub-network generator is performed back. 8. The network model training method according to claim 1, wherein after the step of determining the target prediction model based on the optimal sub-network and the main network, the network model training method further comprises: When the duration after the target prediction model is determined reaches the preset duration, the steps of acquiring training data and verification data corresponding to the preset message are returned to. 9. The network model training method according to any one of claims 1 to 8, wherein the training data is user feature data, the verification data is exposure click data, and the optimal sub-network is based on the And after the step of determining the target prediction model by the main network, the network model training further includes: Obtaining the feature data of the user to be predicted, and inputting the user data to be predicted into the target prediction model to obtain the click probability corresponding to the preset message; When the click probability is greater than a preset threshold, it is determined that the user corresponding to the to-be-predicted user characteristic data is the to-be-delivered user. 10. A network model training device, wherein the network model training device comprises: an acquisition module for acquiring training data and verification data corresponding to the preset message; A generation module, used for generating multiple candidate sub-networks through the deep neural sub-network generator, and determining the training error corresponding to each candidate sub-network based on the training data and the verification data; A determination module, configured to determine an optimal sub-network in each candidate sub-network based on each of the training errors, and then determine a target prediction model based on the optimal sub-network and the main network. 11. A network model training device, characterized in that the network model training device comprises: a memory, a processor, and a network model training program stored on the memory and running on the processor, the network When the model training program is executed by the processor, the steps of the network model training method according to any one of claims 1 to 9 are implemented. 12. A computer-readable storage medium, characterized in that, a network model training program is stored on the computer-readable storage medium, and when the network model training program is executed by a processor, any one of claims 1 to 9 is implemented. The steps of the network model training method described in item.

Images