WO2020164275A1 - 基于预测模型的处理结果预测方法、装置及服务器 - Google Patents

基于预测模型的处理结果预测方法、装置及服务器 Download PDF

Info

Publication number
WO2020164275A1
WO2020164275A1 PCT/CN2019/117992 CN2019117992W WO2020164275A1 WO 2020164275 A1 WO2020164275 A1 WO 2020164275A1 CN 2019117992 W CN2019117992 W CN 2019117992W WO 2020164275 A1 WO2020164275 A1 WO 2020164275A1
Authority
WO
WIPO (PCT)
Prior art keywords
processing result
target network
training period
training
prediction model
Prior art date
Application number
PCT/CN2019/117992
Other languages
English (en)
French (fr)
Inventor
马进
王健宗
Original Assignee
平安科技（深圳）有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 平安科技（深圳）有限公司 filed Critical 平安科技（深圳）有限公司
Publication of WO2020164275A1 publication Critical patent/WO2020164275A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Definitions

  • This application relates to the field of computer technology, and in particular to a method, device and server for predicting processing results based on a predictive model.
  • the embodiments of the present application provide a method, device, and server for predicting processing results based on a predictive model, which can predict the processing results of the network, thereby effectively improving the efficiency of determining the processing results of the network.
  • an embodiment of the present application provides a method for predicting processing results based on a predictive model, including:
  • the processing result set includes t1 processing result data obtained by processing the training data by the target network, and processing the target network using the processing result prediction model
  • the t2 processing result data obtained by the result prediction, the T and t1 are positive integers, and the t2 is a non-negative integer;
  • processing result prediction model to process the processing result data in the processing result set, and predicting the processing result data of the target network in the T+1 training period;
  • the processing result prediction model is constructed based on the network setting information of the target network.
  • an embodiment of the present application provides a processing result prediction device based on a prediction model, the device including:
  • the acquisition unit is used to acquire the processing result set of the target network in the previous T training periods, the processing result set includes t1 processing result data obtained by the target network processing the training data, and the processing result prediction model for all
  • a processing unit configured to use the processing result prediction model to process the processing result data in the processing result set, and predict the processing result data of the target network in the T+1 training period;
  • the processing result prediction model is constructed based on the network setting information of the target network.
  • embodiments of the present application provide a server, including a processor, a communication interface, and a memory.
  • the processor, the communication interface, and the memory are connected to each other, wherein the memory is used to store a computer program,
  • the computer program includes program instructions, and the processor is configured to call the program instructions to perform the following steps:
  • the processing result set includes t1 processing result data obtained by processing the training data by the target network, and processing the target network using the processing result prediction model
  • the t2 processing result data obtained by the result prediction, the T and t1 are positive integers, and the t2 is a non-negative integer;
  • processing result prediction model to process the processing result data in the processing result set, and predicting the processing result data of the target network in the T+1 training period;
  • the processing result prediction model is constructed based on the network setting information of the target network.
  • an embodiment of the present application provides a computer-readable storage medium that stores a computer program.
  • the computer program includes program instructions that, when executed by a processor, cause all The processor executes the processing result prediction method based on the prediction model according to any one of the first aspects described above.
  • the processing result prediction model is used to process the processing result data in the processing result set, and the processing result data of the target network in the next training period can be predicted, thereby effectively improving the efficiency of determining the processing result of the network.
  • FIG. 1 is a schematic flowchart of a method for predicting processing results based on a predictive model provided by an embodiment of the present application
  • FIG. 2 is a schematic structural diagram of a processing result prediction device based on a prediction model provided by an embodiment of the present application
  • Fig. 3 is a schematic structural diagram of a server provided by an embodiment of the present application.
  • FIG. 1 is a schematic flowchart of a method for predicting processing results based on a predictive model provided by an embodiment of the present application.
  • the processing result prediction method may include the following steps:
  • the server obtains the processing result set of the target network in the previous T training periods.
  • the processing result set of the target network in the first T training periods includes: t1 processing result data obtained by the target network processing the training data, and the processing result prediction model for the target network obtained by using the processing result prediction model t2 processing result data.
  • T and t1 are both positive integers, that is, an integer greater than 0;
  • t2 is a non-negative integer, that is, a positive integer and 0.
  • the processing result set only includes t1 processing result data obtained by the target network processing the training data, and does not include the processing result data obtained by using the processing result prediction model to predict the processing result of the target network.
  • each training period of the target network corresponds to one processing result data
  • T is equal to the sum of t1 and t2.
  • the target network processes the training data in t1 training periods in the first T training periods to obtain t1 processing result data.
  • the server uses the processing result prediction model to predict the processing result of the target network, and predicts t2 processing result data corresponding to t2 training periods.
  • the time corresponding to the t2 training period is later than the time corresponding to the t1 training period.
  • the t2 processing result data are predicted by the processing result prediction model based on the t1 processing result data.
  • the processing result data may be used to indicate the performance result of the target network during the training period, and the performance result may be the classification accuracy rate.
  • the target network performs classification processing on the training data during the training period to obtain the classification processing result; then based on the classification processing result, the classification accuracy rate of the target network is determined, and the determined classification accuracy rate is determined as the corresponding processing of the target network during the training period Result data.
  • the server uses the processing result prediction model to predict the classification accuracy of the target network, and determines the predicted classification accuracy as the processing result data corresponding to the target network during the training period.
  • the processing result prediction model is constructed based on the network setting information of the target network.
  • the server obtains network setting information of the target network, and the network setting information includes structural parameters and hyperparameters of the target network.
  • the target network may be a neural network
  • the structural parameters include the convolutional layer parameters and fully connected layer parameters of the target network, etc., and specifically may include the weight parameters of each convolutional layer, the number of convolutional layers, The number of neurons in each convolutional layer, etc.
  • the hyperparameters include the learning rate of the target network, the learning step length, the weight attenuation, the size of the input network picture, whether to use data enhancement (for example, represented by 0 or 1) and so on.
  • the server determines the target characteristic parameter according to the network setting information of the target network, and constructs and obtains the processing result prediction model corresponding to the target network according to the target characteristic parameter.
  • the server may directly determine the structural parameters and hyperparameters included in the network setting information as the target characteristic parameters corresponding to the processing result prediction model.
  • the server can use the regression support vector machine v-SVR model to train the processing result prediction model to obtain the trained processing result prediction model.
  • the regression support vector machine v-SVR model can quickly complete the training of the prediction model for the processing results.
  • the server can use the support vector machine regression LIBSVM software to train the processing result prediction model to obtain the trained processing result prediction model.
  • the SVM type corresponding to the LIBSVM software can be designated as the regression support vector machine v-SVR, and the kernel function corresponding to the LIBSVM software can be designated as the radial basis function (Radial Basis Function, RBF); the category data corresponding to the LIBSVM software is It can be determined based on the network setting information of the target network.
  • the Bayesian method may also be used to train the processing result prediction model to obtain a trained processing result prediction model.
  • the server uses a processing result prediction model to process the processing result data in the processing result set, and predicting the processing result data of the target network in the T+1 training period.
  • the server uses the processing result prediction model to process the target network in the processing result set for t1 processing result data obtained by processing the training data, and uses the processing result prediction model to perform the processing result prediction for the target network t2
  • the processing result data is processed, and the processing result data of the target network in the T+1 training period is predicted.
  • the processing result data is used to indicate the classification accuracy of the target network.
  • the server calls the processing result prediction model to calculate the first mean value of the classification accuracy from the TMth training period to the Tth training period in the processing result set, where M is an integer greater than 1 and less than T, and the TMth in the processing result set is calculated
  • the second mean value of the change in the classification accuracy from the training period to the T-th training period calculate the first difference between the T-th training period and the T-1th training period in the processing result set, and calculate The second difference of the classification accuracy between the T-1 training period and the T-2 training period in the processing result set, and the calculated first mean, second mean, first difference, and second difference Value and the classification accuracy of the T-th training period, predict the classification accuracy of the target network in the T+1 training period, and determine the predicted classification accuracy as the target network in the T+1 training period Process the result data.
  • the change value of the classification accuracy from the TMth training period to the Tth training period may refer to the difference between the classification accuracy of each adjacent training period from the TMth training period to the Tth training period, or It is the value obtained by multiplying the difference between the classification accuracy of each adjacent training period by the corresponding weight.
  • the weight corresponding to the classification accuracy difference in the subsequent adjacent training period is greater than the weight corresponding to the classification accuracy difference in the previous adjacent training period; because the classification accuracy difference in the previous adjacent training period is relatively larger Therefore, in order to improve the accuracy of the subsequent prediction, the weight corresponding to the difference in the classification accuracy of the previous adjacent training period can be set to a value less than 1.
  • the processing result set only includes t1 processing result data obtained by the target network processing the training data.
  • the first T training periods are all using the training data to target the target network. The period of training.
  • the server can use the processing result prediction model to process the t1 processing result data corresponding to the target network in the first T training period, and predict the processing result data of the target network in the T+1 training period; then use the processing result prediction model to The t1 processing result data corresponding to the target network in the first T training period, and the processing result data of the target network predicted by the processing result prediction model in the T+1 training period are processed, and the target network is predicted to be in the T +2 training period processing result data.
  • the processing result data of the predicted target network in the T+3, T+4...T+n training period can be deduced by analogy, and will not be repeated here.
  • n is a positive integer.
  • the processing results of the target network in the next n training periods can be based on the processing result prediction model and the t1
  • One processing result data is predicted. This can not only quickly predict the processing result data of the target network in the next n training periods, but also greatly reduce the number of times that the target network is trained using the training data, thereby reducing the amount of data processing and saving computing resources and time. Effectively improve the efficiency of determining the processing results of the network.
  • the processing result data is used to indicate the classification accuracy of the target network
  • S represents the classification accuracy of the predicted target network in the T+1 training period
  • w, hp, S_mean, S_sd, S', S" all represent the input items
  • w represents the weight information of the target network, which is each The weight of the convolution kernel of the first layer
  • hp represents the hyperparameter of the target network
  • S_mean is the average value of the classification accuracy of the target network in the first T training periods, which can be sorted in the order of training time in the first T training periods.
  • S_sd is the standard deviation of the processing result data of the previous T training epochs;
  • S' is the training time sequence in the previous T training epochs Sorting in order, the mean value of the change in classification accuracy of the preset number of training periods at the end;
  • S" includes the first T training periods in the order of training time, and the training period with the last one is relatively ranked The classification accuracy improved by the second-to-last training period; and the classification accuracy improved by the second-to-last training period relative to the third-to-last training period.
  • the processing result data is used to indicate the classification accuracy of the target network.
  • the server uses the processing result prediction model to process the processing result data in the processing result set, and predicts the target network in the T+1 training period. After processing the result data, it is detected whether the classification accuracy indicated by the processing result data of the predicted target network in the T+1 training period is less than a preset value, which is, for example, 90%; if it is detected that the target network is The classification accuracy indicated by the processing result data of the T+1 training period is less than the preset value, then sample data is obtained, and the target network is trained with the sample data to adjust the parameters of the target network to improve The classification accuracy of the target network.
  • the sample data may be the same as or different from the above-mentioned training data.
  • the server uses the processing result prediction model to process the processing result data in the processing result set, and after predicting the processing result data of the target network in the T+1 training period, it establishes a communication connection with the server
  • the terminal outputs prompt information including the processing result data of the target network in the T+1 training period, and the prompt information is used to prompt the terminal user whether to use the sample data to train the target network. If the terminal detects the confirmation training instruction input by the terminal user for the prompt information, it sends the confirmation training instruction to the server.
  • the server receives and responds to the confirmation training instruction to obtain sample data, and uses the sample data to train the target network to adjust the parameters of the target network to improve the accuracy of the processing result of the target network.
  • the processing result prediction model is used to process the processing result data in the processing result set, and the processing result data of the target network in the next training period can be predicted, thereby effectively improving the efficiency of determining the processing result of the network.
  • FIG. 2 is a schematic structural diagram of a processing result prediction device based on a prediction model provided by an embodiment of the present application.
  • the processing result prediction device of the embodiment of the present application includes a unit for executing the above processing result prediction method.
  • the processing result prediction apparatus 200 of the embodiment of the present application may include: an acquisition unit 201, a processing unit 202, a construction unit 203, and a training unit 204. among them:
  • the acquiring unit 201 is configured to acquire a processing result set of the target network in the previous T training periods, and the processing result set includes t1 processing result data obtained by processing the training data by the target network and using the processing result prediction
  • the t2 pieces of processing result data obtained by the model predicting the processing result of the target network, the T and t1 are positive integers, and the t2 is a non-negative integer;
  • the processing unit 202 is configured to use the processing result prediction model to process the processing result data in the processing result set to predict the processing result data of the target network in the T+1 training period;
  • the processing result prediction model is constructed based on the network setting information of the target network.
  • the processing result data is used to indicate the classification accuracy of the target network, and the processing unit 202 is specifically used to:
  • the classification accuracy of the target network in the T+1 training period is predicted, and the prediction is obtained
  • the classification accuracy of is determined as the processing result data of the target network in the T+1 training period.
  • the obtaining unit 201 is further configured to obtain network setting information of the target network, where the network setting information includes structural parameters and hyperparameters of the target network;
  • the construction unit 203 is configured to determine target characteristic parameters according to the network setting information, and construct and obtain the processing result prediction model according to the target characteristic parameters.
  • the training unit 204 is configured to use support vector machine regression LIBSVM software to train the processing result prediction model to obtain a trained processing result prediction model;
  • the support vector machine SVM type corresponding to the LIBSVM software is a regression support vector machine v-SVR
  • the kernel function corresponding to the LIBSVM software is a radial basis function RBF.
  • the processing result data is used to indicate the classification accuracy of the target network
  • the processing unit 202 is also used to detect the processing of the predicted target network in the T+1 training period Whether the classification accuracy indicated by the result data is less than the preset value
  • the training unit 204 is triggered to obtain sample data, and use the sample data to compare the The target network is trained.
  • the processing unit 202 is further configured to output prompt information including the processing result data of the target network in the T+1 training period, and the prompt information is used to prompt whether to use sample data to The target network is trained;
  • the training unit 204 is triggered to acquire the sample data, and use the sample data to train the target network.
  • each training period of the target network corresponds to one processing result data
  • the T is equal to the sum of the t1 and the t2.
  • the target network is a neural network.
  • the processing result prediction apparatus 200 can implement part or all of the steps in the processing result prediction method in the embodiment shown in FIG. 1 through the above-mentioned units. It should be understood that the embodiments of the present application are device embodiments corresponding to the method embodiments, and the description of the method embodiments is also applicable to the embodiments of the present application.
  • the processing result prediction model is used to process the processing result data in the processing result set, and the processing result data of the target network in the next training period can be predicted, thereby effectively improving the efficiency of determining the processing result of the network.
  • FIG. 3 is a schematic structural diagram of a server provided by an embodiment of the present application.
  • the server is used to execute the aforementioned method for predicting processing results based on the predictive model.
  • the server 300 in this embodiment may include: one or more processors 301 and a memory 302.
  • the server may further include one or more communication interfaces 303.
  • the above-mentioned processor 301, communication interface 303, and memory 302 may be connected through a bus 304, or may be connected in other ways, as illustrated in FIG. 3 by way of a bus.
  • the processor 301 may be a central processing unit (Central Processing Unit, CPU), and the processor may also be other general-purpose processors, digital signal processors (Digital Signal Processor, DSP), and application specific integrated circuits (Application Specific Integrated Circuits). Circuit, ASIC), ready-made programmable gate array (Field-Programmable Gate Array, FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc.
  • the general-purpose processor may be a microprocessor or the processor may also be any conventional processor or the like.
  • the communication interface 303 may be used to exchange information or signaling, and to receive and transmit signals.
  • the communication interface 303 may include a receiver and a transmitter for communicating with other devices.
  • the memory 302 may mainly include a storage program area and a storage data area, where the storage program area can store an operating system and a storage program required by at least one function (such as a text storage function, a location storage function, etc.); the storage data area can store Data (such as image data, text data) created according to the use of the server, etc., and may include application storage programs, etc.
  • the memory 302 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one magnetic disk storage device, a flash memory device, or other volatile solid-state storage devices.
  • the memory 302 is also used to store program instructions.
  • the processor 301 can call the program instructions stored in the memory 302 to implement the method for predicting processing results based on a predictive model as shown in the embodiment of the present application.
  • the processor 301 may be used to call the program instructions to perform the following steps: obtain a processing result set of the target network in the previous T training periods, the processing result set including t1 processing obtained by the target network processing the training data The result data, and t2 processing result data obtained by predicting the processing result of the target network using the processing result prediction model, the T and t1 are positive integers, and the t2 is a non-negative integer; using the processing result prediction model Process the processing result data in the processing result set to predict the processing result data of the target network in the T+1 training period; wherein the processing result prediction model is based on the network settings of the target network Information is constructed.
  • the processing result data is used to indicate the classification accuracy rate of the target network
  • the processor is specifically configured to call the program instructions to execute the following steps: call the processing result prediction model to calculate the processing The first mean value of the classification accuracy from the TMth training period to the Tth training period in the result set, where M is an integer greater than 1 and less than T; calculate the TMth training period to the Tth in the processing result set The second mean value of the variation of the classification accuracy rate during the training period; calculating the first difference between the classification accuracy rates of the T-th training period and the T-1th training period in the processing result set, and calculating the processing result The second difference between the classification accuracy of the T-1 training period and the T-2 training period in the set; according to the first mean, the second mean, the first difference and the first Two difference values, predicting the classification accuracy rate of the target network in the T+1 training period, and determining the predicted classification accuracy rate as the processing result data of the target network in the T+1 training period.
  • the processor 301 may also call the program instructions to perform the following steps: obtain network setting information of the target network, where the network setting information includes structural parameters and hyperparameters of the target network;
  • the network setting information determines the target characteristic parameter, and constructs and obtains the processing result prediction model according to the target characteristic parameter.
  • the processor 301 calls the program instructions to execute the prediction model of the processing result constructed according to the target characteristic parameters, it is further configured to perform the following steps: use the support vector machine regression LIBSVM software to perform the The processing result prediction model is trained to obtain a trained processing result prediction model; wherein, the support vector machine SVM type corresponding to the LIBSVM software is a regression support vector machine v-SVR, and the kernel function corresponding to the LIBSVM software is a path To the basis function RBF.
  • the processing result data is used to indicate the classification accuracy rate of the target network
  • the processor 301 is calling the program instructions to execute the processing result prediction model for the processing result set
  • the processing result data is processed, and after the processing result data of the target network in the T+1 training period is predicted, it is also used to perform the following steps: detecting the predicted target network in the T+1 training period Whether the classification accuracy rate indicated by the processing result data is less than a preset value; if the classification accuracy rate indicated by the processing result data of the target network in the T+1 training period is less than the preset value, then sample data is obtained, And use the sample data to train the target network.
  • the processor 301 calls the program instructions to execute the processing result prediction model using the processing result prediction model to process the processing result data in the processing result set, and it is predicted that the target network is at the T+1 th
  • the processor 301 is also used to perform the following steps: output through the communication interface 303 prompt information including the processing result data of the target network in the T+1 training period, and the prompt information is used for Prompt whether to use sample data to train the target network; if a confirmation training instruction input for the prompt information is detected, obtain the sample data and use the sample data to train the target network.
  • each training period of the target network corresponds to one processing result data
  • the T is equal to the sum of the t1 and the t2.
  • the target network is a neural network.
  • the processor 301 described in the embodiment of the present application, etc. can perform the implementation described in the method embodiment shown in FIG. 1 above, and can also perform the implementation of the units described in FIG. Way, I won’t go into details here.
  • the embodiment of the application uses the processing result prediction model to process the processing result data in the processing result set, which can predict the processing result data of the target network in the next training period, thereby effectively improving the efficiency of determining the processing result of the network.
  • An embodiment of the present application also provides a computer-readable storage medium that stores a computer program, and when the computer program is executed by a processor, the processing result prediction described in the embodiment corresponding to FIG. 1 can be realized. Part or all of the steps in the method can also realize the function of the processing result prediction apparatus in the embodiment shown in FIG. 2 of this application, and can also realize the function of the server in the embodiment shown in FIG. 3 of this application, which will not be repeated here.
  • the embodiments of the present application also provide a computer program product containing instructions, which when run on a computer, cause the computer to execute some or all of the steps in the above-mentioned processing result prediction method.
  • the storage medium may be the processing result prediction device described in the foregoing embodiment or the internal storage unit of the server, such as the processing result prediction device or the hard disk or memory of the server.
  • the storage medium may also be the processing result prediction device or an external storage device of the server, such as the processing result prediction or plug-in hard disk equipped on the server, a smart memory card (Smart Media Card, SMC), a secure digital ( Secure Digital, SD card, Flash Card, etc.
  • the size of the sequence number of the above-mentioned processes does not mean the order of execution.
  • the execution order of each process should be determined by its function and internal logic, and should not correspond to the implementation process of the embodiments of the present application. Constitute any limitation.

Abstract

一种基于预测模型的处理结果预测方法、装置及服务器，该方法包括：服务器获取目标网络在前T个训练时期的处理结果集合，所述处理结果集合包括所述目标网络针对训练数据进行处理得到的t1个处理结果数据、以及利用处理结果预测模型针对所述目标网络进行处理结果预测得到的t2个处理结果数据（S101），所述T、t1为正整数，所述t2为非负整数；所述服务器利用所述处理结果预测模型对所述处理结果集合中的处理结果数据进行处理，预测得到所述目标网络在第T+1个训练时期的处理结果数据，所述处理结果预测模型是基于所述目标网络的网络设置信息构建得到的（S102）。采用所述方法，可以对网络的处理结果进行预测，以有效提高确定网络的处理结果的效率。

Description

基于预测模型的处理结果预测方法、装置及服务器
本申请要求于2019年02月15日提交中国专利局、申请号为201910121267.9、申请名称为“基于预测模型的处理结果预测方法、装置及服务器”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请涉及计算机技术领域，尤其涉及一种基于预测模型的处理结果预测方法、装置及服务器。
背景技术
在设计网络模型时，无论是人工设计网络模型，还是计算机自动设计网络模型(或者说网络架构搜索)，通常需要将大量的样本数据输入网络模型中进行处理后，才能得到网络模型的处理结果，处理结果例如可以用于指示网络模型的分类准确率等。但上述确定网络模型处理结果的方式处理数据量大，需要大量的计算资源和处理时间，导致确定网络模型处理结果的效率低。因此，如何有效提高确定网络模型处理结果的效率是有待解决的问题。
发明内容
本申请实施例提供提供一种基于预测模型的处理结果预测方法、装置及服务器，可以对网络的处理结果进行预测，从而有效提高确定网络的处理结果的效率。
第一方面，本申请实施例提供了一种基于预测模型的处理结果预测方法，包括：
获取目标网络在前T个训练时期的处理结果集合，所述处理结果集合包括所述目标网络针对训练数据进行处理得到的t1个处理结果数据、以及利用处理结果预测模型针对所述目标网络进行处理结果预测得到的t2个处理结果数据，所述T、t1为正整数，所述t2为非负整数；
利用所述处理结果预测模型对所述处理结果集合中的处理结果数据进行处理，预测得到所述目标网络在第T+1个训练时期的处理结果数据；
其中，所述处理结果预测模型是基于所述目标网络的网络设置信息构建得到的。
第二方面，本申请实施例提供了一种基于预测模型的处理结果预测装置，所述装置包括：
获取单元，用于获取目标网络在前T个训练时期的处理结果集合，所述处理结果集合包括所述目标网络针对训练数据进行处理得到的t1个处理结果数据、以及利用处理结果预测模型针对所述目标网络进行处理结果预测得到的t2个处理结果数据，所述T、t1为正整数，所述t2为非负整数；
处理单元，用于利用所述处理结果预测模型对所述处理结果集合中的处理结果数据进行处理，预测得到所述目标网络在第T+1个训练时期的处理结果数据；
其中，所述处理结果预测模型是基于所述目标网络的网络设置信息构建得到的。
第三方面，本申请实施例提供了一种服务器，包括处理器、通信接口和存储器，所述处理器、所述通信接口和所述存储器相互连接，其中，所述存储器用于存储计算机程序，所述计算机程序包括程序指令，所述处理器被配置用于调用所述程序指令，执行以下步骤：
获取目标网络在前T个训练时期的处理结果集合，所述处理结果集合包括所述目标网络针对训练数据进行处理得到的t1个处理结果数据、以及利用处理结果预测模型针对所述目标网络进行处理结果预测得到的t2个处理结果数据，所述T、t1为正整数，所述t2为非负整数；
利用所述处理结果预测模型对所述处理结果集合中的处理结果数据进行处理，预测得到所述目标网络在第T+1个训练时期的处理结果数据；
其中，所述处理结果预测模型是基于所述目标网络的网络设置信息构建得到的。
第四方面，本申请实施例提供了一种计算机可读存储介质，所述计算机可读存储介质存储有计算机程序，所述计算机程序包括程序指令，所述程序指令当被处理器执行时使所述处理器执行上述第一方面任一项所述的基于预测模型的处理结果预测方法。
本申请实施例利用处理结果预测模型对处理结果集合中的处理结果数据进行处理，可预测得到目标网络在下一个训练时期的处理结果数据，从而有效提高确定网络的处理结果的效率。
附图说明
为了更清楚地说明本申请实施例技术方案，下面将对实施例描述中所需要使用的附图进行说明。
图1是本申请实施例提供的一种基于预测模型的处理结果预测方法的流程示意图；
图2是本申请实施例提供的一种基于预测模型的处理结果预测装置的结构示意图；
图3是本申请实施例提供的一种服务器的结构示意图。
具体实施方式
下面将结合本申请实施例中的附图，对本申请实施例中的技术方案进行清楚、完整地描述。
请参见图1，图1是本申请实施例提供的一种基于预测模型的处理结果预测方法的流程示意图。具体的，如图1所示，该处理结果预测方法可以包括以下步骤：
S101、服务器获取目标网络在前T个训练时期的处理结果集合。
本申请实施例中，目标网络在前T个训练时期的处理结果集合包括：目标网络针对训练数据进行处理得到的t1个处理结果数据、以及利用处理结果预测模型针对目标网络进行处理结果预测得到的t2个处理结果数据。其中，T和t1均为正整数，即大于0的整数；t2为非负整数，即正整数和0。当t2为0时，该处理结果集合中只包括目标网络针对训练数据进行处理得到的t1个处理结果数据，不包括利用处理结果预测模型针对目标网络进行处理结果预测得到的处理结果数据。
在一实施方式中，目标网络的每一个训练时期对应一个处理结果数据，则T等于t1与t2的和。目标网络在前T个训练时期中的t1个训练时期，针对训 练数据进行处理得到t1个处理结果数据。可以利用y t1表示目标网络在该t1个训练时期的处理结果数据，对于训练了t1个时期的目标网络，可以记录下y(t1)＝y 1,y 2,y 3,…,y t1。服务器利用处理结果预测模型针对目标网络进行处理结果预测，预测得到t2个训练时期对应的t2个处理结果数据。可以利用x t2表示利用处理结果预测模型预测得到的目标网络在该t2个训练时期的处理结果数据，可以记录下x(t2)＝x 1,x 2,x 3,…,x t2。其中，该t2个训练时期对应的时间晚于该t1个训练时期对应的时间。该t2个处理结果数据是处理结果预测模型基于该t1个处理结果数据预测得到的。
在一实施方式中，处理结果数据可以用于指示目标网络在训练时期的表现结果，该表现结果可以是分类准确率。目标网络在训练时期针对训练数据进行分类处理，得到分类处理结果；然后基于该分类处理结果确定出目标网络的分类准确率，并将确定出的分类准确率确定为目标网络在训练时期对应的处理结果数据。服务器利用处理结果预测模型针对目标网络进行分类准确率预测，并将预测得到的分类准确率确定为目标网络在训练时期对应的处理结果数据。
本申请实施例中，处理结果预测模型是基于目标网络的网络设置信息构建得到的。具体地，服务器获取目标网络的网络设置信息，该网络设置信息包括目标网络的结构参数和超参数。在一实施方式中，目标网络可以是神经网络，该结构参数包括目标网络的卷积层参数和全连接层参数等，具体可以包括每一卷积层的权重参数、卷积层的层数、每一卷积层的神经元个数等。该超参数包括目标网络的学习速率、学习步长、权重衰减、输入网络图片尺寸、是否采用数据增强(例如以0或1表示)等。服务器然后根据目标网络的网络设置信息确定出目标特征参数，并根据该目标特征参数构建得到目标网络对应的处理结果预测模型。在一实施方式中，服务器可以将该网络设置信息包括的结构参数以及超参数直接确定为处理结果预测模型对应的目标特征参数。
进一步地，服务器根据该目标特征参数构建得到目标网络对应的处理结果预测模型之后，可以采用回归型支持向量机v-SVR模型对该处理结果预测模型进行训练，得到训练后的处理结果预测模型。采用回归型支持向量机v-SVR模型可以快速完成针对处理结果预测模型的训练。具体地，服务器可以利用支持向量机回归LIBSVM软件对该处理结果预测模型进行训练，得到训练后的 处理结果预测模型。其中，可以指定LIBSVM软件对应的支持向量机SVM类型为回归型支持向量机v-SVR，指定LIBSVM软件对应的核函数为径向基函数(Radial Basis Function，RBF)；LIBSVM软件对应的类别数据则可以根据目标网络的网络设置信息确定得出。在一实施方式中，也可以利用贝叶斯方法对该处理结果预测模型进行训练，得到训练后的处理结果预测模型。
S102、所述服务器利用处理结果预测模型对所述处理结果集合中的处理结果数据进行处理，预测得到所述目标网络在第T+1个训练时期的处理结果数据。
本申请实施例中，服务器利用处理结果预测模型对处理结果集合中的目标网络针对训练数据进行处理得到的t1个处理结果数据、以及利用处理结果预测模型针对目标网络进行处理结果预测得到的t2个处理结果数据进行处理，预测得到目标网络在第T+1个训练时期的处理结果数据。
在一实施方式中，处理结果数据用于指示目标网络的分类准确率。服务器调用处理结果预测模型计算处理结果集合中第T-M个训练时期到第T个训练时期的分类准确率的第一均值，所述M为大于1小于T的整数，计算处理结果集合中第T-M个训练时期到第T个训练时期的分类准确率的变化值的第二均值，计算处理结果集合中第T个训练时期与第T-1个训练时期的分类准确率的第一差值，以及计算处理结果集合中第T-1个训练时期与第T-2个训练时期的分类准确率的第二差值，以及根据计算出的第一均值、第二均值、第一差值、第二差值以及第T个训练时期的分类准确率，预测得到目标网络在第T+1个训练时期的分类准确率，并将预测得到的分类准确率确定为目标网络在第T+1个训练时期的处理结果数据。其中，第T-M个训练时期到第T个训练时期的分类准确率的变化值可以是指：第T-M个训练时期到第T个训练时期中各相邻训练时期的分类准确率的差值，或者为各相邻训练时期的分类准确率的差值乘以相应权重后得到的数值。处于后面的相邻训练时期的分类准确率差值对应的权重大于处于前面的相邻训练时期的分类准确率差值对应的权重；由于处于前面的相邻训练时期的分类准确率差值相对较大，故为提高后续预测的准确性，可以处于前面的相邻训练时期的分类准确率差值对应的权重设置为小于1的数值。
举例来说，假设t2所指示的数值为0，则表示该处理结果集合中只包括目标网络针对训练数据进行处理得到的t1个处理结果数据，前T个训练时期均是利用训练数据对目标网络进行训练的时期。服务器可以利用处理结果预测模型对目标网络在前T个训练时期对应的t1个处理结果数据进行处理，预测得到目标网络在第T+1个训练时期的处理结果数据；然后利用处理结果预测模型对目标网络在前T个训练时期对应的的t1个处理结果数据、以及利用处理结果预测模型预测得到的目标网络在第T+1个训练时期的处理结果数据进行处理，预测得到目标网络在第T+2个训练时期的处理结果数据。预测目标网络在第T+3、T+4......T+n个训练时期的处理结果数据则可以此类推，此处不再赘述。n为正整数。采用上述方式，只需目标网络在前T个训练时期中针对训练数据进行处理，得到t1个处理结果数据；对于目标网络在后n个训练时期的处理结果则可以基于处理结果预测模型以及该t1个处理结果数据预测得到。这样不仅可以快速预测得到目标网络在后n个训练时期的处理结果数据，还可以大大减少利用训练数据对目标网络进行训练的次数，从而可以降低数据处理量，达到节省计算资源和时间的目的，有效提高确定网络的处理结果的效率。
在一实施方式中，处理结果数据用于指示目标网络的分类准确率，处理结果预测模型所指示的表达式如下：S＝f(w，hp，S_mean，S_sd，S’，S”)。其中，S表示预测的目标网络在第T+1个训练时期的分类准确率，w，hp，S_mean，S_sd，S’，S”均表示输入项；w代表目标网络的权重信息，为目标网络每一层的卷积核权重；hp代表目标网络的超参数；S_mean为目标网络在前T个训练时期的分类准确率的均值，具体可以是前T个训练时期中按照训练时间的先后顺序排序，排在最后的预设数量(例如3)个训练时期的分类准确率的均值；S_sd为前T个训练时期的处理结果数据的的标准差；S’为前T个训练时期中按照训练时间先后顺序排序，排在最后的预设数量个训练时期的分类准确率的变化值的均值；S”包括前T个训练时期中按照训练时间先后顺序排序，排在倒数第一的训练时期相对排在倒数第二的训练时期所提升的分类准确率；以及排在倒数第二的训练时期相对排在倒数第三的训练时期所提升的分类准确率。
在一实施方式中，处理结果数据用于指示目标网络的分类准确率，服务器利用处理结果预测模型对处理结果集合中的处理结果数据进行处理，预测得到 目标网络在第T+1个训练时期的处理结果数据之后，检测预测得到的目标网络在第T+1个训练时期的处理结果数据所指示的分类准确率是否小于预设数值，该预设数值例如是90％；若检测到目标网络在第T+1个训练时期的处理结果数据所指示的分类准确率小于该预设数值，则获取样本数据，并利用该样本数据对目标网络进行训练，以对目标网络的参数进行调整，以提高目标网络的分类准确率。其中，该样本数据可以与上述训练数据相同，也可以与上述训练数据不同。
在另一实施方式中，服务器利用处理结果预测模型对处理结果集合中的处理结果数据进行处理，预测得到目标网络在第T+1个训练时期的处理结果数据之后，通过与服务器建立通信连接的终端输出包括目标网络在第T+1个训练时期的处理结果数据的提示信息，该提示信息用于提示终端用户是否利用样本数据对目标网络进行训练。终端若检测到终端用户针对该提示信息输入的确认训练指令，则向服务器发送该确认训练指令。服务器接收并响应该确认训练指令获取样本数据，并利用该样本数据对目标网络进行训练，以对目标网络的参数进行调整，以提高目标网络的处理结果的准确率。
本申请实施例利用处理结果预测模型对处理结果集合中的处理结果数据进行处理，可预测得到目标网络在下一个训练时期的处理结果数据，从而有效提高确定网络的处理结果的效率。
请参见图2，图2是本申请实施例提供的一种基于预测模型的处理结果预测装置的结构示意图。本申请实施例的处理结果预测装置包括用于执行上述处理结果预测方法的单元。具体的，本申请实施例的处理结果预测装置200可以包括：获取单元201、处理单元202、构建单元203和训练单元204。其中：
所述获取单元201，用于获取目标网络在前T个训练时期的处理结果集合，所述处理结果集合包括所述目标网络针对训练数据进行处理得到的t1个处理结果数据、以及利用处理结果预测模型针对所述目标网络进行处理结果预测得到的t2个处理结果数据，所述T、t1为正整数，所述t2为非负整数；
所述处理单元202，用于利用所述处理结果预测模型对所述处理结果集合中的处理结果数据进行处理，预测得到所述目标网络在第T+1个训练时期的处理结果数据；
其中，所述处理结果预测模型是基于所述目标网络的网络设置信息构建得到的。
在一实施方式中，所述处理结果数据用于指示所述目标网络的分类准确率，所述处理单元202具体用于：
调用所述处理结果预测模型计算所述处理结果集合中第T-M个训练时期到第T个训练时期的分类准确率的第一均值，所述M为大于1小于T的整数；
计算所述处理结果集合中第T-M个训练时期到第T个训练时期的分类准确率的变化值的第二均值；
计算所述处理结果集合中第T个训练时期与第T-1个训练时期的分类准确率的第一差值，以及计算所述处理结果集合中第T-1个训练时期与第T-2个训练时期的分类准确率的第二差值；
根据所述第一均值、所述第二均值、所述第一差值和所述第二差值，预测得到所述目标网络在第T+1个训练时期的分类准确率，并将预测得到的分类准确率确定为所述目标网络在第T+1个训练时期的处理结果数据。
在一实施方式中，所述获取单元201，还用于获取所述目标网络的网络设置信息，所述网络设置信息包括所述目标网络的结构参数和超参数；
所述构建单元203，用于根据所述网络设置信息确定目标特征参数，并根据所述目标特征参数构建得到所述处理结果预测模型。
在一实施方式中，所述训练单元204，用于利用支持向量机回归LIBSVM软件对所述处理结果预测模型进行训练，得到训练后的处理结果预测模型；
其中，所述LIBSVM软件对应的支持向量机SVM类型为回归型支持向量机v-SVR，所述LIBSVM软件对应的核函数为径向基函数RBF。
在一实施方式中，所述处理结果数据用于指示所述目标网络的分类准确率，所述处理单元202，还用于检测预测得到的所述目标网络在第T+1个训练时期的处理结果数据所指示的分类准确率是否小于预设数值；
若所述目标网络在第T+1个训练时期的处理结果数据所指示的分类准确率小于所述预设数值，则触发所述训练单元204获取样本数据，并利用所述样本数据对所述目标网络进行训练。
在一实施方式中，所述处理单元202，还用于输出包括所述目标网络在第 T+1个训练时期的处理结果数据的提示信息，所述提示信息用于提示是否利用样本数据对所述目标网络进行训练；
若检测到针对所述提示信息输入的确认训练指令，则触发所述训练单元204获取所述样本数据，并利用所述样本数据对所述目标网络进行训练。
在一实施方式中，所述目标网络的每一个训练时期对应一个处理结果数据，所述T等于所述t1与所述t2的和。
在一实施方式中，所述目标网络为神经网络。
具体的，该处理结果预测装置200可通过上述单元实现上述图1所示实施例中的处理结果预测方法中的部分或全部步骤。应理解，本申请实施例是对应方法实施例的装置实施例，对方法实施例的描述，也适用于本申请实施例。
本申请实施例利用处理结果预测模型对处理结果集合中的处理结果数据进行处理，可预测得到目标网络在下一个训练时期的处理结果数据，从而有效提高确定网络的处理结果的效率。
请参见图3，图3是本申请实施例提供的一种服务器的结构示意图。该服务器用于执行上述基于预测模型的处理结果预测方法。如图3所示，本实施例中的服务器300可以包括：一个或多个处理器301和存储器302。可选的，该服务器还可包括一个或多个通信接口303。上述处理器301、通信接口303和存储器302可通过总线304连接，或者可以通过其他方式连接，图3中以总线方式进行示例说明。
其中，所述处理器301可以是中央处理单元(Central Processing Unit，CPU)，该处理器还可以是其他通用处理器、数字信号处理器(Digital Signal Processor，DSP)、专用集成电路(Application Specific Integrated Circuit，ASIC)、现成可编程门阵列(Field-Programmable Gate Array，FPGA)或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件等。通用处理器可以是微处理器或者该处理器也可以是任何常规的处理器等。
所述通信接口303可用于收发信息或信令的交互，以及信号的接收和传递，通信接口303可包括接收器和发射器，用于与其他设备进行通信。所述存储器302可主要包括存储程序区和存储数据区，其中，存储程序区可存储操作系统、至少一个功能所需的存储程序(比如文字存储功能、位置存储功能等)；存储 数据区可存储根据服务器的使用所创建的数据(比如图像数据、文字数据)等，并可以包括应用存储程序等。此外，存储器302可以包括高速随机存取存储器，还可以包括非易失性存储器，例如至少一个磁盘存储器件、闪存器件、或其他易失性固态存储器件。
所述存储器302还用于存储程序指令。所述处理器301可以调用上述存储器302存储的程序指令，实现如本申请实施例所示的基于预测模型的处理结果预测方法。
其中，处理器301可用于调用所述程序指令执行以下步骤：获取目标网络在前T个训练时期的处理结果集合，所述处理结果集合包括所述目标网络针对训练数据进行处理得到的t1个处理结果数据、以及利用处理结果预测模型针对所述目标网络进行处理结果预测得到的t2个处理结果数据，所述T、t1为正整数，所述t2为非负整数；利用所述处理结果预测模型对所述处理结果集合中的处理结果数据进行处理，预测得到所述目标网络在第T+1个训练时期的处理结果数据；其中，所述处理结果预测模型是基于所述目标网络的网络设置信息构建得到的。
在一实施方式中，所述处理结果数据用于指示所述目标网络的分类准确率，所述处理器具体用于调用所述程序指令执行以下步骤：调用所述处理结果预测模型计算所述处理结果集合中第T-M个训练时期到第T个训练时期的分类准确率的第一均值，所述M为大于1小于T的整数；计算所述处理结果集合中第T-M个训练时期到第T个训练时期的分类准确率的变化值的第二均值；计算所述处理结果集合中第T个训练时期与第T-1个训练时期的分类准确率的第一差值，以及计算所述处理结果集合中第T-1个训练时期与第T-2个训练时期的分类准确率的第二差值；根据所述第一均值、所述第二均值、所述第一差值和所述第二差值，预测得到所述目标网络在第T+1个训练时期的分类准确率，并将预测得到的分类准确率确定为所述目标网络在第T+1个训练时期的处理结果数据。
在一实施方式中，处理器301还可调用所述程序指令执行以下步骤：获取所述目标网络的网络设置信息，所述网络设置信息包括所述目标网络的结构参数和超参数；根据所述网络设置信息确定目标特征参数，并根据所述目标特征 参数构建得到所述处理结果预测模型。
在一实施方式中，处理器301在调用所述程序指令执行所述根据所述目标特征参数构建得到所述处理结果预测模型之后，还用于执行以下步骤：利用支持向量机回归LIBSVM软件对所述处理结果预测模型进行训练，得到训练后的处理结果预测模型；其中，所述LIBSVM软件对应的支持向量机SVM类型为回归型支持向量机v-SVR，所述LIBSVM软件对应的核函数为径向基函数RBF。
在一实施方式中，所述处理结果数据用于指示所述目标网络的分类准确率，处理器301在调用所述程序指令执行所述利用所述处理结果预测模型对所述处理结果集合中的处理结果数据进行处理，预测得到所述目标网络在第T+1个训练时期的处理结果数据之后，还用于执行以下步骤：检测预测得到的所述目标网络在第T+1个训练时期的处理结果数据所指示的分类准确率是否小于预设数值；若所述目标网络在第T+1个训练时期的处理结果数据所指示的分类准确率小于所述预设数值，则获取样本数据，并利用所述样本数据对所述目标网络进行训练。
在一实施方式中，处理器301在调用所述程序指令执行所述利用所述处理结果预测模型对所述处理结果集合中的处理结果数据进行处理，预测得到所述目标网络在第T+1个训练时期的处理结果数据之后，还用于执行以下步骤：通过所述通信接口303输出包括所述目标网络在第T+1个训练时期的处理结果数据的提示信息，所述提示信息用于提示是否利用样本数据对所述目标网络进行训练；若检测到针对所述提示信息输入的确认训练指令，则获取所述样本数据，并利用所述样本数据对所述目标网络进行训练。
在一实施方式中，所述目标网络的每一个训练时期对应一个处理结果数据，所述T等于所述t1与所述t2的和。
在一实施方式中，所述目标网络为神经网络。
具体实现中，本申请实施例中所描述的处理器301等可执行上述图1所示的方法实施例中所描述的实现方式，也可执行本申请实施例图2所描述的各单元的实现方式，此处不赘述。
本申请实施例利用处理结果预测模型对处理结果集合中的处理结果数据 进行处理，可预测得到目标网络在下一个训练时期的处理结果数据，从而有效提高确定网络的处理结果的效率。
本申请实施例还提供了一种计算机可读存储介质，所述计算机可读存储介质存储有计算机程序，所述计算机程序被处理器执行时可实现图1所对应实施例中描述的处理结果预测方法中的部分或全部步骤，也可实现本申请图2所示实施例的处理结果预测装置的功能，也可实现本申请图3所示实施例的服务器的功能，此处不赘述。
本申请实施例还提供了一种包含指令的计算机程序产品，当其在计算机上运行时，使得计算机执行上述处理结果预测方法中的部分或全部步骤。
所述存储介质可以是前述实施例所述的处理结果预测装置或者服务器的内部存储单元，例如处理结果预测装置或者服务器的硬盘或内存。所述存储介质也可以是所述处理结果预测装置或者服务器的外部存储设备，例如所述处理结果预测或者服务器上配备的插接式硬盘，智能存储卡(Smart Media Card，SMC)，安全数字(Secure Digital，SD)卡，闪存卡(Flash Card)等。
在本申请中，术语“和/或”，仅仅是一种描述关联对象的关联关系，表示可以存在三种关系，例如，A和/或B，可以表示：单独存在A，同时存在A和B，单独存在B这三种情况。另外，本文中字符“/”，一般表示前后关联对象是一种“或”的关系。
在本申请的各种实施例中，上述各过程的序号的大小并不意味着执行顺序的先后，各过程的执行顺序应以其功能和内在逻辑确定，而不应对本申请实施例的实施过程构成任何限定。
以上所述，仅为本申请的部分实施方式，但本申请的保护范围并不局限于此，任何熟悉本技术领域的技术人员在本申请揭露的技术范围内，可轻易想到各种等效的修改或替换，这些修改或替换都应涵盖在本申请的保护范围之内。

Claims (20)

  1. 一种基于预测模型的处理结果预测方法，其特征在于，所述方法包括：
    获取目标网络在前T个训练时期的处理结果集合，所述处理结果集合包括所述目标网络针对训练数据进行处理得到的t1个处理结果数据、以及利用处理结果预测模型针对所述目标网络进行处理结果预测得到的t2个处理结果数据，所述T、t1为正整数，所述t2为非负整数；
    利用所述处理结果预测模型对所述处理结果集合中的处理结果数据进行处理，预测得到所述目标网络在第T+1个训练时期的处理结果数据；
    其中，所述处理结果预测模型是基于所述目标网络的网络设置信息构建得到的。
  2. 根据权利要求1所述的方法，其特征在于，所述处理结果数据用于指示所述目标网络的分类准确率，所述利用所述处理结果预测模型对所述处理结果集合中的处理结果数据进行处理，预测得到所述目标网络在第T+1个训练时期的处理结果数据，包括：
    调用所述处理结果预测模型计算所述处理结果集合中第T-M个训练时期到第T个训练时期的分类准确率的第一均值，所述M为大于1小于T的整数；
    计算所述处理结果集合中第T-M个训练时期到第T个训练时期的分类准确率的变化值的第二均值；
    计算所述处理结果集合中第T个训练时期与第T-1个训练时期的分类准确率的第一差值，以及计算所述处理结果集合中第T-1个训练时期与第T-2个训练时期的分类准确率的第二差值；
    根据所述第一均值、所述第二均值、所述第一差值和所述第二差值，预测得到所述目标网络在第T+1个训练时期的分类准确率，并将预测得到的分类准确率确定为所述目标网络在第T+1个训练时期的处理结果数据。
  3. 根据权利要求1或2所述的方法，其特征在于，所述方法还包括：
    获取所述目标网络的网络设置信息，所述网络设置信息包括所述目标网络的结构参数和超参数；
    根据所述网络设置信息确定目标特征参数，并根据所述目标特征参数构建得到所述处理结果预测模型。
  4. 根据权利要求3所述的方法，其特征在于，所述根据所述目标特征参数构建得到所述处理结果预测模型之后，所述方法还包括：
    利用支持向量机回归LIBSVM软件对所述处理结果预测模型进行训练，得到训练后的处理结果预测模型；
    其中，所述LIBSVM软件对应的支持向量机SVM类型为回归型支持向量机v-SVR，所述LIBSVM软件对应的核函数为径向基函数RBF。
  5. 根据权利要求1至4中任一项所述的方法，其特征在于，所述处理结果数据用于指示所述目标网络的分类准确率，所述利用所述处理结果预测模型对所述处理结果集合中的处理结果数据进行处理，预测得到所述目标网络在第T+1个训练时期的处理结果数据之后，所述方法还包括：
    检测预测得到的所述目标网络在第T+1个训练时期的处理结果数据所指示的分类准确率是否小于预设数值；
    若所述目标网络在第T+1个训练时期的处理结果数据所指示的分类准确率小于所述预设数值，则获取样本数据，并利用所述样本数据对所述目标网络进行训练。
  6. 根据权利要求1至4中任一项所述的方法，其特征在于，所述利用所述处理结果预测模型对所述处理结果集合中的处理结果数据进行处理，预测得到所述目标网络在第T+1个训练时期的处理结果数据之后，所述方法还包括：
    输出包括所述目标网络在第T+1个训练时期的处理结果数据的提示信息，所述提示信息用于提示是否利用样本数据对所述目标网络进行训练；
    若检测到针对所述提示信息输入的确认训练指令，则获取所述样本数据，并利用所述样本数据对所述目标网络进行训练。
  7. 根据权利要求1至4中任一项所述的方法，其特征在于，所述目标网络的每一个训练时期对应一个处理结果数据，所述T等于所述t1与所述t2的和。
  8. 根据权利要求1至4中任一项所述的方法，其特征在于，所述目标网络为神经网络。
  9. 一种基于预测模型的处理结果预测装置，其特征在于，所述装置包括：
    获取单元，用于获取目标网络在前T个训练时期的处理结果集合，所述处 理结果集合包括所述目标网络针对训练数据进行处理得到的t1个处理结果数据、以及利用处理结果预测模型针对所述目标网络进行处理结果预测得到的t2个处理结果数据，所述T、t1为正整数，所述t2为非负整数；
    处理单元，用于利用所述处理结果预测模型对所述处理结果集合中的处理结果数据进行处理，预测得到所述目标网络在第T+1个训练时期的处理结果数据；
    其中，所述处理结果预测模型是基于所述目标网络的网络设置信息构建得到的。
  10. 根据权利要求9所述的装置，其特征在于，所述处理结果数据用于指示所述目标网络的分类准确率，所述处理单元具体用于：
    调用所述处理结果预测模型计算所述处理结果集合中第T-M个训练时期到第T个训练时期的分类准确率的第一均值，所述M为大于1小于T的整数；
    计算所述处理结果集合中第T-M个训练时期到第T个训练时期的分类准确率的变化值的第二均值；
    计算所述处理结果集合中第T个训练时期与第T-1个训练时期的分类准确率的第一差值，以及计算所述处理结果集合中第T-1个训练时期与第T-2个训练时期的分类准确率的第二差值；
    根据所述第一均值、所述第二均值、所述第一差值和所述第二差值，预测得到所述目标网络在第T+1个训练时期的分类准确率，并将预测得到的分类准确率确定为所述目标网络在第T+1个训练时期的处理结果数据。
  11. 根据权利要求9或10所述的装置，其特征在于，所述获取单元，还用于获取所述目标网络的网络设置信息，所述网络设置信息包括所述目标网络的结构参数和超参数；
    所述装置还包括构建单元，所述构建单元用于根据所述网络设置信息确定目标特征参数，并根据所述目标特征参数构建得到所述处理结果预测模型。
  12. 根据权利要求11所述的装置，其特征在于，所述装置还包括：
    训练单元，用于利用支持向量机回归LIBSVM软件对所述处理结果预测模型进行训练，得到训练后的处理结果预测模型；
    其中，所述LIBSVM软件对应的支持向量机SVM类型为回归型支持向量 机v-SVR，所述LIBSVM软件对应的核函数为径向基函数RBF。
  13. 根据权利要求9至12中任一项所述的装置，其特征在于，所述处理结果数据用于指示所述目标网络的分类准确率；
    所述处理单元，还用于检测预测得到的所述目标网络在第T+1个训练时期的处理结果数据所指示的分类准确率是否小于预设数值；
    所述装置还包括训练单元，所述处理单元，还用于若所述目标网络在第T+1个训练时期的处理结果数据所指示的分类准确率小于所述预设数值，则触发所述训练单元获取样本数据，并利用所述样本数据对所述目标网络进行训练。
  14. 根据权利要求9至12中任一项所述的装置，其特征在于，所述处理单元，还用于输出包括所述目标网络在第T+1个训练时期的处理结果数据的提示信息，所述提示信息用于提示是否利用样本数据对所述目标网络进行训练；
    所述装置还包括训练单元，所述处理单元，还用于若检测到针对所述提示信息输入的确认训练指令，则触发所述训练单元获取所述样本数据，并利用所述样本数据对所述目标网络进行训练。
  15. 根据权利要求9至12中任一项所述的装置，其特征在于，所述目标网络的每一个训练时期对应一个处理结果数据，所述T等于所述t1与所述t2的和。
  16. 根据权利要求9至12中任一项所述的装置，其特征在于，所述目标网络为神经网络。
  17. 一种服务器，其特征在于，包括处理器、通信接口和存储器，所述处理器、所述通信接口和所述存储器相互连接，其中，所述存储器用于存储计算机程序，所述计算机程序包括程序指令，所述处理器被配置用于调用所述程序指令，执行以下步骤：
    获取目标网络在前T个训练时期的处理结果集合，所述处理结果集合包括所述目标网络针对训练数据进行处理得到的t1个处理结果数据、以及利用处理结果预测模型针对所述目标网络进行处理结果预测得到的t2个处理结果数据，所述T、t1为正整数，所述t2为非负整数；
    利用所述处理结果预测模型对所述处理结果集合中的处理结果数据进行处理，预测得到所述目标网络在第T+1个训练时期的处理结果数据；
    其中，所述处理结果预测模型是基于所述目标网络的网络设置信息构建得到的。
  18. 根据权利要求17所述的服务器，其特征在于，所述处理结果数据用于指示所述目标网络的分类准确率，所述处理器具体用于调用所述程序指令执行以下步骤：
    调用所述处理结果预测模型计算所述处理结果集合中第T-M个训练时期到第T个训练时期的分类准确率的第一均值，所述M为大于1小于T的整数；
    计算所述处理结果集合中第T-M个训练时期到第T个训练时期的分类准确率的变化值的第二均值；
    计算所述处理结果集合中第T个训练时期与第T-1个训练时期的分类准确率的第一差值，以及计算所述处理结果集合中第T-1个训练时期与第T-2个训练时期的分类准确率的第二差值；
    根据所述第一均值、所述第二均值、所述第一差值和所述第二差值，预测得到所述目标网络在第T+1个训练时期的分类准确率，并将预测得到的分类准确率确定为所述目标网络在第T+1个训练时期的处理结果数据。
  19. 根据权利要求17或18所述的服务器，其特征在于，所述处理器还用于调用所述程序指令执行以下步骤：
    获取所述目标网络的网络设置信息，所述网络设置信息包括所述目标网络的结构参数和超参数；
    根据所述网络设置信息确定目标特征参数，并根据所述目标特征参数构建得到所述处理结果预测模型。
  20. 一种计算机可读存储介质，其特征在于，所述计算机可读存储介质存储有计算机程序，所述计算机程序包括程序指令，所述程序指令当被处理器执行时使所述处理器执行如权利要求1至8中任一项所述的基于预测模型的处理结果预测方法。
PCT/CN2019/117992 2019-02-15 2019-11-13 基于预测模型的处理结果预测方法、装置及服务器 WO2020164275A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910121267.9A CN109919314A (zh) 2019-02-15 2019-02-15 基于预测模型的处理结果预测方法、装置及服务器
CN201910121267.9 2019-02-15

Publications (1)

Publication Number Publication Date
WO2020164275A1 true WO2020164275A1 (zh) 2020-08-20

Family

ID=66961746

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/117992 WO2020164275A1 (zh) 2019-02-15 2019-11-13 基于预测模型的处理结果预测方法、装置及服务器

Country Status (2)

Country Link
CN (1) CN109919314A (zh)
WO (1) WO2020164275A1 (zh)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109919314A (zh) * 2019-02-15 2019-06-21 平安科技（深圳）有限公司 基于预测模型的处理结果预测方法、装置及服务器

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20160028599A1 (en) * 2014-07-23 2016-01-28 Cisco Technology, Inc. Scheduling predictive models for machine learning systems
CN107977754A (zh) * 2017-12-18 2018-05-01 深圳前海微众银行股份有限公司 数据预测方法、系统和计算机可读存储介质
CN108596399A (zh) * 2018-05-04 2018-09-28 国家邮政局邮政业安全中心 快递业务量预测的方法、装置、电子设备及存储介质
CN109919314A (zh) * 2019-02-15 2019-06-21 平安科技（深圳）有限公司 基于预测模型的处理结果预测方法、装置及服务器

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20160028599A1 (en) * 2014-07-23 2016-01-28 Cisco Technology, Inc. Scheduling predictive models for machine learning systems
CN107977754A (zh) * 2017-12-18 2018-05-01 深圳前海微众银行股份有限公司 数据预测方法、系统和计算机可读存储介质
CN108596399A (zh) * 2018-05-04 2018-09-28 国家邮政局邮政业安全中心 快递业务量预测的方法、装置、电子设备及存储介质
CN109919314A (zh) * 2019-02-15 2019-06-21 平安科技（深圳）有限公司 基于预测模型的处理结果预测方法、装置及服务器

Also Published As

Publication number Publication date
CN109919314A (zh) 2019-06-21

Similar Documents

Publication Publication Date Title
TWI777010B (zh) 資訊轉化率的預測、資訊推薦方法和裝置
CN106888236B (zh) 会话管理方法及会话管理装置
WO2020063686A1 (zh) 一种热负荷预测方法、装置、可读介质及电子设备
CN111027428B (zh) 一种多任务模型的训练方法、装置及电子设备
CN111049756B (zh) 请求响应方法、装置、电子设备及计算机可读存储介质
WO2022027913A1 (zh) 目标检测模型生成方法、装置、设备及存储介质
US20200117623A1 (en) Adaptive Interrupt Coalescing
US20140032745A1 (en) Subscription watch lists for event handling
WO2019071883A1 (zh) 一种行为断点的跟进处理方法、存储介质和服务器
TW201935334A (zh) 確定業務對應的決策策略的方法、裝置和電子設備
CN110061930B (zh) 一种数据流量的限制、限流值的确定方法和装置
WO2022166294A1 (zh) 一种目标检测方法和装置
CN109062715A (zh) 内存时钟频率的确定方法、装置及终端
US20230117088A1 (en) Method and device for improving performance of data processing model, storage medium and electronic device
US8780120B2 (en) GPU self throttling
TWI713019B (zh) 資料標籤產生、模型訓練、事件識別方法和裝置
WO2020164275A1 (zh) 基于预测模型的处理结果预测方法、装置及服务器
WO2018166499A1 (zh) 文本分类方法、设备和存储介质
CN112596985A (zh) 一种it资产探测方法、装置、设备和介质
CN110060188B (zh) 核身方式推荐方法、装置及电子设备
CN112530074A (zh) 排队叫号提醒方法、装置、设备及存储介质
CN110443746B (zh) 基于生成对抗网络的图片处理方法、装置及电子设备
US11593014B2 (en) System and method for approximating replication completion time
US11561801B2 (en) Predicting execution duration for user-defined functions on function-as-a-service (FaaS) cloud computing platforms
CN112015924A (zh) 流媒体缓存方法及其装置、计算机可存储介质

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 19915226

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 19915226

Country of ref document: EP

Kind code of ref document: A1