CN110633742A - Method for acquiring characteristic information and computer storage medium - Google Patents
Method for acquiring characteristic information and computer storage medium Download PDFInfo
- Publication number
- CN110633742A CN110633742A CN201910844447.XA CN201910844447A CN110633742A CN 110633742 A CN110633742 A CN 110633742A CN 201910844447 A CN201910844447 A CN 201910844447A CN 110633742 A CN110633742 A CN 110633742A
- Authority
- CN
- China
- Prior art keywords
- deep learning
- learning model
- information
- training
- parameter configuration
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Withdrawn
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- 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
Abstract
The embodiment of the application discloses a method for acquiring feature information and a computer storage medium. The method comprises the following steps: training the deep learning system by using pre-acquired training data to obtain running state information of the deep learning model; according to a pre-acquired information extraction strategy, acquiring characteristic information corresponding to the information extraction strategy from the running state information of the deep learning model; and generating an analysis report of the deep learning model according to the characteristic information.
Description
Technical Field
The present disclosure relates to the field of information processing, and more particularly, to a method for obtaining feature information and a computer storage medium.
Background
At present, artificial intelligence has entered into industries such as safe cities, financial forecasting, intelligent transportation and the like, and can excellently complete various tasks in various fields, which is not separated from the support of deep learning models. With the wide application of the deep learning model, various application scenarios to be dealt with are more and more complex, and the design challenge of the deep learning model is more and more large.
The design of the deep learning model is often an iterative process based on a specific scenario and the existing data volume. A new deep learning model structure is designed each time, training is firstly carried out on a GPU server based on training data to obtain a trained deep learning model, and then performance of the model is tested on a verification set to be used as a judgment basis for judging whether the deep learning model structure is designed.
In the design process of the deep learning model structure, if the effectiveness and the rationality of the designed deep learning model structure are judged only according to the performance on the verification set, on one hand, the time and the labor are wasted, and the design, the training and the verification are needed to be carried out firstly, and then the adjustment, the training and the verification are carried out, so that the cycle is repeated. The training of the deep learning model is a very time-consuming process, which needs several days in short time and several months in long time, and the algorithm development period is prolonged because the structure of the deep learning model is blindly adjusted according to the test result of the verification set. On the other hand, the structure of the deep learning model is modified according to the result of the algorithm on the verification set every time, the characteristics of the model and the operation characteristics of the model in the training process cannot be deeply analyzed, the adjustment of the model structure is unreasonable, and the modification every time is equivalent to the process of trial and error, and the specific problem of the design model cannot be fundamentally found and discovered.
In the related art, the deep learning model feature analysis is generally counted by means of a model structure diagram, and an overall structure of the deep learning model is displayed, including a type of each module, an input and output size of data, and a number of parameters, a size and the like of each module. The deep learning model feature analysis method usually focuses on the structural characteristics of the deep learning model and the theoretical calculation amount of the deep learning model. The optimization method for deep learning model feature analysis is mainly used for analyzing and optimizing the features of a deep learning model based on past experiences in a manual mode on the basis of early statistical results.
The analysis method has low processing efficiency and is a problem to be solved urgently.
Disclosure of Invention
In order to solve any one of the above technical problems, embodiments of the present application provide a method for acquiring feature information and a computer storage medium.
In order to achieve the purpose of the embodiment of the present application, an embodiment of the present application provides a method for acquiring feature information, including:
training the deep learning system by using pre-acquired training data to obtain running state information of the deep learning model;
according to a pre-acquired information extraction strategy, acquiring characteristic information corresponding to the information extraction strategy from the running state information of the deep learning model;
and generating an analysis report of the deep learning model according to the characteristic information.
In an exemplary embodiment, after the deep learning system is trained by using the pre-acquired training data to obtain the operating state information of the deep learning model, the method further includes:
judging whether the operating environment of the deep learning model meets the operating requirement of the deep learning model or not according to the operating state of the deep learning model;
and executing the operation of acquiring the characteristic information after judging that the operating environment of the deep learning model meets the operating requirement of the deep learning model.
In an exemplary embodiment, the training operation of the deep learning system by using pre-acquired training data includes:
obtaining a frame type of the deep learning model;
acquiring training data of the deep learning model according to the frame type;
carrying out single-card and multi-card training operation on the deep learning model by using the training data;
the judging whether the operating environment of the deep learning model meets the operating requirement of the deep learning model according to the operating state of the deep learning model comprises the following steps:
acquiring acceleration information of the deep learning model for executing training operation of a single card and multiple cards;
judging whether the acceleration information of the single card and the multiple cards meets a preset acceleration strategy or not to obtain a judgment result;
and if the judgment result is that the acceleration strategy is met, determining that the operating environment of the deep learning model meets the operating requirement of the deep learning model.
In an exemplary embodiment, after generating the analysis report of the deep learning model according to the feature information, the method further includes:
acquiring prediction accuracy rate information of the deep learning model in a training process;
determining parameter configuration strategies used by the deep learning system under different prediction accuracy rates;
selecting a parameter configuration strategy which meets the judgment condition of the preset optimal configuration from the obtained parameter configuration strategies corresponding to different prediction accuracy rates as a first parameter configuration strategy;
and outputting the first parameter configuration strategy.
In an exemplary embodiment, after generating the analysis report of the deep learning model according to the feature information, the method further includes:
acquiring resource use information of the deep learning model in a training process, wherein the resource use information comprises at least one of the following: the method comprises the steps of calculating amount information of training operation, using information of a memory occupied by the training operation and using information of a processor occupied by the training operation;
acquiring parameter configuration strategies used by the deep learning model in different resource use states;
selecting a parameter configuration strategy which accords with the optimal configuration judgment condition from the corresponding parameter configuration strategies in different resource use states as a second parameter configuration strategy;
and outputting the second parameter configuration strategy.
A computer storage medium comprising a processor and a memory, wherein the memory stores a computer program, the processor invoking the computer program in the memory to implement operations comprising:
training the deep learning system by using pre-acquired training data to obtain running state information of the deep learning model;
according to a pre-acquired information extraction strategy, acquiring characteristic information corresponding to the information extraction strategy from the running state information of the deep learning model;
and generating an analysis report of the deep learning model according to the characteristic information.
In an exemplary embodiment, the processor calls the computer program in the memory to implement the operation of training the deep learning system by using the pre-acquired training data to obtain the running state information of the deep learning model, and then the processor calls the computer program in the memory to implement the following operations, including:
judging whether the operating environment of the deep learning model meets the operating requirement of the deep learning model or not according to the operating state of the deep learning model;
and executing the operation of acquiring the characteristic information after judging that the operating environment of the deep learning model meets the operating requirement of the deep learning model.
In an exemplary embodiment, the processor invokes a computer program in the memory to implement a training operation on the deep learning system using pre-acquired training data, comprising:
obtaining a frame type of the deep learning model;
acquiring training data of the deep learning model according to the frame type;
carrying out single-card and multi-card training operation on the deep learning model by using the training data;
the processor calls a computer program in the memory to realize the operation of judging whether the running environment of the deep learning model meets the running requirement of the deep learning model according to the running state of the deep learning model, and the operation comprises the following steps:
acquiring acceleration information of the deep learning model for executing training operation of a single card and multiple cards;
judging whether the acceleration information of the single card and the multiple cards meets a preset acceleration strategy or not to obtain a judgment result;
and if the judgment result is that the acceleration strategy is met, determining that the operating environment of the deep learning model meets the operating requirement of the deep learning model.
In an exemplary embodiment, after the processor invokes the computer program in the memory to implement the operation of generating the analysis report of the deep learning model according to the feature information, the processor invokes the computer program in the memory to further implement the following operations, including:
acquiring prediction accuracy rate information of the deep learning model in a training process;
determining parameter configuration strategies used by the deep learning system under different prediction accuracy rates;
selecting a parameter configuration strategy which meets the judgment condition of the preset optimal configuration from the obtained parameter configuration strategies corresponding to different prediction accuracy rates as a first parameter configuration strategy;
and outputting the first parameter configuration strategy.
In an exemplary embodiment, after the processor invokes the computer program in the memory to implement the operation of generating the analysis report of the deep learning model according to the feature information, the processor invokes the computer program in the memory to further implement the following operations, including:
acquiring resource use information of the deep learning model in a training process, wherein the resource use information comprises at least one of the following: the method comprises the steps of calculating amount information of training operation, using information of a memory occupied by the training operation and using information of a processor occupied by the training operation;
acquiring parameter configuration strategies used by the deep learning model in different resource use states;
selecting a parameter configuration strategy which accords with the optimal configuration judgment condition from the corresponding parameter configuration strategies in different resource use states as a second parameter configuration strategy;
and outputting the second parameter configuration strategy.
According to the embodiment of the application, the deep learning system is trained by using the pre-acquired training data to obtain the running state information of the deep learning model, the characteristic information corresponding to the information extraction strategy is acquired from the running state information of the deep learning model according to the pre-acquired information extraction strategy, the analysis report of the deep learning model is generated according to the characteristic information, and the automatic extraction of the running characteristics of the deep learning model in the training process greatly reduces the manual operation in the deep learning model design process, shortens the model design period and saves the investment of manpower and time in the model development process.
Additional features and advantages of the embodiments of the application will be set forth in the description which follows, and in part will be obvious from the description, or may be learned by the practice of the embodiments of the application. The objectives and other advantages of the embodiments of the application may be realized and attained by the structure particularly pointed out in the written description and claims hereof as well as the appended drawings.
Drawings
The accompanying drawings are included to provide a further understanding of the embodiments of the present application and are incorporated in and constitute a part of this specification, illustrate embodiments of the present application and together with the examples of the embodiments of the present application do not constitute a limitation of the embodiments of the present application.
Fig. 1 is a flowchart of a method for acquiring feature information according to an embodiment of the present application;
fig. 2 is a flowchart of a method for acquiring feature information of a deep learning model according to an embodiment of the present application.
Detailed Description
In order to make the objects, technical solutions and advantages of the embodiments of the present application more apparent, the embodiments of the present application will be described in detail below with reference to the accompanying drawings. It should be noted that, in the embodiments of the present application, features in the embodiments and the examples may be arbitrarily combined with each other without conflict.
The inventors found that at least the following problems exist in the related art:
the deep learning model feature analysis method in the related technology is mainly based on a manual statistics mode. For different deep learning algorithm models, statistics needs to be carried out independently, the universality is not strong, time and labor are consumed, and the deep grasping and adjusting guiding significance of the network structure of the later deep learning model is not large;
the deep learning algorithm feature analysis method in the related art only analyzes the deep learning network structure and does not consider the software and hardware environment of the deep learning model in the actual operation process, so that the model has deviation in the actual operation process;
the characteristic analysis and optimization of the deep learning model in the related technology completely depend on manual work according to statistical results, in the whole process of developing the deep learning algorithm, human resources cannot be liberated, great waste is caused, meanwhile, more time is spent, and great inconvenience and delay are brought to practical application and production.
Based on the above analysis, the embodiments of the present application provide the following solutions:
fig. 1 is a flowchart of a method for acquiring feature information according to an embodiment of the present application. The method shown in fig. 1, comprising:
101, training the deep learning system by using pre-acquired training data to obtain running state information of the deep learning model;
in an exemplary embodiment, in the process of training a deep learning model, the running state of the deep learning model is recorded, and a data basis is provided for the subsequent management of the deep learning model;
in an exemplary embodiment, the obtaining of the operating state information may be performed by performing targeted extraction on the operating state information of the deep learning model according to a preset obtaining strategy.
102, acquiring characteristic information corresponding to an information extraction strategy from the running state information of the deep learning model according to the information extraction strategy acquired in advance;
in an exemplary embodiment, the information extraction policy may be set according to the received management requirement, for example, if the received management requirement is the number of parameters for optimizing the deep learning model, it is required to obtain the calculation state information of the deep learning system when the number of parameters of different number combinations is used, wherein the calculation state information may be prediction accuracy, hardware resource consumption information, and the like.
And 103, generating an analysis report of the deep learning model according to the characteristic information.
In an exemplary embodiment, the pre-stored output template can be utilized, the obtained feature information is output through the template, the labor consumption of the output operation is simplified, and the processing efficiency is improved.
According to the method provided by the embodiment of the application, the deep learning system is trained by utilizing the pre-acquired training data to obtain the running state information of the deep learning model, the characteristic information corresponding to the information extraction strategy is acquired from the running state information of the deep learning model according to the pre-acquired information extraction strategy, the analysis report of the deep learning model is generated according to the characteristic information, and the automatic extraction of the running characteristics of the deep learning model in the training process greatly reduces the manual operation in the deep learning model design process, shortens the model design period and saves the investment of manpower and time in the model development process.
The method provided by the embodiments of the present application is explained as follows:
in an exemplary embodiment, after the deep learning system is trained by using the pre-acquired training data to obtain the operating state information of the deep learning model, the method further includes:
judging whether the operating environment of the deep learning model meets the operating requirement of the deep learning model or not according to the operating state of the deep learning model;
and executing the operation of acquiring the characteristic information after judging that the operating environment of the deep learning model meets the operating requirement of the deep learning model.
In an exemplary embodiment, after the deep learning model is executed, whether the current operating environment meets the operating requirement of the deep learning model is determined according to the execution result, so as to determine whether the current training operation can be used as data for subsequently adjusting the deep learning model, and reduce data errors caused by the processing speed in the external operating environment.
In an exemplary embodiment, the training operation of the deep learning system by using pre-acquired training data includes:
obtaining a frame type of the deep learning model;
acquiring training data of the deep learning model according to the frame type;
carrying out single-card and multi-card training operation on the deep learning model by using the training data;
the judging whether the operating environment of the deep learning model meets the operating requirement of the deep learning model according to the operating state of the deep learning model comprises the following steps:
acquiring acceleration information of the deep learning model for executing training operation of a single card and multiple cards;
judging whether the acceleration information of the single card and the multiple cards meets a preset acceleration strategy or not to obtain a judgment result;
and if the judgment result is that the acceleration strategy is met, determining that the operating environment of the deep learning model meets the operating requirement of the deep learning model.
By adopting the mode, the detection of the running environment is realized while the training operation is completed, and the processing efficiency is improved.
In an exemplary embodiment, after generating the analysis report of the deep learning model according to the feature information, the method further includes:
acquiring prediction accuracy rate information of the deep learning model in a training process;
determining parameter configuration strategies used by the deep learning system under different prediction accuracy rates;
selecting a parameter configuration strategy which meets the judgment condition of the preset optimal configuration from the obtained parameter configuration strategies corresponding to different prediction accuracy rates as a first parameter configuration strategy;
and outputting the first parameter configuration strategy.
Compared with the prior art, the method provided by the embodiment of the application is based on the obtained data, and obtains the optimal parameter configuration information according to the preset judgment condition to realize the optimization of the deep learning model, wherein the judgment condition of the optimal configuration can be set according to the received configuration requirement, so that different configuration requirements are met. By adopting the steps, the optimization of the use efficiency and the use amount of the parameters can be realized.
In an exemplary embodiment, after generating the analysis report of the deep learning model according to the feature information, the method further includes:
acquiring resource use information of the deep learning model in a training process, wherein the resource use information comprises at least one of the following: the method comprises the steps of calculating amount information of training operation, using information of a memory occupied by the training operation and using information of a processor occupied by the training operation;
acquiring parameter configuration strategies used by the deep learning model in different resource use states;
selecting a parameter configuration strategy which accords with the optimal configuration judgment condition from the corresponding parameter configuration strategies in different resource use states as a second parameter configuration strategy;
and outputting the second parameter configuration strategy.
Compared with the prior art, the method provided by the embodiment of the application is based on the obtained data, and obtains the optimal parameter configuration information according to the preset judgment condition to realize the optimization of the deep learning model, wherein the judgment condition of the optimal configuration can be set according to the received configuration requirement, so that different configuration requirements are met. By adopting the steps, the optimization of the use of the hardware resources can be realized.
The method provided by the embodiments of the present application is explained as follows:
the system for automatically analyzing and optimizing the characteristics of the deep learning model provided by the embodiment of the application comprises:
the hardware environment automatic detection module is used for automatically detecting the hardware environment of the deep learning model;
the software environment automatic detection module is used for automatically detecting the software environment of the deep learning model;
the benchmark performance testing module of the operating environment is used for automatically carrying out benchmark performance testing according to the operating environment of the deep learning model;
the automatic extraction module of the operating characteristic is used for extracting the characteristic in the data generated by the training operation of the deep learning algorithm;
the running characteristic automatic analysis module is used for analyzing the obtained characteristic information according to a pre-obtained analysis strategy;
and the training parameter optimization module is used for optimizing the deep learning model according to the analysis result.
By using the system, the normal operation environment of the deep learning model is ensured by detecting the software and hardware environments of the deep learning model training platform and automatically testing the benchmark performance; through automatic extraction and automatic analysis of the running characteristics of the deep learning model in the training process, the optimization of the training parameters of the deep learning model is realized, so that the efficiency of design and training of the deep learning model is improved, and the application requirements are met.
Fig. 2 is a flowchart of a method for acquiring feature information of a deep learning model according to an embodiment of the present application. As shown in fig. 2, the method comprises:
step 201, acquiring hardware resource information and/or software resource information in an operating environment of a deep learning model;
in an exemplary embodiment, the hardware resource information may be obtained by a preset hardware resource detection tool for hardware configuration information of one or at least two servers used for deploying the deep learning model.
In one exemplary embodiment, the hardware information that may be detected includes at least one of a processor, memory, storage, and network bandwidth; the processor can be a CPU or a GPU, and the type and the number of the processors can be obtained; the memory can acquire the number and capacity of the memory.
Step 202, acquiring software resource information in the operating environment of the deep learning model;
in one exemplary embodiment, the software information includes at least one of a GPU-driven version, a CUDA (computer Unified Device architecture) version, a CuDNN version, a Python version, a framework type of the deep learning model, and a version of the deep learning model.
The CUDA is a general computing architecture based on a new parallel programming model and an instruction set architecture, which is introduced by the Invitta company, and can solve many complex computing tasks more efficiently than a CPU (central processing unit) by utilizing a parallel computing engine of an Invitta GPU. cuDNN is a GPU acceleration library for deep neural networks.
Step 203, performing basic performance test on the deep learning model, specifically including:
a1, acquiring training data of a deep learning model according to the frame type of the deep learning model;
in an exemplary embodiment, single-card and multi-card training data corresponding to a deep learning model are obtained according to the type of a deep learning framework; the deep learning model can be AlexNet, inclusion v3, ResNet50, etc.
Step A2, training the deep learning model by using the training data of the deep learning model to obtain the running state information of the running environment;
in an exemplary embodiment, training operations of the single card and the multiple cards are executed on the deep learning model by using training data of the single card and the multiple cards, and acceleration information of the single card and the multiple cards in the current operating environment is obtained;
step A3, judging whether the current operating environment can meet the operating requirement of the deep learning model according to the operating state information;
in one exemplary embodiment, the acceleration ratio of the single card to the multiple cards is calculated according to the acceleration information of the single card and the acceleration information of the multiple cards; comparing the acceleration ratio with a preset threshold value; if the acceleration ratio is larger than the threshold value, determining that the current operating environment can meet the operating requirement of the deep learning model; otherwise, determining that the current operating environment cannot meet the operating requirement of the deep learning model.
Optionally, when it is determined that the current operating environment cannot meet the operating requirement of the deep learning model, the hardware information and/or the software information and the test result are recorded for reference in subsequent configuration, so that the success rate of the subsequent configuration is improved.
204, when the current operating environment is determined to meet the operating requirement of the deep learning model, extracting characteristic information of the deep learning model in the training process;
in one exemplary embodiment, the feature information includes at least one of basic structure information of the deep learning model, deep learning model statistical information, and resource usage information of the deep learning model; wherein:
the basic structure information of the deep learning model comprises at least one of the following: model network structure connection information, input data size, type and number of convolutional layers, type and number of pooling layers, type and number of activation functions, number of layers of full connection layers and number of neurons in each layer, type and number of data normalization layers, type of loss function, type of optimizer, size of output data and the like;
the deep learning model statistical information may perform statistics on specific information according to a preset statistical strategy, for example, at least one of the following statistical operations may be performed, including: the method comprises the steps of deep learning model parameter quantity, calculated quantity, video memory occupation, feature diagram size, storage size and the like.
The deep learning model runs the resource usage information as at least one of: the method comprises the following steps of CPU usage, memory size usage, memory bandwidth usage, GPU number, video memory size usage, CUDA core usage and the like.
Step 205, generating an analysis report of a deep learning model according to the characteristic information of the deep learning system;
in an exemplary embodiment, the extracted deep learning model training features automatically draw and store a deep learning model structure diagram and a detailed parameter diagram of each layer; or automatically drawing and storing a statistical information table of the deep learning model; automatically drawing a deep learning model operation resource use information graph; or, outputting learning rate and loss statistical information of the training optimizer;
step 206, obtaining the optimization information of the deep learning model according to the analysis report;
in an exemplary embodiment, according to the deep learning model training characteristics, providing parameter quantity of the model and use comparison information of the video memory; according to the accuracy and the parameter quantity in the deep learning model training process, parameter use efficiency and parameter quantity optimization information of the current model are given; and giving optimal batch processing parameter optimization information according to the calculated amount information of the training operation, the use information of the memory occupied by the training operation and the use information of the processor occupied by the training operation.
The method provided by the embodiment of the application is simple in process operation, manual operation in the deep learning model design process is greatly reduced, the model design period is shortened, manpower and time investment in the model development process are saved, and the standardization and the efficiency of the deep learning model design are guaranteed. The practicability is strong, the application range is wide, and the popularization is easy. The method can be conveniently expanded to the automatic analysis process of the deep learning model reasoning characteristics. In addition, the automatic optimization analysis tool for the deep learning model features in the technical scheme can be operated on a single server and can be conveniently popularized to a server cluster for operation.
An embodiment of the present application provides a computer storage medium, including a processor and a memory, where the memory stores a computer program, and the processor calls the computer program in the memory to implement the following operations, including:
training the deep learning system by using pre-acquired training data to obtain running state information of the deep learning model;
according to a pre-acquired information extraction strategy, acquiring characteristic information corresponding to the information extraction strategy from the running state information of the deep learning model;
and generating an analysis report of the deep learning model according to the characteristic information.
In an exemplary embodiment, the processor calls the computer program in the memory to implement the operation of training the deep learning system by using the pre-acquired training data to obtain the running state information of the deep learning model, and then the processor calls the computer program in the memory to implement the following operations, including:
judging whether the operating environment of the deep learning model meets the operating requirement of the deep learning model or not according to the operating state of the deep learning model;
and executing the operation of acquiring the characteristic information after judging that the operating environment of the deep learning model meets the operating requirement of the deep learning model.
In an exemplary embodiment, the processor invokes a computer program in the memory to implement a training operation on the deep learning system using pre-acquired training data, comprising:
obtaining a frame type of the deep learning model;
acquiring training data of the deep learning model according to the frame type;
carrying out single-card and multi-card training operation on the deep learning model by using the training data;
the processor calls a computer program in the memory to realize the operation of judging whether the running environment of the deep learning model meets the running requirement of the deep learning model according to the running state of the deep learning model, and the operation comprises the following steps:
acquiring acceleration information of the deep learning model for executing training operation of a single card and multiple cards;
judging whether the acceleration information of the single card and the multiple cards meets a preset acceleration strategy or not to obtain a judgment result;
and if the judgment result is that the acceleration strategy is met, determining that the operating environment of the deep learning model meets the operating requirement of the deep learning model.
In an exemplary embodiment, after the processor invokes the computer program in the memory to implement the operation of generating the analysis report of the deep learning model according to the feature information, the processor invokes the computer program in the memory to further implement the following operations, including:
acquiring prediction accuracy rate information of the deep learning model in a training process;
determining parameter configuration strategies used by the deep learning system under different prediction accuracy rates;
selecting a parameter configuration strategy which meets the judgment condition of the preset optimal configuration from the obtained parameter configuration strategies corresponding to different prediction accuracy rates as a first parameter configuration strategy;
and outputting the first parameter configuration strategy.
In an exemplary embodiment, after the processor invokes the computer program in the memory to implement the operation of generating the analysis report of the deep learning model according to the feature information, the processor invokes the computer program in the memory to further implement the following operations, including:
acquiring resource use information of the deep learning model in a training process, wherein the resource use information comprises at least one of the following: the method comprises the steps of calculating amount information of training operation, using information of a memory occupied by the training operation and using information of a processor occupied by the training operation;
acquiring parameter configuration strategies used by the deep learning model in different resource use states;
selecting a parameter configuration strategy which accords with the optimal configuration judgment condition from the corresponding parameter configuration strategies in different resource use states as a second parameter configuration strategy;
and outputting the second parameter configuration strategy.
The computer storage medium provided by the embodiment of the application utilizes pre-acquired training data to train the deep learning system to obtain the running state information of the deep learning model, obtains the characteristic information corresponding to the information extraction strategy from the running state information of the deep learning model according to the pre-acquired information extraction strategy, generates the analysis report of the deep learning model according to the characteristic information, greatly reduces the manual operation in the deep learning model design process through the automatic extraction of the running characteristics of the deep learning model in the training process, shortens the model design period, and saves the investment of manpower and time in the model development process.
It will be understood by those of ordinary skill in the art that all or some of the steps of the methods, systems, functional modules/units in the devices disclosed above may be implemented as software, firmware, hardware, and suitable combinations thereof. In a hardware implementation, the division between functional modules/units mentioned in the above description does not necessarily correspond to the division of physical components; for example, one physical component may have multiple functions, or one function or step may be performed by several physical components in cooperation. Some or all of the components may be implemented as software executed by a processor, such as a digital signal processor or microprocessor, or as hardware, or as an integrated circuit, such as an application specific integrated circuit. Such software may be distributed on computer readable media, which may include computer storage media (or non-transitory media) and communication media (or transitory media). The term computer storage media includes volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data, as is well known to those of ordinary skill in the art. Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, Digital Versatile Disks (DVD) or other optical disk storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other medium which can be used to store the desired information and which can accessed by a computer. In addition, communication media typically embodies computer readable instructions, data structures, program modules or other data in a modulated data signal such as a carrier wave or other transport mechanism and includes any information delivery media as known to those skilled in the art.
Claims (10)
1. A method for acquiring feature information is characterized by comprising the following steps:
training the deep learning system by using pre-acquired training data to obtain running state information of the deep learning model;
according to a pre-acquired information extraction strategy, acquiring characteristic information corresponding to the information extraction strategy from the running state information of the deep learning model;
and generating an analysis report of the deep learning model according to the characteristic information.
2. The method according to claim 1, wherein after the deep learning system is trained by using the pre-acquired training data to obtain the operating state information of the deep learning model, the method further comprises:
judging whether the operating environment of the deep learning model meets the operating requirement of the deep learning model or not according to the operating state of the deep learning model;
and executing the operation of acquiring the characteristic information after judging that the operating environment of the deep learning model meets the operating requirement of the deep learning model.
3. The method of claim 2, wherein:
the training operation of the deep learning system by using the pre-acquired training data comprises:
obtaining a frame type of the deep learning model;
acquiring training data of the deep learning model according to the frame type;
carrying out single-card and multi-card training operation on the deep learning model by using the training data;
the judging whether the operating environment of the deep learning model meets the operating requirement of the deep learning model according to the operating state of the deep learning model comprises the following steps:
acquiring acceleration information of the deep learning model for executing training operation of a single card and multiple cards;
judging whether the acceleration information of the single card and the multiple cards meets a preset acceleration strategy or not to obtain a judgment result;
and if the judgment result is that the acceleration strategy is met, determining that the operating environment of the deep learning model meets the operating requirement of the deep learning model.
4. The method according to any one of claims 1 to 3, wherein after generating the analysis report of the deep learning model according to the feature information, the method further comprises:
acquiring prediction accuracy rate information of the deep learning model in a training process;
determining parameter configuration strategies used by the deep learning system under different prediction accuracy rates;
selecting a parameter configuration strategy which meets the judgment condition of the preset optimal configuration from the obtained parameter configuration strategies corresponding to different prediction accuracy rates as a first parameter configuration strategy;
and outputting the first parameter configuration strategy.
5. The method according to any one of claims 1 to 3, wherein after generating the analysis report of the deep learning model according to the feature information, the method further comprises:
acquiring resource use information of the deep learning model in a training process, wherein the resource use information comprises at least one of the following: the method comprises the steps of calculating amount information of training operation, using information of a memory occupied by the training operation and using information of a processor occupied by the training operation;
acquiring parameter configuration strategies used by the deep learning model in different resource use states;
selecting a parameter configuration strategy which accords with the optimal configuration judgment condition from the corresponding parameter configuration strategies in different resource use states as a second parameter configuration strategy;
and outputting the second parameter configuration strategy.
6. A computer storage medium comprising a processor and a memory, wherein the memory stores a computer program, and wherein the processor invokes the computer program in the memory to perform operations comprising:
training the deep learning system by using pre-acquired training data to obtain running state information of the deep learning model;
according to a pre-acquired information extraction strategy, acquiring characteristic information corresponding to the information extraction strategy from the running state information of the deep learning model;
and generating an analysis report of the deep learning model according to the characteristic information.
7. The computer storage medium of claim 6, wherein the processor calls a computer program in the memory to implement the operation of training the deep learning system with pre-acquired training data to obtain the running state information of the deep learning model, and then the processor calls the computer program in the memory to implement the following operations, including:
judging whether the operating environment of the deep learning model meets the operating requirement of the deep learning model or not according to the operating state of the deep learning model;
and executing the operation of acquiring the characteristic information after judging that the operating environment of the deep learning model meets the operating requirement of the deep learning model.
8. The computer storage medium of claim 7, wherein:
the processor calls a computer program in the memory to implement a training operation on the deep learning system using pre-acquired training data, including:
obtaining a frame type of the deep learning model;
acquiring training data of the deep learning model according to the frame type;
carrying out single-card and multi-card training operation on the deep learning model by using the training data;
the processor calls a computer program in the memory to realize the operation of judging whether the running environment of the deep learning model meets the running requirement of the deep learning model according to the running state of the deep learning model, and the operation comprises the following steps:
acquiring acceleration information of the deep learning model for executing training operation of a single card and multiple cards;
judging whether the acceleration information of the single card and the multiple cards meets a preset acceleration strategy or not to obtain a judgment result;
and if the judgment result is that the acceleration strategy is met, determining that the operating environment of the deep learning model meets the operating requirement of the deep learning model.
9. The computer storage medium of any of claims 6 to 8, wherein after the processor invokes the computer program in the memory to perform the operation of generating the analysis report of the deep learning model based on the feature information, the processor invokes the computer program in the memory to further perform operations comprising:
acquiring prediction accuracy rate information of the deep learning model in a training process;
determining parameter configuration strategies used by the deep learning system under different prediction accuracy rates;
selecting a parameter configuration strategy which meets the judgment condition of the preset optimal configuration from the obtained parameter configuration strategies corresponding to different prediction accuracy rates as a first parameter configuration strategy;
and outputting the first parameter configuration strategy.
10. The computer storage medium of any of claims 6 to 8, wherein after the processor invokes the computer program in the memory to perform the operation of generating the analysis report of the deep learning model based on the feature information, the processor invokes the computer program in the memory to further perform operations comprising:
acquiring resource use information of the deep learning model in a training process, wherein the resource use information comprises at least one of the following: the method comprises the steps of calculating amount information of training operation, using information of a memory occupied by the training operation and using information of a processor occupied by the training operation;
acquiring parameter configuration strategies used by the deep learning model in different resource use states;
selecting a parameter configuration strategy which accords with the optimal configuration judgment condition from the corresponding parameter configuration strategies in different resource use states as a second parameter configuration strategy;
and outputting the second parameter configuration strategy.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910844447.XA CN110633742A (en) | 2019-09-06 | 2019-09-06 | Method for acquiring characteristic information and computer storage medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910844447.XA CN110633742A (en) | 2019-09-06 | 2019-09-06 | Method for acquiring characteristic information and computer storage medium |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110633742A true CN110633742A (en) | 2019-12-31 |
Family
ID=68972243
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910844447.XA Withdrawn CN110633742A (en) | 2019-09-06 | 2019-09-06 | Method for acquiring characteristic information and computer storage medium |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110633742A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113342631A (en) * | 2021-07-02 | 2021-09-03 | 厦门美图之家科技有限公司 | Distribution management optimization method and device and electronic equipment |
-
2019
- 2019-09-06 CN CN201910844447.XA patent/CN110633742A/en not_active Withdrawn
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113342631A (en) * | 2021-07-02 | 2021-09-03 | 厦门美图之家科技有限公司 | Distribution management optimization method and device and electronic equipment |
| CN113342631B (en) * | 2021-07-02 | 2022-08-26 | 厦门美图之家科技有限公司 | Distribution management optimization method and device and electronic equipment |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| EP3798846B1 (en) | Operation and maintenance system and method | |
| CN112529023A (en) | Configured artificial intelligence scene application research and development method and system | |
| CN114048055A (en) | Time series data abnormal root cause analysis method and system | |
| CN110377519B (en) | Performance capacity test method, device and equipment of big data system and storage medium | |
| CN114911615A (en) | Method and application for intelligent prediction scheduling during micro-service operation | |
| CN109800776A (en) | Material mask method, device, terminal and computer readable storage medium | |
| CN111522736A (en) | Software defect prediction method and device, electronic equipment and computer storage medium | |
| CN110633742A (en) | Method for acquiring characteristic information and computer storage medium | |
| CN112783508B (en) | File compiling method, device, equipment and storage medium | |
| CN111339072B (en) | User behavior-based change value analysis method and device, electronic equipment and medium | |
| CN112862013A (en) | Problem diagnosis method and device for quantitative transaction strategy | |
| CN111045912B (en) | AI application performance evaluation method, device and related equipment | |
| CN116152609B (en) | Distributed model training method, system, device and computer readable medium | |
| CN114610590A (en) | Method, device and equipment for determining operation time length and storage medium | |
| CN113849484A (en) | Big data component upgrading method and device, electronic equipment and storage medium | |
| CN113986495A (en) | Task execution method, device, equipment and storage medium | |
| CN111625352A (en) | Scheduling method, device and storage medium | |
| CN112232960B (en) | Transaction application system monitoring method and device | |
| CN116341633B (en) | Model deployment method, device, equipment and storage medium | |
| CN116069504B (en) | Scheduling method and device for multi-core processor in automatic driving simulation | |
| CN113837863B (en) | Business prediction model creation method and device and computer readable storage medium | |
| CN111563033B (en) | Simulation data generation method and device | |
| KR102619539B1 (en) | Optimization method of neural network for multi-gpu system and optimization system using the same | |
| CN114185756A (en) | Distributed system state analysis method, device and computer readable storage medium | |
| CN117370001A (en) | Graphics processor scheduling method, system, equipment and storage medium |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| WW01 | Invention patent application withdrawn after publication |
Application publication date: 20191231 |
|
| WW01 | Invention patent application withdrawn after publication |