CN113157453A - Task complexity-based high-energy-efficiency target detection task dynamic scheduling method - Google Patents
Task complexity-based high-energy-efficiency target detection task dynamic scheduling method Download PDFInfo
- Publication number
- CN113157453A CN113157453A CN202110435960.0A CN202110435960A CN113157453A CN 113157453 A CN113157453 A CN 113157453A CN 202110435960 A CN202110435960 A CN 202110435960A CN 113157453 A CN113157453 A CN 113157453A
- Authority
- CN
- China
- Prior art keywords
- task
- complexity
- target detection
- user
- energy consumption
- 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.)
- Granted
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F9/00—Arrangements for program control, e.g. control units
- G06F9/06—Arrangements for program control, e.g. control units using stored programs, i.e. using an internal store of processing equipment to receive or retain programs
- G06F9/46—Multiprogramming arrangements
- G06F9/50—Allocation of resources, e.g. of the central processing unit [CPU]
- G06F9/5005—Allocation of resources, e.g. of the central processing unit [CPU] to service a request
- G06F9/5027—Allocation of resources, e.g. of the central processing unit [CPU] to service a request the resource being a machine, e.g. CPUs, Servers, Terminals
- G06F9/5055—Allocation of resources, e.g. of the central processing unit [CPU] to service a request the resource being a machine, e.g. CPUs, Servers, Terminals considering software capabilities, i.e. software resources associated or available to the machine
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/18—Complex mathematical operations for evaluating statistical data, e.g. average values, frequency distributions, probability functions, regression analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Software Systems (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- Pure & Applied Mathematics (AREA)
- Computational Mathematics (AREA)
- Mathematical Analysis (AREA)
- Mathematical Optimization (AREA)
- Mathematical Physics (AREA)
- Evolutionary Computation (AREA)
- Algebra (AREA)
- Artificial Intelligence (AREA)
- Operations Research (AREA)
- Databases & Information Systems (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Probability & Statistics with Applications (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Life Sciences & Earth Sciences (AREA)
- Medical Informatics (AREA)
- Computer Hardware Design (AREA)
- Geometry (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
- Mobile Radio Communication Systems (AREA)
Abstract
A high-energy-efficiency target detection task dynamic scheduling method based on task complexity is characterized in that a neural network is used for evaluating the complexity of a task, a target detection model and a cloud platform meeting user requirements are judged by weighting the current state of a mobile phone and the requirements of a user, when the user submits the target detection task, the task is unloaded according to the complexity of the task, the current state of the mobile phone and the requirements of the user, and the calculation pressure and energy consumption of a mobile phone end can be greatly reduced by unloading the calculation-intensive target detection task to the cloud for execution under the condition of good network; under the condition that the network is not good, the energy consumption of the mobile phone end can be reduced by calling the local small model of the mobile phone end, and the energy consumption pressure of the mobile phone end can be greatly reduced by utilizing the method and the system under the condition that certain user experience is provided and the precision of the target detection model is kept.
Description
Technical Field
The invention belongs to the technical field of computers, and particularly relates to a task complexity-based dynamic scheduling method for an energy-efficient target detection task.
Background
Most CNN-based target detection models are only applicable to recommendation systems. For example, the process of searching for free parking spaces by a city camera is performed by a slow accurate model, whereas car collision warnings are related to a fast inaccurate model. The accuracy of the real-time target detection models is improved, and the real-time target detection models can be used for prompting to generate a recommendation system and can also be used for independent process management and manual input reduction. Real-time object detector operations on conventional Graphics Processing Units (GPUs) allow them to be used in large quantities at affordable prices. The most accurate modern neural networks cannot operate in real time and require a large number of GPUs for large batches of training. Further accuracy requirements have prompted larger and deeper target detection models. However, the deployment of a large-scale target detection model at the mobile end is not only very expensive in time, but also very expensive in energy consumption. In recent years, many improvements have been proposed for depth target detection models, however these models are still bulky compared to mobile devices with little computational power.
In recent years, lightweight neural networks designed for mobile terminals have become increasingly popular, and designing deep neural network architectures to achieve the best compromise between accuracy and efficiency has been an active area of research. Novel manual structural and algorithmic neural architecture searches play an important role in advancing the field. Through these design techniques, some lightweight neural networks designed for mobile platforms play a tremendous role. However, these lightweight neural networks are difficult to satisfy in terms of accuracy compared to deep neural networks. Furthermore, these neural networks do not guarantee optimal results for any mobile-side architecture.
Therefore, a method is needed which can be applied to any mobile platform while maintaining the accuracy, and can reduce the inference cost of the mobile terminal.
Disclosure of Invention
In order to overcome the defects of the prior art, the invention aims to provide a task complexity-based high-energy-efficiency target detection task dynamic scheduling method, which has the advantage of reducing the energy consumption of a mobile platform while maintaining the precision.
In order to achieve the purpose, the invention provides the following technical scheme:
a high-energy-efficiency target detection task dynamic scheduling method based on task complexity selects a proper task unloading strategy according to the state of a current user and the task complexity when the user submits a target detection task, and is characterized by comprising the following steps:
step S1, acquiring complexity C of the task submitted by the user by using the opencv libraryxAnd through the complexity of the task, the prediction capability C of the supported ith target detection model to the task submitted by the user is giveniAcquiring the predicted fraction S of the model to the task by utilizing the prediction capability and the complexity of the taski;
Step S2, using the network status of the current mobile terminal, modeling the network status, respectively calculating to unload the tasks to different server terminals CloudiGenerated communication durationAnd communication energy consumption
Step S3, calculating the calculation energy consumption E required by running different models on different platformsiAnd calculating the time TiAnd generating different weights according to different requirements of users to generate final weighted scores of different models on different platforms for tasksWherein alpha, beta and gamma are weight values adjusted according to user settings, F is relu function, F is sigmoid function, G is logarithmic function with natural constant as base, and the expected fraction SiThrough step S1 deriving, communication durationEnergy consumption of communicationThe energy consumption E is calculated in step S2iAnd duration TiAnd finally, selecting the strategy with the maximum weighting score as the final strategy through calculation in the step.
The step S1 further includes the following steps:
step S11, respectively obtaining the statistical characteristics F of the tasks submitted by the users by using the opencv library in the local platformiAnd using the obtained different statistical characteristics to model the statistical characteristics by using a neural network, and calculating the complexity C of the submitted taskx;
Step S12, the prediction capabilities of all models supported by the cloud platform are obtained locally, and the prediction capability C of a certain model on the task submitted by the user is calculated by utilizing the built neural network modeli;
Step S13, calculating the weighted value of basic attribute such as complexity of model and prediction ability to obtain the predicted score S of model to taski。
In step S11, a multilayer fully-connected neural network is used as a default for modeling.
The step S2 further includes the following steps:
step S21, reading the device information file of the mobile phone end by using C/C + + language, and acquiring some hardware information related to the current transmission of the mobile phone;
step S22, using the transmission time and transmission energy consumption of some fixed points acquired in advance, making linear prediction on the information acquired in step S21, and calculating the generated communication time lengthAnd communication energy consumption
In step S21, some hardware information includes, by default, the current voltage status, the current network speed, and the maximum bandwidth of the mobile phone.
The step S3 includes the steps of:
step S31, selecting a group of fixed weight combinations alpha, beta and gamma according to different setting strategies of users, and increasing the weight corresponding to a certain demand when the user is biased to the demand;
step S32, calculating energy consumption E required by running different models on different platformsiAnd time TiAnd generating a weighted score based on the data of step S13, step S22 and the present step;
in step S33, all the weighted scores generated in step S32 are sorted, and the selection policy that has the largest weighted score is selected as the final selection policy.
In step S31, the weight combination needs to satisfy α + β + γ of 1 by default.
In step S32, a weighted score is generated by an exponential function with a natural constant as a base as a default.
The invention has the beneficial effects that:
the invention provides a task complexity-based high-energy-efficiency target detection task dynamic scheduling method, and aims at designing a scheduling mode for unloading a computation task of a deep neural network by using a cloud under the background that a target detection requirement possibly proposed by a user can cause a large amount of computation and energy consumption pressure on a self mobile platform under the condition that the computation capability of a mobile terminal is insufficient.
According to the scheduling strategy provided by the invention, by comprehensively considering various factors such as user requirements, task complexity, the current mobile terminal state and the like, the calculation pressure brought to the mobile platform can be reduced as much as possible when the user has the target detection task requirements, so that the user experience is greatly improved. Meanwhile, the energy consumption of the mobile terminal can be reduced to the maximum extent while the calculation pressure of the mobile platform is reduced.
Drawings
Fig. 1 is a schematic view of a usage scenario of the present invention.
Fig. 2 is a schematic view of the working process of the present invention.
Detailed Description
The present invention will be described in further detail with reference to the accompanying drawings, but the present invention is not limited to the following embodiments.
Example 1
As shown in a usage scenario diagram of the invention shown in fig. 1, the method for dynamically scheduling an energy-efficient target detection task based on task complexity provided by the invention designs a scheduling method for offloading a computation task of a deep neural network by using a cloud under the background that a target detection requirement possibly proposed by a user causes a large amount of computation and energy consumption pressure on a mobile platform of the user under the situation that the computing power of a mobile terminal is insufficient. By comprehensively considering various factors such as user requirements, task complexity and the state of the current mobile terminal, the computing pressure on the mobile platform can be reduced as much as possible when the user has the requirement of the target detection task, and therefore the user experience is greatly improved and the energy consumption caused by computing is reduced.
As shown in the workflow diagram of fig. 2, after a user submits a task, some hardware information of the mobile platform of the user, such as network speed, power, etc., is first obtained; next, acquiring some configuration information set by the user, such as the minimum waiting time and the like; and finally, selecting the deployed hardware platform and the deployed model according to the information, and giving an optimal selection strategy.
Referring to fig. 1 and 2, a task complexity-based energy-efficient dynamic scheduling method for target detection tasks selects a proper task offloading strategy according to the state of a current user and the complexity of the task when the user submits a target detection task, and specifically includes the following steps:
step S1, acquiring complexity C of the task submitted by the user by using the opencv libraryxAnd through the complexity of the task, the supported prediction energy of the ith target detection model on the task submitted by the user is givenForce CiAcquiring the predicted fraction S of the model to the task by utilizing the prediction capability and the complexity of the taski。
Step S2, using the network status of the current mobile terminal, modeling the network status, respectively calculating to unload the tasks to different server terminals CloudiGenerated communication durationAnd communication energy consumption
Step S3, calculating the calculation energy consumption E required by running different models on different platformsiAnd calculating the time TiAnd generating different weights according to different requirements of users to generate final weighted scores of different models on different platforms for tasksWherein alpha, beta and gamma are weight values adjusted according to user settings, F is relu function, F is sigmoid function, G is logarithmic function with natural constant as base, and the expected fraction SiThe communication duration is determined by step S1Energy consumption of communicationThe energy consumption E is calculated in step S2iAnd duration TiAnd finally, selecting the strategy with the maximum weighting score as the final strategy through calculation in the step.
The step S1 includes the steps of:
step S11, respectively obtaining the statistical characteristics F of the tasks submitted by the users by using the opencv library in the local platformiAnd using the obtained different statistical characteristics to model the statistical characteristics by using a neural network, and calculating the complexity C of the submitted taskx;
Under the default condition, a multilayer fully-connected neural network is used for modeling;
step S12, the prediction capabilities of all models supported by the cloud platform are obtained locally, and the prediction capability C of a certain model on the task submitted by the user is calculated by utilizing the built neural network modeli;
Step S13, calculating the weighted value of basic attribute such as complexity of model and prediction ability to obtain the predicted score S of model to taski。
The step S2 includes the steps of:
step S21, acquiring some current hardware information of the mobile phone by reading the device file of the mobile phone end;
under the default condition, the hardware information comprises the current voltage state, the current network speed and the maximum bandwidth of the mobile phone;
step S22, using the transmission time and transmission energy consumption of some fixed points acquired in advance, making linear prediction on the information acquired in step S21, and calculating the generated communication time lengthAnd communication energy consumption
The step S3 includes the steps of:
step S31, selecting a group of fixed weight combinations alpha, beta and gamma according to the setting strategy of the user;
by default, the weight combination needs to satisfy α + β + γ ═ 1.
Step S32, calculating energy consumption E required by running different models on different platformsiAnd time TiAnd generating a weighted score based on the data of steps S13, S22 and the step;
generating a weighted score by an exponential function with a natural constant as a base under a default condition;
in step S33, all the weighted scores generated in step S32 are sorted, and the selection policy that has the largest weighted score is selected as the final selection policy.
Claims (8)
1. A high-energy-efficiency target detection task dynamic scheduling method based on task complexity selects a proper task unloading strategy according to the state of a current user and the task complexity when the user submits a target detection task, and is characterized by comprising the following steps:
step S1, acquiring complexity C of the task submitted by the user by using the opencv libraryxAnd through the complexity of the task, the prediction capability C of the supported ith target detection model to the task submitted by the user is giveniAcquiring the predicted fraction S of the model to the task by utilizing the prediction capability and the complexity of the taski;
Step S2, using the network status of the current mobile terminal, modeling the network status, respectively calculating to unload the tasks to different server terminals CloudiGenerated communication durationAnd communication energy consumption
Step S3, calculating the calculation energy consumption E required by running different models on different platformsiAnd calculating the time TiAnd generating different weights according to different requirements of users to generate final weighted scores of different models on different platforms for tasksWherein alpha, beta and gamma are weight values adjusted according to user settings, F is relu function, F is sigmoid function, G is logarithmic function with natural constant as base, and the expected fraction SiThe communication duration is determined by step S1Energy consumption of communicationThe energy consumption E is calculated in step S2iAnd duration TiAnd finally, selecting the strategy with the maximum weighting score as the final strategy through calculation in the step.
2. The method for dynamically scheduling task-complexity-based energy-efficient target detection tasks according to claim 1, wherein the step S1 further comprises the following steps:
step S11, respectively obtaining the statistical characteristics F of the tasks submitted by the users by using the opencv library in the local platformiAnd using the obtained different statistical characteristics to model the statistical characteristics by using a neural network, and calculating the complexity C of the submitted taskx;
Step S12, the prediction capabilities of all models supported by the cloud platform are obtained locally, and the prediction capability C of a certain model on the task submitted by the user is calculated by utilizing the built neural network modeli;
Step S13, calculating the weighted value of basic attribute such as complexity of model and prediction ability to obtain the predicted score S of model to taski。
3. The method for dynamically scheduling task-complexity-based energy-efficient target detection tasks according to claim 2, wherein in step S11, a multi-layer fully-connected neural network is used for modeling by default.
4. The method for dynamically scheduling task-complexity-based energy-efficient target detection tasks according to claim 1, wherein the step S2 further comprises the following steps:
step S21, reading the device information file of the mobile phone end by using C/C + + language, and acquiring some hardware information related to the current transmission of the mobile phone;
5. The method as claimed in claim 4, wherein in step S21, some hardware information includes current voltage status, current network speed, and maximum bandwidth of the mobile phone by default.
6. The method for dynamically scheduling energy-efficient object detection tasks based on task complexity as claimed in claim 1, wherein the step S3 comprises the following steps:
step S31, selecting a group of fixed weight combinations alpha, beta and gamma according to different setting strategies of users, and increasing the weight corresponding to a certain demand when the user is biased to the demand;
step S32, calculating energy consumption E required by running different models on different platformsiAnd time TiAnd generating a weighted score based on the data of step S13, step S22 and the present step;
in step S33, all the weighted scores generated in step S32 are sorted, and the selection policy that has the largest weighted score is selected as the final selection policy.
7. The method for dynamically scheduling task-complexity-based energy-efficient target detection tasks according to claim 6, wherein in step S31, the weight combination needs to satisfy α + β + γ ═ 1 by default.
8. The method for dynamically scheduling energy-efficient object detection tasks according to claim 6, wherein in step S32, the weighting score is generated by an exponential function with a natural constant as a base by default.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110435960.0A CN113157453B (en) | 2021-04-22 | 2021-04-22 | Task complexity-based high-energy-efficiency target detection task dynamic scheduling method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110435960.0A CN113157453B (en) | 2021-04-22 | 2021-04-22 | Task complexity-based high-energy-efficiency target detection task dynamic scheduling method |
Publications (2)
Publication Number | Publication Date |
---|---|
CN113157453A true CN113157453A (en) | 2021-07-23 |
CN113157453B CN113157453B (en) | 2022-11-25 |
Family
ID=76869723
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202110435960.0A Active CN113157453B (en) | 2021-04-22 | 2021-04-22 | Task complexity-based high-energy-efficiency target detection task dynamic scheduling method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN113157453B (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116880401A (en) * | 2023-07-28 | 2023-10-13 | 江苏道达智能科技有限公司 | Automatic stereoscopic warehouse control system and method |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20140289733A1 (en) * | 2013-03-22 | 2014-09-25 | Palo Alto Research Center Incorporated | System and method for efficient task scheduling in heterogeneous, distributed compute infrastructures via pervasive diagnosis |
CN109376012A (en) * | 2018-10-10 | 2019-02-22 | 电子科技大学 | A kind of self-adapting task scheduling method based on Spark for isomerous environment |
CN111756812A (en) * | 2020-05-29 | 2020-10-09 | 华南理工大学 | Energy consumption perception edge cloud cooperation dynamic unloading scheduling method |
WO2021003422A1 (en) * | 2019-07-02 | 2021-01-07 | Northeastern University | Network and method for servicing a computation request |
-
2021
- 2021-04-22 CN CN202110435960.0A patent/CN113157453B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20140289733A1 (en) * | 2013-03-22 | 2014-09-25 | Palo Alto Research Center Incorporated | System and method for efficient task scheduling in heterogeneous, distributed compute infrastructures via pervasive diagnosis |
CN109376012A (en) * | 2018-10-10 | 2019-02-22 | 电子科技大学 | A kind of self-adapting task scheduling method based on Spark for isomerous environment |
WO2021003422A1 (en) * | 2019-07-02 | 2021-01-07 | Northeastern University | Network and method for servicing a computation request |
CN111756812A (en) * | 2020-05-29 | 2020-10-09 | 华南理工大学 | Energy consumption perception edge cloud cooperation dynamic unloading scheduling method |
Non-Patent Citations (4)
Title |
---|
XIN HE: "Offloading Deadline-aware Task in Edge Computing", 《2020 IEEE 13TH INTERNATIONAL CONFERENCE ON CLOUD COMPUTING (CLOUD)》 * |
任杰等: "面向边缘设备的高能效深度学习任务调度策略", 《计算机学报》 * |
汪灿: "面向边缘环境的视频分析任务调度问题研究", 《中国优秀博硕士学位论文全文数据库(硕士)信息科技辑》 * |
马景超: "移动边缘计算场景下任务执行效率优化策略研究", 《中国优秀博硕士学位论文全文数据库(硕士)信息科技辑》 * |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116880401A (en) * | 2023-07-28 | 2023-10-13 | 江苏道达智能科技有限公司 | Automatic stereoscopic warehouse control system and method |
Also Published As
Publication number | Publication date |
---|---|
CN113157453B (en) | 2022-11-25 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110223292B (en) | Image evaluation method, device and computer readable storage medium | |
CN112541124B (en) | Method, apparatus, device, medium and program product for generating a multitasking model | |
CN111985495A (en) | Model deployment method, device, system and storage medium | |
US20230401833A1 (en) | Method, computer device, and storage medium, for feature fusion model training and sample retrieval | |
CN111967608A (en) | Data processing method, device, equipment and storage medium | |
CN114239861A (en) | Model compression method and system based on multi-teacher combined guidance quantification | |
CN112200296A (en) | Network model quantification method and device, storage medium and electronic equipment | |
CN111369430A (en) | Mobile terminal portrait intelligent background replacement method based on mobile deep learning engine | |
CN113157453B (en) | Task complexity-based high-energy-efficiency target detection task dynamic scheduling method | |
CN110503149B (en) | Method and system for classifying local features in image | |
CN112101543A (en) | Neural network model determination method and device, electronic equipment and readable storage medium | |
CN114943335A (en) | Layer-by-layer optimization method of ternary neural network | |
CN113420792A (en) | Training method of image model, electronic equipment, road side equipment and cloud control platform | |
CN112308213A (en) | Convolutional neural network compression method based on global feature relationship | |
CN112183744A (en) | Neural network pruning method and device | |
CN116580251A (en) | Image classification method, device, equipment and computer readable storage medium | |
CN111276138B (en) | Method and device for processing voice signal in voice wake-up system | |
US20220004849A1 (en) | Image processing neural networks with dynamic filter activation | |
CN112561050A (en) | Neural network model training method and device | |
CN113159318A (en) | Neural network quantification method and device, electronic equipment and storage medium | |
CN113361707A (en) | Model compression method, system and computer readable medium | |
CN114139678A (en) | Convolutional neural network quantization method and device, electronic equipment and storage medium | |
CN114386469A (en) | Method and device for quantizing convolutional neural network model and electronic equipment | |
CN112488291A (en) | Neural network 8-bit quantization compression method | |
CN116776926B (en) | Optimized deployment method, device, equipment and medium for dialogue model |
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 | ||
GR01 | Patent grant | ||
GR01 | Patent grant |