An edge computing model training method based on automated machine learning

Abstract

The present invention discloses an edge computing model training method based on automated machine learning, comprising the following steps: obtaining current computing task information and server computing resource conditions before training a computing model on an edge server; learning and constraining hyperparameters of an automated machine learning algorithm by NAS (Neural Architecture Search) according to the computing task information and computing resource conditions; after determining the hyperparameters of the automated machine learning algorithm, automatically training an optimal computing model according to input data by the automated machine learning algorithm. Different from the computing models on current edge servers, the models of this method adopt automated machine learning technology, which can reduce the model development and maintenance costs, and consider the heterogeneity of edge servers, resource limitations, and the problem of computing latency requirements of computing tasks when automatically generating models, making the generated computing models more generalizable.

Description

Technical Field

[0001] The present invention relates to the field of edge computing, and particularly to a method for training an edge computing model based on automated machine learning. Background Art

[0002] In order to overcome problems faced by traditional IT technologies, such as low resource utilization, high production costs, low network utilization, and high energy consumption, cloud computing technology emerged. Enterprises can use distributed computing and virtualization technologies to build data centers or supercomputers, and provide hardware resources (servers, memories, CPUs, etc.) and software resources (application software, integrated development environments, etc.) to technical personnel or enterprise users in the form of free or on-demand rental. Enterprise users only need to select corresponding services according to their actual needs, without having to purchase, own, and maintain physical data devices and databases. Additionally, cloud computing also has advantages such as wide network access, resource pooling, rapid elastic scaling, and metered services.

[0003] However, with the rapid development of 5G technology and Internet of Things technology in recent years and the advent of the era of all things connected, a large number of terminal devices have been connected to the network, resulting in an explosive growth in computing demands. Among them, there are many computing tasks with high requirements for the real-time nature of computing services, such as autonomous driving, VR games, and smart homes. When facing challenges in aspects such as massive data computing, emerging computing scenarios, and real-time processing of small data, the traditional cloud computing architecture limited by factors such as distance and network bandwidth is difficult to meet the requirements of these computing tasks.

[0004] Therefore, it will be an important development trend to sink the capabilities of cloud computing to the edge side and terminal side, and perform unified delivery, operation and maintenance, and management and control through the cloud. With the computing power sinking from the cloud to the edge, a new computing system - the cloud-edge-terminal collaborative computing architecture has emerged. Among them, "cloud" is the traditional cloud data center, which is the management and control end of edge computing; "edge" is the edge network composed of edge nodes between the cloud data center and terminal devices; "terminal" refers to intelligent terminal devices, such as mobile phones, smart home appliances, various sensors, etc. Compared with the traditional cloud computing architecture, the cloud-edge-terminal collaborative computing architecture does not need to upload all the computing tasks and data generated by terminal devices to the cloud. It only needs to hand over small-scale and real-time intelligent analysis tasks to the edge nodes near the terminal, shortening the distance between computing resources and customers, and enabling faster network service responses, meeting the needs of various industries in aspects such as real-time services, application intelligence, security, and privacy protection.

[0005] However, due to reasons such as limited computing resources on edge servers, heterogeneous chips cannot be coordinated and utilized uniformly, and the low utilization rate of computing resources, problems such as high computing latency will actually occur. Therefore, in recent years, technicians have proposed models and methods for optimizing the resource utilization rate of edge servers, protecting data privacy, and reducing computing latency based on various technologies.

[0006] For example, the "Method for Training Neural Network Model Based on Edge Computing" proposed by Beijing Jiaotong University (application number 202210176214.9) improves the model by using non-learning layers, optimizes data transmission between the edge side and terminal devices, and reduces communication latency.

[0007] The "Task Scheduling Method Based on Deep Reinforcement Learning in Hierarchical Edge Computing Environment" proposed by Chongqing University of Technology (application number 202111012837.4) improves the quality of scheduling decisions and can achieve a balance between efficiency and quality.

[0008] The "Task Scheduling Method Based on Edge Computing" proposed by Shanghai Jiao Tong University (application number 202210007565.7) maintains a policy network in each region, updates network parameters online based on the deep reinforcement learning algorithm of federated learning, and each region independently schedules tasks received in real time within its own region to enhance model scalability, protect data privacy, and improve system performance.

[0009] The above model methods have improved the performance of edge servers to a certain extent, but they do not consider the heterogeneity of edge servers and the problem of limited computing resources on edge servers. Therefore, the present invention applies automatic machine learning technology to edge computing, automatically trains a computing model, thereby obtaining an optimal computing model, and reducing the cost of model development and maintenance. At the same time, considering the problems of limited computing resources on edge servers and the requirements of computing tasks for computing latency, by learning and constraining the hyperparameters of the automatic machine learning algorithm according to computing task information and computing resource conditions, the generated computing model has stronger generalization ability. Summary of the Invention

[0010] The purpose of the present invention is to overcome the shortcomings and deficiencies of the prior art, and provide a method for training an edge computing model based on automatic machine learning. This method can consider the heterogeneity of edge servers, resource limitations, and the requirements of computing tasks for computing latency, make the generated computing model have stronger generalization ability, and at the same time reduce the cost of model development and maintenance.

[0011] The purpose of the present invention is achieved by the following technical solutions:

[0012] A method for training an edge computing model based on automatic machine learning, comprising the following steps:

[0013] Step 1: Before training the computing model, the edge server obtains the information matrix Job of each computing task in the current task queue and the computing resource matrix Source of the current server;

[0014] Step 2: Combine the information matrix x1 = Job of the computing task and the computing resource matrix x2 = Source of the server to form X = {x1, x2}, and use X as the input of the subsequent neural network model;

[0015] Step 3: Input X in Step 2 into the already trained neural network model, and finally the neural network model outputs the hyperparameters Hyperparameters of the automated machine learning algorithm;

[0016] Step 4: Set the hyperparameters of the automated machine learning algorithm according to the hyperparameters Hyperparameters in Step 3, and then automatically train the optimal machine learning computing model Best using the input data through the automated machine learning algorithm model 。

[0017] An edge computing model training method based on automated machine learning includes the following steps:

[0018] S1. For the image recognition computing task of the electricity meter reading, deploy a script on the Atlas 200DK edge device in advance that can automatically obtain the information matrix Job of the computing tasks generated when different computing tasks run on the CPU and NPU and the computing resource matrix Source of the server;

[0019] S2. Regularly collect the image recognition tasks of m electricity meter readings and the generated information Job on the Atlas 200DK edge device, where Job = {j1, j2,..., j m}, where j = [Runtime, Energy, Node, T]; and the corresponding historical data information Trace = {Source1, Source2,..., Source m}, where Source = [Fnode, Fmemory, Energy];

[0020] S3. Deploy a NAS automatic training script based on the deep learning framework TensorFlow on the Atlas 200DK edge device;

[0021] S4. Use the NAS automatic training script in step S3 to train a DNN neural network model Model with a maximum number of layers not exceeding 10 and a width of no more than 300 neurons per layer. Its input is X = {Job, Source}, and the output is the preset hyperparameters Hyperparameters = [Layers, Filters, Kernel, Loss, Seed] of the NAS automatic training algorithm for the CNN model that identifies the electricity meter reading. Here, Layers represents the maximum number of layers searched by the CNN, Filters and Kernel represent the maximum number of Filters channels and the maximum convolution kernel size of each convolutional layer of the CNN model, Loss represents the selected loss function, and Seed represents the initial seed number of the CNN;

[0022] S5. Before the Atlas 200DK edge device trains the CNN model for identifying electricity meter readings, obtain the information matrix Job_now of each computing task in the current task queue. Its attribute features include: the approximate time required for the computing task to run on the CPU and NPU, the approximate energy consumption caused by the computing task running on the CPU and NPU, the number of computing nodes required for the computing task to run on the CPU and NPU, and the computing delay requirement of the computing task; and the computing resource matrix Source_now of the current Atlas 200DK edge device. Its attribute features include: the idle computing node situation of the two NPUs and the CPU on the Atlas 200DK edge device, the idle computing memory of the two NPUs and the CPU, and the average energy consumption of the device during the current time period;

[0023] S6. Use the task information matrix Job and the computing resource matrix Source obtained in step S5 as the input X, X = [Job_now, Source_now}, and input them into the DNN neural network model Model obtained in step S4 to obtain the preset hyperparameters Hyperparameters = [6, 64, 10, Softmax, 3] of the NAS automatic training algorithm for the CNN model that identifies the electricity meter reading;

[0024] S7. Uniformly process the m electricity meter reading images regularly collected on the Atlas 200DK edge device into 256*256-sized images and cooperate with the image cutting program to form the input dataset Input for training the CNN model;

[0025] S8. Based on the deep learning framework TensorFlow, use the Hyperparameters obtained in S6 to set the hyperparameters of the NAS algorithm of the CNN model, and use the Input generated in step S7 as the training set to train the optimal CNN model Best model, which includes three convolutional pooling layers, two fully connected layers, and the loss function is Softmax;

[0026] S9. Use the optimal CNN model Best obtained in S8 model to recognize the electricity meter image and obtain the final electricity meter reading.

[0027] The attribute features of the information matrix Job of the computing task include: the approximate time required for the computing task to run on different computing chips, the approximate energy consumption caused by the computing task running on different computing chips, the number of computing nodes required for the computing task to run on different computing chips, and the computing delay requirement of the computing task.

[0028] The attribute features of the computing resource matrix Source of the server include the situation of idle computing nodes on different computing chips, the idle computing memory on different computing chips, and the average energy consumption at the current time period.

[0029] The neural network model takes the information matrix Job of historical computing tasks and the computing resource matrix Source of the edge server as inputs, the hyperparameters Hyperparameters of the automated machine learning algorithm as outputs, uses cross-entropy loss as the loss function, and is automatically trained through NAS using algorithms such as evolutionary algorithms, greedy algorithms, or grid search.

[0030] By artificially initializing the hyperparameters Pipleine and RandomSeed of the NAS, starting from the initial seed, continuously changing the network structure and weight parameters until reaching the constrained model size, various models are obtained during the process, and the optimal model is selected according to the model scores.

[0031] In step 3, the automated machine learning algorithm includes: neural network architecture search, hyperparameter optimization, CASH, automated data mining, automatic reinforcement learning, meta-learning, Bayesian optimization of automated machine learning, evolutionary algorithm of automated machine learning, multi-objective optimization of automated machine learning, automatic artificial intelligence;

[0032] Among them, the characteristic attributes of the hyperparameters of the neural network architecture search include: the hyperparameter Pipeline that constrains the final model structure size and scale; the model initial seed parameter RandomSeed that accelerates model optimization.

[0033] In step 4, the input data is the training set corresponding to different computing models.

[0034] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0035] Before training a computing model on an edge server, the present invention obtains the current computing task information and the server computing resource situation; through NAS (Neural Architecture Search), it learns and constrains the hyperparameters of the automated machine learning algorithm according to the computing task information and the computing resource situation; after determining the hyperparameters of the automated machine learning algorithm, it automatically trains an optimal computing model according to the input data through the automated machine learning algorithm. Different from the computing model on the current edge server, the present invention generates a computing model through automated machine learning technology on the premise of considering the heterogeneity, limited resources of the edge server, and the requirements of computing tasks for computing latency, reducing the development and maintenance costs of the model and making the generated computing model more generalizable. Description of the Drawings

[0036] Figure 1 It is a schematic diagram of a method for training an edge computing model based on automated machine learning.

[0037] Figure 2 It is a schematic diagram of the information matrix of a computing task.

[0038] Figure 3 It is a schematic diagram of the edge server computing resource matrix.

[0039] Figure 4 It is a flowchart of a method for training an edge computing model based on automated machine learning. Detailed Embodiment

[0040] The present invention will be further described in detail below in conjunction with the embodiments and the drawings, but the embodiments of the present invention are not limited thereto.

[0041] As Figures 1-4 , a method for training an edge computing model based on automated machine learning includes the following steps:

[0042] Step 1: An edge server is a computer at the end or "edge" of the network, closer to the user. Before training a computing model, the edge server should obtain the information matrix Job of each computing task in the current task queue, and its attribute features include: the approximate time required for the computing task to run on different computing chips, the approximate energy consumption caused by the computing task running on different computing chips, the number of computing nodes required for the computing task to run on different computing chips, and the computing latency requirement of the computing task; and the computing resource matrix Source of the current server, and its attribute features are the idle computing node situation on different computing chips, the idle computing memory on different computing chips, and the average energy consumption in the current period.

[0043] The information matrix Job of the computing task includes the following attributes:

[0044] Approximate time required for a computing task to run on different computing chips:

[0045] Runtime (1)

[0046] = (CPU runtime , GPU runtime , NPU runtime ……)

[0047] Approximate energy consumption caused by a computing task running on different computing chips:

[0048] Energy (2)

[0049] = (CPU energy , GPU energy , NPU energy ……)

[0050] Number of computing nodes required for a computing task to run on different computing chips:

[0051] Node = (CPU node , GPU node , NPU node ......) (3)

[0052] Computing latency requirement T of the computing task.

[0053] Among them, the approximate time required for a computing task to run on different computing chips, the approximate energy consumption caused by a computing task running on different computing chips, and the number of computing nodes required for a computing task to run on different computing chips can be obtained from historical data.

[0054] The computing resource matrix Source includes the following attributes:

[0055] Idle computing nodes of different computing chips on the server:

[0056] Fnode = (CPU k , GPU k , NPU k ......) (4)

[0057] Idle computing memory of different computing chips on the server:

[0058] Fmemory = (CPU m , GPU m , NPU m ......) (5)

[0059] Average energy consumption energy of the server during the current time period.

[0060] Step 2: Combine the information matrix x1 = Job of the computing task and the computing resource matrix x2 = Source of the server to form X = {x1, x2} as the input for the subsequent neural network model.

[0061] Step 3: Input X in Step 2 into the pre-trained neural network, and finally output the hyperparameters Hyperparameters of the automated machine learning algorithm.

[0062] The 10 techniques covered by the automated machine learning algorithm include:

[0063] Neural Architecture Search (NAS Neural Architecture Search)

[0064] Hyperparameter Optimization (HPO Hyperparameter Optimization)

[0065] CASH (Combined Algorithm Selection and Hyperparameter Optimization)

[0066] Automated Data Mining

[0067] AutoRL Automated Reinforcement Learning

[0068] Meta-Learning and Learning to Learn

[0069] Bayesian Optimization for AutoML

[0070] Evolutionary Algorithm for AutoML

[0071] Multi-Objective Optimization for AutoML

[0072] AutoAI

[0073] Taking NAS (Neural Architecture Search) as an example, the characteristic attributes of its hyperparameters include: the hyperparameter Pipeline that restricts the size and scale of the final model structure; the model initial seed parameter Random Seed that accelerates model optimization.

[0074] Step 4: Set the hyperparameters of the automated machine learning algorithm according to the Hyperparameters in Step 3, and then use the input data to automatically train the optimal machine learning calculation model Best through the automated machine learning algorithm model ; The input data is the training set corresponding to different calculation models.

[0075] The neural network model takes the information matrix Job of historical computing tasks and the computing resource matrix Source of edge servers as inputs, and the hyperparameters Hyperparameters of the automated machine learning algorithm as outputs, uses cross-entropy loss as the loss function, and is automatically trained through NAS using algorithms such as evolutionary algorithms, greedy algorithms, or grid search.

[0076] By artificially initializing the hyperparameters Pipeline and Random Seed of NAS, starting from the initial seed, continuously changing the network structure and weight parameters until reaching the constrained model size, various models are obtained during the process, and the optimal model is selected according to the model scores.

[0077] Apply the method of the present invention in the scenario of intelligent meter reading in the power grid:

[0078] For the image recognition computing task of meter reading, an implementation of an edge computing model training method based on automated machine learning on a Huawei Atlas 200DK edge device includes the following steps:

[0079] S1. Deploy a script on the Atlas 200DK edge device in advance that can automatically obtain the computing task information matrix Job and the server computing resource matrix Source generated when different computing tasks run on the CPU and NPU.

[0080] S2. Regularly collect the image recognition tasks of m meter readings and the generated information Job on the Atlas 200DK edge device, where Job = {j1, j2,..., j m}, where j = [Runtime, Energy, Node, T]; and the corresponding historical data information Trace = {Source1, Source2,..., Source m}, where Source = [Fnode, Fmemory, Energy].

[0081] S3. Deploy a NAS automatic training script based on the deep learning framework TensorFlow on the Atlas 200DK edge device.

[0082] S4. Use the NAS automatic training script in S3 to train a DNN neural network model Model with a maximum number of layers not exceeding 10 and a width of no more than 300 neurons per layer. Its input is X = {Job, Source}, and the output is the preset hyperparameters Hyperparameters = [Layers, Filters, Kernel, Loss, Seed] of the NAS automatic training algorithm of the CNN model for identifying electricity meter readings, where Layers represents the maximum number of layers searched by the CNN, Filters and Kernel represent the maximum number of Filters channels and the maximum convolution kernel size of each convolutional layer of the CNN model, Loss represents the selected loss function, and Seed represents the initial seed number of the CNN.

[0083] S5: Before the Atlas 200DK edge device trains the CNN model for identifying electricity meter readings, obtain the information matrix Job_now of each computing task in the current task queue. Its attribute features include: the approximate time required for the computing task to run on the CPU and NPU, the approximate energy consumption caused by the computing task running on the CPU and NPU, the number of computing nodes required for the computing task to run on the CPU and NPU, and the computing delay requirement of the computing task; and the computing resource matrix Source_now of the current Atlas 200DK edge device. Its attribute features include: the idle computing node situation of the two NPUs and the CPU on the Atlas 200DK edge device, the idle computing memory of the two NPUs and the CPU, and the average energy consumption of the device during the current time period.

[0084] S6: Take the task information matrix Job and the computing resource matrix Source obtained in S5 as the input X, X = {Job_now, Source_now}, and input them into the DNN neural network model Model obtained in S4 to obtain the preset hyperparameters Hyperparameters = [6, 64, 10, Softmax, 3] of the NAS automatic training algorithm of the CNN model for identifying electricity meter readings.

[0085] S7: Uniformly process the m electricity meter reading images regularly collected on the Atlas 200DK edge device into 256*256-sized images and cooperate with the image cutting program to form the input dataset Input for training the CNN model.

[0086] S8: Based on the deep learning framework TensorFlow, use the Hyperparameters obtained in S6 to set the hyperparameters of the NAS algorithm of the CNN model, and use the Input generated in S7 as the training set to train the optimal CNN model Best model, which includes three convolutional pooling layers (filter = [64, 32, 8], kernel = [4, 4, 4], stride and padding are default parameter 1), two fully connected layers ([128, 10]), and the loss function is Softmax.

[0087] S9: Use the optimal CNN model Best obtained in S8 model to recognize the electricity meter image and obtain the final electricity meter reading. Best model Compared with the manually trained CNN model, when maintaining the same inference accuracy of 98.57%, the energy consumption is reduced by 17%, the memory occupancy is reduced by 11%, and the inference speed is increased by 261.25%.

[0088] In summary, the edge computing model training method based on automated machine learning of the present invention has the following innovative points:

[0089] First, applying the automated machine learning technology to edge computing to reduce the model development and maintenance costs.

[0090] Second, when automatically generating the model, considering issues such as the heterogeneity of edge servers, limited resources, and the requirements of computing tasks for computing latency, so that the generated computing model has stronger generalization ability.

[0091] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications made without departing from the spirit and principle of the present invention shall be equivalent replacement methods and are all included in the protection scope of the present invention.

Claims

1. A method for training an edge computing model based on automated machine learning, characterized in that, It includes the following steps: Step 1: Before training the computing model, the edge server obtains the information matrix Job of each computing task in the current task queue and the computing resource matrix Source of the current server; The attribute features of the information matrix Job of the computing task include: the approximate time required for the computing task to run on different computing chips, the approximate energy consumption caused by the computing task running on different computing chips, the number of computing nodes required for the computing task to run on different computing chips, and the computing delay requirement of the computing task; The attribute features of the computing resource matrix Source of the server include the situation of idle computing nodes on different computing chips, the idle computing memory on different computing chips, and the average energy consumption in the current period; Step 2: Combine the information matrix x1 = Job of the computing task and the computing resource matrix x2 = Source of the server to form X = {x1, x2}, and use X as the input of the subsequent neural network model; The neural network model takes the information matrix Job of historical computing tasks and the computing resource matrix Source of the edge server as inputs, and the hyperparameters Hyperparameters of the automatic machine learning algorithm as outputs. It is automatically trained using the cross-entropy loss as the loss function through NAS using evolutionary algorithms, greedy algorithms, or grid search algorithms; By artificially initializing the hyperparameters Pipeline and Random Seed of the NAS, starting from the initial seed, continuously changing the network structure and weight parameters until reaching the constrained model size. Various models are obtained during the process, and the optimal model is selected according to the model scores; Among them, the characteristic attributes of the hyperparameters of the neural network architecture search include: the hyperparameter Pipeline that constrains the final model structure size and scale; the model initial seed parameter Random Seed that accelerates model optimization; Step 3: Input X in Step 2 into the already trained neural network model, and finally the neural network model outputs the hyperparameters Hyperparameters of the automatic machine learning algorithm; Step 4: Set the hyperparameters of the automated machine learning algorithm according to the hyperparameters in Step 3, and then use the input data to automatically train the optimal machine learning calculation model Best through the automated machine learning algorithm model .

2. The edge computing model training method based on automated machine learning according to claim 1, wherein It includes the following steps: S1. For the image recognition computing task of the electricity meter reading, deploy a script on the Atlas 200DK edge device in advance that can automatically obtain the information matrix Job of the computing task generated when different computing tasks run on the CPU and NPU and the computing resource matrix Source of the server; S2. Periodically collect the image recognition tasks of m electricity meter readings and the information Job generated on the Atlas 200DK edge device, where Job = {j1, j2,..., j m}, where j = [Runtime, Energy, Node, T]; and the corresponding historical data information Trace = {Source1, Source2,..., Source m}, where Source = [Fnode, Fmemory, Energy]; S3. Deploy the NAS automatic training script based on the deep learning framework TensorFlow on the Atlas 200DK edge device; S4. Use the NAS automatic training script in step S3 to train a DNN neural network model Model with a maximum number of layers not exceeding 10 and a width of no more than 300 neurons per layer. Its input is X = {Job, Source}, and the output is the preset hyperparameters Hyperparameters = [Layers, Filters, Kernel, Loss, Seed] of the NAS automatic training algorithm of the CNN model for identifying electricity meter readings. Among them, Layers represents the maximum number of layers searched by the CNN, Filters and Kernel represent the maximum number of Filters channels and the maximum convolution kernel size of each convolutional layer of the CNN model, Loss represents the selected loss function, and Seed represents the initial seed number of the CNN; S5. Before the Atlas 200DK edge device trains the CNN model for identifying electricity meter readings, obtain the information matrix Job_now of each computing task in the current task queue. Its attribute features include: the approximate time required for the computing task to run on the CPU and NPU, the approximate energy consumption caused by the computing task running on the CPU and NPU, the number of computing nodes required for the computing task to run on the CPU and NPU, and the computing delay requirement of the computing task; and the computing resource matrix Source_now of the current Atlas 200DK edge device. Its attribute features include: the idle computing node situation of the two NPUs and the CPU on the Atlas 200DK edge device, the idle computing memory of the two NPUs and the CPU, and the average energy consumption of the device during the current time period; S6. Take the task information matrix Job and the computing resource matrix Source obtained in step S5 as the input X, X = {Job_now, Source_now}, and input them into the DNN neural network model Model obtained in step S4 to obtain the preset hyperparameters Hyperparameters = [6, 64, 10, Softmax, 3] of the NAS automatic training algorithm of the CNN model for identifying electricity meter readings; S7. Uniformly process m electricity meter reading images regularly collected on the Atlas 200DK edge device into 256*256-sized images and cooperate with the image cutting program to form the input dataset Input for training the CNN model; S8. Based on the deep learning framework TensorFlow, use the Hyperparameters obtained in S6 to set the hyperparameters of the NAS algorithm for the CNN model, and use the Input generated in step S7 as the training set to train the optimal CNN model Best model , which includes three convolutional pooling layers, two fully connected layers, and the loss function is Softmax; S9. Use the optimal CNN model Best obtained in S8 model to identify the electricity meter image and obtain the final electricity meter reading.

3. The edge computing model training method based on automated machine learning according to claim 1, wherein In step 3, the automatic machine learning algorithm includes: neural network architecture search, hyperparameter optimization, CASH, automated data mining, automatic reinforcement learning, meta-learning, Bayesian optimization of automatic machine learning, evolutionary algorithm of automatic machine learning, multi-objective optimization of automatic machine learning, automatic artificial intelligence.

4. The method for training an edge computing model based on automated machine learning according to claim 1, wherein In step 4, the input data is the training set corresponding to different computing models.

Images