End-side collaborative graph neural network architecture search and deployment method and device

Abstract

The embodiment of the invention discloses a framework searching and deploying method and device for an end-edge collaborative graph neural network, and belongs to the technical field of computers. The architecture searching method comprises the following steps: acquiring target application scene information of a target collaborative graph neural network; obtaining a plurality of candidate architectures of the target collaborative graph neural network based on the target application scene information and a collaborative graph neural network hypernetwork; wherein each layer in the collaborative graph neural network super network comprises each operation in a collaborative search space and a function corresponding to each operation, the operation comprises a communication operation, and each candidate architecture is obtained by randomly selecting a target operation and a target function corresponding to the target operation in each layer of the collaborative graph neural network super network; and selecting a target architecture of the target collaborative graph neural network from the plurality of candidate architectures based on collaborative reasoning performance information of each candidate architecture for the target application identification information under the target end edge environment information.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a method and device for architecture search and deployment of an end-edge collaborative graph neural network. Background Art

[0002] Graph Neural Network (GNN) is a neural network model that processes non-Euclidean spatial data. It has demonstrated strong learning capabilities in many applications and has strong application potential in research fields such as social network analysis, knowledge graphs, recommendation systems, and point cloud processing. In practical applications, a graph neural network with a suitable architecture can be selected, trained, and deployed on a device so that the graph neural network can perform reasoning on the device.

[0003] The essence of graph neural networks is graph + neural network. Although it has powerful learning capabilities, its calculations are relatively complex. A single device often has limited resources, making it difficult for graph neural networks to efficiently perform reasoning on it. In order to enable graph neural networks to perform reasoning efficiently, the graph neural network can be deployed in end-edge collaboration so that the graph neural network can perform end-edge collaborative reasoning. In related technologies, for end-edge collaborative graph neural networks, a suitable architecture can be manually designed based on expert experience. It is difficult to take into account all bottlenecks when manually designing the architecture of end-edge collaborative graph neural networks, resulting in poor reasoning performance of the designed graph neural architecture after end-edge collaborative deployment. Summary of the invention

[0004] The purpose of the embodiments of the present application is to provide an architecture search and deployment method and device for an end-edge collaborative graph neural network, so as to improve the end-edge collaborative reasoning performance of the end-edge collaboratively deployed graph neural network.

[0005] In a first aspect, an embodiment of the present application provides an architecture search method for a device-edge collaborative graph neural network, the method comprising: Obtain target application scenario information of the target collaborative graph neural network; wherein the target application scenario information includes target application identification information and target end-side environment information; Based on the target application scenario information and the collaborative graph neural network supernet, multiple candidate architectures of the target collaborative graph neural network are obtained; wherein the collaborative graph neural network supernet is constructed based on the collaborative search space of the end-edge collaborative graph neural network, and each layer in the collaborative graph neural network supernet includes each operation in the collaborative search space and a function corresponding to each operation, and the operation includes a communication operation, and the communication operation indicates that a device switching is performed between an end device and an edge device, and each candidate architecture is obtained by randomly selecting a target operation and a target function corresponding to the target operation in each layer of the collaborative graph neural network supernet; Based on the collaborative reasoning performance information of each candidate architecture for the target application identification information under the target end-edge environment information, a target architecture of the target collaborative graph neural network is selected from the multiple candidate architectures.

[0006] In a second aspect, an embodiment of the present application provides a method for deploying an architecture of a device-edge collaborative graph neural network, the method comprising: Based on the first end-edge environment information of the end-edge system for deploying the target collaborative graph neural network, a first target architecture is obtained from a target architecture warehouse; wherein the target architecture warehouse is an architecture warehouse of the target collaborative graph neural network, which stores multiple target architectures corresponding to multiple end-edge environment information of the target collaborative graph neural network, and the target architecture corresponding to each end-edge environment information is obtained by taking each end-edge environment information as the target end-edge environment information in the target application scenario information, and performing the architecture search method as described in the first aspect of the embodiment of the present application; the first target architecture is a target architecture corresponding to the first end-edge environment information; Deploying the first target architecture to the end-edge system to perform end-edge collaborative reasoning on the end-edge system based on the first target architecture; During the process of performing edge-end collaborative reasoning on the edge-end system based on the first target architecture, monitoring changes in the edge-end environment of the edge-end system; When the edge device environment changes, a second target architecture is acquired from the target architecture repository; wherein the second target architecture is a target architecture corresponding to the second edge device environment information, and the second edge device environment information is the changed edge device environment information; The second target architecture is deployed on the end-edge system to perform end-edge collaborative reasoning on the end-edge system based on the second target architecture.

[0007] In a third aspect, an embodiment of the present application further provides an architecture search device for a device-edge collaborative graph neural network, the device comprising: A scenario information acquisition module, used to acquire target application scenario information of a target collaborative graph neural network; wherein the target application scenario information includes target application identification information and target end-side environment information; A candidate architecture search module, used to obtain multiple candidate architectures of the target collaborative graph neural network based on the target application scenario information and the collaborative graph neural network supernet; wherein the collaborative graph neural network supernet is constructed based on the collaborative search space of the end-edge collaborative graph neural network, each layer in the collaborative graph neural network supernet includes each operation in the collaborative search space and a function corresponding to each operation, the operation includes a communication operation, and the communication operation indicates that a device switching is performed between an end device and an edge device, and each candidate architecture is obtained by randomly selecting a target operation and a target function corresponding to the target operation in each layer of the collaborative graph neural network supernet; A target architecture selection module is used to select a target architecture of the target collaborative graph neural network from the multiple candidate architectures based on the collaborative reasoning performance information of each candidate architecture for the target application identification information under the target terminal edge environment information.

[0008] In a fourth aspect, an embodiment of the present application further provides an architecture deployment device for a device-edge collaborative graph neural network, the device comprising: An architecture acquisition module, configured to acquire a first target architecture from a target architecture warehouse based on first end-edge environment information of an end-edge system for deploying a target collaborative graph neural network; wherein the target architecture warehouse is an architecture warehouse of the target collaborative graph neural network, which stores a plurality of target architectures corresponding to a plurality of end-edge environment information of the target collaborative graph neural network, and the target architecture corresponding to each end-edge environment information is obtained by using each end-edge environment information as the target end-edge environment information in the target application scenario information, and performing an architecture search method according to any one of claims 1 to 5; the first target architecture is a target architecture corresponding to the first end-edge environment information; a deployment module, configured to deploy the first target architecture to the end-edge system, so as to perform end-edge collaborative reasoning on the end-edge system based on the first target architecture; An environment monitoring module, used to monitor changes in the edge environment of the edge system during edge-to-edge collaborative reasoning; The architecture acquisition module is further configured to acquire a second target architecture from the target architecture repository when the edge device environment changes; wherein the second target architecture is a target architecture corresponding to second edge device environment information, and the second edge device environment information is the changed edge device environment information; The deployment module is further used to deploy the second target architecture to the end-edge system to perform end-edge collaborative reasoning on the end-edge system based on the second target architecture.

[0009] In a fifth aspect, an embodiment of the present application provides an electronic device, comprising a processor and a memory, wherein the memory stores programs or instructions that can be run on the processor, and when the programs or instructions are executed by the processor, the steps of the method described in the first aspect or the second aspect of the embodiment of the present application are implemented.

[0010] In an embodiment of the present application, a candidate architecture of the target collaborative graph neural network is obtained by randomly selecting a target operation and an objective function corresponding to the target operation in each layer of a collaborative graph neural network supernet including communication operations. Based on the collaborative reasoning performance information of the candidate architecture for the target application identification information under the target end-edge environment information, the target architecture of the target collaborative graph neural network is selected. The end-edge collaborative deployment scheme can be considered to obtain a graph neural architecture with better collaborative reasoning performance after the end-edge collaborative deployment, thereby improving the end-edge collaborative reasoning performance of the graph neural network with end-edge collaborative deployment. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments of the present application will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0012] Figure 1 A schematic diagram of a process flow of an architecture search method for a device-edge collaborative graph neural network provided in some embodiments of the present application; Figure 2 Examples of information related to collaborative graph neural network supernets provided for some embodiments of the present application; Figure 3 A schematic diagram of a search process for an architecture search strategy provided in some embodiments of the present application; Figure 4 An example of obtaining collaborative reasoning speed information provided in some embodiments of the present application; Figure 5 Search examples of architecture search strategies provided for some embodiments of the present application; Figure 6 A schematic diagram of a process flow of a method for deploying an architecture of a device-edge collaborative graph neural network provided in some embodiments of the present application; Figure 7 An example of an architecture deployment process provided for some embodiments of the present application; Figure 8 An example of the overall process of architecture search deployment provided for some embodiments of the present application; Fig. 9 A schematic diagram of a module of an architecture search device for a device-edge collaborative graph neural network provided in some embodiments of the present application; Fig.10 A schematic diagram of a module of an architecture deployment device for a device-edge collaborative graph neural network provided in some embodiments of the present application; Fig.11 A schematic diagram of the structure of an electronic device provided for some embodiments of the present application. DETAILED DESCRIPTION

[0013] Embodiments of the present application provide a method and device for architecture search and deployment of an end-edge collaborative graph neural network.

[0014] In order to enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in the embodiments of this application will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of this application.

[0015] See also Figure 1 , is a flow chart of a method for searching an architecture of a device-edge collaborative graph neural network provided in some embodiments of the present application. Figure 1 As shown, the architecture search method of the end-edge collaborative graph neural network provided in the embodiment of the present application includes the following steps: S101, obtaining target application scenario information of the target collaborative graph neural network.

[0016] The target collaborative graph neural network is an end-edge collaborative graph neural network of a network architecture to be determined. The end-edge collaborative graph neural network is a graph neural network oriented to end-edge collaboration, such as a graph neural network that needs to be deployed for end-edge collaboration and perform end-edge collaborative reasoning.

[0017] The target application scenario information is information related to the target collaborative graph neural network application scenario, which may include: target application identification information and target end-side environment information.

[0018] The target application identification information is used to indicate the reasoning tasks that the target collaborative graph neural network needs to perform or the problems that need to be solved. For example, it is the application name of the target collaborative graph neural network, such as point cloud processing, text processing, or the application identification of the target collaborative graph neural network, such as Application 1, Application 2, Application 3.

[0019] The target end-edge environment information is the system environment information of the end-edge system used to deploy the target collaborative graph neural network, which may include: device information of the end-end device (i.e., end-side device) in the end-edge system used to deploy the target collaborative graph neural network, such as Raspberry Pi, Jetson, etc.; device information of the edge device (i.e., edge-side device), such as Intel CPU, Nvidia GPU, etc.; network bandwidth information between end-edge devices, such as 40mbps, 20mbps, 10mbps, etc.

[0020] In actual applications, the target application scenario information can be input by the user or obtained according to user demand information.

[0021] S102, based on the target application scenario information and the collaborative graph neural network supernet, obtain multiple candidate architectures of the target collaborative graph neural network.

[0022] The collaborative graph neural network supernet is constructed based on the collaborative search space of the end-edge collaborative graph neural network, and each layer in the collaborative graph neural network supernet contains each operation of the collaborative search space and the function corresponding to each operation. Each candidate architecture is obtained by randomly selecting the target operation and the target function corresponding to the target operation in each layer of the collaborative graph neural network supernet.

[0023] The collaborative search space is a search space specially established for the edge-to-edge collaborative graph neural network. Figure 2 As shown in (a). Figure 2 As shown in (a), the collaborative search space includes an operation space and a function space. The operation space includes all types of operations in the edge-to-end collaborative graph neural network, such as sampling operations, communication operations, aggregation operations, combination operations, global pooling operations, and direct connection operations. The function space includes specific attributes or operation functions corresponding to each type of operation, such as the KNN function and sample function corresponding to the sampling operation, 10Mbps and 40Mbps corresponding to the communication operation, the add function, mean function, and max function corresponding to the aggregation operation, 16, 32, 64, and 128 corresponding to the combination operation, the sum function, mean function, and max function corresponding to the global pooling operation, and the Skip function corresponding to the direct connection operation.

[0024] The collaborative graph neural network supernet is a collaborative search space organized in the form of a supernet. Its structure is as follows: Figure 2 (b). Figure 2 As shown in (b), the collaborative graph neural network supernet contains multiple layers, each of which contains all operations in the collaborative search space and all functions corresponding to each operation. By randomly selecting the target operation and the target function corresponding to the target operation in each layer of the collaborative graph neural network supernet, a candidate architecture can be obtained, such as Figure 2(c) shown.

[0025] It should be noted that the communication operation in the embodiment of the present application is an operation specially constructed for the end-edge collaborative graph neural network, which is obtained by abstracting the communication process between the end device and the edge device during the end-edge collaborative reasoning of the end-edge collaborative graph neural network. A communication operation represents a device switch between the end device and the edge device. The operations in the collaborative search space include communication operations, and each layer in the collaborative graph neural network supernet established based on the collaborative search space also includes communication operations. Since the candidate architecture is obtained by randomly selecting the target operation in each layer of the collaborative graph neural network supernet, the target operation of some layers of the candidate architecture may be a communication operation. Since a communication operation represents a device switch between the end device and the edge device, each operation in the candidate architecture is divided into a part deployed on the end device and a part deployed on the edge device by its communication operation, so that the obtained candidate architecture is an architecture that includes an end-edge collaborative deployment solution.

[0026] For example, assuming that the operations of the third layer and the fifth layer in the candidate architecture are communication operations, then the operations of the first layer and the second layer before the third layer in the candidate architecture are operations deployed on the edge device, the operations of the fourth layer between the third layer and the fifth layer are operations deployed on the edge device, and the operations after the fifth layer are operations deployed on the end device. The candidate architecture is an architecture that includes an end-edge deployment solution.

[0027] S103, selecting a target architecture of the target collaborative graph neural network from the multiple candidate architectures based on the collaborative reasoning performance information of each candidate architecture for the target application identification information under the target terminal edge environment information.

[0028] Among them, the collaborative reasoning performance information is used to indicate the reasoning performance information when performing end-edge collaborative deployment in the target end-edge environment indicated by the target end-edge environment information, and performing end-edge collaborative reasoning for the target application indicated by the target application identification information, which may include: any one or more of the collaborative reasoning speed information and the collaborative reasoning accuracy information.

[0029] For example, the collaborative reasoning performance information may be collaborative reasoning speed information, collaborative reasoning accuracy information, or collaborative reasoning speed information and collaborative reasoning accuracy information.

[0030] Corresponding to the collaborative reasoning performance information, the target architecture may be an optimal candidate architecture in any one or more of the following dimensions: collaborative reasoning speed dimension and collaborative reasoning accuracy dimension.

[0031] For example, the target architecture may be the optimal candidate architecture in terms of collaborative reasoning speed, that is, the candidate architecture with the fastest collaborative reasoning speed; it may also be the optimal candidate architecture in terms of collaborative reasoning accuracy, that is, the candidate architecture with the highest collaborative reasoning accuracy; or, the candidate architecture with the best comprehensive reasoning performance in terms of both collaborative reasoning speed and collaborative reasoning accuracy, that is, the candidate architecture with the fastest collaborative reasoning speed and the highest collaborative reasoning accuracy.

[0032] The architecture search method for an end-edge collaborative graph neural network provided in an embodiment of the present application obtains a candidate architecture of a target collaborative graph neural network by randomly selecting a target operation and an objective function corresponding to the target operation in each layer of a collaborative graph neural network supernet including communication operations, and selects a target architecture of a target collaborative graph neural network based on the collaborative reasoning performance information of the candidate architecture for the target application identification information under the target end-edge environment information. The method can consider the end-edge deployment plan and optimize the architecture of the graph neural network and its end-edge deployment plan at the same time, so as to obtain a graph neural network architecture with better collaborative reasoning performance after the end-edge collaborative deployment, thereby improving the collaborative reasoning performance of the graph neural network after the end-edge deployment.

[0033] The multiple candidate architectures in the above embodiments can be obtained through appropriate architecture search strategies based on target application scenario information and collaborative graph neural network supernet. Figure 3 Schematic diagram of the search process of the architecture search strategy provided by some embodiments of the present application. Figure 3 As shown, in some embodiments of the present application, step S102 in the above embodiment specifically includes the following steps: S301, specifying an initial objective function for each operation in the collaborative graph neural network supernet to obtain an initial architecture of the target collaborative graph neural network.

[0034] Among them, for the communication operations of each layer of the collaborative graph neural network supernet, an initial objective function can be specified for each communication operation based on the network bandwidth information between the end-edge devices in the target end-edge environment information. For other operations of each layer of the collaborative graph neural network supernet, an initial objective function can be specified for each of the other operations based on the function corresponding to each operation in the existing graph neural network architecture. Among them, the existing graph neural network architecture can be an optimized graph neural network architecture, which can include all operations other than the communication operations in the collaborative graph neural network supernet and the functions corresponding to each operation.

[0035] For example, if the network bandwidth information between the end-edge devices in the target end-edge environment is 40mbps, the function corresponding to the sampling operation in the existing graph neural network is: KNN, and the function corresponding to the aggregation operation is add, then the initial objective function of the communication operation in each layer of the collaborative graph neural network supernet can be specified as 40mbps, the initial objective function of the sampling operation in each layer can be specified as KNN, and the function corresponding to the aggregation operation in each layer can be specified as add.

[0036] S302, pre-training the initial architecture based on the target application data set to obtain a pre-trained architecture of the target collaborative graph neural network.

[0037] The target application data set is an application data set corresponding to the target application identification information, that is, a data set used for training the target application. Training the initial architecture based on the target application data set can optimize the weight of the initial architecture for the target application, thereby obtaining a pre-trained architecture suitable for the target application.

[0038] In a specific implementation, the target application dataset can be obtained from an application dataset warehouse based on the target application identification information.

[0039] S303: Obtain multiple first candidate architectures based on the pre-trained architecture.

[0040] Each first candidate architecture is obtained by randomly selecting a target operation in each layer of the pre-trained architecture.

[0041] In specific implementation, the objective function of each layer operation of the pre-trained architecture can be fixed as the initial objective function, and then, an operation is randomly selected in each layer as the target operation of the layer, thereby obtaining a first candidate architecture. By iterating the above process, multiple first candidate architectures can be obtained.

[0042] S304: Select an operation search phase architecture from the plurality of first candidate architectures based on the collaborative reasoning speed information and the collaborative reasoning accuracy information of each first candidate architecture for the target application identification information under the target terminal edge environment information.

[0043] The operation search phase architecture is the architecture with the best comprehensive reasoning performance in terms of the two dimensions of collaborative reasoning speed and collaborative reasoning accuracy among multiple first candidate architectures. For example, among multiple first candidate architectures, the first candidate architecture with the highest first architecture score. The first architecture score is a score obtained based on the collaborative reasoning speed information and the collaborative reasoning accuracy information, which is, for example, the difference between the collaborative reasoning accuracy and the collaborative reasoning speed, or the weighted sum of the collaborative reasoning accuracy and the collaborative reasoning speed.

[0044] In a specific implementation, the collaborative reasoning speed information and the collaborative reasoning accuracy information of each first candidate architecture for the target application identification information under the target end-side environment information can be obtained. Then, based on the collaborative reasoning speed information and the collaborative reasoning accuracy information of each first candidate architecture for the target application identification information under the target end-side environment information, the first architecture score of each first candidate architecture can be calculated. After that, according to the first architecture score of each first candidate architecture, the first candidate architecture with the highest first architecture score can be selected from the multiple first candidate architectures, and the first candidate architecture with the highest first architecture score can be used as the operation search stage architecture.

[0045] Among them, the collaborative reasoning speed information and the collaborative reasoning accuracy information of the first candidate architecture for the target application identification information under the target terminal edge environment information can be obtained in an appropriate manner.

[0046] For example, the first candidate architecture will inherit all weights in the pre-trained architecture, and the first candidate architecture can be deployed end-to-end in a target end-to-end environment set based on the target end-to-end environment information, so that it can perform collaborative reasoning tests on the test data set of the target application indicated by the target application identification information, thereby obtaining the collaborative reasoning accuracy and collaborative reasoning speed of the first candidate architecture through testing.

[0047] S305: Based on the operation search phase architecture, obtain multiple second candidate architectures.

[0048] Each second candidate architecture is obtained by selecting a lighter objective function than the initial objective function for each target operation of the first target architecture, and the objective function of the communication operation is fixed to the initial objective function.

[0049] Specifically, the operation search phase architecture obtained through the above-mentioned S301-304 processing has a target operation at each layer, and each target operation has an initial objective function. The target operation of each layer in the operation search phase architecture is the optimal operation combination selected based on the inference speed information and the inference accuracy information. In step S305, the target operation of each layer of the operation search phase architecture can be fixed, and then, for each target operation, a function that is lighter than the initial objective function corresponding to the target operation is randomly selected from the multiple functions corresponding to the target operation, such as a function with lower computational complexity and fewer required computing resources, as the objective function of the target operation, thereby obtaining a second candidate architecture. By iterating the above-mentioned processing, multiple second candidate architectures can be obtained.

[0050] Among them, since the initial objective function of the communication operation is set based on the bandwidth information between the edge devices in the target edge environment when the initial objective function is set, it is already the optimal function setting under the bandwidth information. Therefore, when selecting the objective function corresponding to the target operation, for the communication operation, the corresponding objective function can remain unchanged, that is, the objective function of the communication operation can be kept as its initial objective function.

[0051] In an embodiment of the present application, the initial objective function of each operation in the collaborative graph neural network supernet is first specified, and then, when the initial objective function is fixed, the target operation of each layer is randomly selected to obtain a first candidate architecture, and then, based on the reasoning speed and accuracy of each first candidate architecture, the operation search phase architecture is selected from multiple first candidate architectures. Afterwards, when the target operation of each layer of the operation search phase architecture is fixed, a function downward search is performed in multiple functions corresponding to each target operation, and a function lighter than its initial objective function is selected as the objective function for each target operation to obtain a second candidate architecture. Since the objective function lighter than the initial objective function is more efficient than the initial objective function, the collaborative reasoning speed of each second candidate architecture is higher than the operation search phase architecture. Therefore, when selecting a target architecture from multiple second candidate architectures, the collaborative reasoning speed of each second candidate architecture may no longer be evaluated, thereby improving the efficiency of architecture search.

[0052] Therefore, further, step S103 in the above embodiment may include the following steps: S306: Select a target architecture from the multiple second candidate architectures based on the collaborative reasoning accuracy information of each second candidate architecture for the target application identification information under the target terminal edge environment information.

[0053] Among them, the target architecture can be the architecture with the highest collaborative reasoning accuracy among multiple second candidate architectures, for example, the second candidate architecture with the highest second architecture score, wherein the second architecture score is obtained based on the collaborative reasoning accuracy information of the second candidate architecture for the target application identification information under the target end-side environment information.

[0054] In a specific implementation, for each second candidate architecture, the collaborative reasoning accuracy of the second candidate architecture for the target application identification information under the target end-side environment information can be obtained, and then, based on the collaborative reasoning accuracy of the second candidate architecture for the target application identification information under the target end-side environment information, the second architecture score of the second candidate architecture can be determined, and then, according to the second architecture score of each second candidate architecture, the architecture with the highest second architecture score can be selected as the target architecture among multiple second candidate architectures.

[0055] The accuracy of the collaborative reasoning of the second candidate architecture for the target application identification information under the target edge environment information can be obtained in an appropriate manner. For example, the second candidate architecture can be deployed in an edge-to-edge collaborative manner in a target edge environment set based on the target edge environment information, so that the collaborative reasoning test is performed on the test data set of the target application indicated by the target application identification information, thereby obtaining the accuracy through testing.

[0056] In the above embodiment, for the collaborative reasoning speed information of the first candidate architecture for the target application identification information under the target end-side environment information, the first candidate architecture can be deployed to the target end-side environment for testing. The first candidate architecture is actually deployed to the target end-side environment for testing, and the collaborative reasoning speed of the first candidate architecture is obtained through testing, which consumes a lot of time, resulting in low efficiency of architecture search. In order to improve the efficiency of architecture search, in some embodiments of the present application, the collaborative reasoning speed of the first candidate architecture can be obtained by estimation or prediction.

[0057] For example, in some embodiments of the present application, the collaborative reasoning feature data of each first candidate architecture can be input into a target collaborative reasoning speed predictor for collaborative reasoning speed prediction to obtain the first collaborative reasoning speed of each first candidate architecture for the target application identification information under the target end-side environment information.

[0058] Among them, the target collaborative reasoning speed predictor is a collaborative reasoning speed predictor corresponding to the target application identification information, which is obtained based on a target collaborative reasoning speed prediction data set; the target collaborative reasoning speed prediction data set is a collaborative reasoning speed prediction data set corresponding to the target application identification information, which includes collaborative reasoning feature data and collaborative reasoning speed data of the sample architecture; the target collaborative reasoning speed predictor is obtained by supervised training with the collaborative reasoning feature data of the sample architecture as input and the collaborative reasoning speed data of the sample architecture as output label; the sample architecture can be searched based on a method similar to the first candidate architecture, the collaborative reasoning feature data of the sample architecture can include architecture feature information and end-edge device information, and the collaborative reasoning speed data of the sample architecture can be obtained by performing end-edge collaborative deployment of the sample architecture in an end-edge environment set based on the end-edge environment information corresponding to its application information, and performing end-edge collaborative reasoning testing.

[0059] In specific implementation, for multiple applications, based on the collaborative reasoning speed prediction data set corresponding to each application, a collaborative reasoning speed predictor corresponding to each application can be pre-trained. Afterwards, a target collaborative reasoning speed predictor corresponding to the target application is obtained based on the target application information. After obtaining a first candidate structure, the collaborative reasoning feature data of the candidate architecture can be input into the target collaborative reasoning speed predictor, and the target collaborative reasoning speed predictor predicts the collaborative reasoning speed of the candidate architecture for the target application identification information under the target end-side environment information, and outputs the predicted collaborative reasoning speed, which is the first collaborative reasoning speed.

[0060] Among them, it can be understood that the structure of the first candidate architecture is similar to the structure of the sample architecture in the collaborative reasoning speed prediction data set corresponding to the target application identification information, and its collaborative reasoning feature data contains architecture feature information and end-edge device information. The architecture feature information is used to describe architecture-related features, which may include graph adjacency matrix information and node feature matrix information of the directed acyclic graph corresponding to the architecture. Among them, the directed acyclic graph corresponding to the architecture can be abstracted by taking the target operation of each layer in the architecture as a node, the connection between operations as an edge, and connecting all nodes to a global node.

[0061] For example, for the first candidate architecture, the target operation of each layer in the first candidate architecture can be used as a node, the connection between operations can be used as an edge, and all nodes can be connected to a global node, so as to abstract the first candidate architecture into a directed acyclic graph. Afterwards, the graph adjacency matrix information and the node feature matrix information can be obtained based on the directed acyclic graph, and then, the architecture feature information of the first candidate architecture can be obtained based on the graph adjacency matrix information and the node feature information. Afterwards, the collaborative reasoning feature data of the first candidate architecture can be obtained based on the architecture feature information of the first candidate architecture and the edge device information in the target edge environment information.

[0062] For example, in some embodiments of the present application, the second collaborative reasoning speed of each first candidate architecture for the target application identification information under the target end-side environment information can be determined based on the sum of the operation delay times corresponding to the target operations of each layer of each first candidate architecture; wherein the reasoning delay time corresponding to each target operation is obtained by searching in the operation delay time lookup table corresponding to the target application scenario information.

[0063] Specifically, an operation delay lookup table corresponding to each application scenario information can be established in advance. An operation delay lookup table corresponding to an application scenario information stores the operation delay time of each operation under the application scenario information, and the delay time of each operation can be collected by testing in the application scenario. The operation delay time represents the time elapsed from the start of the operation to the end of the operation. After obtaining a first candidate architecture, for the target operation in each layer of the candidate architecture, the operation delay time corresponding to each target operation can be found from the operation delay time lookup table corresponding to the target application scenario information, and then the operation delay time corresponding to each target operation can be superimposed to obtain the total operation delay time of the target operation in the candidate architecture, and then, based on the total operation delay time, the end-edge collaborative reasoning speed of the candidate architecture can be estimated. For example, the end-edge collaborative reasoning speed of the candidate architecture can be estimated as: 1 / total operation delay time.

[0064] It can be understood that in the aforementioned example embodiment, the second collaborative reasoning speed data obtained by superimposing the operation delay time is estimated data, which has poor accuracy compared to the collaborative reasoning speed data predicted by the collaborative reasoning speed predictor, but the operation speed is faster. In practical applications, a suitable collaborative reasoning speed determination method can be selected according to application requirements. For example, in some application scenarios, it is necessary to quickly obtain the target architecture, which requires a faster architecture search speed and does not require high accuracy of the architecture. At this time, the operation delay time of each target operation in the first candidate architecture can be superimposed to estimate the end-edge collaborative reasoning speed of the first candidate architecture. For another example, in some applications, the accuracy requirements for the architecture are high and the architecture search speed requirements are low. At this time, a collaborative reasoning speed predictor can be trained to predict the collaborative reasoning speed of the first candidate architecture based on the collaborative reasoning speed predictor.

[0065] Generally, the collaborative reasoning speed data predicted by the collaborative reasoning speed predictor is highly accurate, but in some cases, the collaborative reasoning speed predictor may also make incorrect predictions. For example, the collaborative reasoning delay time of a candidate architecture is at least 100ms, but the collaborative reasoning delay time predicted by the collaborative reasoning speed predictor is 80ms, which is an error in the prediction. The target architecture selected based on the incorrect collaborative reasoning delay time may have serious problems when applied.

[0066] The sum of the operation delay times of all operations in the architecture is the minimum inference delay time of the architecture, which indicates the fastest inference speed of the architecture, where the inference delay time is the time from the start of inference to the end of inference. In some embodiments of the present application, the estimated collaborative inference speed data can be used to verify the predicted collaborative inference speed data to ensure the reliability of the target architecture finally obtained.

[0067] Specifically, the collaborative reasoning feature data of each first candidate architecture can be input into the target collaborative reasoning speed predictor for collaborative reasoning speed prediction to obtain the first collaborative reasoning speed of each first candidate architecture for the target application identification information under the target end-side environment information; based on the sum of the operation delay times corresponding to each layer of target operations of each first candidate architecture, the second collaborative reasoning speed of each first candidate architecture for the target application identification information under the target end-side environment information is obtained; then, based on the first collaborative reasoning speed and the second collaborative reasoning speed of each first candidate architecture, the multiple first candidate architectures are screened; thereafter, for the screened first candidate architectures, based on the first collaborative reasoning speed and reasoning accuracy of each first candidate architecture for the target application identification information under the target end-side environment information, the operation search phase architecture is selected from the screened first candidate architectures.

[0068] For example, assuming that there are three first candidate architectures, the first collaborative reasoning speed of the first candidate architecture 1 is 13ms, and the second collaborative reasoning speed is 10ms; the first collaborative reasoning speed of the first candidate architecture 2 is 14ms, and the second collaborative reasoning speed is 11ms; the first collaborative reasoning speed of the first candidate architecture 3 is 9ms, and the second collaborative reasoning speed is 10ms. The first collaborative reasoning speeds of the first candidate architecture 1 and the first candidate architecture 2 are slower than the second collaborative reasoning speed, and the first collaborative reasoning speed of the first candidate architecture 3 is faster than the second collaborative reasoning speed, which is an incorrect collaborative reasoning speed. During screening, the first candidate architecture 3 can be screened out, leaving the first candidate architecture 1 and the first candidate architecture 2, thereby obtaining the screened first candidate architecture, and then, based on the first collaborative reasoning speeds of the screened first candidate architecture 1 and the first collaborative reasoning speeds of the first candidate architecture 2, the operation search phase architecture can be selected from the screened first candidate architecture 1 and the first candidate architecture 2.

[0069] In practical applications, the above-mentioned target collaborative reasoning speed predictor can be implemented as a system performance predictor. The above-mentioned processing of estimating the collaborative reasoning speed of the architecture for the target application identification information under the target end-side environment information based on the sum of the operation delay times corresponding to each operation in the architecture can be implemented by the system performance estimator. The above-mentioned collaborative reasoning feature data can be implemented as data composed of the end-side device configuration data, the edge-side device configuration data, the graph adjacency matrix information and node feature information (obtained by concatenating the unique hot encoding and delay data of the node) of the directed acyclic graph obtained by architecture abstraction, and the operation delay lookup table. The above-mentioned first candidate architecture can be implemented as a GNN architecture carrying a deployment plan. The above-mentioned collaborative reasoning speed data for the target application identification information under the target end-side environment information can be implemented as collaborative reasoning delay data on the target end-side system.

[0070] like Figure 4 As shown, in a practical application of a specific implementation, after obtaining a GNN architecture carrying a deployment plan (i.e., the first candidate architecture), the architecture can be graph abstracted to obtain a directed acyclic graph corresponding to the candidate architecture. Then, based on the directed acyclic graph and the operation delay data corresponding to each operation, the enhanced node feature data of each node is obtained, and based on the directed acyclic graph, the graph adjacency matrix is ​​obtained. Afterwards, the end-side device configuration data, the edge-side device configuration data, the graph adjacency matrix information corresponding to the architecture, the node feature data of each node, and the operation delay lookup table in the target end-edge environment information can be selected and input into any one or both of the system performance estimator and the system performance predictor, so as to obtain the collaborative reasoning delay data of the architecture on the target end-edge system.

[0071] Of course, it is understandable that in actual applications, in order to ensure that the screened target architecture is available, in addition to screening the first candidate architecture based on the collaborative reasoning speed as described in the above embodiment, the first candidate architecture can also be screened based on other conditions, and then, the operation search stage architecture is selected from the screened first candidate architecture.

[0072] For example, the first candidate architecture may be screened for validity based on an architecture validity screening strategy. The architecture validity screening strategy is, for example: if the target operation of two consecutive layers in the architecture is a communication operation, the architecture is screened out. The target operation of two consecutive layers being a communication operation means that the switch from one device to another is immediately switched back, and no actual calculation is performed, which is an invalid configuration.

[0073] For another example, the first candidate architecture may be screened based on the collaborative reasoning speed requirement information corresponding to the target application identification information. If the collaborative reasoning speed of the first candidate architecture is slower than the required collaborative reasoning speed, the candidate architecture is screened out.

[0074] For another example, the first candidate architecture may be screened based on the collaborative reasoning accuracy requirement information corresponding to the target application identification information. If the collaborative reasoning accuracy of the first candidate architecture is lower than the required collaborative reasoning accuracy, the candidate architecture is screened out.

[0075] For another example, the first candidate architecture may be screened based on the end-side device energy consumption limit information corresponding to the target application identification information. If the end-side device energy consumption of the first candidate architecture is higher than the end-side device energy consumption limit, the candidate architecture is screened out.

[0076] In specific implementation, Figure 5 As shown, when the architecture search starts, the operation search phase can be entered. In the operation search phase, the initial function setting can be selected in the supernet, and the supernet can be pre-trained. After that, based on the pre-trained architecture, N first candidate architectures and the architecture score of each first candidate architecture are obtained through multiple iterations. After that, the operation search phase architecture can be selected from the N first candidate architectures based on the architecture score of each first candidate architecture. Among them, in one iteration, the operation space is sampled to obtain a first candidate architecture, that is, the target operation is randomly selected in each layer of the pre-trained architecture to obtain a first candidate architecture. After that, the candidate architecture can be subjected to various screenings, such as validity check, collaborative reasoning speed requirement screening, collaborative reasoning accuracy requirement screening, etc. After the candidate architecture passes all screenings, the architecture score of the candidate architecture can be calculated based on the collaborative reasoning speed and collaborative reasoning accuracy of the candidate architecture. When the candidate architecture does not pass any screening, the next iteration can be performed. After the operation search phase architecture is selected, the function search phase can be entered. In the function search phase, based on the initial operation search phase architecture determined in the operation search phase, M second candidate architectures and the architecture score of each second candidate architecture can be obtained through multiple iterations. Then, based on the architecture score of each second candidate architecture, the optimal target architecture of the edge-to-edge collaborative graph neural network is selected from the M second candidate architectures and output. In one iteration, the function space is downsampled to obtain a second candidate architecture, that is, a function that is lighter than its initial target function is selected from the functions corresponding to each layer of the target operation of the operation search phase architecture to obtain a second candidate architecture. Then, the collaborative reasoning accuracy of the candidate architecture can be tested, and then, based on the collaborative reasoning speed of the candidate architecture, the architecture score of the candidate architecture is calculated.

[0077] The above is the architecture search method of the end-edge collaborative graph neural network provided by the embodiment of the present application. Based on the architecture search method of the end-edge collaborative graph neural network provided by the above embodiment, the embodiment of the present application also provides an architecture deployment method of the end-edge collaborative graph neural network. Figure 6A schematic diagram of a process flow of a method for deploying an architecture of a device-edge collaborative graph neural network provided in some embodiments of the present application. Figure 6 As shown, in some embodiments of the present application, the architecture deployment method of the end-edge collaborative graph neural network includes the following steps: S601, based on first edge environment information of an edge system for deploying a target collaborative graph neural network, obtain a first target architecture from a target architecture repository.

[0078] Among them, the first target architecture is the target architecture corresponding to the first end-edge environment information. The first end-edge environment information is the end-edge environment information before the end-edge system changes, that is, the current end-edge environment information of the end-edge system before the end-edge environment changes. The target architecture warehouse is the architecture warehouse of the target collaborative graph neural network, which stores multiple target architectures corresponding to multiple end-edge environment information of the target collaborative graph neural network. The target architecture corresponding to each end-edge environment information is obtained by taking each end-edge environment information as the target end-edge environment information in the target application scenario information, and performing the architecture search method provided in the aforementioned embodiment of the present application.

[0079] For example, for the target collaborative graph neural network, the architecture search method of the end-edge collaborative graph neural network provided in the embodiment of the present application can be iterated multiple times to obtain multiple target architectures of the target collaborative graph neural network, and the multiple target architectures are stored in the architecture warehouse. Among them, in each iteration, the target end-edge environment information in the target application scenario information can be updated, thereby obtaining the target architecture under different end-edge environment information, and storing the target architecture under different end-edge environment information in the architecture warehouse, so as to obtain the target architecture warehouse.

[0080] S602: Deploy the first target architecture to the edge-end system to perform edge-end collaborative reasoning on the edge-end system based on the first target architecture.

[0081] Among them, based on the communication operations in the target architecture, each operation in the target architecture can be deployed on the end device or the edge device of the end-edge system respectively.

[0082] For example, if the target architecture has six layers, and the target operations of the third and fifth layers are communication operations, then based on the location of the communication operations, the target operations of the first and second layers before the third layer can be deployed to the end device, the target operations of the fourth layer between the third and fifth layers can be deployed to the edge device, and the target operations of the sixth layer after the fifth layer can be deployed to the end device.

[0083] S603: During the process of performing the edge-end collaborative reasoning, monitor the edge-end environment changes of the edge-end system.

[0084] The change in the edge environment includes: a change in the edge network situation and / or a change in the edge device.

[0085] S604: When the edge environment changes, obtain a second target architecture from the target architecture repository.

[0086] The second target architecture is a target architecture corresponding to the second edge-end environment information, and the second edge-end environment information is the changed edge-end environment information, that is, the current edge-end environment information of the edge-end system after the edge-end environment changes.

[0087] For example, the edge environment information may include: network bandwidth information between edge devices and edge device information. If the network situation information in the edge environment changes, the network bandwidth changes from 40 mbps to 20 mbps, so that the edge environment information changes from edge environment information 1 to edge environment information 2, then the target architecture corresponding to edge environment information 2 can be selected from the architecture warehouse. For another example, if the edge device in the edge environment information changes from device 1 to device 2, so that the edge environment information changes from edge environment information 1 to edge environment information 3, then the target architecture corresponding to edge environment information 3 can be selected from the architecture warehouse.

[0088] It should be noted that, since the target architecture in the target architecture warehouse is obtained based on the architecture search method of the end-edge collaborative graph neural network provided in the embodiment of the present application, a target architecture corresponding to an end-edge environment information includes any one or more of the following: the optimal architecture in the dimension of collaborative reasoning accuracy, the optimal architecture in the dimension of collaborative reasoning speed, and the optimal architecture in the dimension of collaborative reasoning accuracy and the dimension of collaborative reasoning speed. In other words, for a type of end-edge environment information, it may correspond to multiple target architectures. In practical applications, the appropriate target architecture corresponding to the end-edge environment information can be selected according to the application scenario of the target collaborative graph neural network. For example, in application scenarios with high requirements for collaborative reasoning accuracy, the optimal target architecture in the dimension of collaborative reasoning accuracy can be selected. In application scenarios with high requirements for collaborative reasoning speed, the optimal target architecture in the dimension of collaborative reasoning speed can be selected. In general application scenarios, the optimal target architecture in the dimension of collaborative reasoning accuracy and the dimension of collaborative reasoning speed can be selected.

[0089] S605: Deploy the second target architecture to the edge-end system to perform edge-end collaborative reasoning on the edge-end system based on the second target architecture.

[0090] Among them, based on the communication operations in the second target architecture, each operation in the target architecture can be deployed on the end device or the edge device of the end-edge system respectively.

[0091] In the embodiment of the present application, since the target architecture corresponding to the edge environment information in the target architecture warehouse is the optimal reasoning performance architecture for the edge environment information after considering the edge deployment scheme, when the edge environment changes, the corresponding target architecture is dynamically selected for deployment based on the changed edge environment information, which can provide the optimal collaborative reasoning performance when the edge environment changes.

[0092] Among them, in order to further improve the performance of collaborative reasoning, the target architecture can perform pipelined end-edge collaborative reasoning on the end-edge system. For example, after the target architecture is deployed on the end-edge system for end-edge collaborative reasoning, the architecture part on the end device can start processing the second set of data immediately after processing the first set of data and sending the first set of data to the edge device, without waiting for the architecture part on the edge device to finish processing the first set of data before starting to process the second set of data. In this way, the communication delay between the end-edge devices can be concealed, further improving the efficiency of end-edge collaborative reasoning.

[0093] In practical applications, the above-mentioned pipelined edge-to-end collaborative reasoning can be supported by a pipelined collaborative reasoning engine. The above-mentioned processing of dynamically deploying the target architecture according to the changes in the edge-to-end environment can be implemented by a runtime dynamic tuner. The above-mentioned target architecture warehouse can be implemented as a GNN warehouse.

[0094] like Figure 7 As shown, in a specific practical application, the runtime dynamic tuner can monitor the changes in the network environment and equipment of the end-edge system, select the appropriate target architecture from the GNN warehouse according to the changes in the network environment and equipment, and deploy the target architecture on the end-edge system for end-edge collaboration.

[0095] The above-mentioned embodiments provide an architecture search method and an architecture deployment method for an end-edge collaborative graph neural network, which obtains a candidate architecture of the target collaborative graph neural network by randomly selecting a target operation and an objective function corresponding to the target operation in each layer of a collaborative graph neural network supernet including communication operations, and selects a target architecture of the target collaborative graph neural network based on the collaborative reasoning performance information of the candidate architecture for target application identification information under target end-edge environment information. Thereafter, a target architecture warehouse is obtained based on the target architecture, and during reasoning, the corresponding target architecture is dynamically selected from the target architecture warehouse according to changes in the end-edge environment information for end-edge collaborative deployment and end-edge collaborative reasoning, which can improve the end-edge collaborative reasoning performance of the graph neural network during end-edge deployment as a whole.

[0096] In practical applications, the target application scenario information in the above embodiments can be implemented as user requirements and system configuration information, which includes, for example: optimization target information (such as inference delay target, inference accuracy target, inference energy consumption target), target application information (such as point cloud processing, text processing information), network status information (such as network bandwidth information), end-edge device information (such as device type information used on the end side, device type information used on the edge side). The collaborative graph neural network supernet in the above embodiments can be implemented as a GNN supernet built based on the GNN collaborative reasoning search space. The above embodiments Figure 3 The architecture search strategy shown can be implemented as a constraint-based random search strategy. Acquiring the collaborative reasoning speed in the above embodiment can be implemented as an inference efficiency evaluation. Acquiring the collaborative reasoning accuracy in the above embodiment can be implemented as a task accuracy evaluation. The candidate architecture in the above embodiment can be implemented as a candidate GNN architecture. Acquiring the collaborative reasoning accuracy information of the candidate architecture for the target application identification information under the target end-side environment information in the above embodiment can be implemented as system performance perception of the candidate GNN architecture.

[0097] like Figure 8 As shown, in a specific practical application, based on user needs and system configuration information, an operation search can be performed on the GNN collaborative reasoning search space. For the candidate GNN architectures obtained through the operation search, the reasoning efficiency can be evaluated based on system performance perception, and the task accuracy can be evaluated. After that, the optimal candidate GNN architecture in the operation search phase can be selected based on the evaluated reasoning efficiency and task accuracy. After that, based on the optimal candidate GNN architecture, function search is performed through function downward tuning. For the candidate GNN architectures obtained through the function search, the task accuracy can be evaluated. After that, based on the evaluated task accuracy, a GNN architecture is selected from the candidate GNN architectures obtained through the function search. Based on the above method, multiple GNN architectures can be obtained, and a GNN architecture warehouse can be obtained by storing multiple GNN architectures in the architecture warehouse. After that, the runtime dynamic tuner can monitor the changes in the network environment and device changes of the end-edge system, and dynamically select a suitable GNN architecture from the GNN architecture warehouse according to the changes in the network environment and device changes, and deploy the GNN architecture to the edge side device and the end side device, so that the GNN architecture performs end-edge collaborative reasoning on the end side device and the edge side device.

[0098] Based on the same inventive concept as the architecture search method of the end-edge collaborative graph neural network provided in the embodiment of the present application, the embodiment of the present application also provides an architecture search device of the end-edge collaborative graph neural network, such as Fig. 9 As shown in Figure 9, the architecture search device 900 of the end-edge collaborative graph neural network includes: The scenario information acquisition module 901 is used to acquire target application scenario information of the target collaborative graph neural network; wherein the target application scenario information includes target application identification information and target end-side environment information; A candidate architecture search module 902 is used to obtain multiple candidate architectures of the target collaborative graph neural network based on the target application scenario information and the collaborative graph neural network supernet; wherein the collaborative graph neural network supernet is constructed based on the collaborative search space of the end-edge collaborative graph neural network, and each layer in the collaborative graph neural network supernet includes each operation in the collaborative search space and a function corresponding to each operation, and the operation includes a communication operation, and the communication operation indicates that a device switching is performed between an end device and an edge device, and each candidate architecture is obtained by randomly selecting a target operation and a target function corresponding to the target operation in each layer of the collaborative graph neural network supernet; The target architecture selection module 903 is used to select the target architecture of the target collaborative graph neural network from the multiple candidate architectures based on the collaborative reasoning performance information of each candidate architecture for the target application identification information under the target terminal edge environment information.

[0099] In some embodiments of the present application, the candidate architecture search module 902 is specifically used to: Specifying an initial objective function for each operation in the collaborative graph neural network supernet to obtain an initial architecture of the target collaborative graph neural network; wherein the initial objective function of the communication operation in the collaborative graph neural network supernet is specified based on the network bandwidth information between the end-edge devices in the target end-edge environment information; Pre-training the initial architecture based on a target application data set to obtain a pre-trained architecture of the target collaborative graph neural network; wherein the target application data set is an application data set corresponding to the target application identification information; Obtaining a plurality of first candidate architectures based on the pre-trained architecture; wherein each first candidate architecture is obtained by randomly selecting a target operation in each layer of the pre-trained architecture; Selecting an operation search phase architecture from the plurality of first candidate architectures based on collaborative reasoning speed information and collaborative reasoning accuracy information of each first candidate architecture for the target application identification information under the target end-side environment information; Based on the operation search phase architecture, a plurality of second candidate architectures are obtained; wherein each second candidate architecture is obtained by selecting a target function that is more efficient than its initial target function for each target operation in the operation search phase architecture, and the target function of the communication operation remains its initial target function; The target architecture selection module 903 is specifically used to: Based on the collaborative reasoning accuracy information of each second candidate architecture for the target application identification information under the target terminal edge environment information, a target architecture of the target collaborative graph neural network is selected from the multiple second candidate architectures.

[0100] In some embodiments of the present application, the architecture search device 900 further includes: A collaborative reasoning speed prediction module, used for inputting the collaborative reasoning feature data of each first candidate architecture into a target collaborative reasoning speed predictor to perform collaborative reasoning speed prediction, and obtaining a first collaborative reasoning speed of each first candidate architecture for the target application identification information under the target terminal edge environment information; Among them, the target collaborative reasoning speed predictor is a collaborative reasoning speed predictor corresponding to the target application identification information, which is obtained based on a target collaborative reasoning speed prediction data set; the target collaborative reasoning speed prediction data set is a collaborative reasoning speed prediction data set corresponding to the target application identification information, which includes collaborative reasoning feature data and collaborative reasoning speed data of a sample architecture; the target collaborative reasoning speed predictor is obtained through supervised training with the collaborative reasoning feature data of the sample architecture as input and the collaborative reasoning speed data of the sample architecture as output label; the collaborative reasoning feature data includes architecture feature information and edge device information.

[0101] In some embodiments of the present application, the architecture search device 900 further includes: A collaborative reasoning speed estimation module, configured to obtain a second collaborative reasoning speed for each first candidate architecture for target application identification information under target terminal edge environment information based on a sum of operation delay times corresponding to each layer of target operations of each first candidate architecture; The inference delay time corresponding to each target operation is obtained by searching in an operation delay time lookup table corresponding to the target application scenario information.

[0102] In some embodiments of the present application, the architecture search device 900 further includes: A collaborative reasoning speed estimation module, configured to obtain a second collaborative reasoning speed for each first candidate architecture for target application identification information under target terminal edge environment information based on a sum of operation delay times corresponding to each layer of target operations of each first candidate architecture; A screening module, configured to screen the plurality of first candidate architectures based on a first collaborative reasoning speed and a second collaborative reasoning speed of each first candidate architecture; The target architecture selection module 903 is specifically used to: For the screened first candidate architectures, based on the first collaborative reasoning speed and reasoning accuracy of each first candidate architecture for the target application identification information under the target end-edge environment information, an operation search phase architecture is selected from the screened first candidate architectures.

[0103] It should be noted that the embodiment of the architecture search device for the end-edge collaborative graph neural network in the present application and the embodiment of the architecture search method for the end-edge collaborative graph neural network in the present application are based on the same inventive concept. Therefore, the specific implementation of this embodiment can refer to the implementation of the aforementioned method, and the repeated parts will not be repeated.

[0104] Based on the same inventive concept as the architecture search method of the end-edge collaborative graph neural network provided in the embodiment of the present application, the embodiment of the present application also provides an architecture deployment device of the end-edge collaborative graph neural network, such as Fig.10 As shown in 10, the architecture deployment device 1000 of the end-edge collaborative graph neural network includes: The architecture acquisition module 1001 is used to acquire a first target architecture from a target architecture warehouse based on first end-edge environment information of an end-edge system for deploying a target collaborative graph neural network; wherein the target architecture warehouse is an architecture warehouse of the target collaborative graph neural network, which stores multiple target architectures corresponding to multiple end-edge environment information of the target collaborative graph neural network, and the target architecture corresponding to each end-edge environment information is obtained by using each end-edge environment information as the target end-edge environment information in the target application scenario information, and performing the architecture search method according to any one of claims 1 to 5; the first target architecture is a target architecture corresponding to the first end-edge environment information; A deployment module 1002 is configured to deploy the first target architecture to the end-edge system, so as to perform end-edge collaborative reasoning on the end-edge system based on the first target architecture; The environment monitoring module 1003 is used to monitor the edge environment changes of the edge system during the edge-to-edge collaborative reasoning process; The architecture acquisition module 1001 is further configured to acquire a second target architecture from the target architecture repository when the edge device environment changes; wherein the second target architecture is a target architecture corresponding to the second edge device environment information, and the second edge device environment information is the changed edge device environment information; The deployment module 1002 is further configured to deploy the second target architecture to the end-edge system, so as to perform end-edge collaborative reasoning on the end-edge system based on the second target architecture.

[0105] In some embodiments of the present application, the edge environment change includes: edge network situation change and / or edge device change.

[0106] It should be noted that the embodiment of the architecture deployment device of the end-edge collaborative graph neural network in the present application and the embodiment of the architecture deployment method of the end-edge collaborative graph neural network in the present application are based on the same inventive concept. Therefore, the specific implementation of this embodiment can refer to the implementation of the aforementioned method, and the repeated parts will not be repeated.

[0107] Based on the same inventive concept as the architecture search method or architecture deployment method of the end-edge collaborative graph neural network provided in the embodiment of the present application, the embodiment of the present application also provides an electronic device, such as Fig.11 shown.

[0108] The electronic device may provide a terminal device or a server, etc. for the above-mentioned embodiments.

[0109] The electronic device 1100 may have relatively large differences due to different configurations or performances, and may include one or more processors 1101 and memory 1102, and the memory 1102 may store one or more storage applications or data. Among them, the memory 1102 may be a short-term storage or a persistent storage. The application stored in the memory 1102 may include one or more modules (not shown in the figure), and each module may include a series of computer executable instructions in the processing device of the model. Furthermore, the processor 1101 may be configured to communicate with the memory 1102 to execute a series of computer executable instructions in the memory 1102 on the electronic device.

[0110] Specifically, in some embodiments, the electronic device includes a memory and one or more programs, wherein the one or more programs are stored in the memory, and the one or more programs may include one or more modules, and each module may include a series of computer executable instructions in the processing device of the model, and the program or instructions are configured to implement the following process when executed by the processor: Obtain target application scenario information of the target collaborative graph neural network; wherein the target application scenario information includes target application identification information and target end-side environment information; Based on the target application scenario information and the collaborative graph neural network supernet, multiple candidate architectures of the target collaborative graph neural network are obtained; wherein the collaborative graph neural network supernet is constructed based on the collaborative search space of the end-edge collaborative graph neural network, and each layer in the collaborative graph neural network supernet includes each operation in the collaborative search space and a function corresponding to each operation, and the operation includes a communication operation, and the communication operation indicates that a device switching is performed between an end device and an edge device, and each candidate architecture is obtained by randomly selecting a target operation and a target function corresponding to the target operation in each layer of the collaborative graph neural network supernet; Based on the collaborative reasoning performance information of each candidate architecture for the target application identification information under the target end-edge environment information, a target architecture of the target collaborative graph neural network is selected from the multiple candidate architectures.

[0111] In some embodiments of the present application, when the program or instruction is executed by the processor, the following process is specifically implemented: Specifying an initial objective function for each operation in the collaborative graph neural network supernet to obtain an initial architecture of the target collaborative graph neural network; wherein the initial objective function of the communication operation in the collaborative graph neural network supernet is specified based on the network bandwidth information between the end-edge devices in the target end-edge environment information; Pre-training the initial architecture based on a target application data set to obtain a pre-trained architecture of the target collaborative graph neural network; wherein the target application data set is an application data set corresponding to the target application identification information; Obtaining a plurality of first candidate architectures based on the pre-trained architecture; wherein each first candidate architecture is obtained by randomly selecting a target operation in each layer of the pre-trained architecture; Selecting an operation search phase architecture from the plurality of first candidate architectures based on collaborative reasoning speed information and collaborative reasoning accuracy information of each first candidate architecture for the target application identification information under the target end-side environment information; Based on the operation search phase architecture, a plurality of second candidate architectures are obtained; wherein each second candidate architecture is obtained by selecting a target function that is more efficient than its initial target function for each target operation in the operation search phase architecture, and the target function of the communication operation remains its initial target function; Based on the collaborative reasoning accuracy information of each second candidate architecture for the target application identification information under the target terminal edge environment information, a target architecture of the target collaborative graph neural network is selected from the multiple second candidate architectures.

[0112] In some embodiments of the present application, in the collaborative reasoning speed information and the collaborative reasoning accuracy information for the target application identification information based on each first candidate architecture under the target end edge environment information, before the multiple first candidate architectures select an operation search phase architecture, the program or instruction further implements the following process when executed by the processor: Inputting the collaborative reasoning feature data of each first candidate architecture into a target collaborative reasoning speed predictor to perform collaborative reasoning speed prediction, and obtaining a first collaborative reasoning speed of each first candidate architecture for the target application identification information under the target terminal edge environment information; Among them, the target collaborative reasoning speed predictor is a collaborative reasoning speed predictor corresponding to the target application identification information, which is obtained based on a target collaborative reasoning speed prediction data set; the target collaborative reasoning speed prediction data set is a collaborative reasoning speed prediction data set corresponding to the target application identification information, which includes collaborative reasoning feature data and collaborative reasoning speed data of a sample architecture; the target collaborative reasoning speed predictor is obtained through supervised training with the collaborative reasoning feature data of the sample architecture as input and the collaborative reasoning speed data of the sample architecture as output label; the collaborative reasoning feature data includes architecture feature information and edge device information.

[0113] In some embodiments of the present application, in the collaborative reasoning speed information and the collaborative reasoning accuracy information for the target application identification information based on each first candidate architecture under the target end edge environment information, before the multiple first candidate architectures select an operation search phase architecture, the program or instruction further implements the following process when executed by the processor: Based on the sum of the operation delay times corresponding to the target operations of each layer of each first candidate architecture, obtaining a second collaborative reasoning speed for the target application identification information of each first candidate architecture under the target terminal edge environment information; The inference delay time corresponding to each target operation is obtained by searching in an operation delay time lookup table corresponding to the target application scenario information.

[0114] In some embodiments of the present application, in the collaborative reasoning speed information and the collaborative reasoning accuracy information for the target application identification information based on each first candidate architecture under the target end edge environment information, before the multiple first candidate architectures select an operation search phase architecture, the program or instruction further implements the following process when executed by the processor: Based on the sum of the operation delay times corresponding to the target operations of each layer of each first candidate architecture, obtaining a second collaborative reasoning speed for the target application identification information of each first candidate architecture under the target terminal edge environment information; Screening the plurality of first candidate architectures based on a first collaborative reasoning speed and a second collaborative reasoning speed of each first candidate architecture; For the screened first candidate architectures, based on the first collaborative reasoning speed and reasoning accuracy of each first candidate architecture for the target application identification information under the target end-side environment information, an operation search phase architecture is selected from the screened first candidate architectures.

[0115] Specifically in some other embodiments, the electronic device includes a memory and one or more programs, wherein the one or more programs are stored in the memory, and the one or more programs may include one or more modules, and each module may include a series of computer executable instructions in the processing device of the model, and the program or instructions are configured to implement the following process when executed by the processor: Based on the first end-edge environment information of the end-edge system for deploying the target collaborative graph neural network, a first target architecture is obtained from a target architecture warehouse; wherein the target architecture warehouse is an architecture warehouse of the target collaborative graph neural network, which stores multiple target architectures corresponding to multiple end-edge environment information of the target collaborative graph neural network, and the target architecture corresponding to each end-edge environment information is obtained by taking each end-edge environment information as the target end-edge environment information in the target application scenario information, and performing the process of the architecture search method described in the above embodiment; the first target architecture is a target architecture corresponding to the first end-edge environment information; Deploying the first target architecture to the end-edge system to perform end-edge collaborative reasoning on the end-edge system based on the first target architecture; During the process of performing edge-end collaborative reasoning on the edge-end system based on the first target architecture, monitoring changes in the edge-end environment of the edge-end system; When the edge device environment changes, a second target architecture is acquired from the target architecture repository; wherein the second target architecture is a target architecture corresponding to the second edge device environment information, and the second edge device environment information is the changed edge device environment information; The second target architecture is deployed on the end-edge system to perform end-edge collaborative reasoning on the end-edge system based on the second target architecture.

[0116] In some embodiments of the present application, the edge environment change includes: edge network situation change and / or edge device change.

[0117] It should be noted that the embodiments of the electronic device in the present application and the architecture search method of the end-edge collaborative graph neural network in the present application, or the embodiment of the architecture deployment method of the end-edge collaborative graph neural network are based on the same inventive concept. Therefore, the specific implementation of this embodiment can refer to the implementation of the aforementioned method, and the repeated parts will not be repeated.

[0118] The above describes specific embodiments of the present application. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0119] In the 1990s, it was very clear whether the improvement of a technology was hardware improvement (for example, improvement of the circuit structure of diodes, transistors, switches, etc.) or software improvement (improvement of the method flow). However, with the development of technology, many improvements of the method flow today can be regarded as direct improvements of the hardware circuit structure. Designers almost always obtain the corresponding hardware circuit structure by programming the improved method flow into the hardware circuit. Therefore, it cannot be said that the improvement of a method flow cannot be implemented with a hardware entity module. For example, a programmable logic device (PLD) (such as a field programmable gate array (FPGA)) is such an integrated circuit whose logical function is determined by the user's programming of the device. Designers can "integrate" a digital system on a PLD by programming themselves, without having to ask chip manufacturers to design and make dedicated integrated circuit chips. Moreover, nowadays, instead of manually making integrated circuit chips, this kind of programming is mostly implemented by "logic compiler" software, which is similar to the software compiler used when developing and writing programs, and the original code before compilation must also be written in a specific programming language, which is called hardware description language (HDL). There is not only one kind of HDL, but many kinds, such as ABEL (Advanced Boolean Expression Language), AHDL (Altera Hardware Description Language), Confluence, CUPL (Cornell University Programming Language), HDCal, JHDL (Java Hardware Description Language), Lava, Lola, MyHDL, PALASM, RHDL (Ruby Hardware Description Language), etc. The most commonly used ones are VHDL (Very-High-Speed ​​Integrated Circuit Hardware Description Language) and Verilog. Those skilled in the art should also know that it is only necessary to program the method flow slightly in the above-mentioned hardware description languages ​​and program it into the integrated circuit, and then it is easy to obtain the hardware circuit that implements the logic method flow.

[0120] The controller may be implemented in any suitable manner, for example, the controller may take the form of a microprocessor or processor and a computer-readable medium storing a computer-readable program code (e.g., software or firmware) executable by the (micro)processor, a logic gate, a switch, an application-specific integrated circuit (ASIC), a programmable logic controller, and an embedded microcontroller, examples of which include but are not limited to the following microcontrollers: ARC 625D, Atmel AT91SAM, Microchip PIC18F26K20, and Silicone Labs C8051F320, and the memory controller may also be implemented as part of the control logic of the memory. It is also known to those skilled in the art that, in addition to implementing the controller in a purely computer-readable program code manner, the controller may be implemented in the form of a logic gate, a switch, an application-specific integrated circuit, a programmable logic controller, and an embedded microcontroller by logically programming the method steps. Therefore, such a controller may be considered as a hardware component, and the devices for implementing various functions included therein may also be considered as structures within the hardware component. Or even, the devices for implementing various functions may be considered as both software modules for implementing the method and structures within the hardware component.

[0121] The systems, devices, modules or units described in the above embodiments may be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, the computer may be, for example, a personal computer, a laptop computer, a cellular phone, a camera phone, a smart phone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or a combination of any of these devices.

[0122] For the convenience of description, the above devices are described in terms of functions and are divided into various units. Of course, when implementing one or more embodiments of the present application, the functions of each unit can be implemented in the same or multiple software and / or hardware.

[0123] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, one or more embodiments of the present application may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, one or more embodiments of the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0124] The embodiments of the present application are described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable fraud case serial and parallel device to produce a machine, so that the instructions executed by the processor of the computer or other programmable fraud case serial and parallel device generate instructions for implementing the processes in the process. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0125] These computer program instructions may also be stored in a computer readable memory capable of directing a computer or other programmable fraud case serial and parallel device to operate in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0126] These computer program instructions may also be loaded onto a computer or other programmable device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, whereby the instructions executed on the computer or other programmable device provide for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0127] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0128] Memory may include non-permanent storage in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. Memory is an example of a computer-readable medium.

[0129] Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined in this article, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.

[0130] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0131] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems or computer program products. Therefore, one or more embodiments of the present application may take the form of a complete hardware embodiment, a complete software embodiment or an embodiment combining software and hardware. Moreover, one or more embodiments of the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0132] One or more embodiments of the present application may be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform specific tasks or implement specific abstract data types. One or more embodiments of the present application may also be practiced in distributed computing environments, in which tasks are performed by remote processing devices connected through a communication network. In a distributed computing environment, program modules may be located in local and remote computer storage media, including storage devices.

[0133] Each embodiment in this application is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0134] The above is only an embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the scope of the claims of the present application.

Claims

1. A method for searching an architecture of an edge-end collaborative graph neural network, characterized in that: The method comprises: Obtain target application scenario information of the target collaborative graph neural network; wherein the target application scenario information includes target application identification information and target end-side environment information; Based on the target application scenario information and the collaborative graph neural network supernet, multiple candidate architectures of the target collaborative graph neural network are obtained; wherein the collaborative graph neural network supernet is constructed based on the collaborative search space of the end-edge collaborative graph neural network, and each layer in the collaborative graph neural network supernet includes each operation in the collaborative search space and a function corresponding to each operation, and the operation includes a communication operation, and the communication operation indicates that a device switching is performed between an end device and an edge device, and each candidate architecture is obtained by randomly selecting a target operation and a target function corresponding to the target operation in each layer of the collaborative graph neural network supernet; Based on the collaborative reasoning performance information of each candidate architecture for the target application identification information under the target end-edge environment information, a target architecture of the target collaborative graph neural network is selected from the multiple candidate architectures.

2. The architecture search method according to claim 1, characterized in that: The step of obtaining multiple candidate architectures of the target collaborative graph neural network based on the target application scenario information and the collaborative graph neural network supernet includes: Specifying an initial objective function for each operation in the collaborative graph neural network supernet to obtain an initial architecture of the target collaborative graph neural network; wherein the initial objective function of the communication operation in the collaborative graph neural network supernet is specified based on the network bandwidth information between the end-edge devices in the target end-edge environment information; Pre-training the initial architecture based on a target application data set to obtain a pre-trained architecture of the target collaborative graph neural network; wherein the target application data set is an application data set corresponding to the target application identification information; Obtaining a plurality of first candidate architectures based on the pre-trained architecture; wherein each first candidate architecture is obtained by randomly selecting a target operation in each layer of the pre-trained architecture; Selecting an operation search phase architecture from the plurality of first candidate architectures based on collaborative reasoning speed information and collaborative reasoning accuracy information of each first candidate architecture for the target application identification information under the target end-side environment information; Based on the operation search phase architecture, a plurality of second candidate architectures are obtained; wherein each second candidate architecture is obtained by selecting a target function that is more efficient than its initial target function for each target operation in the operation search phase architecture, and the target function of the communication operation remains its initial target function; The selecting a target architecture of the target collaborative graph neural network from the multiple candidate architectures based on the collaborative reasoning performance information of each candidate architecture for the target application identification information under the target terminal edge environment information includes: Based on the collaborative reasoning accuracy information of each second candidate architecture for the target application identification information under the target terminal edge environment information, a target architecture of the target collaborative graph neural network is selected from the multiple second candidate architectures.

3. The architecture search method according to claim 2, wherein before selecting an architecture in an operation search phase based on the collaborative reasoning speed information and the collaborative reasoning accuracy information of each first candidate architecture for the target application identification information under the target end-side environment information, the method further comprises: Inputting the collaborative reasoning feature data of each first candidate architecture into a target collaborative reasoning speed predictor to perform collaborative reasoning speed prediction, and obtaining a first collaborative reasoning speed of each first candidate architecture for the target application identification information under the target terminal edge environment information; Among them, the target collaborative reasoning speed predictor is a collaborative reasoning speed predictor corresponding to the target application identification information, which is obtained based on a target collaborative reasoning speed prediction data set; the target collaborative reasoning speed prediction data set is a collaborative reasoning speed prediction data set corresponding to the target application identification information, which includes collaborative reasoning feature data and collaborative reasoning speed data of a sample architecture; the target collaborative reasoning speed predictor is obtained through supervised training with the collaborative reasoning feature data of the sample architecture as input and the collaborative reasoning speed data of the sample architecture as output label; the collaborative reasoning feature data includes architecture feature information and edge device information.

4. The architecture search method according to claim 2, wherein before selecting an operation search phase architecture based on the collaborative reasoning speed information and the collaborative reasoning accuracy information of each first candidate architecture for the target application identification information under the target end-side environment information, the method further comprises: Based on the sum of the operation delay times corresponding to the target operations of each layer of each first candidate architecture, obtaining a second collaborative reasoning speed for the target application identification information of each first candidate architecture under the target terminal edge environment information; The inference delay time corresponding to each target operation is obtained by searching in an operation delay time lookup table corresponding to the target application scenario information.

5. The architecture search method according to claim 3, wherein before selecting an architecture in an operation search phase based on the collaborative reasoning speed information and the collaborative reasoning accuracy information of each first candidate architecture for the target application identification information under the target end-side environment information, the method further comprises: Based on the sum of the operation delay times corresponding to the target operations of each layer of each first candidate architecture, obtaining a second collaborative reasoning speed for the target application identification information of each first candidate architecture under the target terminal edge environment information; Screening the plurality of first candidate architectures based on a first collaborative reasoning speed and a second collaborative reasoning speed of each first candidate architecture; The selecting an operation search phase architecture from the plurality of first candidate architectures based on the inference speed information and the inference accuracy information of each first candidate architecture for the target application identification information under the target terminal edge environment information comprises: For the screened first candidate architectures, based on the first collaborative reasoning speed and reasoning accuracy of each first candidate architecture for the target application identification information under the target end-edge environment information, an operation search phase architecture is selected from the screened first candidate architectures.

6. A method for deploying an architecture of a device-edge collaborative graph neural network, characterized in that: The method comprises: Based on the first end-edge environment information of the end-edge system for deploying the target collaborative graph neural network, a first target architecture is obtained from a target architecture warehouse; wherein the target architecture warehouse is an architecture warehouse of the target collaborative graph neural network, which stores multiple target architectures corresponding to multiple end-edge environment information of the target collaborative graph neural network, and the target architecture corresponding to each end-edge environment information is obtained by taking each end-edge environment information as the target end-edge environment information in the target application scenario information, and performing the architecture search method according to any one of claims 1 to 5; the first target architecture is a target architecture corresponding to the first end-edge environment information; Deploying the first target architecture to the end-edge system to perform end-edge collaborative reasoning on the end-edge system based on the first target architecture; During the process of performing the end-edge collaborative reasoning, monitoring the end-edge environment changes of the end-edge system; When the edge device environment changes, a second target architecture is acquired from the target architecture repository; wherein the second target architecture is a target architecture corresponding to the second edge device environment information, and the second edge device environment information is the changed edge device environment information; The second target architecture is deployed on the end-edge system to perform end-edge collaborative reasoning on the end-edge system based on the second target architecture.

7. The architecture deployment method according to claim 6, characterized in that: The edge environment change includes: edge network situation change and / or edge device change.

8. An architecture search device for an edge-to-edge collaborative graph neural network, characterized in that: The device comprises: A scenario information acquisition module, used to acquire target application scenario information of a target collaborative graph neural network; wherein the target application scenario information includes target application identification information and target end-side environment information; A candidate architecture search module, used to obtain multiple candidate architectures of the target collaborative graph neural network based on the target application scenario information and the collaborative graph neural network supernet; wherein the collaborative graph neural network supernet is constructed based on the collaborative search space of the end-edge collaborative graph neural network, each layer in the collaborative graph neural network supernet includes each operation in the collaborative search space and a function corresponding to each operation, the operation includes a communication operation, and the communication operation indicates that a device switching is performed between an end device and an edge device, and each candidate architecture is obtained by randomly selecting a target operation and a target function corresponding to the target operation in each layer of the collaborative graph neural network supernet; A target architecture selection module is used to select a target architecture of the target collaborative graph neural network from the multiple candidate architectures based on the collaborative reasoning performance information of each candidate architecture for the target application identification information under the target terminal edge environment information.

9. An architecture deployment device for a device-edge collaborative graph neural network, characterized in that: The device comprises: An architecture acquisition module, configured to acquire a first target architecture from a target architecture warehouse based on first end-edge environment information of an end-edge system for deploying a target collaborative graph neural network; wherein the target architecture warehouse is an architecture warehouse of the target collaborative graph neural network, which stores a plurality of target architectures corresponding to a plurality of end-edge environment information of the target collaborative graph neural network, and the target architecture corresponding to each end-edge environment information is obtained by using each end-edge environment information as the target end-edge environment information in the target application scenario information, and performing an architecture search method according to any one of claims 1 to 5; the first target architecture is a target architecture corresponding to the first end-edge environment information; a deployment module, configured to deploy the first target architecture to the end-edge system, so as to perform end-edge collaborative reasoning on the end-edge system based on the first target architecture; An environment monitoring module, used to monitor changes in the edge environment of the edge system during edge-to-edge collaborative reasoning; The architecture acquisition module is further configured to acquire a second target architecture from the target architecture repository when the edge device environment changes; wherein the second target architecture is a target architecture corresponding to second edge device environment information, and the second edge device environment information is the changed edge device environment information; The deployment module is further used to deploy the second target architecture to the end-edge system to perform end-edge collaborative reasoning on the end-edge system based on the second target architecture.

10. An electronic device, characterized in that: The method comprises a processor and a memory, wherein the memory stores a program or instruction that can be run on the processor, and when the program or instruction is executed by the processor, the steps of the method according to any one of claims 1 to 5 or 6 to 7 are implemented.

Images