Model matching method and device and electronic equipment
By selecting small models on resource-constrained edge devices and working in conjunction with adapted large models, the problem that edge devices cannot run large models is solved, and efficient data processing and application on resource-constrained devices is achieved.
Patent Information
- Application Number
- CN202411945444.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-25
- Publication Date
- 2025-04-25
AI Technical Summary
Resource-constrained edge devices cannot host and run large models, causing them to be unable to apply on these devices.
Through the model matching method, select small models and adapted large models, use small models to process data on edge devices, and input the processing results into the adapted large models for secondary analysis, realizing the joint work of small models and large models.
It realizes that small models run on edge devices with limited resources, and large models run on equipment with rich resources, avoiding the defects of resource-constrained devices being unable to bear large models, and improving application flexibility and efficiency.
Smart Images
Figure CN119884779A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to artificial intelligence technology, and in particular to a model matching method, device and electronic device. Background Art
[0002] A large model refers to a machine learning model with large-scale parameters and complex computing structures. It usually consists of millions to billions of parameters and requires a large amount of data and computing resources for training and reasoning. Large models have powerful representation and generalization capabilities in the field of deep learning and can be applied to various tasks such as speech recognition, natural language processing, and computer vision.
[0003] Currently, in many applications, large models are directly used for processing such as speech recognition. However, for resource-constrained edge devices, they may not be able to carry and run large models due to resource constraints, which makes the large models impossible to apply. Summary of the invention
[0004] The present application provides a model matching method, device and electronic device to achieve the joint operation of a small model and an adapted large model.
[0005] The present invention provides a model matching method, which includes:
[0006] For each operation node in the business orchestration view corresponding to the current business demand, select a first tag from all tags supported by the small model selected for the operation node; the small model refers to a model whose complexity meets the set small model requirements;
[0007] Based on the small model and the first label, obtain N large models selected for the operation node that are adapted to the small model and the first label, and at least one target label that is adapted to the first label and supported by each of the N large models; N is greater than or equal to 1; each target label supported by any of the large models is different from the first label; all target labels supported by different large models are different; any of the large models refers to a model whose complexity meets the requirements of the set large model,
[0008] Determine the operation type of the operation that the small model needs to perform for the first tag; and determine the operation type of the operation that the large model needs to perform based on at least one target tag;
[0009] The small model is controlled to perform a corresponding operation based on the operation type of the operation to be performed by the first tag, and the large model is controlled to perform a corresponding operation based on the operation type of the operation to be performed by at least one target tag.
[0010] The present application provides a model matching device, which includes:
[0011] A label selection unit, configured to select, for each operation node in the business arrangement view corresponding to the current business demand, a first label from all labels supported by a small model selected for the operation node; the small model refers to a model whose complexity meets the set small model requirement;
[0012] A model adaptation unit is used to obtain, based on the small model and the first label, N large models selected for the operation node that are adapted to the small model and the first label, and at least one target label that is adapted to the first label and supported by each of the N large models; N is greater than or equal to 1; each target label supported by any of the large models is different from the first label; all target labels supported by different large models are different; any of the large models refers to a model whose complexity meets the requirements of the set large model,
[0013] An operation unit, configured to determine an operation type of an operation to be performed by the small model for the first tag; and determine an operation type of an operation to be performed by the large model based on at least one target tag;
[0014] A control unit is used to control the small model to perform a corresponding operation based on the operation type of the operation to be performed by the first tag, and to control the large model to perform a corresponding operation based on the operation type of the operation to be performed by at least one target tag.
[0015] The embodiment of the present application also provides an electronic device. The electronic device includes: a processor and a machine-readable storage medium;
[0016] The machine-readable storage medium stores machine-executable instructions that can be executed by the processor;
[0017] The processor is used to execute machine executable instructions to implement the steps of the above disclosed method.
[0018] It can be seen from the above technical scheme that in the present application, a small model is first selected based on the demand and a label (recorded as the first label) is selected from multiple labels executed by the small model. Then, based on the small model and the first label, N large models selected for the operation node that are adapted to the small model and the first label and at least one target label supported by each of the N large models that is adapted to the first label are obtained, where N is greater than or equal to 1; then, the small model can be controlled to perform corresponding operations based on the operation type of the operation to be performed based on the first label, and the large model can be controlled to perform corresponding operations based on the operation type of the operation to be performed based on at least one target label. This achieves that the large model is not directly applied for processing, but the small model and the adapted large model work together, so that the small model can be run on edge devices such as resource-constrained devices, while the large model can be run on other resource-rich devices, avoiding the defect that the large model cannot be applied because the resource-constrained device cannot carry and run the large model.
[0019] Furthermore, this embodiment does not blindly combine the small model and the large model to work together, but selects a small model based on demand and selects a label (recorded as the first label) from multiple labels executed by the small model, and then selects N adapted large models based on the first label and at least one target label supported by each large model that is adapted to the first label. This avoids the problem of mismatch between the small model and the large model when the existing small model and the large model work together, such as label mismatch. BRIEF DESCRIPTION OF THE DRAWINGS
[0020] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present disclosure and, together with the description, serve to explain the principles of the present disclosure.
[0021] Figure 1 A flow chart of the method provided in the embodiment of the present application;
[0022] Figure 2 A schematic diagram of a service arrangement view provided in an embodiment of the present application;
[0023] Figure 3 A schematic diagram of a model interface provided in an embodiment of the present application;
[0024] Figure 4 A structural diagram of a device provided in an embodiment of the present application;
[0025] Figure 5 A structural diagram of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION
[0026] In order to enable those skilled in the art to better understand the technical solutions provided by the embodiments of the present application and to make the above-mentioned purposes, features and advantages of the embodiments of the present application more obvious and understandable, the technical solutions in the embodiments of the present application are further described in detail below in conjunction with the accompanying drawings.
[0027] In order to avoid the resource-constrained edge devices being unable to apply the large model, this embodiment proposes a secondary filtering solution. The secondary filtering solution is specifically: first use a small model to process data, such as data analysis, and then send the processing results, such as key data (generally alarms or valid information) completed by the analysis, to the large model for secondary analysis and correction.
[0028] Here, small models and large models are independent and pre-trained models. As described above, large models usually refer to models whose complexity meets the requirements of large models, such as deep learning models with hundreds of millions to hundreds of billions of parameters. Such models have complex structures and can handle very complex tasks, such as natural language processing, computer vision, and recommendation systems. Large models undergo a special training process and can perform complex processing on large amounts of data, which usually requires the support of high-performance computing resources, such as standard GPUs or cloud clusters.
[0029] Compared with large models, small models refer to models whose complexity meets the requirements of the set small model, such as deep neural network models with fewer parameters. Such models usually run faster and are more lightweight, suitable for devices or scenarios with limited computing resources and storage space, such as mobile devices or embedded devices. Small models are suitable for solving some simple, small-scale problems, such as credit card fraud detection, and can run on low-power devices.
[0030] However, since the small model and the large model are trained independently, there may often be incompatibility issues in specific applications, such as inconsistent labels supported by the large model and the small model. This incompatibility makes it impossible to implement intelligent secondary filtering.
[0031] To solve the above technical problems, this embodiment proposes Figure 1 The method shown:
[0032] See also Figure 1 , Figure 1 The method flow chart provided in the embodiment of the present application is applied to an electronic device. The electronic device here is, for example, a terminal, a server, etc., which is not specifically limited in this embodiment.
[0033] like Figure 1 As shown, the process may include the following steps:
[0034] Step 101 : for each operation node in the business orchestration view corresponding to the current business demand, select a first tag from all tags supported by the small model selected for the operation node.
[0035] The service orchestration view is specifically orchestrated according to service requirements, which is not specifically limited in this embodiment. Figure 2 An example of a service orchestration view is shown.
[0036] Optionally, the service orchestration view is visible to a user, and may also be referred to as a user view.
[0037] like Figure 2 The business orchestration view shown is composed of multiple operation nodes. For example, Figure 2 The operation node 1 (construction site clothing standard detection), operation node 2 (personnel information identification), etc. are shown.
[0038] As an embodiment, this embodiment can output the interface corresponding to each operation node in the business arrangement view after selecting a corresponding small model for the operation node. Figure 2 Taking the operation node 1 (construction site wear standard detection) as an example, the small model selected for the operation node 1 is the hat wearing detection model. Figure 3 The following example shows the interface corresponding to the hat wearing detection model. This interface displays all the tags supported by the hat wearing detection model, such as wearing a hard hat, wearing a duckbill cap, etc.
[0039] After that, the selected event in the interface can be responded to to obtain the selected tag (recorded as the first tag) based on the selected event. Figure 3 Example of wearing a hard hat.
[0040] Step 102, based on the above-mentioned small model and the above-mentioned first label, obtain N large models selected for the operation node that are adapted to the above-mentioned small model and the above-mentioned first label, and at least one target label adapted to the first label supported by each of the N large models; N is greater than or equal to 1.
[0041] As an embodiment, this embodiment can select a large model that can support the first label based on the small model and the first label, and the selected large model is a large model that is compatible with the small model and the first label.
[0042] At least one label (recorded as the target label) supported by any selected large model is different from the first label. For example, any target label supported by any selected large model is a subset of the first label, or any target label supported by any selected large model is associated with the first label, etc., which is not specifically limited in this embodiment. For example: If the small model selected for the operation node is model 1, all labels supported by model 1 are: label 1): human body; label 2): vehicle. If label 2 ("vehicle") supported by model 1 is selected as the first label, a large model that supports label 1 ("small vehicle"), such as model 2, can be selected as a large model that is compatible with the above-mentioned small model and the first label. That is, the following correspondence is finally established: model 1, label 2 corresponds to model 2, label 1.
[0043] By selecting N large models that are compatible with the small model and the first label for the operation node, the intelligent effect can be enhanced. For example, as in the above example, small vehicles can be enhanced, and other vehicles will be directly considered as false alarms and filtered out.
[0044] Step 103, determining the operation type of the operation that the small model needs to perform for the first tag; and determining the operation type of the operation that each large model needs to perform based on at least one target tag.
[0045] For example, the operation type of the operation that the small model needs to perform for the first tag is a detection type, and the operation type of the operation that the large model needs to perform based on at least one target tag is a filtering type. Of course, the operation type of the operation that the small model needs to perform for the first tag can also be a filtering type, and the operation type of the operation that the large model needs to perform based on at least one target tag is a detection type. This embodiment is not specifically limited.
[0046] Step 104, controlling the small model to perform a corresponding operation based on the operation type of the operation to be performed by the first tag, and controlling the large model to perform a corresponding operation based on the operation type of the operation to be performed by at least one target tag.
[0047] As an embodiment, if the above N is 1, the operation type of the operation to be performed by the above first tag is a detection type, and the operation type of the operation to be performed by the above large model based on at least one target tag is a filtering type, then:
[0048] Obtain an operation result obtained by the small model performing a corresponding operation based on the operation type of the operation to be performed by the first tag; input the operation result into the large model, so that the large model performs a corresponding operation on the input data based on at least one target tag to obtain an output result.
[0049] Still taking the above association relationship: model 1, label 2 corresponds to model 2, label 1 as an example, the vehicle detection result output by model 1 is input into model 2. Model 2 identifies whether the input vehicle detection result matches label 1 "small vehicle" based on label 1 "small vehicle". If so, it is determined that a vehicle exists. If not, it is determined that a vehicle does not exist. This means that other vehicles will be directly considered as false alarms and filtered out.
[0050] As an embodiment, if the above N is 1, the operation type of the operation to be performed by the above first tag is a detection type, the operation type of the operation to be performed by the above large model based on at least one target tag is a filtering type, and there is a cascade relationship between the N large models, then:
[0051] Obtain the operation result obtained by the small model performing the corresponding operation based on the operation type of the operation to be performed by the first tag; input the operation result to the first large model of the cascade to trigger N large models to perform corresponding operations on the input data based on at least one supported target tag; for the second large model of the cascade to the last large model, its input data is the output result of the previous large model of the cascade;
[0052] Get the output result of the last model in the cascade of N large models.
[0053] Here, for the second large model to the last large model in the cascade, the input data is the output result of the previous large model in the cascade.
[0054] In this embodiment, any large model has a corresponding template graph. The template graph has a corresponding operator table, which is used to determine the connection relationship and data transfer requirements between operators in the template graph. Based on this, for each large model, data can be transferred between operators based on the corresponding operator table.
[0055] If the above N is greater than 1, there is a cascade relationship between at least two of the above N large models; the cascade relationship between any two large models includes: there is a connection relationship between one of the operators in the template graph corresponding to one of the large models and an operator in the template graph corresponding to another large model.
[0056] As an embodiment, when there is a connection relationship between two operators, the connection relationship has a corresponding weight, wherein the weight corresponding to the connection relationship can be set depending on the data transmission requirement.
[0057] When an operator has a connection relationship with at least two different operators, the connection relationship with the highest weight is preferred. For example, when data needs to be transferred from the starting end to the termination end, if operator A is on the path, if operator B and operator C can be passed from operator A to the termination end, then the weight of the connection relationship between operator A and operator B, and the weight of the connection relationship between operator A and operator C can be used to determine whether to select operator B or operator C. Generally speaking, if operator B is on the shortest path from operator A to the termination end, the weight of the connection relationship between operator A and operator B is high. Similarly, if operator C is on the shortest path from operator A to the termination end, the weight of the connection relationship between operator A and operator C is high.
[0058] So far, completed Figure 1 The process shown.
[0059] pass Figure 1 It can be seen from the shown process that in this embodiment, a small model is first selected based on the demand and a label (recorded as the first label) is selected from multiple labels executed by the small model. Then, based on the small model and the first label, N large models selected for the operation node that are adapted to the small model and the first label and at least one target label supported by each of the N large models and adapted to the first label are obtained, where N is greater than or equal to 1; then, the small model can be controlled to perform corresponding operations based on the operation type of the operation to be performed based on the first label, and the large model can be controlled to perform corresponding operations based on the operation type of the operation to be performed based on at least one target label. This achieves that the large model is not directly applied for processing, but the small model and the adapted large model work together, so that the small model can be run on edge devices with limited resources, such as resource-constrained devices, while the large model can be run on other resource-rich devices, avoiding the defect that the large model cannot be applied because the resource-constrained device cannot carry and run the large model.
[0060] Furthermore, this embodiment does not blindly combine the small model and the large model to work together, but selects a small model based on demand and selects a label (recorded as the first label) from multiple labels executed by the small model, and then selects N adapted large models based on the first label and at least one target label supported by each large model that is adapted to the first label. This avoids the problem of mismatch between the small model and the large model when the existing small model and the large model work together, such as label mismatch.
[0061] The method provided in the embodiment of the present application is described above. The device provided in the embodiment of the present application is described below:
[0062] See also Figure 4 , Figure 4 This is a diagram of the structure of the device provided in the embodiment of the present application. The device includes:
[0063] A label selection unit, configured to select, for each operation node in the business arrangement view corresponding to the current business demand, a first label from all labels supported by a small model selected for the operation node; the small model refers to a model whose complexity meets the set small model requirement;
[0064] A model adaptation unit is used to obtain, based on the small model and the first label, N large models selected for the operation node that are adapted to the small model and the first label, and at least one target label that is adapted to the first label and supported by each of the N large models; N is greater than or equal to 1; each target label supported by any of the large models is different from the first label; all target labels supported by different large models are different; any of the large models refers to a model whose complexity meets the requirements of the set large model,
[0065] An operation unit, configured to determine an operation type of an operation to be performed by the small model for the first tag; and determine an operation type of an operation to be performed by the large model based on at least one target tag;
[0066] A control unit is used to control the small model to perform a corresponding operation based on the operation type of the operation to be performed by the first tag, and to control the large model to perform a corresponding operation based on the operation type of the operation to be performed by at least one target tag.
[0067] Optionally, controlling the small model to perform a corresponding operation based on an operation type of an operation to be performed by the first tag, and controlling the large model to perform a corresponding operation based on an operation type of an operation to be performed by at least one target tag includes:
[0068] If N is 1, the operation type of the operation to be performed by the first tag is a detection type, and the operation type of the operation to be performed by the large model based on at least one target tag is a filtering type, then:
[0069] Obtaining an operation result obtained by the small model performing a corresponding operation based on the operation type of the operation to be performed by the first tag; inputting the operation result into the large model, so that the large model performs a corresponding operation on the input data based on at least one target tag to obtain an output result;
[0070] If N is greater than 1, the operation type of the operation to be performed by the first tag is a detection type, the operation type of the operation to be performed by the large model based on at least one target tag is a filtering type, and there is a cascade relationship between the N large models, then:
[0071] Obtain the operation result obtained by the small model performing the corresponding operation based on the operation type of the operation to be performed by the first tag; input the operation result to the first large model of the cascade to trigger N large models to perform corresponding operations on the input data based on at least one supported target tag; for the second large model of the cascade to the last large model, the input data is the output result of the previous large model of the cascade;
[0072] Get the output result of the last model in the cascade of N large models.
[0073] Optionally, any of the large models has a corresponding template graph; the template graph has a corresponding operator table, and the operator table is used to determine the connection relationship and data transfer requirements between the operators in the template graph;
[0074] If N is greater than 1, there is a cascade relationship between at least two of the N large models; the cascade relationship between any two large models includes: there is a connection relationship between an operator in the template graph corresponding to one of the large models and an operator in the template graph corresponding to another large model;
[0075] When there is a connection relationship between two operators, the connection relationship has a corresponding weight;
[0076] When an operator has a connection relationship with at least two different operators, the connection relationship with the highest weight is preferred.
[0077] Optionally, selecting the first label from all labels supported by the small model selected for the node comprises:
[0078] In response to a selection event in the interface corresponding to the mini-model, the interface displays all tags supported by the mini-model;
[0079] The first selected tag is obtained based on the selection event.
[0080] The present application also provides a hardware structure of the above device. Figure 5 , Figure 5 This is a structural diagram of an electronic device provided in an embodiment of the present application. Figure 5 As shown, the hardware structure may include: a processor and a machine-readable storage medium, the machine-readable storage medium storing machine-executable instructions that can be executed by the processor; the processor is used to execute the machine-executable instructions to implement the method disclosed in the above example of this application.
[0081] Based on the same application concept as the above method, an embodiment of the present application also provides a machine-readable storage medium, on which a number of computer instructions are stored. When the computer instructions are executed by a processor, the method disclosed in the above example of the present application can be implemented.
[0082] Exemplarily, the above-mentioned machine-readable storage medium can be any electronic, magnetic, optical or other physical storage device that can contain or store information, such as executable instructions, data, etc. For example, the machine-readable storage medium can be: RAM (Radom Access Memory), volatile memory, non-volatile memory, flash memory, storage drive (such as hard disk drive), solid state drive, any type of storage disk (such as CD, DVD, etc.), or similar storage medium, or a combination thereof.
[0083] 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 model matching method, characterized in that: The method includes: For each operation node in the business orchestration view corresponding to the current business demand, select a first tag from all tags supported by the small model selected for the operation node; the small model refers to a model whose complexity meets the set small model requirements; Based on the small model and the first label, obtain N large models selected for the operation node that are adapted to the small model and the first label, and at least one target label that is adapted to the first label and supported by each of the N large models; N is greater than or equal to 1; each target label supported by any of the large models is different from the first label; all target labels supported by different large models are different; any of the large models refers to a model whose complexity meets the requirements of the set large model, Determine the operation type of the operation that the small model needs to perform for the first tag; and determine the operation type of the operation that the large model needs to perform based on at least one target tag; The small model is controlled to perform a corresponding operation based on the operation type of the operation to be performed by the first tag, and the large model is controlled to perform a corresponding operation based on the operation type of the operation to be performed by at least one target tag.
2. The method according to claim 1, characterized in that The controlling the small model to perform a corresponding operation based on the operation type of the operation to be performed by the first tag, and the controlling the large model to perform a corresponding operation based on the operation type of the operation to be performed by at least one target tag include: If N is 1, the operation type of the operation to be performed by the first tag is a detection type, and the operation type of the operation to be performed by the large model based on at least one target tag is a filtering type, then: Obtain an operation result obtained by the small model performing a corresponding operation based on the operation type of the operation to be performed by the first tag; input the operation result into the large model, so that the large model performs a corresponding operation on the input data based on at least one target tag to obtain an output result.
3. The method according to claim 1, characterized in that The controlling the small model to perform a corresponding operation based on the operation type of the operation to be performed by the first tag, and the controlling the large model to perform a corresponding operation based on the operation type of the operation to be performed by at least one target tag include: If N is greater than 1, the operation type of the operation to be performed by the first tag is a detection type, the operation type of the operation to be performed by the large model based on at least one target tag is a filtering type, and there is a cascade relationship between the N large models, then: Obtain the operation result obtained by the small model performing the corresponding operation based on the operation type of the operation to be performed by the first tag; input the operation result to the first large model of the cascade to trigger N large models to perform corresponding operations on the input data based on at least one supported target tag; for the second large model of the cascade to the last large model, the input data is the output result of the previous large model of the cascade; Get the output result of the last model in the cascade of N large models.
4. The method according to any one of claims 1 to 3, characterized in that: Any of the large models has a corresponding template graph; the template graph has a corresponding operator table, and the operator table is used to determine the connection relationship and data transfer requirements between the operators in the template graph; If N is greater than 1, there is a cascade relationship between at least two of the N large models; the cascade relationship between any two large models includes: there is a connection relationship between one of the operators in the template graph corresponding to one of the large models and an operator in the template graph corresponding to another large model.
5. The method according to claim 4, characterized in that When there is a connection relationship between two operators, the connection relationship has a corresponding weight; When an operator has a connection relationship with at least two different operators, the connection relationship with the highest weight is preferred.
6. The method according to claim 1, characterized in that The selecting the first label from all labels supported by the small model selected for the node comprises: In response to a selection event in the interface corresponding to the mini-model, the interface displays all tags supported by the mini-model; The first selected tag is obtained based on the selection event.
7. A model matching device, characterized in that: The device includes: A label selection unit, configured to select, for each operation node in the business arrangement view corresponding to the current business demand, a first label from all labels supported by a small model selected for the operation node; the small model refers to a model whose complexity meets the set small model requirement; A model adaptation unit is used to obtain, based on the small model and the first label, N large models selected for the operation node that are adapted to the small model and the first label, and at least one target label that is adapted to the first label and supported by each of the N large models; N is greater than or equal to 1; each target label supported by any of the large models is different from the first label; all target labels supported by different large models are different; any of the large models refers to a model whose complexity meets the requirements of the set large model, An operation unit, configured to determine an operation type of an operation to be performed by the small model for the first tag; and determine an operation type of an operation to be performed by the large model based on at least one target tag; A control unit is used to control the small model to perform a corresponding operation based on the operation type of the operation to be performed by the first tag, and to control the large model to perform a corresponding operation based on the operation type of the operation to be performed by at least one target tag.
8. The device according to claim 7, characterized in that The controlling the small model to perform a corresponding operation based on the operation type of the operation to be performed by the first tag, and the controlling the large model to perform a corresponding operation based on the operation type of the operation to be performed by at least one target tag include: If N is 1, the operation type of the operation to be performed by the first tag is a detection type, and the operation type of the operation to be performed by the large model based on at least one target tag is a filtering type, then: Obtaining an operation result obtained by the small model performing a corresponding operation based on the operation type of the operation to be performed by the first tag; inputting the operation result into the large model, so that the large model performs a corresponding operation on the input data based on at least one target tag to obtain an output result; If N is greater than 1, the operation type of the operation to be performed by the first tag is a detection type, the operation type of the operation to be performed by the large model based on at least one target tag is a filtering type, and there is a cascade relationship between the N large models, then: Obtain the operation result obtained by the small model performing the corresponding operation based on the operation type of the operation to be performed by the first tag; input the operation result to the first large model of the cascade to trigger N large models to perform corresponding operations on the input data based on at least one supported target tag; for the second large model of the cascade to the last large model, the input data is the output result of the previous large model of the cascade; Get the output result of the last model in the cascade of N large models.
9. The device according to claim 7 or 8, characterized in that Any of the large models has a corresponding template graph; the template graph has a corresponding operator table, and the operator table is used to determine the connection relationship and data transfer requirements between the operators in the template graph; If N is greater than 1, there is a cascade relationship between at least two of the N large models; the cascade relationship between any two large models includes: there is a connection relationship between an operator in the template graph corresponding to one of the large models and an operator in the template graph corresponding to another large model; When there is a connection relationship between two operators, the connection relationship has a corresponding weight; When an operator has a connection relationship with at least two different operators, the connection relationship with the highest weight is preferred.
10. An electronic device, characterized in that: The electronic device includes: a processor and a machine-readable storage medium; The machine-readable storage medium stores machine-executable instructions that can be executed by the processor; The processor is used to execute machine executable instructions to implement the method steps of any one of claims 1-6.