Distillation extraction method, device and storage medium of router storage model

Abstract

The application provides a distillation extraction method and device of a router storage model and a storage medium. The method comprises the following steps: a router in an intelligent network obtains a model popularity table corresponding to a plurality of demand types; a predetermined number of models are stored based on the ordering of the model popularity of each demand type in the model popularity table, and a confidence or satisfaction index of the stored model for each demand type is recorded; a non-maximum suppression algorithm is used to select a model with the optimal confidence or satisfaction index in the stored model corresponding to each demand type as the optimal model of the demand type; the demand type of a user is obtained, the optimal model corresponding to the demand type of the user is used for distillation extraction, a user-specific model is obtained, and the user-specific model is sent to the user. The application can provide an optimal model for specific user demand to improve the performance of the intelligent network.

Description

Technical Field

[0001] This invention relates to the field of semantic communication technology, and in particular to a method and apparatus for distillation extraction of router storage models. Background Technology

[0002] Compared to traditional network design concepts that primarily enhance network capabilities through resource stacking, intelligent simplified networks enable the representation and transmission of semantic information. Given the presence of intelligent nodes within these networks, intelligent simplified networks can transform traditional data sharing between nodes into neural network model sharing by highly integrating neural networks and nodes. Simultaneously, the node function shifts from storing data to storing neural network models. Based on this, intelligent simplified networks autonomously learn and update models according to internal and external network demands, achieving cognition and learning based on network-native intelligence, forming a highly autonomous network with swarm intelligence. External demands primarily include user-initiated requirements; internal demands mainly refer to the evolving needs of the intelligent simplified network itself, such as the increasing amount of data being forwarded while hardware resources are limited, potentially requiring lightweight model updates. The characteristic of intelligent simplified networks is the reduced data transmission volume, resulting in higher network throughput at the same transmission rate. This is due to the ability of intelligent simplified networks to fully utilize the computational processing power of intelligent nodes, transmitting not just the information itself, but its semantic features and a deep learning model for semantic decoding. The task of decoding and recovering the transmitted information is then handled by the deep learning model at the receiving end. The core of intelligent simplified network propagation is the propagation model, not the propagated content. In intelligent simplified networks, nodes can store models and share their stored models as needed.

[0003] A smart router, or router in a smart network, is a bridge node for data or model communication within the network. While forwarding data or models, it can read relevant information, retain data or models that can improve its own model performance, increase its own knowledge base, expand the stored dataset, and then train the model through incremental learning, transfer learning, federated learning, and other methods to improve its own model performance, thereby improving the performance of surrounding node models and reducing network model transmission loss.

[0004] However, the current intelligent simplified network only proposes a semantic communication method based on the propagation model, and has not yet proposed a specific implementation plan to meet user needs. If the model stored by the router itself is directly fed back to the user, it will not only lack model specificity, but also affect the performance of the intelligent simplified network. Therefore, how to provide an optimal model for the specific needs of users to improve the performance of the intelligent simplified network, thereby speeding up the processing of user requests, ensuring the accuracy of information transmission, and improving the matching degree between the optimal model and the specific needs of users is a problem to be solved. Summary of the Invention

[0005] In view of this, embodiments of the present invention provide a method and apparatus for distilling and extracting router storage models to provide the best user-specific models for specific user needs, thereby improving the performance of models in intelligent and simplified networks.

[0006] One aspect of the present invention provides a distillation extraction method for a router storage model, the method comprising the following steps:

[0007] In the intelligent and simplified network, the router obtains a model popularity table corresponding to multiple demand types;

[0008] Based on the ranking of model popularity for each demand type in the model popularity table, a predetermined number of models are stored as candidate models, and the confidence or satisfaction index of the candidate models for each demand type is recorded.

[0009] The nonmaximum suppression algorithm is used to rank the candidate models corresponding to each demand type by confidence or satisfaction index. The model with the best confidence or satisfaction index among the candidate models corresponding to each demand type is taken as the best model for that demand type.

[0010] Obtain the user's demand type, perform distillation extraction based on the best model corresponding to the user's demand type to obtain a user-specific model, and send the user-specific model to the user.

[0011] In some embodiments of the present invention, the method further includes: a router in the intelligent simplified network obtaining a demand popularity table based on the demand history of network nodes;

[0012] Based on the model popularity ranking for each demand type in the model popularity table, a predetermined number of models are stored as candidate models, and the confidence or satisfaction index of the candidate models for each demand type is recorded, including:

[0013] Select a set number of demand types based on the popularity ranking of each demand type in the demand popularity table;

[0014] Based on the popularity ranking of each model corresponding to each demand type in the model popularity table, a predetermined number of models are selected as candidate models for the corresponding demand type and stored.

[0015] Record the confidence or satisfaction index of the candidate model for each demand type.

[0016] In some embodiments of the present invention, the predetermined number of candidate models stored by the router is set based on the router's hardware storage resources;

[0017] The types of requirements include: dataset requirements, latency requirements, and recovery accuracy requirements.

[0018] In some embodiments of the present invention, the confidence or satisfaction index of the candidate model for each demand type includes the confidence or satisfaction score of the model performance for each demand type and the confidence or satisfaction score of the model size for each demand type.

[0019] The nonmaximum suppression algorithm was used to rank the candidate models for each demand type based on confidence or satisfaction indices, including:

[0020] Based on resource constraints, the confidence or satisfaction scores of model performance for each demand type and the confidence or satisfaction scores of model size for each demand type are weighted and summed to obtain the candidate model confidence or satisfaction scores for each demand type.

[0021] The nonmaximum suppression algorithm was used to rank the confidence or satisfaction scores of the candidate models corresponding to each demand type.

[0022] In some embodiments of the present invention, the routers in the intelligent simplified network obtain the model popularity tables corresponding to multiple demand types in the following manner:

[0023] In the intelligent and simplified network, the router obtains a model popularity table corresponding to multiple demand types based on the model transmission history of each demand type of its associated network node.

[0024] In a simplified network, routers obtain model popularity tables corresponding to multiple demand types from routers within the communication range; or

[0025] In the intelligent and simplified network, the router obtains model popularity tables corresponding to multiple demand types based on the model transmission history of each demand type of its associated network node and the model popularity table broadcast by the router within the communication range.

[0026] In some embodiments of the present invention, when the user's demand types include two or more demand types, distillation extraction is performed based on the optimal model corresponding to the user's demand type to obtain a user-specific model, including:

[0027] When there are two or more demand types containing different demand types with the same attribute, the router simultaneously distills the best model corresponding to the different demand types with the same attribute, extracts multiple models, distills the multiple models again into a model that satisfies the same attribute for different demand types, and uses the model obtained by the second distillation as the user-specific model.

[0028] When there are two or more demand types containing different demand types with different attributes, the router obtains multiple user-specific models corresponding to the different demand types with different attributes by distilling the optimal models corresponding to the different demand types with different attributes.

[0029] In some embodiments of the present invention, the user's demand type and the user-specific model are broadcast between routers in the intelligent simplified network.

[0030] In some embodiments of the present invention, the user's demand type is obtained, and distillation extraction is performed based on the optimal model corresponding to the user's demand type to obtain a user-specific model, including:

[0031] Obtain the user's need type, and then perform distillation extraction based on the best model corresponding to that user's need type to obtain a user-specific model; or

[0032] Obtain the user's demand type, and simultaneously perform distillation extraction based on the best model corresponding to the user's demand type and at least one suboptimal model to obtain a user-specific model.

[0033] Another aspect of the present invention provides a distillation extraction apparatus for a router storage model, the apparatus comprising: a processor and a memory, the memory storing computer instructions, the processor executing the computer instructions stored in the memory, and the apparatus implementing the steps of the method described in any of the above embodiments when the computer instructions are executed by the processor.

[0034] Another aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of the method described in any of the above embodiments.

[0035] The method and apparatus for distilling and extracting router storage models of the present invention sorts the candidate models stored in the router of the intelligent and simplified network using a non-maximum suppression algorithm, and distills and extracts the best model corresponding to the user's needs to obtain a user-specific model. This can speed up the processing of user requests, ensure the accuracy of information transmission, and improve the matching degree between the best model and the user's specific needs.

[0036] Additional advantages, objects, and features of the invention will be set forth in part in the description which follows, and will also become apparent in part to those skilled in the art upon studying the description, or may be learned by practice of the invention. The objects and other advantages of the invention can be realized and obtained by means of the structures specifically pointed out in the description and drawings.

[0037] Those skilled in the art will understand that the objectives and advantages achievable with the present invention are not limited to those specifically described above, and that the above and other objectives achievable with the present invention will become clearer from the following detailed description. Attached Figure Description

[0038] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, are not intended to limit the scope of the invention. In the drawings:

[0039] Figure 1 This is a schematic flowchart of a distillation extraction method for a router storage model according to an embodiment of the present invention.

[0040] Figure 2 This is a schematic diagram illustrating the specific steps of distillation extraction in a router storage model according to a specific embodiment of the present invention. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the embodiments and accompanying drawings. Here, the illustrative embodiments and descriptions of this invention are used to explain the invention, but are not intended to limit the invention.

[0042] It should also be noted that, in order to avoid obscuring the invention with unnecessary details, only the structures and / or processing steps closely related to the solution according to the invention are shown in the accompanying drawings, while other details that are not closely related to the invention are omitted.

[0043] It should be emphasized that the term "including or comprising" as used herein refers to the presence of a feature, element, step, or component, but does not exclude the presence or addition of one or more other features, elements, steps, or components.

[0044] To simplify the description, unless otherwise specified, the router mentioned below refers to the SmartSimplified Router.

[0045] To address the problems existing in the prior art, this invention provides a model distillation extraction method for router storage models based on model distillation technology. Model distillation is an advantageous method for compressing models. Model distillation aims to use a pre-trained, large, and effective teacher model to guide the training of a lightweight student model with fewer parameters, reducing model size and computational resources while maintaining the accuracy of the student model as close as possible to that of the teacher model. In embodiments of this invention, the optimal model for a user's specific needs is obtained, and distillation extraction technology is used to provide the user with a more targeted and accurate user-specific model, thereby reducing the pressure on the server cloud, reducing the amount of data transmitted during communication, and accelerating the processing speed of user requests. Embodiments of the invention will be described below with reference to the accompanying drawings. In the drawings, the same reference numerals represent the same or similar parts or the same or similar steps.

[0046] Figure 1This is a schematic flowchart of a distillation extraction method for a router storage model according to an embodiment of the present invention. Figure 1 As shown, the method may specifically include steps S110 to S140.

[0047] Step S110: The routers in the intelligent simplified network obtain the model popularity tables corresponding to multiple demand types.

[0048] In some embodiments of the present invention, the routers in the intelligent simplified network obtain the model popularity tables corresponding to multiple demand types in the following manner:

[0049] In a simplified intelligent network, routers obtain multiple model popularity tables corresponding to various demand types based on the model transmission history of their associated network nodes for each demand type; routers in a simplified intelligent network obtain multiple model popularity tables corresponding to various demand types from routers within their communication range; or routers in a simplified intelligent network obtain multiple model popularity tables corresponding to various demand types based on the model transmission history of their associated network nodes for each demand type and the model popularity tables broadcast by routers within their communication range. This improves the accuracy of model popularity tables in a simplified intelligent network.

[0050] To simplify the description, unless otherwise specified, all nodes mentioned below refer to network nodes.

[0051] The broadcast range is defined during router deployment. In other words, each router can store the addresses of other routers within a certain radius of its communication range during deployment. The specific number of addresses to store can be preset. The router's broadcast switching frequency can be set to periodic, and the switching method can follow a defined transmission protocol, i.e., adding a specific frame header, before sending its information to routers within the communication range.

[0052] As an example, routers mainly transmit information through broadcasting, but the present invention is not limited to this and may also include other information transmission methods, which can be set according to actual needs.

[0053] In some embodiments of the present invention, the requirement types may include: dataset requirements, latency requirements, and recovery accuracy requirements. However, the present invention is not limited to these and may also include other requirement types, which can be set based on actual user needs.

[0054] In some embodiments of the present invention, the router in the intelligent simplified network obtains a demand popularity table based on the demand history of network nodes.

[0055] More specifically, in the simplified intelligent network, the network nodes associated with the router report received user-initiated requests to the simplified intelligent router. The router then stores the various request types uploaded by each node as a request history. The request types uploaded by different nodes in the simplified intelligent network are not necessarily the same. The frequency of historical requests is selected as an indicator to measure the popularity of a request type; the more historical requests a request type has, the higher its popularity. The request type popularity table for the router in the simplified intelligent network can be obtained from the request history of its associated network nodes; or from router broadcasts within the communication range; or from both the request history of its associated network nodes and router broadcasts within the communication range. Network nodes are associated with the simplified intelligent router based on spatial distance, and associated network nodes can be pre-configured during router deployment.

[0056] In some embodiments of the present invention, the router in the intelligent simplified network can also obtain a model popularity table based on the demand history of network nodes and the model transmission history corresponding to each demand type.

[0057] Step S120: The intelligent router stores a predetermined number of models as candidate models based on the ranking of model popularity for each demand type in the model popularity table, and records the confidence or satisfaction index of the candidate models for each demand type.

[0058] As an example, the Smart Router selects multiple demand types based on the popularity ranking of demand types in the demand popularity table, and then selects a predetermined number of models corresponding to these demand types for storage based on the model popularity ranking in the model popularity table, and records the confidence or satisfaction index of the stored models for each demand type.

[0059] In some embodiments of the present invention, the intelligent simplified router stores a predetermined number of models as candidate models based on the ranking of model popularity for each demand type in the model popularity table, and records the confidence or satisfaction index of the candidate models for each demand type, including:

[0060] Step S121: The Smart Router selects a set number of demand types based on the popularity ranking of each demand type in the demand popularity table.

[0061] Furthermore, based on the demand popularity table, the Smart Router selects a set number of demand types according to their popularity from high to low.

[0062] Step S122: The intelligent router selects a predetermined number of models as candidate models for each requirement type based on the popularity ranking of each model corresponding to each requirement type in the model popularity table and stores them.

[0063] As an example, the Smart Router selects a demand type from a set number of demand types selected based on the demand popularity table, sorts them by demand popularity, and selects all models corresponding to the demand type in the model popularity table as candidate models for the demand type based on the popularity ranking of the models corresponding to each demand type in the model popularity table. The next demand type is selected from high to low according to the popularity ranking of the selected set number of demand types, and the above steps are repeated until the router stores a predetermined number of models.

[0064] As an example, in the process of storing a predetermined number of models in the Smart Router, if a model in the Model Popularity Table corresponds to any of the required types in the set number of required types, then it is stored as the model for that required type. If they do not correspond, the model is skipped, and the next model in the Model Popularity Table is determined based on its popularity to see if it corresponds to any of the required types in the selected set number of required types. This step is repeated until the predetermined number of stored models is reached.

[0065] As an example, if there is no model in the model popularity table that corresponds to a certain number of demand types, the Smart Router will report the user's demand to the cloud server. The cloud will then retrieve the best model corresponding to that demand type, send the model to the router, and store it.

[0066] In some embodiments of the present invention, the predetermined number of candidate models stored by the router is set based on the router's hardware storage resources.

[0067] Furthermore, the hardware storage resources of the intelligent simplified router include memory and hard disk. Memory is mainly used for forwarding data and data caching during model distillation, while hard disk is mainly responsible for storing models in the network. In addition, routers in different locations in the intelligent simplified network may be equipped with hard disks of different sizes due to different costs or needs. For example, if the hard disk size is 1GB and the popular model size in the network is around 20MB, then it can store about 40 models; if the hard disk is 5GB, it can store more and larger models.

[0068] Step S123: The intelligent router records the confidence or satisfaction index of the candidate model for each demand type.

[0069] In some embodiments of the present invention, the confidence or satisfaction index of the candidate model for each demand type includes the confidence or satisfaction score of model performance for each demand type and the confidence or satisfaction score of model size for each demand type.

[0070] The confidence or satisfaction score is defined by a fixed protocol in the transmission frame header or frame tail. That is, after a model is trained on the cloud server, a confidence or satisfaction test is performed on it, and the score and the model itself are encapsulated together to form a frame and sent to the intelligent simplified network.

[0071] Step S130: The intelligent router uses the non-maximum suppression algorithm to rank the candidate models corresponding to each demand type according to their confidence or satisfaction index, and selects the model with the best confidence or satisfaction index among the candidate models corresponding to each demand type as the best model for that demand type.

[0072] Non-Maximum Suppression (NMS) is based on the idea of ​​suppressing non-maximum values ​​and searching for local maxima.

[0073] In some embodiments of the present invention, the intelligent simplified router uses a non-maximum suppression algorithm to rank the candidate models corresponding to each demand type based on confidence or satisfaction indices, including:

[0074] Due to hardware resource limitations, the Smart Router performs a weighted sum of the confidence or satisfaction scores of model performance for each demand type and the confidence or satisfaction scores of model size for each demand type to obtain the candidate model confidence or satisfaction scores for each demand type.

[0075] The Smart Router uses a nonmaximum suppression algorithm to rank the confidence or satisfaction scores of candidate models for each type of demand.

[0076] More specifically, the Smart Router uses the weighted sum of the confidence or satisfaction scores of the model performance and model size for each demand type as the confidence or satisfaction index of the model corresponding to each demand type. After sorting the confidence or satisfaction indices of the models corresponding to each demand type, the model with the best confidence or satisfaction index (the largest weighted sum of confidence and satisfaction scores) is selected as the best model for that demand type.

[0077] As an example, if the model performance score weight is set to α = 0.6 in the Smart Router, then the model size weight is 1 - α = 0.4. For ease of calculation, all parameters can be normalized, i.e., the range is controlled between 0 and 1. Finally, the models corresponding to each requirement type are sorted according to the score result of α * model performance score + (1 - α) * model size, and the model corresponding to each requirement type with the highest score is selected as the best model for each requirement type.

[0078] Here, the use of model performance and model size confidence or satisfaction scores for each demand type as indicators to measure the confidence or satisfaction of the model corresponding to each demand type is merely an example. However, this invention is not limited to this and can also measure the confidence or satisfaction scores of the model corresponding to each demand type based on other model parameters.

[0079] Step S140: The Smart Router obtains the user's demand type, performs distillation extraction based on the best model corresponding to the user's demand type to obtain a user-specific model, and sends the user-specific model to the user.

[0080] More specifically, users in the intelligent simplified network send requests to the nearest network node, and the node uploads the user requests to the intelligent simplified router. The router matches the user requests with the demand types stored in its database, selects the best model for that demand type while other models for that demand type play an auxiliary role, and obtains a user-specific model that satisfies the user's request through model distillation, and then feeds it back to the user.

[0081] In some embodiments of the present invention, the intelligent simplified router obtains the user's demand type, performs distillation extraction based on the optimal model corresponding to the user's demand type, and obtains a user-specific model, including:

[0082] The intelligent router obtains the user's need type and, based on the optimal model corresponding to that need type, performs distillation extraction to obtain a user-specific model; or

[0083] The intelligent router obtains the user's demand type, and performs distillation extraction based on the best model corresponding to the user's demand type and at least one suboptimal model to obtain a user-specific model.

[0084] In this invention, the suboptimal model refers to other router storage models besides the optimal model that correspond to this type of requirement.

[0085] As an example, the intelligent router performs distillation extraction based on the best model corresponding to the user's need type and at least one second-best model simultaneously to obtain a user-specific model. This can include: the intelligent router selecting only the best model for distillation, or selecting the two models with the highest confidence or satisfaction indices among the stored models corresponding to each need type for simultaneous distillation extraction, or randomly selecting two models from the four models with the highest confidence or satisfaction indices among the stored models corresponding to each need type for simultaneous distillation extraction. Then, based on the scoring model built into the router, the confidence or satisfaction score of the user model obtained using different distillation methods for the user's needs is estimated, and the distilled model with the highest confidence or satisfaction score is selected as the user-specific model. The model distillation methods mentioned above are merely examples, but the invention is not limited to them and can be set based on actual user needs.

[0086] In some embodiments of the present invention, when a user's needs include two or more types of needs, the intelligent simplified router performs distillation extraction based on the optimal model corresponding to the user's needs to obtain a user-specific model, including:

[0087] When two or more demand types contain different demand types with the same attributes, the router simultaneously distills the optimal models corresponding to the different demand types with the same attributes. After extracting multiple models, the intelligent router distills these multiple models again into a single model that satisfies the same attributes across different demand types, and uses this re-distilled model as the user-specific model. For example, if the intelligent router trains teacher model A and teacher model B using two datasets from different domains, it can simultaneously distill a student model that can recognize both datasets, thereby integrating and transferring data from two different domain datasets.

[0088] When there are two or more demand types containing different demand types with different attributes, the Smart Router obtains multiple user-specific models corresponding to different demand types with different attributes by distilling the optimal models corresponding to the different demand types with different attributes.

[0089] As an example, the attributes of the demand type may include: targeting the same dataset, accuracy, and latency, etc., but the present invention is not limited to this. It may also include other attributes of the demand type or determine whether a user-specific model can be obtained by re-distillation based on whether the basic architecture of the best model corresponding to different user demand types is the same.

[0090] In some embodiments of the present invention, user demand types and user-specific models are broadcast between routers in the intelligent simplified network.

[0091] As an example, if multiple nodes simultaneously or at short intervals send the same user request to different routers, these routers do not all need to perform model distillation. They can simply select the router with the best hardware configuration or the most available hardware resources, and the other routers can obtain the distilled model from that responding router. During this process, the routers can transmit information via broadcast.

[0092] As an example, the information transmitted between routers may include: demand history or demand popularity table, model transmission history or model popularity table corresponding to each demand type, demand type and the model corresponding to each demand type, but the present invention is not limited to this and may also include other information.

[0093] The following is a specific example to describe the distillation extraction method of the router storage model, such as... Figure 2 As shown, the specific steps of this method are as follows:

[0094] 1) A smart router 1 located in the smart network stores the demand history of associated nodes and the model forwarding and transmission history of each demand type. It can exchange historical data with a router (smart router 2) within its communication range to obtain a demand popularity table and a model popularity table for each demand type. Based on the popularity table, and following the principle of prioritizing the storage of models corresponding to demand types with high popularity, the smart router selects a predetermined number of demand types based on demand popularity. Then, based on model popularity and the predetermined number of demand types, it stores a predetermined number of models to local router 1. The stored models are classified according to demand type, and the confidence or satisfaction index of the stored models for each demand type is recorded.

[0095] 2) Using the nonmaximum suppression method, the intelligent router 1 sorts the models according to the confidence scores of each demand type, retains and stores the model with the best confidence score for each demand type as the best model for that demand type, for subsequent distillation calls.

[0096] 3) If a user A in the Smart Simplified Network is not satisfied with the recovery accuracy of their local model (dataset 1), they can request a model with higher recovery accuracy from the network. After user A triggers the priority recovery accuracy request for dataset 1 and it is acquired by Smart Simplified Router 1, Router 1 selects the best model (model 2) corresponding to the user's recovery accuracy request from the stored models, performs distillation extraction to obtain model 2-mini, and then sends the lightweight model (model 2-mini) that meets user A's request to user A.

[0097] Based on the router storage model distillation extraction method described above, this invention utilizes the ability of a smart router to store models and employs a non-maximum suppression algorithm to rank models corresponding to each demand type. For each specific user demand, the optimal model corresponding to that demand type is selected, while other models corresponding to that demand type play an auxiliary role. Through model distillation, a user-specific model is obtained and fed back to the user. This invention provides favorable support for accelerating user request processing, ensuring the accuracy of transmitted information, and improving the matching degree between the optimal model and the user's specific needs.

[0098] Compared to random storage models, models selected for router storage based on demand popularity and model popularity are more likely to match the best model that meets user needs. Through model distillation, larger models are fed back to users as lightweight models while ensuring model accuracy, effectively reducing the amount of data in the communication process and speeding up the processing of user requests. By using nonmaximum suppression methods to rank the storage models based on the confidence or satisfaction index of each demand type, the matching degree between the best model and the user's specific needs is improved.

[0099] Corresponding to the above method, the present invention also provides a distillation extraction apparatus for a router storage model. The apparatus includes a computer device, which includes a processor and a memory. The memory stores computer instructions, and the processor is used to execute the computer instructions stored in the memory. When the computer instructions are executed by the processor, the apparatus performs the steps of the method described above.

[0100] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the aforementioned edge computing server deployment method. The computer-readable storage medium can be a tangible storage medium, such as random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, floppy disks, hard disks, removable storage disks, CD-ROMs, or any other form of storage medium known in the art.

[0101] Those skilled in the art will understand that the exemplary components, systems, and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Whether implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention. When implemented in hardware, it can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this invention are programs or code segments used to perform the desired tasks. The programs or code segments can be stored in a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried in a carrier wave.

[0102] It should be clarified that the present invention is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of the present invention.

[0103] In this invention, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with or in place of features of other embodiments.

[0104] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations of the embodiments of the present invention are possible. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A distillation extraction method of a router storage model, characterized by, The method comprises the following steps: The router in the smart network obtains a model popularity table corresponding to a plurality of demand types; A predetermined number of models are stored as candidate models based on the ranking of the model popularity of each demand type in the model popularity table, and a confidence or satisfaction index of the candidate models for each demand type is recorded; A non-maximum suppression algorithm is used to sort the confidence or satisfaction index of the candidate models corresponding to each demand type, and the model with the optimal confidence or satisfaction index in the candidate models corresponding to each demand type is taken as the optimal model of the demand type; The demand type of the user is obtained, and the optimal model corresponding to the demand type of the user is distilled and extracted to obtain a user-specific model, and the user-specific model is sent to the user.

2. The method of claim 1, wherein, The method further comprises: the router in the smart network obtaining a demand popularity table according to the demand history of the network node; The storing of a predetermined number of models as candidate models based on the ranking of the model popularity of each demand type in the model popularity table, and the recording of a confidence or satisfaction index of the candidate models for each demand type, comprises: A set number of demand types are selected based on the popularity ranking of each demand type in the demand popularity table; A predetermined number of models are selected as candidate models corresponding to each selected demand type based on the popularity ranking of each model corresponding to the demand type in the model popularity table, and the candidate models are stored; The confidence or satisfaction index of the candidate models for each demand type is recorded.

3. The method of claim 1, wherein, The predetermined number of candidate models stored by the router is set based on the hardware storage resources of the router; The demand types include: dataset demand, latency demand, and recovery accuracy demand.

4. The method of claim 1, wherein, The confidence or satisfaction index of the candidate models for each demand type includes a confidence or satisfaction score of model performance for each demand type and a confidence or satisfaction score of model size for each demand type; The sorting of the confidence or satisfaction index of the candidate models corresponding to each demand type by using a non-maximum suppression algorithm comprises: Based on resource constraints, the confidence or satisfaction score of model performance for each demand type and the confidence or satisfaction score of model size for each demand type are weighted and summed to obtain a candidate model confidence or satisfaction score corresponding to each demand type; The candidate model confidence or satisfaction score corresponding to each demand type is sorted by using a non-maximum suppression algorithm.

5. The method of claim 1, wherein, The router in the smart network obtains a plurality of model popularity tables corresponding to a plurality of demand types in the following manner: The router in the smart network obtains a plurality of model popularity tables corresponding to a plurality of demand types based on the model transmission history of each demand type of the network node associated with the router; The router in the smart network obtains a plurality of model popularity tables corresponding to a plurality of demand types from the routers within the communication range; Or The router in the smart network obtains a plurality of model popularity tables corresponding to a plurality of demand types based on the model transmission history of each demand type of the network node associated with the router and the model popularity tables broadcast by the routers within the communication range.

6. The method of claim 3, wherein, In the case that the demand type of the user includes two or more demand types, the distillation of the optimal model corresponding to the demand type of the user is performed to obtain a user-specific model, including: In the case that the two or more demand types include different demand types with the same attribute, the router simultaneously distills the respective optimal models corresponding to the different demand types with the same attribute to obtain a plurality of models, and the obtained plurality of models are distilled again into a model satisfying the different demand types with the same attribute, and the model obtained by the distillation again is taken as the user-specific model; In the case that the two or more demand types include different demand types with different attributes, the router obtains a plurality of user-specific models corresponding to the different demand types with different attributes by distilling the respective optimal models corresponding to the different demand types with different attributes.

7. The method of claim 1, wherein, The demand type of the user and the user-specific model are broadcasted between the routers in the wisdom simple network.

8. The method of claim 1, wherein, The method for obtaining the demand type of the user and the user-specific model obtained by distillation of the optimal model corresponding to the demand type of the user includes: The demand type of the user is obtained, and a user-specific model is obtained by distillation of the optimal model corresponding to the demand type of the user alone; or The demand type of the user is obtained, and a user-specific model is obtained by distillation of the optimal model corresponding to the demand type of the user and at least one suboptimal model simultaneously.

9. A distillation extraction device of a router storage model, comprising a processor and a memory, characterized in that, The memory stores computer instructions, and the processor is configured to execute the computer instructions stored in the memory, so that the device implements the steps of the method according to any one of claims 1 to 8.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to implement the steps of the method according to any one of claims 1 to 8.

Images