Model edge deployment method and system, storage medium and electronic equipment
By using progressive optimization and web assembly technology, highly adaptable target intelligent models are generated, solving the bottlenecks in deployment efficiency and running performance in edge computing, and realizing efficient, cross-platform intelligent model deployment and fast loading.
Patent Information
- Application Number
- CN202511724255.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-23
- Publication Date
- 2026-02-17
AI Technical Summary
In existing technologies, the deployment efficiency and operating performance of intelligent models in edge computing scenarios are limited by the storage and loading requirements of large container images, resulting in increased hardware costs and operational burdens.
The initial intelligent model is optimized through a progressive optimization strategy. Combined with dynamic dual-source fusion and adaptive weight adjustment, a target intelligent model is generated. Web assembly technology is then used to convert it into a format highly adapted to edge devices, which can then be deployed directly on different devices.
It improves the adaptability and performance of intelligent models in edge computing scenarios, solves compatibility issues, and achieves efficient, cross-platform unified deployment and fast loading, meeting low latency requirements.
Smart Images
Figure CN121541893A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of model deployment technology, and more specifically, to a model edge deployment method, system, storage medium, and electronic device. Background Technology
[0002] In today's digital age, intelligent models are widely used in numerous fields. Traditional intelligent models are typically trained and deployed on cloud servers. With the rise of the Internet of Things (IoT), more and more devices are connecting to the network, and the amount of data is growing explosively. The traditional centralized cloud processing model is gradually revealing some limitations, thus giving rise to edge computing.
[0003] Edge computing is an architecture that pushes computing power and data storage to the network edge, aiming to reduce data transmission latency, alleviate the burden on cloud servers, and improve the real-time performance and reliability of the system. Current technologies deploy intelligent models and their runtime environments by packaging them into container images using containerization solutions. Because container images need to include all dependencies, configuration files, and some operating system components required for model operation, the images are quite large. When deployed on resource-constrained edge devices, storing and loading these large images consumes significant amounts of storage space and memory, increasing hardware costs and operational burdens, thus reducing the deployment efficiency and performance of intelligent models in edge computing scenarios.
[0004] Therefore, how to improve the deployment efficiency and operating performance of intelligent models in edge computing scenarios is an urgent problem to be solved in this application. Summary of the Invention
[0005] In view of this, this application discloses a model edge deployment method, system, storage medium and electronic device, aiming to improve the deployment efficiency and operating performance of target intelligent models in edge computing scenarios.
[0006] To achieve the above objectives, the disclosed technical solution is as follows:
[0007] The first aspect of this application discloses a method for deploying models at the edge, the method comprising:
[0008] Acquire standardized datasets and real-time data streams uploaded by edge devices to be deployed;
[0009] The standardized dataset and the data stream are fused using a dynamic dual-source fusion mechanism;
[0010] The fused data is dynamically corrected based on the characteristics of the edge devices to be deployed.
[0011] Accuracy testing is performed on the edge devices to be deployed, and the test results are obtained.
[0012] If the detection results meet the training conditions, the dynamically corrected data will be used to train the initial intelligent model.
[0013] The target intelligent model is obtained by optimizing the initial intelligent model after training through a progressive optimization strategy.
[0014] The target intelligent model is compiled and converted using web assembly technology;
[0015] The compiled and transformed target intelligent model is deployed at the edge on the edge device to be deployed.
[0016] Preferably, the accuracy test of the edge device to be deployed, to obtain the test result, includes:
[0017] Obtain the data labeling accuracy of the edge devices to be deployed;
[0018] During continuous training cycles, the accuracy of the data annotation is compared with a preset accuracy to complete the accuracy detection process and obtain the detection result.
[0019] Preferably, if the detection result meets the training conditions, training the initial intelligent model with the dynamically corrected data includes:
[0020] If the detection result indicates that the data labeling accuracy is greater than the preset accuracy within a continuous training period, the detection result is determined to meet the training conditions.
[0021] If the detection results meet the training conditions, activate the mixed precision training mode.
[0022] In mixed-precision training mode, the initial intelligent model is trained using dynamically corrected data.
[0023] Preferably, if the detection result does not meet the training conditions, the loss weights are adjusted through an adaptive weight adjustment mechanism, and the initial intelligent model is trained and adjusted using the adjusted loss weights, including:
[0024] If the detection result indicates that the data labeling accuracy is less than or equal to the preset accuracy within a continuous training period, it is determined that the detection result does not meet the training conditions.
[0025] If the detection result does not meet the training conditions, the loss weight is increased to a preset multiple of the benchmark value to obtain the adjusted loss weight.
[0026] The adjusted loss weights guide the initial intelligent model to shift towards edge scene features, thus completing the model training adjustment process.
[0027] Preferably, the step of optimizing the initial intelligent model after training using a progressive optimization strategy to obtain the target intelligent model includes:
[0028] Based on the preset weight storage method and preset calculation precision, the training of the initial intelligent model after training is accelerated.
[0029] During the training acceleration process, the changes in the L2 norm of each layer are tracked in real time to obtain the gradient values of the convolutional layers.
[0030] When the gradient value of the convolutional layer meets the preset conditions, the structure pruning algorithm is activated, and the contributing filter is dynamically removed through the structure pruning algorithm.
[0031] Within a predetermined number of layers in the backbone network, adaptive reasoning for device perception is performed by learning gating mechanisms and backbone prediction results to complete the process of optimizing the initial intelligent model after training, thereby obtaining the target intelligent model.
[0032] Preferred options also include:
[0033] If the detection results do not meet the training conditions, the loss weights are adjusted through an adaptive weight adjustment mechanism, and the initial intelligent model is trained and adjusted using the adjusted loss weights.
[0034] Preferably, the step of deploying the compiled and converted target intelligent model on the edge device to be deployed includes:
[0035] Design security mechanisms and perform lifecycle verification on the compiled and converted target intelligent model;
[0036] The designed and validated target intelligent model is deployed at the edge on the edge device to be deployed.
[0037] A second aspect of this application discloses a model edge deployment system, the system comprising:
[0038] The acquisition unit is used to acquire standardized datasets and real-time data streams uploaded by edge devices to be deployed.
[0039] The fusion unit is used to fuse the standardized dataset and the data stream through a dynamic dual-source fusion mechanism;
[0040] The dynamic correction unit is used to dynamically correct the fused data according to the characteristics of the edge device to be deployed;
[0041] The detection unit is used to perform accuracy testing on the edge devices to be deployed and obtain the detection results.
[0042] The training unit is used to train the initial intelligent model with dynamically corrected data if the detection result meets the training conditions.
[0043] The optimization unit is used to optimize the initial intelligent model after training using a progressive optimization strategy to obtain the target intelligent model.
[0044] The compilation and conversion unit is used to compile and convert the target intelligent model using web assembly technology;
[0045] The edge deployment unit is used to deploy the compiled and transformed target intelligent model on the edge device to be deployed.
[0046] A third aspect of this application discloses a storage medium comprising stored instructions, wherein, when the instructions are executed, the device in which the storage medium resides executes the model edge deployment method as described in any one of the first aspects.
[0047] The fourth aspect of this application discloses an electronic device including a memory and one or more instructions, wherein one or more instructions are stored in the memory and configured to be executed by one or more processors using the model edge deployment method as described in any of the first aspects.
[0048] As can be seen from the above technical solution, this application discloses a model edge deployment method, system, storage medium, and electronic device. It acquires a standardized dataset and a real-time data stream uploaded by the edge device to be deployed. The standardized dataset and the data stream are fused through a dynamic dual-source fusion mechanism. The fused data is dynamically corrected according to the characteristics of the edge device to be deployed. Accuracy testing is performed on the edge device to be deployed to obtain the detection result. If the detection result meets the training conditions, the dynamically corrected data is used to train an initial intelligent model. A progressive optimization strategy is used to optimize the trained initial intelligent model to obtain a target intelligent model. The target intelligent model is compiled and converted using web assembly technology. The compiled and converted target intelligent model is then deployed on the edge device to be deployed.
[0049] This solution employs a progressive optimization strategy to optimize the initial trained intelligent model, resulting in the target intelligent model. This progressive optimization strategy enables device-aware adaptive inference, further enhancing the adaptability and performance of the target intelligent model on diverse edge devices and improving its operational performance in edge computing scenarios. Furthermore, web assembly technology precisely converts the target intelligent model into a dedicated format highly compatible with the edge device runtime environment, ensuring stable operation on various devices and effectively resolving compatibility issues without requiring container image packaging. The compiled and converted target intelligent model can be directly and uniformly deployed across different edge devices. Web assembly is a high-efficiency binary instruction format that runs in web browsers, offering advantages such as high performance, cross-platform compatibility, and security. It enables unified deployment and efficient execution across different edge devices, with fast loading speeds, thereby improving model deployment efficiency and meeting low-latency requirements. Attached Figure Description
[0050] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0051] Figure 1 This is a flowchart illustrating a model edge deployment method disclosed in an embodiment of this application;
[0052] Figure 2 This is a schematic diagram of the data processing disclosed in the embodiments of this application;
[0053] Figure 3 This is an architecture diagram of the target intelligent model disclosed in the embodiments of this application;
[0054] Figure 4 This is a schematic diagram illustrating the training of the target intelligent model disclosed in an embodiment of this application;
[0055] Figure 5 This is a schematic diagram illustrating the conversion and compilation process of the target intelligent model disclosed in the embodiments of this application;
[0056] Figure 6 This is a schematic diagram of the structure of a model edge deployment system disclosed in an embodiment of this application;
[0057] Figure 7 This is a schematic diagram of the structure of the electronic device disclosed in the embodiments of this application. Detailed Implementation
[0058] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0059] In this application, the terms "comprising," "including," or any other variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0060] As the background technology shows, container images need to contain all the dependencies, configuration files and some operating system components required for the model to run, resulting in a large image size. When deployed on resource-constrained edge devices, storing and loading these large images will consume a lot of storage space and memory, thereby increasing the hardware cost and maintenance burden of the device, which in turn reduces the deployment efficiency and running performance of intelligent models in edge computing scenarios.
[0061] To address the aforementioned issues, this application discloses a model edge deployment method, system, storage medium, and electronic device. This solution employs a progressive optimization strategy to optimize the initial trained intelligent model, yielding a target intelligent model. This progressive optimization strategy enables device-aware adaptive inference, further enhancing the adaptability and performance of the target intelligent model on diverse edge devices and improving its operational performance in edge computing scenarios. Furthermore, web assembly technology precisely converts the target intelligent model into a dedicated format highly compatible with the edge device runtime environment, ensuring stable operation on various devices and effectively resolving compatibility issues without requiring container image packaging. The compiled and converted target intelligent model can be directly and uniformly deployed across different edge devices. Web assembly is a high-efficiency binary instruction format that can run in web browsers, offering advantages such as high performance, cross-platform compatibility, and security. It enables unified deployment and efficient execution across different edge devices, with fast loading speeds, thereby improving model deployment efficiency and meeting low-latency requirements. Specific implementation details are provided in the following embodiments.
[0062] It should be noted that the model edge deployment method, system, storage medium and electronic device provided in this application can be used in the technical field of model deployment, etc. The above is only an example and does not limit the application field of the model edge deployment method, system, storage medium and electronic device provided in this application.
[0063] refer to Figure 1 The diagram shown is a flowchart illustrating a model edge deployment method disclosed in an embodiment of this application. The model edge deployment method mainly includes the following steps:
[0064] S101: Obtain standardized datasets and real-time data streams uploaded by edge devices to be deployed.
[0065] The standardized basic dataset is a collection of basic data that has been standardized in terms of data elements, coding rules, quality rules, and other dimensions.
[0066] Edge devices refer to physical devices with computing, storage, and communication capabilities, including but not limited to sensors, smartphones, and routers.
[0067] S102: The standardized dataset and data stream are fused through a dynamic dual-source fusion mechanism.
[0068] A dynamic dual-source fusion mechanism is adopted to jointly train the model using a standardized base dataset and real-time data streams uploaded from edge devices. The base layer of the model loads a standardized dataset in a unified format, while the real-time layer receives data streams from edge devices via a message queue, enabling real-time data acquisition and processing.
[0069] The dynamic dual-source fusion mechanism is used to fuse standardized basic datasets with real-time data streams from edge devices, and introduces an adaptive weight adjustment mechanism and data augmentation transformation to improve the model's generalization ability and robustness.
[0070] S103: Dynamically correct the fused data based on the characteristics of the edge devices to be deployed.
[0071] During operation, the stream processing engine dynamically adjusts its settings based on the characteristics of different edge devices to be deployed, ensuring the accuracy and consistency of the data input.
[0072] S104: Perform accuracy testing on the edge devices to be deployed and obtain the test results.
[0073] In S104, the data annotation accuracy of the edge device to be deployed is obtained. During the continuous training cycle, the data annotation accuracy is compared with the preset accuracy to complete the accuracy detection process and obtain the detection result.
[0074] The continuous training cycle can be 3 hours, 4 hours, etc., and the continuous training cycle can be set according to the actual situation. This application does not make specific limitations.
[0075] The preset accuracy rate can be 95%, 97%, etc. The preset accuracy rate can be set according to the actual situation, and this application does not make specific limitations.
[0076] S105: If the detection results meet the training conditions, the dynamically corrected data will be used to train the initial intelligent model.
[0077] Specifically, if the detection results meet the training conditions, the process of training the initial intelligent model with the dynamically corrected data is shown in A1-A3.
[0078] A1: If the test result indicates that the data labeling accuracy is greater than the preset accuracy within a continuous training period, the test result is determined to meet the training conditions.
[0079] For example, when the data labeling accuracy of a certain type of edge device is greater than 95% within 3 hours of continuous training, the detection results are determined to meet the training conditions.
[0080] A2: If the test results meet the training conditions, enable the mixed precision training mode.
[0081] A3: In mixed precision training mode, the dynamically corrected data is used to train the initial intelligent model.
[0082] If the detection results do not meet the training conditions, the loss weights are adjusted through an adaptive weight adjustment mechanism, and the initial intelligent model is then trained and adjusted using the adjusted loss weights. See details in B1-B3.
[0083] B1: If the test result indicates that the data labeling accuracy is less than or equal to the preset accuracy within a continuous training period, it is determined that the test result does not meet the training conditions.
[0084] For example, if the data labeling accuracy of a certain type of edge device is less than or equal to 95% within 3 hours of continuous training, the detection result is determined to be inconsistent with the training conditions.
[0085] B2: If the detection results do not meet the training conditions, increase the loss weight to a preset multiple of the baseline value to obtain the adjusted loss weight.
[0086] B3: The adjusted loss weights guide the initial intelligent model to shift towards edge scene features, thus completing the model training adjustment process.
[0087] An adaptive weight adjustment mechanism is introduced. For example, when the data labeling accuracy of a certain type of edge device exceeds 95% in a continuous training cycle, the loss weight of that type of data is automatically increased to 1.3 times the benchmark value (the multiple is set according to the actual situation, and this application does not make a specific limit), guiding the model to shift towards high-frequency edge scene features.
[0088] In addition, the system also incorporates data augmentation transformations such as vibration waveform and phase noise, tailored to the characteristics of the equipment, to further enhance the model's generalization ability and robustness. The overall data processing flow is as follows: Figure 2 As shown.
[0089] Figure 2 In terms of data processing, this solution employs a dynamic dual-source fusion mechanism, jointly training the standardized base dataset with real-time data streams uploaded from edge devices. The base layer loads a standardized dataset in a unified format, while the real-time layer accesses the data streams from edge devices via a message queue, enabling real-time data acquisition and processing. During operation, the stream processing engine dynamically adjusts based on the characteristics of different devices to ensure the accuracy and consistency of the input data.
[0090] Meanwhile, the system introduces an adaptive weight adjustment mechanism. When the data labeling accuracy of a certain type of edge device exceeds 95% in consecutive training cycles, the loss weight of that type of data is automatically increased to 1.3 times the baseline value, guiding the model to shift towards high-frequency edge scene features. In addition, the system also injects data augmentation transformations such as vibration waveforms and phase noise to further improve the model's generalization ability and robustness, tailored to device characteristics.
[0091] S106: The initial intelligent model after training is optimized using a progressive optimization strategy to obtain the target intelligent model.
[0092] The progressive model optimization strategy refers to using a mixed precision training mode during the training process, combining gradient monitoring and structural pruning algorithms, and introducing a lightweight device encoder module to improve model adaptability and performance.
[0093] A target intelligence model is a program model with intelligent judgment and decision-making capabilities. By learning from large amounts of data, it can make corresponding predictions or decisions on new input data, such as identifying objects in an image or predicting future trends.
[0094] The specific process of obtaining the target intelligent model is shown in C1-C4.
[0095] C1: Accelerate the training of the initial intelligent model after training based on the preset weight storage method and preset calculation precision.
[0096] The preset weight storage method includes, but is not limited to, FP16 weight storage. In this application, FP16 weight storage is preferred as the preset weight storage method.
[0097] The preset calculation precision includes, but is not limited to, TF32 calculation precision. In this application, TF32 calculation precision is preferred as the preset calculation precision.
[0098] It should be noted that, in terms of model architecture, this solution implements a progressive optimization strategy. First, during the training process, a mixed precision training mode is enabled, which combines FP16 weight storage with TF32 calculation precision to effectively reduce memory usage and accelerate the training process.
[0099] C2: During the training acceleration process, the changes in the L2 norm of each layer are tracked in real time to obtain the gradient values of the convolutional layers.
[0100] C3: When the gradient value of the convolutional layer meets the preset conditions, the structure pruning algorithm is activated, and the contributing filter is dynamically removed through the structure pruning algorithm.
[0101] The preset condition can be set to three consecutive epochs with a value below 10⁻. 5 The specific preset conditions should be set according to the actual situation; this application does not impose specific limitations.
[0102] When the gradient value of the convolutional layer is below 10⁻ for three consecutive epochs 5 At that time, it is determined that the gradient value of the convolutional layer meets the preset conditions.
[0103] The system incorporates a gradient monitoring module to track changes in the L2 norm of each layer in real time. When the gradient value of a convolutional layer is detected to be below 10⁻ for three consecutive epochs, the module will detect the change. 5 At that time, the structure pruning algorithm is automatically activated to dynamically remove low-contribution filters, and the maximum pruning intensity can reach 30%.
[0104] C4: In the preset number of layers of the backbone network, adaptive reasoning for device perception is performed by learning the gating mechanism and the backbone prediction results, so as to complete the process of optimizing the initial intelligent model after training and obtain the target intelligent model.
[0105] The preset layer number is the second to last layer.
[0106] A lightweight device encoder module is introduced in the penultimate layer of the backbone network. This module takes the feature parameters of the edge device as input and outputs a 32-dimensional feature vector. It then fuses the feature vector with the backbone prediction results through a learnable gating mechanism to achieve adaptive inference for device perception. This further improves the adaptability and performance of the target intelligent model on diverse edge devices. A schematic diagram of the target intelligent model's architecture is shown below. Figure 3 As shown.
[0107] Figure 3 In the training phase of the target intelligent model, this solution constructs a heterogeneous training cluster in the cloud, equipped with high-performance hardware (such as Graphics Processing Units (GPUs)) and software to support large-scale parallel training tasks. It integrates various acceleration tools to improve training speed and inference performance, ensuring efficient inference and decision-making for the model. To adapt to the diversity of edge device hardware architectures, virtualization components are pre-installed in the training image, supporting dynamic simulation of multiple instruction set architectures.
[0108] Furthermore, the system maps pre-trained model weights and real-time data streams uploaded from edge devices to a unified namespace via shared storage volumes, forming an efficient data channel and providing fundamental support for subsequent incremental training and continuous learning. The training architecture of the target intelligent model is as follows: Figure 4 As shown.
[0109] Figure 4 In the training process of the target intelligent model, the system triggers an edge device verification process every 50 training rounds, performing end-to-end performance testing in a simulated edge device environment. If the inference latency exceeds a preset threshold, an exponentially increasing penalty term is immediately injected into the loss function, and a power constraint component is activated. During backpropagation, a power gradient bias is superimposed, guiding the model optimization towards lower latency and lower power consumption. Through this real-time feedback-driven mechanism, the system can dynamically adjust the model structure and parameters during training, ensuring that the final generated model has good real-time performance and energy efficiency on edge devices.
[0110] S107: Compile and convert the target intelligent model using web assembly technology.
[0111] In S107, the target intelligent model is exported according to a standardized format, and the standardized target intelligent model is compiled and converted using web assembly technology.
[0112] Web assembly is a programming technique that enables specific programs to run efficiently on multiple devices. It acts like a universal "adapter," allowing different types of devices to better understand and execute programs quickly.
[0113] Web assembly technology is essentially a low-level assembly-like language represented in binary format. It is highly platform-independent, and can run on almost all operating systems and devices as long as there is a corresponding web assembly runtime environment. This makes it an ideal solution to the compatibility issues caused by the diversity of edge devices.
[0114] Existing container images are large, consuming significant space and resources for storage and loading, leading to high hardware costs, heavy operational burdens, and long container startup times, failing to meet the low-latency requirements of edge scenarios. This solution utilizes web assembly technology to compile standardized target intelligent models into web assembly format, eliminating the need for container image packaging. It enables unified deployment and efficient execution across different edge devices, resulting in faster loading speeds, significantly improved deployment efficiency, and compliance with low-latency requirements.
[0115] After the target intelligent model is trained, the system exports it in a standardized format and compiles it using web assembly technology. This compilation is adapted to different edge device operating environments, ensuring stable operation of the target intelligent model on various devices and effectively solving compatibility issues. The conversion process generates platform-independent low-level bytecode. The compiler configures a linear memory space pre-allocation strategy based on different edge device architectures and automatically detects operator features in the intermediate representation. If vectorizable operations are found, computationally intensive operations such as convolution are converted into parallel instruction sequences, significantly improving the model's execution efficiency.
[0116] Furthermore, by converting the target intelligent model into web assembly format, updates only require retraining and replacing the old module, eliminating the need to rebuild the container image, thus simplifying maintenance and update processes and improving maintenance efficiency. Compiling the target intelligent model into web assembly format enables efficient adaptation to the runtime environment of edge devices, improving deployment efficiency and compatibility.
[0117] By converting the target intelligent model into web assembly format, the model can be easily updated. Simply retrain the model and recompile it into a web assembly module, then distribute it to edge devices to replace the old module, simplifying model maintenance and updates.
[0118] Furthermore, the system performs fixed-point conversion on the generated bytecode, replacing floating-point operation instructions with integer opcodes, and combines this with a quantization parameter calibration algorithm to generate a low-entropy data distribution to optimize subsequent compression efficiency. Simultaneously, the module's import table is updated to dynamically link it with the processor-specific acceleration library, ensuring accurate invocation of the device's native instruction set at runtime. Based on the dependency graph, the system dynamically constructs a tight coupling mechanism between the hardware and the runtime environment, further enhancing the model's security and portability. The overall process is as follows: Figure 5 As shown. Figure 5 A schematic diagram illustrating the conversion and compilation process of the target intelligent model is shown.
[0119] Figure 5After the target intelligent model is trained, the system exports it in a standardized format and compiles it using web assembly technology. During this conversion, platform-independent low-level bytecode is generated. The compiler configures a linear memory space pre-allocation strategy based on different edge device architectures and automatically detects operator features in the intermediate representation. If vectorizable operations are found, computationally intensive operations such as convolution are converted into parallel instruction sequences, significantly improving the model's execution efficiency. Furthermore, the system performs fixed-point conversion on the generated bytecode, replacing floating-point instructions with integer opcodes, and combines quantization parameter calibration algorithms to generate a low-entropy data distribution to optimize subsequent compression efficiency. Simultaneously, the module's import table is updated to dynamically link with the processor-specific acceleration library, ensuring accurate invocation of the device's native instruction set at runtime. Based on the dependency graph, the system dynamically constructs a tight coupling mechanism between hardware and the runtime environment, further enhancing the security and portability of the target intelligent model.
[0120] S108: Deploy the compiled and transformed target intelligent model on the edge device to be deployed.
[0121] In S108, a security mechanism is designed and lifecycle verification is performed on the compiled and transformed target intelligent model. The designed and verified target intelligent model is then deployed on the edge device to be deployed.
[0122] Edge deployment involves placing applications or models on network edge devices close to the data source or user. For example, deploying models on terminal devices such as mobile phones or computers to implement intelligent functions can reduce data transmission latency and improve response speed.
[0123] This solution, based on web assembly technology, is an intelligent model edge deployment method designed to effectively address key issues encountered by traditional intelligent models during edge deployment, such as poor compatibility, performance optimization difficulties, and inconvenient updates and maintenance. By utilizing web assembly technology, traditional backend intelligent models are precisely converted into a dedicated format highly adapted to the edge device runtime environment. Efficient network communication protocols enable accurate model distribution, reliable loading, and rapid initialization, while supporting dynamic model replacement and seamless reloading. This significantly improves the deployment efficiency, operational performance, and maintenance convenience of the target intelligent model in edge computing scenarios.
[0124] In this embodiment, a progressive optimization strategy is used to optimize the initial intelligent model after training to obtain the target intelligent model. This progressive optimization strategy enables device-aware adaptive inference, further improving the adaptability and performance of the target intelligent model on diverse edge devices and enhancing its operational performance in edge computing scenarios. Furthermore, web assembly technology is used to accurately convert the target intelligent model into a dedicated format highly compatible with the edge device runtime environment, ensuring stable operation of the target intelligent model on various devices and effectively solving compatibility issues without requiring container image packaging. The compiled and converted target intelligent model can be directly and uniformly deployed on different edge devices. Web assembly is a high-efficiency binary instruction format that can run in a web browser, offering advantages such as high performance, cross-platform compatibility, and security. It enables unified deployment and efficient execution on different edge devices, with fast loading speeds, thereby improving model deployment efficiency and meeting low-latency requirements.
[0125] Based on the above embodiments Figure 1 The disclosed method for deploying models at the edge also includes a corresponding system for deploying models at the edge, as described in this application. Figure 6 As shown, the edge deployment system of this model includes:
[0126] Acquisition unit 601 is used to acquire standardized datasets and real-time data streams uploaded by edge devices to be deployed;
[0127] The fusion unit 602 is used to fuse the standardized dataset and the data stream through a dynamic dual-source fusion mechanism;
[0128] The dynamic correction unit 603 is used to dynamically correct the fused data according to the characteristics of the edge device to be deployed;
[0129] The detection unit 604 is used to perform accuracy detection on the edge device to be deployed and obtain the detection results;
[0130] The training unit 605 is used to train the initial intelligent model with dynamically corrected data if the detection results meet the training conditions.
[0131] The optimization unit 606 is used to optimize the initial intelligent model after training through a progressive optimization strategy to obtain the target intelligent model.
[0132] The compilation and conversion unit 607 is used to compile and convert the target intelligent model using web assembly technology;
[0133] Edge deployment unit 608 is used to deploy the compiled and transformed target intelligent model on the edge device to be deployed.
[0134] Furthermore, the detection unit 604 includes:
[0135] The acquisition module is used to obtain the data labeling accuracy of the edge devices to be deployed;
[0136] The comparison module is used to compare the data annotation accuracy with a preset accuracy during continuous training cycles to complete the accuracy detection process and obtain the detection result.
[0137] Furthermore, training unit 605 includes:
[0138] The first determining module is used to determine that the detection result meets the training conditions if the detection result indicates that the data labeling accuracy rate is greater than the preset accuracy rate in the continuous training period.
[0139] The enable module is used to enable the mixed precision training mode when the detection results meet the training conditions.
[0140] The training module is used to train the initial intelligent model with dynamically corrected data in mixed-precision training mode.
[0141] Furthermore, the model edge deployment system also includes:
[0142] The adjustment unit is used to adjust the loss weights through an adaptive weight adjustment mechanism if the detection results do not meet the training conditions, and then use the adjusted loss weights to train and adjust the initial intelligent model.
[0143] Furthermore, the adjustment unit includes:
[0144] The second determining module is used to determine that the detection result does not meet the training conditions if the detection result indicates that the data labeling accuracy rate is less than or equal to the preset accuracy rate within a continuous training period.
[0145] The enhancement module is used to increase the loss weight to a preset multiple of the baseline value when the detection results do not meet the training conditions, thus obtaining the adjusted loss weight.
[0146] The offset module is used to guide the initial intelligent model to offset towards edge scene features through the adjusted loss weights, thereby completing the model training adjustment process.
[0147] Furthermore, the optimization unit 606 includes:
[0148] The training acceleration module is used to accelerate the training of the initial intelligent model after training, based on the preset weight storage method and preset calculation precision.
[0149] The real-time tracking module is used to track the changes in the L2 norm of each layer in real time during the training acceleration process, and obtain the gradient values of the convolutional layers.
[0150] The activation removal module is used to activate the structure pruning algorithm when the gradient value of the convolutional layer meets the preset conditions, and to dynamically remove the contributing filter through the structure pruning algorithm.
[0151] The adaptive inference module is used to perform adaptive inference for device perception in a preset number of layers of the backbone network by learning the gating mechanism and the backbone prediction results, so as to complete the process of optimizing the initial intelligent model after training and obtain the target intelligent model.
[0152] Furthermore, the edge deployment unit 608 includes:
[0153] The design and verification module is used to design security mechanisms and verify the lifecycle of the compiled and converted target intelligent model.
[0154] The edge deployment module is used to deploy the designed and verified target intelligent model on the edge device to be deployed.
[0155] In this embodiment, a progressive optimization strategy is used to optimize the initial intelligent model after training to obtain the target intelligent model. This progressive optimization strategy enables device-aware adaptive inference, further improving the adaptability and performance of the target intelligent model on diverse edge devices and enhancing its operational performance in edge computing scenarios. Furthermore, web assembly technology is used to accurately convert the target intelligent model into a dedicated format highly compatible with the edge device runtime environment, ensuring stable operation of the target intelligent model on various devices and effectively solving compatibility issues without requiring container image packaging. The compiled and converted target intelligent model can be directly and uniformly deployed on different edge devices. Web assembly is a high-efficiency binary instruction format that can run in a web browser, offering advantages such as high performance, cross-platform compatibility, and security. It enables unified deployment and efficient execution on different edge devices, with fast loading speeds, thereby improving model deployment efficiency and meeting low-latency requirements.
[0156] This application embodiment also provides a storage medium, the storage medium including stored instructions, wherein, when the instructions are executed, the device where the storage medium is located is controlled to perform the model edge deployment method as described above.
[0157] This application also provides an electronic device, the structural schematic diagram of which is shown below. Figure 7 As shown, it specifically includes a memory 701 and one or more instructions 702, wherein one or more instructions 702 are stored in the memory 701 and configured to be executed by one or more processors 703 to perform the above-described model edge deployment method.
[0158] For the foregoing method embodiments, in order to simplify the description, they are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, because according to this application, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to this application.
[0159] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For system-type embodiments, since they are basically similar to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.
[0160] The steps in the methods of the various embodiments of this application can be adjusted, combined, or deleted according to actual needs.
[0161] Finally, it should be noted that in this paper, relational terms such as first and second are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations.
[0162] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
[0163] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.
Claims
1. A model edge deployment method, characterized by, The method comprises: acquiring a standardized data set and a data stream uploaded in real time by an edge device to be deployed; fusing the standardized data set and the data stream through a dynamic dual-source fusion mechanism; performing dynamic correction on the fused data according to the characteristics of the edge device to be deployed; performing accuracy detection on the edge device to be deployed to obtain a detection result; training an initial intelligent model using the dynamically corrected data if the detection result meets a training condition; optimizing the trained initial intelligent model through a progressive optimization strategy to obtain a target intelligent model; compiling and converting the target intelligent model through a web assembly technology; performing edge deployment of the compiled and converted target intelligent model on the edge device to be deployed.
2. The method of claim 1, wherein, The accuracy detection on the edge device to be deployed to obtain a detection result comprises: acquiring a data labeling accuracy of the edge device to be deployed; comparing the data labeling accuracy with a preset accuracy in a continuous training period to complete the accuracy detection process and obtain a detection result.
3. The method of claim 2, wherein, The training of the initial intelligent model using the dynamically corrected data if the detection result meets a training condition comprises: determining that the detection result meets the training condition if the detection result indicates that the data labeling accuracy is greater than the preset accuracy in the continuous training period; starting a mixed precision training mode if the detection result meets the training condition; training the initial intelligent model using the dynamically corrected data in the mixed precision training mode.
4. The method of claim 2, wherein, Further comprising: adjusting a loss weight through an adaptive weight adjustment mechanism and adjusting the model training of the initial intelligent model using the adjusted loss weight if the detection result does not meet the training condition.
5. The method of claim 4, wherein, The adjustment of the loss weight through the adaptive weight adjustment mechanism and the model training adjustment of the initial intelligent model using the adjusted loss weight if the detection result does not meet the training condition comprises: determining that the detection result does not meet the training condition if the detection result indicates that the data labeling accuracy is less than or equal to the preset accuracy in the continuous training period; increasing the loss weight to a baseline value preset multiple to obtain an adjusted loss weight if the detection result does not meet the training condition; guiding the initial intelligent model to deviate towards edge scene characteristics through the adjusted loss weight to complete the model training adjustment process.
6. The method of claim 1, wherein, The model optimization of the trained initial intelligent model through the progressive optimization strategy to obtain a target intelligent model comprises: training the trained initial intelligent model according to a preset weight storage method and a preset calculation precision; tracking the L2 norm changes of each layer in real time to obtain a convolution layer gradient value during the training acceleration process; activating a structure pruning algorithm and dynamically removing a contribution filter through the structure pruning algorithm when the convolution layer gradient value meets a preset condition; performing adaptive inference of device perception through a learning gate mechanism and a backbone prediction result in a preset number of layers of the backbone network to complete the model optimization of the trained initial intelligent model to obtain a target intelligent model.
7. The method of claim 1, wherein, The edge deployment of the target intelligent model after the compiling conversion includes: The target intelligent model after the compiling conversion is subjected to security mechanism design and life cycle verification; The target intelligent model after the design and verification is subjected to edge deployment on the edge device to be deployed.
8. A model edge deployment system, characterized by, The system includes: An acquisition unit is configured to acquire a standardized data set and a data stream uploaded by an edge device to be deployed in real time; A fusion unit is configured to fuse the standardized data set and the data stream through a dynamic double-source fusion mechanism; A dynamic correction unit is configured to dynamically correct the fused data according to characteristics of the edge device to be deployed; A detection unit is configured to detect accuracy of the edge device to be deployed to obtain a detection result; A training unit is configured to train an initial intelligent model with the data after the dynamic correction if the detection result meets a training condition; An optimization unit is configured to optimize the initial intelligent model after the training through a progressive optimization strategy to obtain a target intelligent model; A compiling conversion unit is configured to compile and convert the target intelligent model through a web assembly technology; An edge deployment unit is configured to perform edge deployment of the target intelligent model after the compiling conversion on the edge device to be deployed.
9. A storage medium, characterized by The storage medium includes stored instructions, wherein the instructions, when executed, control a device where the storage medium is located to perform the model edge deployment method according to any one of claims 1 to 7.
10. An electronic device, comprising: The device includes a memory and one or more instructions, wherein the one or more instructions are stored in the memory and configured to be executed by one or more processors to perform the model edge deployment method according to any one of claims 1 to 7.