A method and system for adapting a deep learning framework based on edge computing of a signal creation host

Abstract

The application provides a method and system for adapting a deep learning framework for a Xiongmao host based on edge computing, wherein the method comprises: converting an original deep learning model in the deep learning framework into a format supported by the Xiongmao host by an adaptation layer in the Xiongmao host to obtain a second deep learning model; performing runtime optimization on the second deep learning model by the adaptation layer to obtain a third deep learning model; deploying the third deep learning model on the Xiongmao host; collecting data from a data source by an edge computing node, and preprocessing the collected data to generate data conforming to an input data format of the third deep learning model; receiving the preprocessed data from the edge computing node by the Xiongmao host, inputting the data into the third deep learning model for inference, and generating a prediction result. The application can realize efficient adaptation of the Xiongmao host to the deep learning framework and improve the application capability of the Xiongmao host in an edge computing environment.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and specifically to a method and system for adapting a domestically developed host computer based on edge computing to a deep learning framework. Background Technology

[0002] In edge computing scenarios, such as smart security, industrial quality inspection, and intelligent transportation, systems typically need to meet requirements for low latency, high bandwidth, and real-time performance. These scenarios rely on deep learning models to perform localized inference tasks, such as real-time object detection from cameras and anomaly identification from sensor data. The successful execution of these tasks depends on the model loading and inference capabilities of deep learning frameworks.

[0003] However, with the advancement of national information technology application innovation (IT innovation), the application of IT innovation-based mainframes in China is gradually increasing. IT innovation-based mainframes typically use domestically produced hardware architectures and operating systems, which cannot be directly adapted to existing deep learning frameworks, thus limiting their application in edge computing scenarios.

[0004] Current edge computing and deep learning technologies mainly focus on data acquisition, preprocessing, and hardware acceleration. However, a comprehensive solution is still lacking in how to systematically adapt domestically developed servers to run deep learning frameworks. This lack of adaptation hinders the widespread application of domestically developed servers in edge computing scenarios, especially in applications requiring high-efficiency inference capabilities.

[0005] Therefore, developing a systematic solution to enable deep learning frameworks to run efficiently on domestically developed servers is of great significance for promoting the application of domestically developed servers in edge computing scenarios. This will not only improve the performance and efficiency of domestically developed servers, but also meet the stringent requirements of edge computing scenarios for real-time performance and low latency. Summary of the Invention

[0006] This invention provides a system and method for adapting domestically developed IT hosts to deep learning frameworks based on edge computing, which can achieve efficient adaptation of domestically developed IT hosts to deep learning frameworks and improve their application capabilities in edge computing environments.

[0007] In a first aspect, this application provides a method for adapting a domestically developed IT host to a deep learning framework based on edge computing, comprising the following steps:

[0008] S1. Deep Learning Model Conversion: The adaptation layer in the domestically developed host converts the first deep learning model into a format supported by the domestically developed host. The conversion includes deep learning model structure conversion and parameter adaptation to obtain the second deep learning model. The first deep learning model is the original deep learning model in the deep learning framework.

[0009] S2. Runtime Optimization: The adaptation layer performs runtime optimization on the second deep learning model to obtain the third deep learning model; runtime optimization includes computation graph optimization, memory management optimization, and parallel computing optimization.

[0010] S3. Deployment of deep learning models: Deploy the third deep learning model on the domestically developed host;

[0011] S4. Data Acquisition and Preprocessing: Edge computing nodes collect data from data sources and preprocess the collected data to generate data that conforms to the input data format of the third deep learning model.

[0012] S5. Deep Learning Model Inference: The domestically developed host receives preprocessed data from the edge computing node, inputs it into the third deep learning model for inference, and generates prediction results.

[0013] In one possible implementation, the method further includes: S6, result feedback: the domestically developed host feeds back the inference result to the edge computing node for further processing or direct output by the edge computing node.

[0014] In one possible implementation, in step S1, the deep learning model structure conversion includes: first, parsing the structure of the first deep learning model, which includes network layers, connection relationships, and activation functions; second, mapping the parsed structure of the first deep learning model to a format supported by the domestically developed host, specifically converting some layers or operations in the first deep learning model into equivalent layers or operations available on the domestically developed host.

[0015] In one possible implementation, in step S1, parameter adaptation includes: first, extracting model parameters from the first deep learning model, the parameters including weights and biases; second, converting the extracted model parameters into a format supported by the domestically developed host; and third, optimizing the converted model parameters to achieve lightweighting of the first deep learning model.

[0016] In one possible implementation, step S1 further includes model verification, which includes: after the second deep learning model is converted, performing functional verification on it by running test data on the domestically developed host and comparing the results with the output of the first deep learning model running the same test data to ensure that the behavior of the second deep learning model on the domestically developed host is consistent with that of the first deep learning model; verifying the performance of the second deep learning model on the domestically developed host, including inference speed and resource consumption, to ensure that the second deep learning model meets the requirements of practical applications.

[0017] In one possible implementation, in step S2: computation graph optimization optimizes the computation graph of the second deep learning model, specifically by merging some operations in the computation graph, eliminating redundant computations, or using pruning techniques to remove unimportant connections or neurons in the second deep learning model; memory management optimization optimizes the memory usage of the second deep learning model; and parallel computing optimization utilizes the multi-core processor of the domestically developed host for parallel computing by decomposing the task of the second deep learning model into multiple subtasks and executing them in parallel on multiple processor cores.

[0018] In one possible implementation, the adaptation layer supports conversion of multiple first deep learning models and multiple runtime optimizations for second deep learning models.

[0019] One possible implementation also includes a dynamic resource allocation step: during the training and inference phases of the third deep learning model, the adaptation layer dynamically allocates computing resources based on the hardware resources of the domestically developed host and the real-time running status of the model.

[0020] One possible implementation also includes an adaptive learning mechanism step: introducing an adaptive learning mechanism to adaptively optimize the third deep learning model to improve its robustness in dynamic environments; the adaptive learning mechanism in the training phase includes: dynamically adjusting the learning rate according to training progress and changes in data distribution, and / or using an adaptive optimization algorithm to automatically adjust the learning rate of each parameter according to gradient changes, and / or dynamically optimizing the network structure according to changes in data distribution; the adaptive learning mechanism in the inference phase includes: continuously learning from new data, fine-tuning model parameters to adapt to changes in data distribution, and / or dynamically adjusting the model's compression and quantization strategies according to resource constraints of the inference environment and changes in data distribution to maintain inference efficiency and accuracy.

[0021] Secondly, this application provides a system for adapting a domestically developed host to a deep learning framework based on edge computing, including a domestically developed host and edge computing nodes. The domestically developed host runs a domestically developed operating system and hardware architecture, serving as a computing node in the edge computing environment. An adaptation layer is deployed on the domestically developed host, which is used to implement the deep learning model conversion and runtime optimization in the above method. The edge computing nodes are deployed close to the data source and are used to perform data acquisition and preprocessing in the above method. The domestically developed host and the edge computing nodes work together to realize the operation of the deep learning framework on the domestically developed host, and complete the deployment, inference, and result feedback of the deep learning model.

[0022] This application achieves efficient operation of deep learning frameworks on domestically developed servers through the collaboration of the domestically developed server adaptation layer and edge computing nodes, and has the following beneficial effects:

[0023] 1. Solve compatibility challenges and expand application scenarios for domestically developed mainframes. Overcome the limitations of domestically developed mainframe hardware architecture (such as the domestic ARM architecture) and domestic operating systems. Through an adaptation layer, complete the structural conversion and parameter adaptation of deep learning models, enabling mainstream deep learning frameworks (such as TensorFlow and PyTorch) that were previously unable to run directly to be implemented on domestically developed mainframes. Fill the technical gap in adapting domestically developed mainframes to deep learning frameworks in edge computing scenarios, enabling domestically developed mainframes to be applied to scenarios with high requirements for low latency and real-time performance, such as intelligent security, industrial quality inspection, and intelligent transportation.

[0024] 2. Optimize operational performance and improve edge computing efficiency. By lightweighting models (such as parameter quantization and pruning) and optimizing runtime (computation graph merging, parallel computing, and dynamic memory management), the computational load and memory consumption of the domestically developed host are reduced, improving model inference speed and meeting the real-time requirements of edge scenarios. A dynamic resource allocation and adaptive learning mechanism is introduced, which can adjust resource and model strategies according to changes in the hardware load and data distribution of the domestically developed host, further improving resource utilization and avoiding hardware overload or idleness.

[0025] 3. Meets security requirements and ensures data autonomy and controllability. The domestically developed host itself uses domestically produced hardware and systems. When paired with the adaptation solution of this application, it does not rely on foreign software and hardware components, thus meeting the security requirements of the National Information Technology Application Innovation (Domestic Innovation) Program. Edge computing nodes collect and preprocess data locally, while the domestically developed host performs inference locally, reducing data transmission links, lowering the risk of data leakage, and ensuring data security and autonomy and controllability in edge scenarios.

[0026] 4. It possesses flexible scalability and adapts to diverse business needs. The adaptation layer supports the conversion of various original deep learning models and is compatible with multiple runtime optimization methods. Different deep learning frameworks and optimization strategies can be flexibly selected according to business scenarios (such as object detection and anomaly recognition). In the system architecture, the domestically developed host and edge computing nodes work collaboratively. The number of nodes can be flexibly expanded according to the number of data sources and the complexity of inference tasks, adapting to edge computing services of different scales. Attached Figure Description

[0027] Figure 1 This is a flowchart of a method in one embodiment of this application. Detailed Implementation

[0028] To enable those skilled in the art to better understand the present application, the technical solution of the present application will be further described in detail below with reference to the embodiments and accompanying drawings.

[0029] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this application described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0030] In this embodiment of the disclosure, the character " / " indicates that the objects before and after it are in an "or" relationship. For example, A / B means: A or B.

[0031] The term "and / or" describes an association between objects, indicating that three relationships can exist. For example, A and / or B means: A or B, or A and B.

[0032] Specific embodiments according to this application will now be described with reference to the accompanying drawings.

[0033] Example 1:

[0034] This application provides a method for adapting a domestically developed IT host to a deep learning framework based on edge computing, including:

[0035] S1. Deep Learning Model Conversion: The adaptation layer in the domestically developed host converts the first deep learning model into a format supported by the domestically developed host. The conversion includes deep learning model structure conversion and parameter adaptation to obtain the second deep learning model. The first deep learning model refers to the original deep learning model in the deep learning framework.

[0036] In some embodiments, deep learning model structure transformation includes: firstly, parsing the structure of a first deep learning model, including network layers, connections, activation functions, etc.; then mapping the parsed structure to a format supported by the domestically developed host. For example, converting certain layers or operations into equivalent operations available on the domestically developed host. This could involve converting complex convolutional layers into a series of simple convolutional layers, or converting certain activation functions into activation functions supported by the domestically developed host.

[0037] In some embodiments, parameter adaptation includes: first, extracting the model parameters, including weights, biases, etc., from the first deep learning model; then, converting the extracted parameters into a format supported by the domestically developed host. For example, this may involve converting the parameter data type (e.g., from 32-bit floating-point numbers to 16-bit or 8-bit integers) or rearranging the parameters. Finally, optimizing the parameters, for example, by using quantization techniques to convert high-precision floating-point data (e.g., weights) into low-precision data (e.g., 8-bit integers), thereby achieving a lightweight deep learning model, reducing the computational load and storage requirements of the model, and simultaneously reducing memory usage.

[0038] In some embodiments, the deep learning model conversion further includes: model verification.

[0039] In some embodiments, the model verification includes: after the transformation is completed, performing functional verification on the second deep learning model to ensure that its behavior on the domestically developed host is consistent with that of the first deep learning model. For example, this is achieved by running some test data on the domestically developed host and comparing it with the output of the first deep learning model.

[0040] In some embodiments, the model verification further includes: verifying the performance of the second deep learning model on the domestically developed host, including inference speed, resource consumption, etc., to ensure that it meets the needs of practical applications.

[0041] S2. Runtime optimization: The adaptation layer performs runtime optimization on the second deep learning model to obtain the third deep learning model; the runtime optimization includes computation graph optimization, memory management optimization and parallel computing optimization.

[0042] Computational graph optimization refers to optimizing the computational graph of the second deep learning model to improve computational efficiency. Examples include merging certain operations and eliminating redundant computations. For instance, pruning techniques can be used to remove unimportant connections or neurons from the second deep learning model to reduce model complexity and computational cost, while also reducing memory usage. Another example is using knowledge distillation techniques to train a small model to mimic the behavior of a larger model, thereby reducing model size and computational cost.

[0043] Memory management optimization refers to optimizing the memory usage of the second deep learning model and reducing memory consumption. On one hand, a dynamic memory pool mechanism can be used to uniformly manage model memory allocation and deallocation. First, based on the hardware characteristics of the domestically developed host, the memory pool is initialized and partitioned. Combining the cache characteristics of the ARM architecture CPU in the domestically developed host (such as the 64MB L3 cache of the Phytium FT-2000+ / 64), the memory pool is divided into multiple partitions, such as a high-frequency small memory area (1KB-64KB, adapted to the frequently used convolutional kernel weights and bias parameters in the model), a mid-frequency memory area (64KB-4MB, adapted to feature map tensors), and a low-frequency large memory area (4MB and above, adapted to the model's input and output tensors). Each partition pre-allocates memory according to the memory block size and usage frequency to avoid fragmented allocation. Second, a memory allocation and reuse mechanism is used to allocate and match memory blocks on demand. When the third deep learning model is loaded, the adaptation layer intercepts the memory allocation requests from the deep learning framework. Based on the requested memory size, it matches the corresponding free memory block in the partition. For example, when requesting 32KB of memory, it allocates directly from the high-frequency, small memory area, avoiding frequent memory requests to the operating system and reducing system call overhead. For tensors of the same size that appear repeatedly during model inference (such as the output feature map of a convolutional layer being the same size as the input feature map of the next layer), the adaptation layer records the memory block usage trajectory. After inference, the memory block is not placed, but marked as reusable. The next time a tensor of the same size is requested, this memory block is directly reused to improve memory reuse and reduce redundant memory allocation. On the other hand, parameter loading and caching optimization mechanisms can be adopted. For large-sized models (such as the Vision Transformer model exceeding 2GB), the adaptation layer does not load all weights into memory at once, but loads them in batches according to the inference process (such as loading the feature extraction layer weights first, placing them after inference, and then loading the classification layer weights), reducing peak memory usage. Frequently accessed weight parameters (such as convolutional layer kernel weights) are preferentially stored in the L3 cache of the domestically developed host CPU (such as the 64MB L3 cache of Phytium CPU). The cache pages are locked through the cache management submodule of the adaptation layer to prevent them from being preempted by other processes, thereby reducing the weight read latency and reducing memory bandwidth usage.

[0044] Parallel computing optimization refers to using the multi-core processors of domestically developed mainframes for parallel computing to improve computation speed. For example, the task of the second deep learning model is decomposed into multiple subtasks and executed in parallel on multiple processor cores.

[0045] In some embodiments, the adaptation layer supports multiple first deep learning model conversions and multiple second deep learning model runtime optimizations.

[0046] In some embodiments, the deep learning framework includes, but is not limited to, mainstream deep learning frameworks such as TensorFlow and PyTorch.

[0047] In some embodiments, deep learning models can be converted and optimized using Open Neural Network Exchange (ONNX), TensorRT, and TVM. ONNX is an open format for representing deep learning models. The adapter layer can utilize ONNX to convert models into a format supported by domestically developed host systems. NVIDIA's TensorRT is a high-performance deep learning inference optimizer that can be used to optimize and deploy deep learning models. Apache TVM is an open-source deep learning compiler stack that can be used for cross-platform model optimization and deployment.

[0048] S3. Deployment of deep learning models: Deploy the third deep learning model on the domestically developed host;

[0049] S4. Data Acquisition and Preprocessing: Edge computing nodes collect data from data sources and preprocess the collected data to generate data that conforms to the input data format of the third deep learning model.

[0050] S5. Deep Learning Model Inference: The domestically developed host receives preprocessed data from the edge computing node, inputs it into the third deep learning model for inference, and generates prediction results.

[0051] In some embodiments, the method further includes: S6, result feedback: the domestically developed host feeds back the inference result to the edge computing node for further processing or direct output by the edge computing node.

[0052] In some embodiments, the method further includes: dynamically allocating resources through an adaptation layer during the training and inference phases of the third deep learning model.

[0053] In some embodiments, the dynamic resource allocation includes: dynamically allocating hardware resources based on the hardware resources of the domestically developed host and the real-time running status of the model, improving resource utilization, and ensuring the efficiency and stability of the deep learning model's operation. For example, real-time monitoring of the hardware resources of the domestically developed host, including CPU utilization, memory utilization, storage space, network bandwidth, etc.; monitoring the running status of the deep learning model, including inference time, throughput, latency, etc.; analyzing the model's resource requirements, including computational load, memory usage, I / O requirements, etc.; determining task priorities based on the real-time nature and importance of the tasks; dynamically allocating tasks based on the load of each computing node to avoid overloading any node; reserving certain hardware resources for high-priority tasks to ensure their performance and real-time performance; and dynamically increasing or decreasing resources based on the model's resource requirements.

[0054] In some embodiments, an adaptive learning mechanism is introduced to adaptively optimize the third deep learning model, thereby improving the robustness of the third deep learning model in dynamic environments.

[0055] In some embodiments, the adaptive learning mechanism includes: dynamically adjusting the learning rate during the training phase based on training progress and changes in data distribution to accelerate convergence or avoid overfitting; using adaptive optimization algorithms such as Adam and RMSprop to automatically adjust the learning rate of each parameter based on gradient changes; and dynamically and automatically optimizing the network structure based on changes in data distribution, such as adding or removing layers or adjusting the number of neurons.

[0056] In some embodiments, the adaptive learning mechanism includes: during the model inference phase, continuing to learn from new data and fine-tuning model parameters to adapt to changes in data distribution; and dynamically adjusting the model's compression and quantization strategies based on resource constraints of the inference environment and changes in data distribution to maintain inference efficiency and accuracy.

[0057] Example 2:

[0058] This application provides a system for adapting a domestically developed host to a deep learning framework based on edge computing, including a domestically developed host and edge computing nodes. The system uses the method described in Embodiment 1 to enable the domestically developed host to adapt to the deep learning framework for inference.

[0059] In this application, the domestically developed host computer runs a domestically produced operating system and hardware architecture, serving as a computing node in an edge computing environment. An adaptation layer is deployed on the domestically developed host computer. This adaptation layer is used to adapt the domestically developed host computer to deep learning frameworks, enabling deep learning model conversion, runtime optimization, and resource management. Edge computing nodes are deployed close to the data source and are responsible for data acquisition and preprocessing. All parts of this application work together to achieve efficient operation of deep learning frameworks on the domestically developed host computer.

[0060] Taking an intelligent security monitoring system as an example, edge computing nodes are deployed near surveillance cameras to collect video stream data and perform preprocessing. The adaptation layer converts the TensorFlow-trained object detection deep learning model into a format supported by the domestically developed host computer, resulting in a second deep learning model. The adaptation layer then performs runtime optimization on the second deep learning model to obtain a third deep learning model. This third deep learning model is deployed on the domestically developed host computer, and the preprocessed video stream data is input into the second deep learning model for real-time inference to detect target objects within the monitored area. The inference results are fed back to the edge computing nodes for further analysis or direct output of alarm information.

[0061] This application can achieve the following technical effects:

[0062] Highly efficient adaptation: It solves the problem that domestically developed servers cannot be directly adapted to deep learning frameworks, and improves the application capabilities of domestically developed servers in edge computing environments.

[0063] Performance optimization: By lightweighting the deep learning model and optimizing runtime, the inference speed and resource utilization of the domestically developed host have been improved.

[0064] Security: The use of domestically developed servers complies with national information security requirements, improving data security and independent controllability in edge computing environments.

[0065] Flexibility: The adaptation layer supports multiple deep learning frameworks, which can be flexibly selected and switched according to actual needs.

[0066] It should be understood that the numbers S1 to S6 above are only used to distinguish and facilitate the expression of different steps, and do not necessarily constitute a restriction on the execution order between the steps.

[0067] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for adapting a domestically developed IT host to a deep learning framework based on edge computing, characterized in that, Includes the following steps: S1. Deep learning model conversion: The adaptation layer in the domestically developed host converts the first deep learning model into a format supported by the domestically developed host. The conversion includes deep learning model structure conversion and parameter adaptation to obtain a second deep learning model. The first deep learning model is the original deep learning model in the deep learning framework. S2. Runtime Optimization: The adaptation layer performs runtime optimization on the second deep learning model to obtain the third deep learning model. Runtime optimization includes computation graph optimization, memory management optimization, and parallel computing optimization. Memory management optimization optimizes the memory usage of the second deep learning model to reduce memory consumption, including two aspects: First, a dynamic memory pool mechanism is used to uniformly manage model memory allocation and reclamation. Firstly, based on the hardware characteristics of the domestically developed host, the memory pool is initialized and partitioned. Combining the cache characteristics of the ARM architecture CPU of the domestically developed host, the memory pool is divided into multiple partitions, and each partition pre-allocates memory according to memory block size and usage frequency. Secondly, memory allocation and reuse are implemented. The mechanism allocates and matches memory blocks on demand: When the third deep learning model is loaded, the adaptation layer intercepts the memory allocation request of the deep learning framework and matches the free memory block of the corresponding partition according to the requested memory size; for tensors of the same size that appear repeatedly during the model inference process, the adaptation layer records the memory block usage trajectory, and does not place the memory block after inference is completed, but marks it as reusable, so that the memory block can be reused directly when the same size tensor is allocated next time; on the other hand, a parameter loading and caching optimization mechanism is adopted. For large-sized models, the adaptation layer loads the model weights into memory in batches according to the inference process; the frequently accessed weight parameters are preferentially stored in the L3 cache of the domestically developed host CPU, and the cache page is locked through the cache management submodule of the adaptation layer; S3. Deployment of deep learning model: Deploy the third deep learning model on the domestically developed host; S4. Data Acquisition and Preprocessing: Edge computing nodes acquire data from data sources and preprocess the acquired data to generate data that conforms to the input data format of the third deep learning model. S5. Deep learning model inference: The domestically developed host receives preprocessed data from the edge computing node, inputs it into the third deep learning model for inference, and generates prediction results.

2. The method according to claim 1, characterized in that, The method further includes: S6, result feedback: the domestically developed host feeds back the inference result to the edge computing node for further processing or direct output by the edge computing node.

3. The method according to claim 1, characterized in that, In step S1, the deep learning model structure conversion includes: first, parsing the structure of the first deep learning model, the structure including network layers, connection relationships and activation functions; second, mapping the parsed structure of the first deep learning model to the format supported by the domestically developed host, specifically converting some layers or operations in the first deep learning model into equivalent layers or equivalent operations available on the domestically developed host.

4. The method according to claim 1, characterized in that, In step S1, the parameter adaptation includes: first, extracting model parameters from the first deep learning model, the parameters including weights and biases; second, converting the extracted model parameters into a format supported by the domestically developed host; and third, optimizing the converted model parameters to achieve lightweighting of the first deep learning model.

5. The method according to claim 1, characterized in that, Step S1 further includes model verification, which includes: after the second deep learning model is converted, performing functional verification on it by running test data on the domestically developed host and comparing the running results with the output results of the first deep learning model running the same test data, to ensure that the behavior of the second deep learning model on the domestically developed host is consistent with that of the first deep learning model; verifying the performance of the second deep learning model on the domestically developed host, the performance including inference speed and resource consumption, to ensure that the second deep learning model meets the actual application requirements.

6. The method according to claim 1, characterized in that, In step S2: the computation graph optimization is to optimize the computation graph of the second deep learning model, specifically by merging some operations in the computation graph, eliminating redundant computations, or using pruning techniques to remove unimportant connections or neurons in the second deep learning model; the parallel computing optimization is to use the multi-core processor of the domestically developed host to perform parallel computing, by decomposing the task of the second deep learning model into multiple sub-tasks and executing them in parallel on multiple processor cores.

7. The method according to claim 1, characterized in that, The adaptation layer supports various conversions of the first deep learning model and various runtime optimizations for the second deep learning model.

8. The method according to claim 1, characterized in that, It also includes a dynamic resource allocation step: during the training and inference phases of the third deep learning model, the adaptation layer dynamically allocates computing resources based on the hardware resources of the domestically developed host and the real-time running status of the model.

9. The method according to claim 1, characterized in that, It also includes an adaptive learning mechanism step: introducing an adaptive learning mechanism to adaptively optimize the third deep learning model in order to improve the robustness of the third deep learning model in dynamic environments; The adaptive learning mechanism includes the following during the training phase: dynamically adjusting the learning rate based on training progress and changes in data distribution, and / or using an adaptive optimization algorithm to automatically adjust the learning rate of each parameter based on gradient changes, and / or dynamically optimizing the network structure based on changes in data distribution; the adaptive learning mechanism includes the following during the inference phase: continuously learning from new data, fine-tuning model parameters to adapt to changes in data distribution, and / or dynamically adjusting the model's compression and quantization strategies based on resource constraints of the inference environment and changes in data distribution to maintain inference efficiency and accuracy.

10. A system for adapting a deep learning framework to a domestically developed IT host based on edge computing, characterized in that, The system includes a domestically developed host computer and edge computing nodes. The domestically developed host computer runs a domestically developed operating system and hardware architecture, serving as a computing node in the edge computing environment. An adaptation layer is deployed on the domestically developed host computer, which is used to implement the deep learning model conversion and runtime optimization in any one of the methods described in claims 1-9. The edge computing nodes are deployed close to the data source and are used to perform data acquisition and preprocessing in any one of the methods described in claims 1-9. The domestically developed host computer and the edge computing nodes work together to enable the deep learning framework to run on the domestically developed host computer, and to complete the deployment, inference, and result feedback of the deep learning model.

Images