An edge computing model compression acceleration method in an unstable network environment

Abstract

The application discloses a model compression acceleration method for edge computing in an unstable network environment, and belongs to the technical field of edge computing and deep learning model optimization. The method performs multi-strategy hybrid pruning of artificial experience pruning, regularization pruning and reinforcement learning intelligent search pruning fusion on an original intelligent model in sequence to obtain a set of light-weight intelligent models. Soft label and hard label joint knowledge distillation is performed on the light-weight models as students and the original model as a teacher to realize secondary compression of the model. INT8 low-bit-width quantization and precision compensation are combined with the hardware characteristics of a domestic MLU200 deep learning processor to complete parallel acceleration optimization and deployment to an edge embedded device. The adaptive adjustment module is used to realize real-time sensing of network bandwidth, automatically reduce data transmission and model calculation when the bandwidth is reduced, guarantee low delay of distributed computing, and realize double improvement of model compression ratio and inference acceleration effect.

Description

Technical Field

[0001] This invention belongs to the field of edge computing and deep learning model optimization technology, specifically relating to a model compression and acceleration method for edge computing in unstable network environments. Background Technology

[0002] Edge computing, as a computing mode closer to the data generation end, can effectively reduce data transmission latency and reduce core network bandwidth consumption, becoming a core technical means for data processing in unstable network environments. However, due to the large number of parameters and high computational complexity, deep neural network models generally suffer from high inference latency and insufficient computing power when deployed on edge embedded devices with limited computing and storage resources, making it difficult to meet the low-latency deterministic requirements of edge distributed real-time networking.

[0003] Existing technologies for compression acceleration of deep learning models include model pruning, knowledge distillation, and model quantization. However, each of these methods has significant limitations when used individually: model pruning relies heavily on human experience or simple rules, resulting in low pruning efficiency, poor generalization, and easy loss of model accuracy; student model design for knowledge distillation lacks specificity and is difficult to adapt to the hardware characteristics of edge devices, leading to limited secondary compression effects; and model quantization suffers a sharp increase in accuracy loss when using low bit-width quantization, and it does not incorporate customized optimizations based on the characteristics of domestic edge chips, resulting in minimal improvement in inference efficiency after deployment.

[0004] Meanwhile, existing technologies lack an integrated model compression acceleration framework. The various compression methods are independent of each other and cannot form a synergistic effect, making it difficult to achieve a dual improvement in compression ratio and acceleration effect. Furthermore, they do not combine model compression with adaptive adjustment of unstable networks. When the link bandwidth decreases, they cannot automatically adjust the data transmission and model computation load, and cannot guarantee the low-latency determinism of edge distributed computing. Summary of the Invention

[0005] To address the aforementioned issues, this invention proposes a model compression acceleration method for edge computing in unstable network environments. This method achieves a dual improvement in model compression ratio and inference acceleration, exhibits strong real-time performance and high robustness in unstable network environments, and can adapt to the computing power and storage constraints of edge embedded devices. It is suitable for weak network edge computing scenarios such as industrial IoT, emergency communication, and battlefield awareness.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A method for accelerating edge computing model compression in unstable network environments includes the following steps: S1. Perform multi-strategy hybrid pruning on the original intelligent model, which combines manual experience pruning, rule-based pruning, and reinforcement learning intelligent search pruning, to obtain a set of lightweight intelligent models with compression ratios that meet the requirements of edge computing. S2. Using the lightweight intelligent model set as the student model and the original intelligent model as the teacher model, perform joint knowledge distillation of soft and hard labels based on edge business data to complete secondary compression and obtain the enhanced lightweight intelligent model set. S3. Combining the hardware characteristics of the domestic MLU200 deep learning processor, perform model quantization acceleration processing on the enhanced lightweight intelligent model set to obtain the quantized lightweight intelligent model. S4. Perform parallel acceleration optimization on the quantized lightweight intelligent model by model parallelism, data parallelism, operator customization and fusion, and deploy it to edge embedded devices to obtain a low-latency lightweight intelligent model. S5. The low-latency lightweight intelligent model operation submodule, the adaptive computing environment perception submodule, and the adaptive transmission link submodule together form an adaptive adjustment module. The adaptive adjustment module senses the edge distributed network bandwidth in real time. When the bandwidth decreases, it automatically reduces the amount of data transmission and model calculation to ensure low latency in distributed computing.

[0007] Preferably, in step S1, the specific process of regularized pruning is as follows: First, the original intelligent model is trained using L2 regularization to learn the importance of network connections. Connections with weights below a threshold are pruned to obtain a sparse network. The initial weights of the surviving connections are retained for retraining. Based on the Dropout rate during retraining, the pruning-retraining process is iterated until the target compression ratio is achieved. Neurons with zero input or zero-output connections are automatically deleted. The formula for calculating the Dropout rate during retraining is: In the formula, The Dropout rate during retraining; This is the original Dropout rate; Let i be the number of connections in the i-th layer of the original network; This represents the number of connections in the i-th layer of the retrained network; i is the network layer number.

[0008] Preferably, in step S1, the specific process of reinforcement learning intelligent search pruning is as follows: Using the AMC framework with deep deterministic policy gradients as agents, the pre-trained original intelligent model is processed layer by layer. The optimal sparsity rate of each network layer is automatically searched through a continuous compression ratio control strategy, supporting both resource-constrained compression and accuracy-guaranteed compression protocols. The loss function for reinforcement learning is: , In the formula, The loss function for reinforcement learning; N is the number of samples; The target Q value; This is the output of the current Q network; The state features of the i-th layer; The compression ratio action for the i-th layer; These are the parameters of the Q-network; b is the reward value; b is the base reward. This is the discount factor; The state features of the (i+1)th layer; The optimal action output by the policy network; These are the parameters of the policy network; The Q-value is evaluated for the next state and the optimal action.

[0009] Preferably, in step S2, the specific process of knowledge distillation is as follows: Edge business data is simultaneously input into both the teacher model and the student model. The output of the teacher model is processed by a Softmax(T) function with a temperature parameter to generate soft labels. The student model is simultaneously processed by a Softmax(T) function with a temperature parameter and a Softmax function to output soft prediction results and hard prediction results, respectively. The soft prediction results are used to calculate the distillation loss with the output of the teacher model; the hard prediction results are used to calculate the true loss with the true labels. Then, the student model is jointly trained after fusing the total loss according to preset weights to complete the secondary compression. The specific calculation formula is as follows: The formula for generating soft labels for the teacher model is: ， In the formula, The probability of the soft label of class d output by the teacher model; This is the raw output of the last layer of the teacher model; T is the temperature parameter. Total number of tags; The formula for generating the soft prediction results of the student model is: In the formula, The soft prediction probability of class d output by the student model; This is the raw output of the last layer of the student model; For the last layer of the student model Original output; The formula for generating hard prediction results for the student model is: In the formula, The hard prediction probability of class d output by the student model; The formula for calculating distillation loss is: In the formula, This is due to distillation losses; The total number of categories; The formula for calculating the actual loss is: In the formula, This is a real loss; This is the one-hot encoding of the d-th class of real labels; The formula for calculating the total loss is: In the formula, Total loss; These are the weighting coefficients. .

[0010] Preferably, in step S2, the student model is a lightweight network based on AttoNet, employing a human-machine collaborative architecture design strategy. First, an initial prototype is built based on rules for increasing network depth and adding shortcut connections. Then, module-level and micro-architecture optimizations are completed through a machine-driven generative synthesizer and DenseNAS to adapt to the computing power constraints of edge embedded devices. The module-level loss function completed by DenseNAS is: In the formula, Complete the module-level loss function for DenseNAS; These are the model weight parameters; Choose weights for the operation; Assign weights to the modules; Cross-entropy loss; The penalty coefficient is... Temperature coefficient; This refers to the computational cost of the network architecture.

[0011] Preferably, in step S3, the model quantization acceleration processing is INT8 low-bit-width quantization, and a precision compensation method based on generative adversarial neural networks and knowledge distillation is introduced during the quantization process. Specifically, the original intelligent model feature map with full precision is used as the real sample, and the quantized model feature map is used as the fake sample. The robustness of the quantized model is improved through game-like training between the generator and the discriminator. Simultaneously, JS divergence is introduced to construct a structural loss function, which, combined with the quantized model's own loss function, forms a comprehensive loss function to guide the update of the quantized model's weight parameters and compensate for the quantization precision loss. The specific calculation formula is as follows: The formula for calculating KL divergence is: In the formula, Let KL divergence be a metric. The probability distribution of the original intelligent model feature map at full precision; This represents the probability distribution of the quantized model feature map; The formula for calculating JS divergence is: In the formula, is the JS divergence, with a value range of [0,1]. It takes a value of 0 when the probability distributions P and Q are exactly the same, and a value of 1 when they are completely opposite. The formula for calculating the structural loss function is: In the formula, For structural loss functions; The formula for calculating the self-loss function is: In the formula, It is its own loss function; One-hot encoding for the real label; The model predicts the output; The formula for calculating the comprehensive loss function is: In the formula, For comprehensive loss functions; Represents the self-loss function; F This is the scaling factor, with a value range of [0,1].

[0012] Preferably, in step S5, the low-latency lightweight intelligent model operation submodule is used to load the matching lightweight intelligent model, dynamically reduce the model's computational load, and ensure low latency in distributed computing; the adaptive computing environment awareness submodule is used to collect information on network bandwidth, edge node computing power, and service data volume; and the adaptive transmission link submodule is used to adjust the data transmission resolution and transmission frame rate according to bandwidth changes.

[0013] By adopting the above technical solution, the present invention has the following beneficial effects: 1. This invention adopts a hybrid pruning method that combines manual experience pruning, rule-based pruning, and reinforcement learning intelligent search pruning. This method overcomes the shortcomings of single pruning methods that rely on manual labor and suffer from significant accuracy loss. It automatically searches for the optimal sparsity rate, and while meeting the computing power constraints of edge devices, it preserves the model accuracy to the greatest extent. The multi-strategy collaborative compression achieves a balance between compression efficiency and accuracy.

[0014] 2. This invention constructs a knowledge distillation structure using a lightweight model as the student and the original model as the teacher. It adopts a joint training method of soft labels and hard labels to effectively inherit the feature expression ability of the original model. While further compressing the model size, it improves the generalization and accuracy of the lightweight model, adapts to the deployment requirements of edge embedded devices, and has a significant secondary compression effect.

[0015] 3. This invention leverages the hardware characteristics of the domestically produced MLU200 deep learning processor to perform INT8 low-bit-width quantization, and introduces generative adversarial networks and JS divergence structural loss for accuracy compensation. This reduces computational load and storage consumption while avoiding the sharp drop in accuracy caused by low-bit-width quantization, significantly improving edge inference speed and hardware adaptability. Furthermore, this invention fully exploits the computing power potential of edge computing processors through model parallelism, data parallelism, operator customization, and operator fusion optimization, reducing model computational complexity and improving the model's execution efficiency and development convenience on edge embedded devices.

[0016] 4. This invention deeply integrates model compression acceleration with network environment awareness. Through an adaptive adjustment module, it senses bandwidth fluctuations in real time and automatically and dynamically adjusts the amount of data transmission and model computation. Even in unstable networks and scenarios with sudden bandwidth drops, it can still ensure low latency and high reliability of edge distributed computing. At the same time, it constructs an integrated acceleration framework of pruning-distillation-quantization-deployment-adaptive adjustment, with each link working together to enhance efficiency. It can be widely adapted to various deep learning models and edge hardware, meeting the engineering application needs of highly dynamic and real-time edge computing scenarios such as industrial IoT, emergency communication, and battlefield awareness. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0019] like Figure 1 As shown, a model compression acceleration method for edge computing in unstable network environments includes the following steps: S1. Perform multi-strategy hybrid pruning on the original intelligent model, which combines manual experience pruning, rule-based pruning, and reinforcement learning intelligent search pruning, to obtain a set of lightweight intelligent models with compression ratios that meet the requirements of edge computing. In step S1, the specific process of regularized pruning is as follows: First, the original intelligent model is trained using L2 regularization to learn the importance of network connections. Connections with weights below a threshold are pruned to obtain a sparse network. The initial weights of the surviving connections are retained for retraining. Based on the Dropout rate during retraining, the pruning-retraining process is iterated until the target compression ratio is achieved. Neurons with zero input or fractional-zero output connections are automatically deleted. The formula for calculating the Dropout rate during retraining is: In the formula, The Dropout rate during retraining; This is the original Dropout rate; Let i be the number of connections in the i-th layer of the original network; This represents the number of connections in the i-th layer of the retrained network; i is the network layer number. In step S1, the specific process of reinforcement learning intelligent search pruning is as follows: Using the AMC framework with deep deterministic policy gradients as the agent, the pre-trained original intelligent model is processed layer by layer. The optimal sparsity rate of each network layer is automatically searched through a continuous compression ratio control strategy, supporting both resource-constrained compression and accuracy-guaranteed compression protocols. The loss function for reinforcement learning is: , In the formula, The loss function for reinforcement learning; N is the number of samples; The target Q value; This is the output of the current Q network; The state features of the i-th layer; The compression ratio action for the i-th layer; These are the parameters of the Q-network; b is the reward value; b is the base reward. This is the discount factor; The state features of the (i+1)th layer; The optimal action output by the policy network; These are the parameters of the policy network; Evaluate the Q-value for the next state and the optimal action; S2. Using the lightweight intelligent model set as the student model and the original intelligent model as the teacher model, perform joint knowledge distillation of soft and hard labels based on edge business data to complete secondary compression and obtain the enhanced lightweight intelligent model set. In step S2, the specific process of knowledge distillation is as follows: Edge business data is simultaneously input into both the teacher model and the student model. The output of the teacher model is processed by a Softmax(T) function with a temperature parameter to generate soft labels. The student model is simultaneously processed by a Softmax(T) function with a temperature parameter and a Softmax function to output soft prediction results and hard prediction results, respectively. The soft prediction results are used to calculate the distillation loss with the output of the teacher model; the hard prediction results are used to calculate the true loss with the true labels. Then, the total loss is fused according to preset weights, and the student model is jointly trained to complete the secondary compression. The specific calculation formula is as follows: The formula for generating soft labels for the teacher model is: ， In the formula, The probability of the soft label of class d output by the teacher model; This is the raw output of the last layer of the teacher model; T is the temperature parameter. Total number of tags; The formula for generating the soft prediction results of the student model is: In the formula, The soft prediction probability of class d output by the student model; This is the raw output of the last layer of the student model; For the last layer of the student model Original output; The formula for generating hard prediction results for the student model is: In the formula, The hard prediction probability of class d output by the student model; The formula for calculating distillation loss is: In the formula, This is due to distillation losses; The total number of categories; The formula for calculating the actual loss is: In the formula, This is a real loss; This is the one-hot encoding of the d-th class of real labels; The formula for calculating the total loss is: In the formula, Total loss; These are the weighting coefficients. ; In step S2, the student model is a lightweight network based on AttoNet, employing a human-machine collaborative architecture design strategy. First, an initial prototype is built based on rules for increasing network depth and adding shortcut connections. Then, module-level and micro-architecture optimizations are completed through a machine-driven generative synthesizer and DenseNAS, adapting to the computing power constraints of edge embedded devices. The module-level loss function completed by DenseNAS is: In the formula, Complete the module-level loss function for DenseNAS; These are the model weight parameters; Choose weights for the operation; Assign weights to the modules; Cross-entropy loss; The penalty coefficient is... Temperature coefficient; The computational cost of network architecture; S3. Combining the hardware characteristics of the domestic MLU200 deep learning processor, perform model quantization acceleration processing on the enhanced lightweight intelligent model set to obtain the quantized lightweight intelligent model. In step S3, the model quantization acceleration process is INT8 low-bit-width quantization, and a precision compensation method based on generative adversarial neural networks and knowledge distillation is introduced during the quantization process. Specifically, the original intelligent model feature map with full precision is used as the real sample, and the quantized model feature map is used as the fake sample. The robustness of the quantized model is improved through game-like training between the generator and the discriminator. Simultaneously, JS divergence is introduced to construct a structural loss function, which, combined with the quantized model's own loss function, forms a comprehensive loss function to guide the update of the quantized model's weight parameters and compensate for the quantization precision loss. The specific calculation formula is as follows: The formula for calculating KL divergence is: In the formula, Let KL divergence be a metric. The probability distribution of the original intelligent model feature map at full precision; This represents the probability distribution of the quantized model feature map; The formula for calculating JS divergence is: In the formula, is the JS divergence, with a value range of [0,1]. It takes a value of 0 when the probability distributions P and Q are exactly the same, and a value of 1 when they are completely opposite. The formula for calculating the structural loss function is: In the formula, For structural loss functions; The formula for calculating the self-loss function is: In the formula, It is its own loss function; One-hot encoding for the real label; The model predicts the output; The formula for calculating the comprehensive loss function is: In the formula, For comprehensive loss functions; Represents the self-loss function; F This is the scaling factor, with a value range of [0,1]. S4. Perform parallel acceleration optimization on the quantized lightweight intelligent model by model parallelism, data parallelism, operator customization and fusion, and deploy it to edge embedded devices to obtain a low-latency lightweight intelligent model. S5. The low-latency lightweight intelligent model operation submodule, the adaptive computing environment perception submodule, and the adaptive transmission link submodule are combined to form an adaptive adjustment module. The adaptive adjustment module senses the edge distributed network bandwidth in real time. When the bandwidth decreases, it automatically reduces the amount of data transmission and model calculation to ensure low latency in distributed computing. In step S5, the low-latency lightweight intelligent model operation submodule is used to load the matching lightweight intelligent model, dynamically reduce the model's computational load, and ensure low latency in distributed computing; the adaptive computing environment awareness submodule is used to collect information on network bandwidth, edge node computing power, and business data volume; and the adaptive transmission link submodule is used to adjust the data transmission resolution and transmission frame rate according to bandwidth changes.

[0020] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for accelerating edge computing model compression in unstable network environments, characterized in that, Includes the following steps: S1. Perform multi-strategy hybrid pruning on the original intelligent model, which combines manual experience pruning, rule-based pruning, and reinforcement learning intelligent search pruning, to obtain a set of lightweight intelligent models with compression ratios that meet the requirements of edge computing. S2. Using the lightweight intelligent model set as the student model and the original intelligent model as the teacher model, perform joint knowledge distillation of soft and hard labels based on edge business data to complete secondary compression and obtain the enhanced lightweight intelligent model set. S3. Combining the hardware characteristics of the domestic MLU200 deep learning processor, perform model quantization acceleration processing on the enhanced lightweight intelligent model set to obtain the quantized lightweight intelligent model. S4. Perform parallel acceleration optimization on the quantized lightweight intelligent model by model parallelism, data parallelism, operator customization and fusion, and deploy it to edge embedded devices to obtain a low-latency lightweight intelligent model. S5. The low-latency lightweight intelligent model operation submodule, the adaptive computing environment perception submodule, and the adaptive transmission link submodule together form an adaptive adjustment module. The adaptive adjustment module senses the edge distributed network bandwidth in real time. When the bandwidth decreases, it automatically reduces the amount of data transmission and model calculation to ensure low latency in distributed computing.

2. The model compression acceleration method for edge computing in an unstable network environment as described in claim 1, characterized in that, In step S1, the specific process of regularized pruning is as follows: First, the original intelligent model is trained using L2 regularization to learn the importance of network connections. Connections with weights below a threshold are pruned to obtain a sparse network. The initial weights of the surviving connections are retained for retraining. Based on the Dropout rate during retraining, the pruning-retraining process is iterated until the target compression ratio is achieved. Neurons with zero input or fractional-zero output connections are automatically deleted. The formula for calculating the Dropout rate during retraining is: In the formula, The Dropout rate during retraining; This is the original Dropout rate; Let i be the number of connections in the i-th layer of the original network; This represents the number of connections in the i-th layer of the retrained network; i is the network layer number.

3. The model compression and acceleration method for edge computing in an unstable network environment as described in claim 2, characterized in that, In step S1, the specific process of reinforcement learning intelligent search pruning is as follows: Using the AMC framework with deep deterministic policy gradients as the agent, the pre-trained original intelligent model is processed layer by layer. The optimal sparsity rate of each network layer is automatically searched through a continuous compression ratio control strategy, supporting both resource-constrained compression and accuracy-guaranteed compression protocols. The loss function for reinforcement learning is: , In the formula, The loss function for reinforcement learning; N is the number of samples; The target Q value; This is the output of the current Q network; The state features of the i-th layer; The compression ratio action for the i-th layer; These are the parameters of the Q-network; b is the reward value; b is the base reward. This is the discount factor; The state features of the (i+1)th layer; The optimal action output by the policy network; These are the parameters of the policy network; The Q-value is evaluated for the next state and the optimal action.

4. The model compression acceleration method for edge computing in an unstable network environment as described in claim 3, characterized in that, In step S2, the specific process of knowledge distillation is as follows: Edge business data is simultaneously input into both the teacher model and the student model. The output of the teacher model is processed by a Softmax(T) function with a temperature parameter to generate soft labels. The student model is simultaneously processed by a Softmax(T) function with a temperature parameter and a Softmax function to output soft prediction results and hard prediction results, respectively. The soft prediction results are used to calculate the distillation loss with the output of the teacher model; the hard prediction results are used to calculate the true loss with the true labels. Then, the total loss is fused according to preset weights, and the student model is jointly trained to complete the secondary compression. The specific calculation formula is as follows: The formula for generating soft labels for the teacher model is: ， In the formula, The probability of the soft label of class d output by the teacher model; This is the raw output of the last layer of the teacher model; T is the temperature parameter. Total number of tags; The formula for generating the soft prediction results of the student model is: In the formula, The soft prediction probability of class d output by the student model; This is the raw output of the last layer of the student model; For the last layer of the student model Original output; The formula for generating hard prediction results for the student model is: In the formula, The hard prediction probability of class d output by the student model; The formula for calculating distillation loss is: In the formula, This is due to distillation losses; The total number of categories; The formula for calculating the actual loss is: In the formula, This is a real loss; This is the one-hot encoding of the d-th class of real labels; The formula for calculating the total loss is: In the formula, Total loss; These are the weighting coefficients. .

5. The model compression acceleration method for edge computing in an unstable network environment as described in claim 4, characterized in that, In step S2, the student model is a lightweight network based on AttoNet, employing a human-machine collaborative architecture design strategy. First, an initial prototype is built based on rules for increasing network depth and adding shortcut connections. Then, module-level and micro-architecture optimizations are completed through a machine-driven generative synthesizer and DenseNAS, adapting to the computing power constraints of edge embedded devices. The module-level loss function completed by DenseNAS is: In the formula, Complete the module-level loss function for DenseNAS; These are the model weight parameters; Choose weights for the operation; Assign weights to the modules; Cross-entropy loss; The penalty coefficient is... Temperature coefficient; This refers to the computational cost of the network architecture.

6. The model compression acceleration method for edge computing in an unstable network environment as described in claim 5, characterized in that, In step S3, the model quantization acceleration process is INT8 low-bit-width quantization, and a precision compensation method based on generative adversarial neural networks and knowledge distillation is introduced during the quantization process. Specifically, the original intelligent model feature map with full precision is used as the real sample, and the quantized model feature map is used as the fake sample. The robustness of the quantized model is improved through game-like training between the generator and the discriminator. Simultaneously, JS divergence is introduced to construct a structural loss function, which, combined with the quantized model's own loss function, forms a comprehensive loss function to guide the update of the quantized model's weight parameters and compensate for the quantization precision loss. The specific calculation formula is as follows: The formula for calculating KL divergence is: In the formula, Let KL divergence be a metric. The probability distribution of the original intelligent model feature map at full precision; This represents the probability distribution of the quantized model feature map; The formula for calculating JS divergence is: In the formula, is the JS divergence, with a value range of [0,1]. It takes a value of 0 when the probability distributions P and Q are exactly the same, and a value of 1 when they are completely opposite. The formula for calculating the structural loss function is: In the formula, For structural loss functions; The formula for calculating the self-loss function is: In the formula, It is its own loss function; One-hot encoding for the real label; The model predicts the output; The formula for calculating the comprehensive loss function is: In the formula, For comprehensive loss functions; Represents the self-loss function; F This is the scaling factor, with a value range of [0,1].

7. The model compression and acceleration method for edge computing in an unstable network environment as described in claim 6, characterized in that: In step S5, the low-latency lightweight intelligent model operation submodule is used to load the matching lightweight intelligent model, dynamically reduce the model's computational load, and ensure low latency in distributed computing; the adaptive computing environment awareness submodule is used to collect information on network bandwidth, edge node computing power, and business data volume; and the adaptive transmission link submodule is used to adjust the data transmission resolution and transmission frame rate according to bandwidth changes.

Images