Outlet selection and partitioning deployment strategy generation method for efficient reasoning of multi-outlet neural networks

Abstract

The present invention discloses a method for generating an outlet selection and partition deployment strategy for efficient reasoning of a multi-outlet neural network, the method comprising the following steps: step S1, multi-outlet DNN operation performance analysis and modeling; step S2, performance analysis of each outlet of the multi-outlet DNN; step S3, optimization of the multi-outlet DNN partition deployment scheme; step S4, optimization of the multi-outlet DNN outlet selection scheme. The method is suitable for supporting multiple different complex multi-outlet DNNs and distributed heterogeneous edge computing environments, by analyzing the computing performance, outlet accuracy, etc. of the multi-outlet DNN, combining genetic algorithms with hierarchical reinforcement learning algorithms to select and partition the DNN outlet, generating a suitable partition deployment scheme for the multi-outlet DNN, and optimizing the energy consumption of the terminal device under delay constraints, thereby achieving efficient DNN reasoning.

Description

Technical Field

[0001] The present invention belongs to the field of computer service technology, and relates to a method for generating an outlet selection and partition deployment strategy for efficient reasoning of a multi-outlet neural network, and specifically to a deep neural network partition deployment algorithm based on reinforcement learning and a multi-outlet deep neural network outlet selection algorithm based on a genetic algorithm. Background Art

[0002] With the continuous development of the Internet of Things and Deep Neural Network (DNN), deep learning solutions have been widely used in various industrial IoT applications, including predictive maintenance, product quality inspection, asset tracking, and collaborative robotics, and their performance is often better than manual methods. These applications require complex DNN reasoning to be performed in real time, with high computational load and sensitivity to latency. IoT devices with limited computing resources find it difficult to complete these complex reasoning tasks independently, while cloud computing platforms with a large amount of computing resources often cause significant latency problems due to long distances. To resolve this contradiction, researchers adopted the edge computing paradigm, using edge servers close to IoT devices for collaborative reasoning, making full use of the computing power of edge servers and effectively reducing latency.

[0003] The edge computing paradigm can reduce latency and enhance privacy protection by bringing computing closer to data producers. The current mainstream approach is to split DNN into partitions of different sizes, with shallow partitions processed on IoT devices and deep partitions offloaded to edge servers to complete the remaining reasoning. This approach avoids high-latency network transmission between cloud servers and end devices and can improve overall performance. However, dynamic changes in the edge environment may affect the performance of collaborative reasoning. Specifically, when the network environment is poor, in order to reduce the communication burden, the best partitioning decision may tend to let IoT devices take on more computing tasks, which undoubtedly increases the computing burden of IoT devices; on the contrary, when the network environment is good, the best decision may tend to offload all reasoning calculations to edge servers, but this may cause edge servers to be overloaded and increase computing costs. In addition, DNNs generally have the problem of overthinking, which can lead to misclassification of some samples and bring unnecessary computing overhead.

[0004] To address these challenges, researchers have proposed a multi-exit DNN. Unlike the single-output structure of a general DNN, a multi-exit DNN adds multiple classifiers as the exits of the shallow neural network layer. Some reasoning tasks can be completed ahead of time with high confidence through shallow exits with acceptable accuracy. In addition, a multi-exit DNN can also avoid unnecessary data transmission between end devices and edge servers. However, too many exits will increase tail latency, resulting in a decrease in service quality for some users.

[0005] Based on the above research background, it can be found that by selecting appropriate outlets, partitioning and deploying multi-export DNNs, and using edge servers for collaborative reasoning, while meeting delay constraints, the computing load on the end device can be effectively reduced, thereby reducing the computing energy consumption of the end device, increasing the life of the end device, and ultimately achieving efficient DNN reasoning. Summary of the invention

[0006] In order to solve the problems existing in the background technology, the present invention provides a method for generating an outlet selection and partition deployment strategy for efficient reasoning of multi-outlet neural networks. This method is suitable for supporting multiple different complex multi-outlet DNNs and distributed heterogeneous edge computing environments. By analyzing the computing performance, outlet accuracy, etc. of multi-outlet DNNs, combining genetic algorithms with hierarchical reinforcement learning algorithms to select and partition DNN outlets, a suitable partition deployment scheme for multiple-outlet DNNs is generated, and under the condition of delay constraints, the energy consumption of the terminal device is optimized to achieve efficient DNN reasoning.

[0007] The objective of the present invention is achieved through the following technical solutions:

[0008] A method for generating an outlet selection and partitioning deployment strategy for efficient reasoning of a multi-outlet neural network comprises the following steps:

[0009] Step S1: Multi-export DNN operation performance analysis and modeling:

[0010] Step S1.1, collect multi-export DNN inference operation data to obtain key information at all levels;

[0011] Step S1.2: Analyze the multi-export DNN, including the device-edge collaborative computing environment, DNN inference delay, IoT device energy consumption, and edge server cost. On this basis, formalize the multi-dimensional energy consumption optimization problem of export selection, model partitioning, and resource allocation. The specific steps are as follows:

[0012] Step S1.2.1, Analyze the device-edge collaborative computing environment:

[0013] Definition 1. IOT = {I1, I2, …, I n}, represents a set of IoT devices, where q i Indicates the floating point computing capability of IoT devices; Indicates the power of IoT devices when they are running; Indicates the transmission power of IoT devices when sending data through the network; Indicates the resting power of IoT devices; Bw i Denotes the bandwidth between IoT devices and edge servers; i Represents the connection latency between IoT devices and edge servers;

[0014] Definition 2: Edge e ,C e ,M e >, indicating an edge server, where: q e Indicates the floating point computing capability of the edge server; C e Indicates the number of CPU cores of the edge server; M e Indicates the memory size of the edge server;

[0015] Step S1.2.2: Analyze multi-export DNN:

[0016] Definition 3: The set of DNN models running on each IoT device For IoT devices i Model D running on i For D i Defined as a four-tuple, namely D i = <L i ,E i ,W i ,M i ,P i >, where: Indicates D i The backbone network layer set, l ij = <fl ij ,ml ij >, fl ij Indicates the floating point calculation amount of this layer, ml ij Indicates the memory requirements of this layer; Indicates D i An early export collection, e ij = <fe ij ,me ij >,fe ij Indicates the floating point calculation amount of this layer, me ij Indicates the memory requirements of this layer; Indicates D i The output data size of each layer in the backbone network; Indicates D i The exit rate of each early exit,

[0017]

[0018] Step S1.2.3, Analyze DNN latency: ​

[0019] Calculate the inference latency of each DNN model, including IoT device computation latency, based on model partitioning decisions, exit point selection decisions, and resource allocation decisions Network transmission delay Calculating latency with edge devices in:

[0020]

[0021] Step S1.2.4: Analyze the energy consumption of IoT devices:

[0022] Define the energy consumption of IoT devices and the operating power of known devices Transmission power And resting power Based on the computing time of the task in the IoT device and network transmission time The computing energy consumption and standby energy consumption of computing equipment and the transmission energy consumption of the network, namely:

[0023]

[0024] Step S2: Performance analysis of each outlet of multi-export DNN:

[0025] Collect multi-exit DNN input data on the terminal device, set different entropy thresholds for each exit, and test the exit probability and accuracy of each exit under different settings. The specific steps are as follows:

[0026] Step S2.1, collect non-IID input data on heterogeneous devices to capture the diversity and complexity of data from each device;

[0027] Step S2.2, design and deploy a DNN model with multiple exits, each exit corresponding to a different level, to balance computational complexity and inference accuracy;

[0028] Step S2.3, set a series of different entropy thresholds for each exit, and by testing these thresholds, record the exit probability and classification accuracy of each exit under different thresholds;

[0029] Step S2.4, analyzing the test results and selecting the best entropy threshold to achieve the best export selection on different devices;

[0030] Step S2.5: Develop and deploy an export selection strategy, and continuously optimize and iterate based on actual application feedback to improve the overall performance of the model under various device and data complexity conditions;

[0031] Step S3: Optimization of multi-export DNN partition deployment scheme:

[0032] Step S31: Analyze the target DNN model according to the multi-export DNN model operation performance analysis and modeling in step S1 and the multi-export DNN export performance analysis in step S2 to understand its structure, parameter scale, hierarchical relationship, etc., and extract the computational complexity and memory requirements of each level;

[0033] Step S32: construct an optimization problem and decompose the problem into two sub-problems, namely, a partition deployment solution optimization problem and an exit selection problem;

[0034] Step S33: Using a deep reinforcement learning algorithm, the partition deployment scheme is continuously optimized under the premise of satisfying performance constraints, including the selection of partition points and resource allocation on edge servers;

[0035] Step S4: Optimization of multi-export DNN export selection scheme:

[0036] The partition deployment scheme obtained in step S3 is used to optimize the exit selection scheme using a genetic algorithm. The population individuals of the genetic algorithm represent an exit selection scheme. Through crossover mutation, a better exit selection scheme is obtained within a limited number of iterations. Step S3 is repeated to further optimize the partition deployment scheme. Repeated evaluation and adjustment are performed until the optimal partition deployment scheme and exit selection scheme that meet the performance requirements are found.

[0037] Compared with the prior art, the present invention has the following advantages:

[0038] 1. Support the optimization of large-scale multi-export DNN partition deployment and export selection, reduce the resource consumption of end devices, improve the resource utilization of end devices, and maximize the running time under limited resource conditions.

[0039] 2. Various delay constraints are taken into account during optimization, which not only meets the user's average response delay and ensures the stability of the user's overall service quality, but also takes into account the user's tail delay and ensures the lower limit of the user's service quality.

[0040] 3. The complex coupling problem was solved by using the exit selection algorithm based on genetic algorithm and the partitioning scheme optimization algorithm based on reinforcement learning. First, the exit selection was decomposed from the original problem, and then the exit selection algorithm based on genetic algorithm was designed. On the basis of the exit selection algorithm, the partitioning deployment scheme optimization algorithm based on reinforcement learning was designed to optimize the model partitioning and resource allocation decisions. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a flow chart of a method for generating a multi-exit neural network outlet selection and partition deployment strategy;

[0042] Figure 2It is a flow chart of the partition deployment scheme optimization algorithm;

[0043] Figure 3 It is a multi-export DNN application case in the Industrial Internet. DETAILED DESCRIPTION

[0044] The technical solution of the present invention is further described below in conjunction with the accompanying drawings, but is not limited thereto. Any modification or equivalent replacement of the technical solution of the present invention without departing from the spirit and scope of the technical solution of the present invention should be included in the protection scope of the present invention.

[0045] The present invention provides a method for generating an outlet selection and partitioning deployment strategy for high-efficiency reasoning of a multi-outlet neural network, such as Figure 1 As shown, the method comprises the following steps:

[0046] Step S1: Multi-export DNN operation performance analysis:

[0047] Step S1.1: To support the subsequent optimization of the export selection and partition deployment plan, the multi-export DNN inference operation data is collected to obtain key information at all levels. The key information includes: (1) in-depth analysis of the neural network structure, collection of model levels, parameter scale, computational complexity and resource requirements; (2) detailed description of the partition structure, including module division and communication overhead; (3) understanding of the deployment structure, including hardware resource configuration, device load and computing power; (4) measurement of end-to-end delay, including model inference time, communication time and data transmission time.

[0048] Step S1.2: Analyze and model the multi-export DNN. As the mathematical basis of the optimization algorithm, modeling requires a detailed analysis of the problem scenario, focusing on the device-edge collaborative computing environment, the delay model of DNN, the energy consumption model of IoT devices, and the cost model of edge servers. Then, a formal description of the multi-dimensional energy consumption optimization problem of export selection, model partitioning, and resource allocation is given. Specifically, it includes the following contents:

[0049] Step S1.2.1, Analyze the device-edge collaborative computing environment: This step requires defining the IoT device set and edge server, including their basic attributes such as computing power, network bandwidth, and power:

[0050] Definition 1. IOT = {I1, I2, …, I n}, represents a set of IoT devices, where q i Indicates the floating point computing capability of IoT devices, in FLOPS; Indicates the power of the IoT device when it is running, in w / s; It indicates the transmission power of IoT devices when sending data through the network, in w / s; Indicates the resting power of IoT devices in W / s; Bw i Indicates the bandwidth between the IoT device and the edge server, in Mbps; i Indicates the connection delay between the IoT device and the edge server, in ms.

[0051] Definition 2: Edge e ,C e ,M e >, indicating an edge server, where: q e Indicates the floating point computing capability of the edge server, in FLOPS; C e Indicates the number of CPU cores of the edge server, in units; M e Indicates the memory size of the edge server, in GB;

[0052] Step S1.2.2, analyze multi-export DNN: This step requires defining the set of DNN models running on each IoT device, and describing in detail the backbone network structure of each model and the intermediate output size of the early export structure.

[0053] Definition 3: The set of DNN models running on each IoT device For IoT devices i Model D running on i For D i Defined as a four-tuple, namely D i = <L i ,E i ,W i ,M i ,P i >, where: Indicates D i The backbone network layer set, l ij = <fl ij ,ml ij >, fl ij Indicates the floating point calculation amount of this layer, ml ij Represents the memory requirement of this layer. The memory requirement of the network mainly comes from the memory overhead of its weight vector and activation vector, which can be calculated by the size and bit width of its weight vector and activation vector; Indicates D i An early export collection, e ij = <fe ij ,me ij >,fe ij Indicates the floating point calculation amount of this layer, me ij ​It represents the memory requirement of this layer. Here, we do not need to care about the network structure of the early exit. The present invention regards the entire exit as a complete unit. Indicates D i The output data size of each layer in the backbone network; Indicates D i The exit rate of each early exit,

[0054] Step S1.2.3, Analyze DNN latency: This step requires calculating the inference latency of each DNN model, including the IoT device computation latency, based on the model partitioning decision, exit point selection decision, and resource allocation decision. Network transmission delay Calculating latency with edge devices in:

[0055]

[0056] Step S1.2.4: Analyze the energy consumption of IoT devices: This step requires defining the energy consumption of IoT devices and the operating power of known devices. Transmission power And resting power Based on the computing time of the task in the IoT device and network transmission time The computing energy consumption and standby energy consumption of the device and the transmission energy consumption of the network can be calculated, that is:

[0057]

[0058] This step mainly focuses on the export selection and partition deployment optimization requirements required for high-efficiency reasoning of multi-export DNNs. It provides modeling of multi-export DNNs and standardizes the description of neural network structure, computing environment, computing delay and other factors.

[0059] Step S2: Performance analysis of each outlet of multi-export DNN:

[0060] Since there is a certain intrinsic connection between exit selection and inference performance, different exit settings will bring different performance changes. However, the data collected by heterogeneous devices are generally non-IID data, which leads to significant differences in data complexity. Therefore, when selecting exits for multi-exit DNNs on different end devices, their unique exit performance conditions need to be considered. First, collect the multi-exit DNN input data on the end device, and then set different entropy thresholds for each exit to test the exit probability and accuracy of each exit under different settings. The specific steps are as follows:

[0061] Step S2.1, collect non-IID input data on heterogeneous devices to capture the diversity and complexity of data from each device;

[0062] Step S2.2, design and deploy a DNN model with multiple exits, each exit corresponding to a different level, to balance computational complexity and inference accuracy;

[0063] Step S2.3, set a series of different entropy thresholds for each exit, and by testing these thresholds, record the exit probability and classification accuracy of each exit under different thresholds;

[0064] Step S2.4, analyzing the test results and selecting the best entropy threshold to achieve the best export selection on different devices;

[0065] Step S2.5: Develop and deploy an export selection strategy, and continuously optimize and iterate based on actual application feedback to improve the overall performance of the model under various device and data complexity conditions.

[0066] Step S3: Optimization of multi-export DNN partition deployment scheme:

[0067] Step S31: According to the performance analysis and modeling of the multi-export DNN model in step S1 and the export performance analysis of the multi-export DNN in step S2, the target DNN model is analyzed to understand its structure, parameter scale, hierarchical relationship, etc., and the computational complexity and memory requirements of each level are extracted.

[0068] Step S32: construct an optimization problem and decompose the problem into two sub-problems, namely, a partition deployment solution optimization problem and an export selection problem.

[0069] Step S33: Using the deep reinforcement learning algorithm, the partition deployment plan is continuously optimized under the premise of meeting performance constraints, including the selection of partition points and resource allocation on edge servers.

[0070] This step constructs a joint optimization problem of partition deployment and outlet selection based on the analytical modeling of the multi-export DNN in step S1 and the performance analysis of each outlet of the multi-export DNN in step S2. In order to solve this problem, the outlet selection problem is decoupled, and the partition deployment scheme is first optimized through reinforcement learning. In this step, the partition deployment problem is expressed as a Markov decision process (Markov Decision Process, MDP), and the deep reinforcement learning (Deep Reinforcement Learning, DRL) algorithm is used to determine the optimal solution in the continuous space. In view of the fact that large action dimensions will complicate resource allocation, the present invention adopts the PPO algorithm. PPO (Proximal Policy Optimization) is a widely used reinforcement learning algorithm. It is a policy-based reinforcement learning algorithm off-policy, proposed by OpenAI in 2017. Its core idea is to limit the difference between the new strategy and the old strategy in the process of policy optimization by introducing a clipping mechanism, so as to maintain an appropriate distance between the new and old strategies to prevent excessive policy updates and cause a significant decline in performance. The specific process is as follows: Figure 2 As shown in the figure. For the partition deployment plan obtained by reinforcement learning, the exit selection algorithm based on genetic algorithm is used to find the best exit selection plan under the request, and it is fed back to the reinforcement learning algorithm as a reward. Through continuous iteration, the deployment plan can be further optimized to obtain the best partition deployment plan and the best partition selection plan.

[0071] Step S4: Optimization of multi-export DNN export selection scheme:

[0072] The partition deployment scheme obtained in step S3 is used to optimize the export selection scheme using a genetic algorithm. The population individuals of the genetic algorithm represent an export selection scheme. Through crossover mutation, a better export selection scheme is obtained within a limited number of iterations. Step S3 is repeated to further optimize the partition deployment scheme. Repeated evaluation and adjustment are performed until the optimal partition deployment scheme and export selection scheme that meet the performance requirements are found.

[0073] This step uses a genetic algorithm to optimize the exit selection scheme. An important consideration is to determine how the partitioning scheme is represented in terms of genes and to define the crossover and mutation operations in the genetic algorithm. For a DNN, the arrangement of its exits is linear and the number is limited, because too many exits will impose a great burden on training and reasoning. Therefore, this step uses a 0 / 1 array to represent the exit selection scheme, that is, a gene. The crossover operation uses a two-point crossover, that is, randomly selects two crossover points and exchanges the exits between two genes in the same multi-exit DNN. The mutation operation includes two types of mutations: randomly adding a new exit and randomly deleting an exit.

[0074] Figure 3 This is a multi-export DNN application case in the industrial Internet. The model is split and then the partitions are deployed on IoT devices and edge servers to achieve collaborative reasoning. Figure 3 As shown in the figure, bearings with less wear and concentrated wear can be easily identified and reasoned through shallower exits, while bearings with larger wear and dispersed wear require deeper exits for processing. By setting early exits, some tasks can be reasoned on IoT devices, which greatly reduces the number of tasks that need to be offloaded to edge servers, avoiding unnecessary transmission between IoT devices and edge servers, and reducing the computing load of both parties.

Claims

1. A method for generating an outlet selection and partitioning deployment strategy for efficient reasoning of multi-outlet neural networks, characterized in that The method comprises the following steps: Step S1: Multi-export DNN operation performance analysis and modeling: Step S1.1, collect multi-export DNN inference operation data to obtain key information at all levels; Step S1.2: Analyze the multi-exit DNN, including the device-edge collaborative computing environment, the inference delay of DNN, the energy consumption of IoT devices, and the cost of edge servers. On this basis, formally describe the multi-dimensional energy consumption optimization problem of outlet selection, model partitioning, and resource allocation; Step S2: Performance analysis of each outlet of multi-export DNN: Collect multi-exit DNN input data on the terminal device, set different entropy thresholds for each exit, and test the exit probability and accuracy of each exit under different settings; Step S3: Optimization of multi-export DNN partition deployment scheme: Step S31: Analyze the target DNN model according to the multi-export DNN model operation performance analysis and modeling in step S1 and the multi-export DNN export performance analysis in step S2 to understand its structure, parameter scale, and hierarchical relationship, and extract the computational complexity and memory requirements of each level; Step S32: construct an optimization problem and decompose the problem into two sub-problems, namely, a partition deployment solution optimization problem and an exit selection problem; Step S33: Using a deep reinforcement learning algorithm, the partition deployment scheme is continuously optimized under the premise of satisfying performance constraints, including the selection of partition points and resource allocation on edge servers; Step S4: Optimization of multi-export DNN export selection scheme: The partition deployment scheme obtained in step S3 is used to optimize the exit selection scheme using a genetic algorithm. The population individuals of the genetic algorithm represent an exit selection scheme. Through crossover mutation, a better exit selection scheme is obtained within a limited number of iterations. Step S3 is repeated to further optimize the partition deployment scheme. Repeated evaluation and adjustment are performed until the optimal partition deployment scheme and exit selection scheme that meet the performance requirements are found.

2. The method for generating an outlet selection and partitioning deployment strategy for high-efficiency reasoning of a multi-outlet neural network according to claim 1 is characterized in that In step S1.1, the key information includes: (1) in-depth analysis of the neural network structure, collecting the model's hierarchy, parameter scale, computational complexity, and resource requirements; (2) a detailed description of the partition structure, including module division and communication overhead; (3) understanding of the deployment structure, including hardware resource configuration, device load, and computing power; (4) measurement of end-to-end latency, including model inference time, communication time, and data transmission time.

3. The method for generating an outlet selection and partitioning deployment strategy for high-efficiency reasoning of a multi-outlet neural network according to claim 1, characterized in that The specific steps of step S1.2 are as follows: Step S1.2.1, Analyze the device-edge collaborative computing environment: Definition 1. IOT = {I1, I2, …, I n }, represents a set of IoT devices, where q i Indicates the floating point computing capability of IoT devices; Indicates the power of IoT devices when they are running; Indicates the transmission power of IoT devices when sending data through the network; Indicates the resting power of IoT devices; Bw i Denotes the bandwidth between IoT devices and edge servers; i Represents the connection latency between IoT devices and edge servers; Definition 2: Edge e ,C e ,M e >, indicating an edge server, where: q e Indicates the floating point computing capability of the edge server; C e Indicates the number of CPU cores of the edge server; M e Indicates the memory size of the edge server;​ Step S1.2.2: Analyze multi-export DNN: Definition 3: The set of DNN models running on each IoT device For IoT devices i Model D running on i For D i Defined as a four-tuple, namely D i = <L i ,E i ,W i ,M i ,P i >, where: Indicates D i The backbone network layer set, l ij = <fl ij ,ml ij >, fl ij Indicates the floating point calculation amount of this layer, ml ij Indicates the memory requirement of this layer; E i = Indicates D i An early export collection, e ij = <fe ij ,me ij >,fe ij Indicates the floating point calculation amount of this layer, me ij Indicates the memory requirements of this layer; Indicates D i The output data size of each layer in the backbone network; Indicates D i The exit rate of each early exit, Step S1.2.3, Analyze DNN latency: Calculate the inference latency of each DNN model, including IoT device computation latency, based on model partitioning decisions, exit point selection decisions, and resource allocation decisions Network transmission delay Calculating latency with edge devices in: Step S1.2.4: Analyze the energy consumption of IoT devices: Define the energy consumption of IoT devices and the operating power of known devices Transmission power And resting power Based on the computing time of the task in the IoT device and network transmission time The computing energy consumption and standby energy consumption of computing equipment and the transmission energy consumption of the network, namely:

4. The method for generating an outlet selection and partitioning deployment strategy for high-efficiency reasoning of a multi-outlet neural network according to claim 1, characterized in that The specific steps of step S2 are as follows: Step S2.1, collect non-IID input data on heterogeneous devices to capture the diversity and complexity of data from each device; Step S2.2, design and deploy a DNN model with multiple exits, each exit corresponding to a different level, to balance computational complexity and inference accuracy; Step S2.3, set a series of different entropy thresholds for each exit, and by testing these thresholds, record the exit probability and classification accuracy of each exit under different thresholds; Step S2.4, analyzing the test results and selecting the best entropy threshold to achieve the best export selection on different devices; Step S2.5: Develop and deploy an export selection strategy, and continuously optimize and iterate based on actual application feedback to improve the overall performance of the model under various device and data complexity conditions.

5. The method for generating an outlet selection and partitioning deployment strategy for high-efficiency reasoning of a multi-outlet neural network according to claim 1, characterized in that In step S4, a 0 / 1 array is used to represent the export selection scheme, that is, a gene.

6. The method for generating an outlet selection and partitioning deployment strategy for high-efficiency reasoning of a multi-outlet neural network according to claim 1, characterized in that In step S4, the crossover operation adopts a two-point crossover, that is, two crossover points are randomly selected to exchange outlets between two genes in the same multi-outlet DNN.

7. The method for generating an outlet selection and partitioning deployment strategy for high-efficiency reasoning of a multi-outlet neural network according to claim 1, characterized in that In step S4, the mutation operation includes two types of mutations: randomly adding a new exit and randomly deleting an exit.

Images