Dynamic energy-saving unloading method and system for inference task of deep neural network

Abstract

The invention provides a dynamic energy-saving unloading method and system for a deep neural network reasoning task, and belongs to the technical field of network computing resource allocation. Terminal equipment generates the deep neural network reasoning task and obtains current state data of the terminal equipment; sending the deep neural network reasoning task and the current self state data to a server; the server calculates and pre-allocates an unloading scheme by maximizing task completion precision on the premise of avoiding sudden drop of electric quantity, namely keeping an energy queue of the equipment stable, based on information acquired by the equipment. According to the invention, through intelligent unloading and dynamic fine tuning strategies, unnecessary data transmission and calculation overhead are reduced, and the overall energy efficiency is improved; key data are transmitted preferentially based on feature map importance sorting, bandwidth occupation is reduced, and reasoning precision is improved; based on a server-side real-time evaluation and transmission strategy self-adaptive adjustment mechanism, a transmission strategy is dynamically optimized, and the data transmission stability and the task completion precision are ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of network computing resource configuration, and in particular to a dynamic energy-saving offloading method and system for deep neural network reasoning tasks. Background Art

[0002] With the recent development of artificial intelligence technology, the demand for AI services, represented by Deep Neural Network (DNN) reasoning tasks, has surged. In practical applications, network edge devices (such as autonomous vehicles and industrial quality inspection sensors) are often limited by local computing power and battery life, making it difficult to independently support high-load task processing for long periods of time. While edge servers possess strong computing power, relying entirely on centralized computing will lead to response timeouts or reduced reliability due to wireless channel latency jitter and bandwidth resource contention. A viable solution at this stage is to use mobile edge computing platforms to provide computing task offloading services for devices to reduce computing latency and energy consumption on terminals.

[0003] In real-world applications, edge intelligence systems must continuously perform a variety of critical tasks, such as object detection and traffic sign recognition in autonomous driving, and visual defect detection in industrial quality inspection. These tasks often face uncertainties such as rapidly changing environmental conditions and dynamically evolving data distributions. This leads to the emergence of a class of sudden, infrequent, yet highly valuable samples during the inference process. These high-value samples often carry critical scenario information, such as abnormal conditions, extreme conditions, or rare events. For example, rare traffic signs and unexpected pedestrians in autonomous driving, as well as low-probability but critical structural cracks in industrial quality inspection, all represent sudden, high-value data requiring immediate response. These samples place higher demands on model performance and system response time. These samples not only directly impact inference accuracy but also pose challenges to the transmission scheduling and energy management of offloading strategies.

[0004] Existing offloading solutions for DNN tasks often employ static resource allocation mechanisms, often focusing on short-term energy optimization or fixed accuracy constraints, and lack in-depth modeling of dynamic environments and data heterogeneity. In actual terminal operation, devices not only face continuously changing channel conditions and long-term battery consumption, but also often need to distinguish between the value differences of task samples. Transmission strategies based on the assumption of homogeneous data struggle to implement targeted protection and transmission scheduling for high-value samples, leading to problems such as reduced inference accuracy and resource waste. Existing energy-saving offloading methods for DNN inference tasks lack a dynamic mechanism to link device long-term battery life and inference performance in terms of energy management, making it impossible to adaptively adjust energy allocation strategies based on task requirements to balance battery life and accuracy goals. Existing offloading strategies often rely on static resource scheduling schemes (such as fixed energy thresholds or single-task optimization). While these strategies can achieve short-term energy control, they struggle to cope with dynamic, time-varying wireless channel fluctuations and the continuous degradation of device batteries, easily leading to sudden battery drops or task interruptions. Furthermore, most solutions design transmission rules based on the assumption of homogeneous data, lacking adaptive transmission protection mechanisms for unusual scenarios or rare events. This fails to guarantee transmission reliability when the channel degrades or when sample complexity fluctuates significantly, leading to reduced model accuracy or additional energy consumption due to retransmissions. Existing technologies have yet to effectively address the problem of co-optimizing dynamic resource allocation and task reliability under long-term energy constraints, limiting the potential application of deep neural networks in complex and dynamic edge scenarios. Summary of the Invention

[0005] The purpose of the present invention is to provide a dynamic energy-saving offloading method and system for deep neural network reasoning tasks to solve at least one technical problem existing in the above-mentioned background technology.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] In a first aspect, the present invention provides a dynamic energy-saving offloading method for deep neural network inference tasks, comprising:

[0008] The terminal device generates a deep neural network inference task, obtains its current state data, and sends the deep neural network inference task and its current state data to the server.

[0009] Based on the information obtained from the device, the server calculates and pre-allocates the offloading plan by maximizing the accuracy of task completion while avoiding sudden power drops and maintaining a stable device energy queue.

[0010] As a further limitation of the first aspect of the present invention, it includes: the terminal device executes DNN local sub-model inference according to the allocated offloading scheme; the terminal device analyzes the intermediate layer feature map of the DNN segmentation point position and sorts it according to importance.

[0011] As a further limitation of the first aspect of the present invention, after the local inference of the terminal device is completed, the channel status is monitored and updated in real time and uploaded to the server; the server dynamically adjusts the device transmission power pre-allocated before local calculation based on the channel status and task value; the terminal device unloads part of the feature map to the server within the time window when the channel status is stable based on the optimized transmission scheme.

[0012] As a further limitation of the first aspect of the present invention, the server receives the feature maps unloaded by the device and evaluates the inference performance indicators of the current cumulative received data in real time; the server determines whether the terminal device needs to transmit more feature maps based on the sample complexity evaluation result: if the cumulative feature maps meet the preset performance threshold, or the cumulative number of feature maps has reached the maximum number of feature maps that can be transmitted, the feature map transmission is terminated; otherwise, the server requests the device to upload more feature maps.

[0013] As a further limitation of the first aspect of the present invention, the server returns the evaluation result to the device, and the device continues or stops transmission according to the instruction; if transmission continues, the server will continuously monitor the device channel status and adjust the transmission strategy in real time.

[0014] As a further limitation of the first aspect of the present invention, edge inference, result feedback and energy queue update include: the server executes DNN edge sub-model inference; the server completes the final inference and returns the inference result to the terminal device; the terminal device updates the energy queue according to its own energy consumption status of this task to optimize the offloading plan for future tasks.

[0015] In a second aspect, the present invention provides a dynamic energy-saving offloading system for deep neural network reasoning tasks, comprising:

[0016] The terminal device is configured to: generate a deep neural network inference task, obtain current state data of the terminal device; and send the deep neural network inference task and current state data of the terminal device to the server;

[0017] The server is configured to calculate and pre-allocate an offloading plan based on the information obtained from the device by maximizing the accuracy of task completion while avoiding a sudden drop in power, that is, maintaining a stable energy queue of the device.

[0018] In a third aspect, the present invention provides a non-transitory computer-readable storage medium, which is used to store computer instructions. When the computer instructions are executed by a processor, the dynamic energy-saving offloading method of the deep neural network inference task as described in the first aspect is implemented.

[0019] In a fourth aspect, the present invention provides a computer device comprising a memory and a processor, wherein the processor and the memory communicate with each other, the memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to execute the dynamic energy-saving unloading method for deep neural network inference tasks as described in the first aspect.

[0020] In a fifth aspect, the present invention provides an electronic device comprising: a processor, a memory, and a computer program; wherein the processor is connected to the memory, and the computer program is stored in the memory. When the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to execute instructions for implementing the dynamic energy-saving unloading method for deep neural network inference tasks as described in the first aspect.

[0021] The beneficial effects of this invention include: reducing unnecessary data transmission and computing overhead and improving overall energy efficiency through intelligent offloading and dynamic fine-tuning strategies; prioritizing the transmission of critical data based on the importance of feature graphs, reducing bandwidth usage and improving inference accuracy; dynamically optimizing transmission strategies based on server-side real-time evaluation and adaptive adjustment of transmission strategies to ensure data transmission stability and task completion accuracy; and dynamically updating the energy queue on the device side to ensure long-term energy consumption optimization, extend the system's continuous working time, and improve reliability. This solution is suitable for scenarios such as autonomous driving and industrial quality inspection, and can effectively improve the execution efficiency and robustness of edge intelligent inference tasks, with important application value in the field of edge intelligent computing.

[0022] Additional advantages of the present invention will be more clearly given in the following description or learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0024] Figure 1 This is a diagram of the mobile edge computing architecture described in an embodiment of the present invention.

[0025] Figure 2 This is a system architecture diagram of the deep neural network task energy-saving offloading method described in an embodiment of the present invention.

[0026] Figure 3 This is an overall flow chart of the energy-saving offloading method for deep neural network tasks according to an embodiment of the present invention.

[0027] Figure 4This is a flow chart of information transmission according to an embodiment of the present invention. DETAILED DESCRIPTION

[0028] The embodiments of the present invention are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention and are not to be construed as limiting the present invention.

[0029] Those skilled in the art will understand that unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by those skilled in the art to which this invention belongs.

[0030] It should also be understood that terms, such as those defined in commonly used dictionaries, should be understood to have a meaning consistent with their meaning in the context of the prior art and will not be interpreted in an idealized or overly formal sense unless as defined herein.

[0031] Those skilled in the art will appreciate that, unless otherwise stated, the singular forms "a," "an," "said," and "the" used herein may also include plural forms. It should be further understood that the term "comprising" used in the specification of the present invention refers to the presence of the stated features, integers, steps, operations, elements, and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, and / or groups thereof.

[0032] In the description of this specification, reference to the terms "one embodiment," "some embodiments," "examples," "specific examples," or "some examples" means that the specific features, structures, materials, or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. Moreover, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples. Those skilled in the art may combine and integrate different embodiments or examples described in this specification, as well as features of different embodiments or examples, unless otherwise contradictory.

[0033] To facilitate understanding of the present invention, the present invention is further explained below with reference to specific embodiments in conjunction with the accompanying drawings. However, the specific embodiments do not constitute a limitation on the embodiments of the present invention.

[0034] Those skilled in the art should understand that the drawings are merely schematic diagrams of embodiments, and the components in the drawings are not necessarily necessary for implementing the present invention.

[0035] The present invention proposes a dynamic energy-saving offloading scheme for deep neural network inference tasks, aiming to achieve a dynamic balance between inference accuracy and high-value sample transmission efficiency under the premise of long-term device energy consumption stability. The present invention introduces a sample value-aware adaptive offloading mechanism, adapts to task samples of different complexities through progressive feature transmission and real-time inference feedback, and improves the recognition and response capabilities of important samples. It also constructs a device battery energy queue to dynamically allocate offloading resources, stabilize long-term energy consumption levels, and avoid sudden power drops. This solution synergistically optimizes data scheduling and energy consumption regulation, providing a solution path with long-term performance guarantees for edge intelligent inference in dynamic environments.

[0036] Example 1

[0037] In mobile edge computing, smart devices can offload AI services to nearby edge servers (such as base stations) through wireless networks to reduce terminal computing delays and energy consumption. Its architecture is as follows: Figure 1 As shown. Within this framework, edge intelligent systems need to perform neural network reasoning in real time, including key tasks such as target detection in autonomous driving and visual defect recognition in industrial quality inspection. Therefore, in a dynamic environment, how to strike a dynamic balance between the long-term energy consumption stability of the equipment and the inference accuracy, while achieving progressive and accurate transmission of high-value data, is the core challenge facing the design of edge computing offloading strategies. Traditional transmission strategies based on static statistical characteristics and isomorphism assumptions are difficult to adapt to the efficient perception and energy-efficient scheduling of sudden, high-value samples in complex scenarios. Therefore, building a long-term energy consumption-accuracy dynamic balance mechanism that integrates data value perception and progressive offloading has become the key to achieving efficient and energy-saving offloading decisions.

[0038] In this embodiment, a dynamic energy-saving offloading method for deep neural network reasoning tasks is proposed. The specific system architecture is as follows: Figure 2 As shown, it includes an edge server and a terminal device. The edge server is responsible for edge reasoning, resource scheduling and reasoning performance monitoring, while the terminal device is responsible for task generation, environmental detection, local reasoning and energy queue perception. The terminal device communicates with the server through a wireless network, shares real-time device information and selectively uploads intermediate feature maps. The system monitors the device status (battery power, channel, etc.) when the task arrives in real time, evaluates the current energy consumption of the device by building an energy queue, and allocates resources under the premise of stabilizing the long-term average energy consumption of the device. At the same time, an innovative feature map number adaptive transmission mechanism is established for tasks of different complexity, thereby significantly improving the system's adaptability to environmental changes and rare samples while stabilizing the energy consumption of the device.

[0039] The system design introduces an adaptive hierarchical and progressive sample transmission mechanism based on task value to optimize communication and energy consumption while ensuring inference accuracy. Specifically, after the device generates a task, the server first analyzes the current device environment and task requirements, performs resource estimation, and deploys an adaptive DNN model solution. After the device performs preliminary inference, the intermediate layer feature maps are gradually unloaded to the server side in layers according to the importance of the feature maps. The server comprehensively considers the uncertainty of inference, task requirements, and other factors to evaluate the value of the received feature maps in real time and judge the importance of the current sample. For high-value tasks, the server will appropriately extend the sample unloading time or request to supplement and unload more features to obtain more sufficient information to support the accuracy requirements; conversely, for samples whose inference results are relatively certain or the task value is low, the server can notify the device in advance to stop feature map unloading after the judgment threshold is met, thereby saving communication resources and reducing energy consumption.

[0040] The overall workflow of the dynamic energy-saving offloading method for deep neural network reasoning tasks described in this embodiment is as follows: Figure 3 As shown in the figure, this approach includes all steps, from task arrival, resource scheduling, local inference, feature map offloading, edge inference, to inference result transmission. This method relies on the computing power of the terminal device and network technology (such as 5G, Wi-Fi, etc.). The terminal device generates DNN inference tasks randomly or at a fixed frequency (e.g., every second, every 100 milliseconds). After task generation, the system analyzes the device's real-time channel environment and energy queue and pre-assigns an offloading plan (including transmission power, DNN partitioning scheme, etc.) to the device and server. The device performs local inference for the task based on the pre-assigned plan. After completing local inference, it sorts the DNN intermediate layer feature maps based on importance and offloads them to the edge server. After each reception, the server evaluates the inference performance achieved by the received feature map in real time, determines whether more feature maps are needed, and fine-tunes the device offloading strategy based on channel conditions. After transmission is complete, the server performs edge inference, and the inference results are returned to the device via the network. Finally, the device updates the energy queue based on energy usage, allowing it to adjust the offloading plan when the next task arrives, achieving long-term energy optimization.

[0041] The specific process is as follows:

[0042] Step 1: Data collection and offload strategy pre-allocation

[0043] Step 11: The terminal device generates a deep neural network inference task. Specifically, this task is usually automatically generated by device-side sensors or business systems based on a preset trigger mechanism. For example, in autonomous driving scenarios, cameras periodically capture images at a fixed frame rate and initiate detection requests, or industrial quality inspection systems generate detection tasks in real time based on items entering the inspection area on the assembly line, or in augmented reality (AR) and virtual reality (VR) applications, perception tasks are triggered in real time based on changes in user perspective. The above task generation method does not require user intervention and has periodic or event-driven characteristics, which can continuously and stably drive the subsequent inference process.

[0044] Step 12: The device obtains its current state data, taking into account the following factors: real-time channel state (including channel noise coefficient, device transmission gain, communication bandwidth, etc.); the device's accumulated task energy queue backlog, which is dynamically updated based on the energy consumption during the previous task execution; the device's computing power (such as the number of floating-point operations per second of the CPU and GPU); system resource constraints (such as maximum transmission power); the amount of computation and feature map dimensions of each layer of the DNN network; and task latency requirements.

[0045] Step 13: The device reports information and task requirements (such as whether it is an urgent task) to the server. The server formulates an offload strategy based on the information obtained from the device. This process is essentially an optimization problem. Its goal is to maximize the inference accuracy of each task under the premise of long-term energy consumption constraints. During this optimization process, the server comprehensively considers the following two key variables:

[0046] The split point location of the DNN model determines the number of neural network layers executed by the terminal device, affecting its local computing energy consumption;

[0047] Transmit power when offloading feature maps on the device: This affects the number of offloaded feature maps and the energy consumption on the device, which in turn affects inference accuracy.

[0048] The server needs to weigh the gain and resource consumption of feature map transmission, combine the task importance and the remaining energy of the device, and achieve a compromise between device energy consumption and task inference accuracy by selecting the appropriate DNN segmentation point location and transmission power to achieve the optimal balance of overall performance.

[0049] Step 14: The server sends the offloading plan to the device and deploys the sub-models after the DNN split. Specifically, the terminal device and edge server activate and run all network layers before the DNN split point according to the pre-offloading plan, and offload the generated intermediate feature maps to the edge server. The server then executes the remaining network layers after the split point.

[0050] In summary, through reasonable data collection and pre-allocation of task offloading schemes, the system can balance task accuracy, computing load, and energy consumption requirements in complex environments, optimize the utilization of computing and communication resources, and help to promptly detect emergencies such as sudden drops in device power or deterioration of channel status, thereby improving task execution efficiency and energy efficiency and achieving long-term energy saving.

[0051] Step 2: Local Inference and Feature Map Sorting

[0052] Step 21: The device performs DNN local sub-model inference according to the assigned offloading scheme.

[0053] Step 22: The device sorts the intermediate layer feature maps at the DNN segmentation point location according to the importance index. The sorting can be based on the following importance indexes: prediction uncertainty (the model has low confidence in the prediction of the feature map), gradient amplitude (the size of the gradient generated by the feature map), and the importance of the convolution kernel corresponding to the feature map (measured by indicators such as L1 norm and L2 norm, important convolution kernels output important feature maps).

[0054] In summary, local reasoning combined with the importance ranking strategy of feature graphs can screen out the feature graphs that contribute most to the reasoning results for subsequent data transmission and edge reasoning, enabling the device to prioritize the processing and transmission of these feature graphs under limited computing and communication resources, reducing redundant data transmission, improving communication bandwidth utilization, and improving the overall efficiency of the system.

[0055] Step 3: Fine-tune the transfer strategy and implement progressive uninstallation

[0056] Step 31: After the device completes local inference, it monitors and updates the channel status in real time and uploads it to the server. Based on the channel status and task value (including task urgency and sample complexity), the server dynamically adjusts the device transmission power pre-allocated before local calculation.

[0057] Step 32: Based on the optimized transmission scheme, the device unloads part of the feature map to the server within the time window when the channel state is stable.

[0058] Step 33: The server receives the feature maps offloaded from the device and evaluates the inference performance metrics of the currently received cumulative data in real time. This evaluation method can utilize a pre-trained small MLP prediction network. To train this prediction network, the system simultaneously collects training data while the DNN is training for the target downstream task. Specifically, a sample is selected from the training set, and its intermediate feature map after passing through the DNN model split point (i.e., the last layer executed on the device) is recorded as input features. Simultaneously, this sample is fed into the full DNN model and inference is completed. The inference uncertainty for this sample is calculated from the final inference output. This process forms the feature map after the split point in the training set and its corresponding inference uncertainty. Inference uncertainty can be measured using the probability distribution of the model's final output. Specifically, after inference, the model assigns a predicted probability to each possible category, indicating the likelihood that the input sample belongs to that category. If the model is highly confident in a category, the predicted probability will be concentrated on one option; if the model is uncertain, the probability distribution will be more dispersed. Generally speaking, the more dispersed the predictions are, the higher the uncertainty is, which means the inference result is more uncertain; conversely, if the predictions are highly concentrated on a certain category, it means that the model is more confident about the prediction.

[0059] The MLP network is a shallow, fully connected neural network that outputs a scalar value, used to predict the expected uncertainty of the feature map for the current task. The network is trained using the Mean Sequential Error (MSE) loss, with the optimization objective of minimizing the error between the predicted value output by the MLP and the true uncertainty. After training, the network is deployed on the server. In practical applications, uncertainty predictions can be made for collected samples of varying complexity (complex samples have greater uncertainty, and more data may be required during transmission to achieve a preset performance threshold).

[0060] The server determines whether the device needs to transmit more feature maps based on the sample complexity evaluation results: if the cumulative feature map meets the preset performance threshold, or the cumulative number of feature maps has reached the maximum number of feature maps that can be transmitted, the feature map transmission is terminated; if the above is not met, the server requests the device to upload more feature maps

[0061] Step 34: The server returns the evaluation results to the device, which then proceeds to transmit or stops transmission based on the instructions. If transmission continues, the server continuously monitors the device's channel status and adjusts the transmission strategy in real time.

[0062] In summary, through real-time dynamic fine-tuning of the transmission strategy, the system can effectively compensate for the non-real-time nature of transmission power decisions during the initial offloading strategy pre-allocation phase while maintaining stable energy consumption, thereby more accurately adapting to changes in the current channel environment. Preset thresholds can be dynamically adjusted based on the importance of the application scenario and resource constraints, reducing data transmission redundancy, reducing the system's communication overhead, and improving the system's robustness to complex samples. This adaptive adjustment mechanism not only improves transmission efficiency but also maintains superior task completion accuracy despite fluctuations in channel quality and sample complexity.

[0063] Step 4: Edge inference, result feedback, and energy queue update

[0064] Step 41: The server performs inference on the DNN edge sub-model. Specifically, after receiving the intermediate feature map uploaded by the device, the server uses it as input data, loads the corresponding subsequent sub-model (i.e., from the DNN segmentation layer to the last layer), and performs forward propagation calculations. This inference process includes performing network operations such as convolution, normalization, activation, and pooling layer by layer until the final prediction result is output.

[0065] Step 42: The server completes the final inference and returns the inference result to the device.

[0066] Step 43: The device updates its energy queue based on the current task's energy consumption to optimize future task offloading strategies. By dynamically updating the device's energy status, the system can adaptively adjust task offloading strategies, achieving long-term energy optimization. This prevents task failures due to energy depletion or excessive charging during device operation, extending overall system operation time and improving reliability.

[0067] In this embodiment, the five key information exchange steps required to implement the above method are as follows: Figure 4 As shown in the figure, the device uploads status information to the server, the server analyzes and sends the uninstallation plan, the device performs local inference and uninstalls the feature map, and updates the real-time status, the server evaluates the inference performance, analyzes the device status and adjusts the transmission strategy, and the server completes the inference and returns the result.

[0068] Example 2

[0069] This embodiment 2 provides a non-transitory computer-readable storage medium, which is used to store computer instructions. When the computer instructions are executed by a processor, the dynamic energy-saving offloading method of the deep neural network inference task as described above is implemented.

[0070] Example 3

[0071] This embodiment 3 provides a computer device, including a memory and a processor, the processor and the memory communicate with each other, the memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to execute the dynamic energy-saving unloading method for deep neural network inference tasks as described above.

[0072] Example 4

[0073] This embodiment 4 provides an electronic device, including: a processor, a memory, and a computer program; wherein the processor is connected to the memory, and the computer program is stored in the memory. When the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to execute instructions for implementing the dynamic energy-saving offloading method for deep neural network inference tasks as described above.

[0074] It will be understood by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0075] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0076] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0077] These computer program instructions can also be loaded onto a computer or other programmable data processing device, and a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, so that the instructions executed on the computer or other programmable device provide the functions for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0078] Although the above describes the specific embodiments of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solutions disclosed in the present invention without the need for creative work should be included in the scope of protection of the present invention.

Claims

1. A dynamic energy-saving offloading method for deep neural network reasoning tasks, characterized in that: include: The terminal device generates a deep neural network inference task and obtains its current state data; And send the deep neural network inference task and current state data to the server; Based on the information obtained from the device, the server calculates and pre-allocates the offloading plan by maximizing the accuracy of task completion while avoiding sudden power drops and maintaining a stable device energy queue.

2. The dynamic energy-saving offloading method for deep neural network reasoning tasks according to claim 1 is characterized in that: include: The terminal device executes DNN local sub-model inference according to the assigned offloading plan; The terminal device analyzes the intermediate layer feature maps of the DNN segmentation point locations and sorts them according to importance.

3. The dynamic energy-saving offloading method for deep neural network reasoning tasks according to claim 2 is characterized in that: After the local inference of the terminal device is completed, the channel status is monitored and updated in real time and uploaded to the server; based on the channel status and task value, the server dynamically adjusts the device transmission power pre-allocated before local calculation; based on the optimized transmission plan, the terminal device unloads some feature maps to the server within the time window when the channel status is stable.

4. The dynamic energy-saving offloading method for deep neural network reasoning tasks according to claim 3 is characterized in that: The server receives the feature maps unloaded by the device and evaluates the inference performance indicators of the current cumulative received data in real time; the server determines whether the terminal device needs to transmit more feature maps based on the sample complexity evaluation results: if the cumulative feature map meets the preset performance threshold, or the cumulative number of feature maps has reached the maximum number of feature maps that can be transmitted, the feature map transmission is terminated; otherwise, the server requests the device to upload more feature maps.

5. The dynamic energy-saving offloading method for deep neural network reasoning tasks according to claim 4 is characterized in that: The server returns the evaluation results to the device, and the device continues or stops transmission according to the instructions; if transmission continues, the server will continuously monitor the device channel status and adjust the transmission strategy in real time.

6. The dynamic energy-saving offloading method for deep neural network reasoning tasks according to claim 5 is characterized in that: Edge inference, result feedback, and energy queue update include: the server executes DNN edge sub-model inference; the server completes the final inference and returns the inference result to the terminal device; the terminal device updates the energy queue based on its own energy consumption status of this task to optimize the offloading plan for future tasks.

7. A dynamic energy-saving offloading system for deep neural network reasoning tasks, characterized by: include: The terminal device is configured to: generate a deep neural network inference task and obtain its current state data; And send the deep neural network inference task and current state data to the server; The server is configured to calculate and pre-allocate an offloading plan based on the information obtained from the device by maximizing the accuracy of task completion while avoiding a sudden drop in power, that is, maintaining a stable energy queue of the device.

8. A non-transitory computer-readable storage medium, characterized in that The non-transitory computer-readable storage medium is used to store computer instructions. When the computer instructions are executed by the processor, the dynamic energy-saving offloading method for the deep neural network reasoning task according to any one of claims 1 to 6 is implemented.

9. A computer device, characterized in that: It includes a memory and a processor, the processor and the memory communicate with each other, the memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to execute the dynamic energy-saving unloading method for the deep neural network reasoning task as described in any one of claims 1 to 6.

10. An electronic device, characterized in that: include: A processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to enable the electronic device to execute instructions for implementing the dynamic energy-saving offloading method for deep neural network inference tasks as described in any one of claims 1-6.

Images