Server-agnostic continual learning system for multi-model video analytics

Abstract

This invention discloses a server-agnostic continuous learning system for multi-model video analysis, comprising at least a data drift detector, a retraining planner, and a resource harvesting scheduler. In the data drift detector, an accuracy distribution model within a non-drift window is constructed offline. During online inference, the detection window is dynamically adjusted and compared with the accuracy distribution model to accurately identify data drift. In the retraining planner, based on multi-model correlation analysis, the retraining priority is determined by analyzing the correlation of inference accuracy among multiple models, prioritizing the model with the greatest accuracy gain for retraining. Finally, in the resource harvesting scheduler, the keep-alive period resources of the server-agnostic function are fully utilized to dynamically schedule retraining tasks, and the efficient sharing of model weights is optimized through a memory file system, achieving high accuracy of service requests and high efficiency in resource utilization.

Description

Technical Field

[0001] This invention belongs to the technical field of computer information computing, and particularly relates to a multi-model continuous learning technology, mainly involving a server-insensitive continuous learning system for multi-model video analysis. Background Technology

[0002] As video gradually becomes the mainstream media format in network and IoT applications, multi-model video analytics has been widely applied in fields such as multimedia search, advertising recommendation, and real-time monitoring. These applications utilize directed acyclic graphs composed of deep neural network models to achieve complex tasks such as object detection, behavior recognition, and classification. However, changes in the distribution of input data in dynamic environments, i.e., data drift, often lead to a significant decline in model performance, necessitating effective countermeasures. Continuous learning techniques, which address the data drift problem by retraining models, have become an important research direction in recent years. Furthermore, server-insensitive computing, due to its on-demand resource allocation and flexible scalability, offers new opportunities for continuous learning, demonstrating significant potential in improving resource efficiency and coping with dynamic load fluctuations.

[0003] Current research progress mainly focuses on data drift detection and model retraining optimization. Traditional methods use fixed time windows or adaptive thresholds to detect data drift and trigger the retraining process of affected models. By responding promptly to drift data, these methods have improved the accuracy of video analysis to some extent. However, the potential of utilizing idle resources to perform retraining tasks in server-agnostic computing scenarios has not been fully explored. Furthermore, research has found significant inter-model correlations in multi-model directed acyclic graphs (DAGs), meaning that retraining some models can significantly restore overall performance. However, existing methods typically employ a strategy of retraining all models affected by data drift, neglecting optimization opportunities through correlation analysis. These findings lay the foundation for improving continuous learning efficiency and resource utilization.

[0004] Despite these challenges, existing technologies still face numerous obstacles. First, traditional data drift detection methods are susceptible to input fluctuations, resulting in high false positive rates and wasted resources or retraining delays. Second, retraining strategies lack analytical methods for the correlation of directed acyclic graphs (DAGs) across multiple models, leading to unnecessary increases in resource consumption. Furthermore, resource scheduling is inadequate in dynamic request environments, failing to fully utilize the server-insensitive keep-alive resources for task optimization. Finally, as the complexity of DAG structures increases, existing frameworks struggle to coordinate model priorities and retraining order, making it difficult to balance computational cost and accuracy. These issues reveal the bottlenecks of continuous learning in practical applications and also point the way for further optimization of future technologies. Summary of the Invention

[0005] This invention addresses the shortcomings of existing technologies by providing a server-agnostic continuous learning system for multi-model video analysis, comprising at least a data drift detector, a retraining planner, and a resource harvesting scheduler. In the data drift detector, a precision distribution model is constructed based on historical data without data drift, and the detection window is dynamically adjusted to accurately identify data drift, significantly reducing the false positive rate. In the retraining planner, based on multi-model correlation analysis, the retraining priority is determined by analyzing the correlation of inference precision among multiple models, prioritizing the model with the greatest precision gain for retraining, thus avoiding resource waste caused by retraining all models affected by data drift. Finally, in the resource harvesting scheduler, resource collection is central, fully utilizing the keep-alive period resources of the server-agnostic function to dynamically schedule retraining tasks, and optimizing the efficient sharing of model weights through a memory file system, achieving accurate service requests and efficient resource utilization.

[0006] To achieve the above objectives, the technical solution adopted by this invention is: a server-insensitive continuous learning system for multi-model video analysis, comprising at least a data drift detector, a retraining planner, and a resource harvesting scheduler.

[0007] The data drift detector: It determines changes in input content by dynamically adjusting the detection window size and monitoring accuracy fluctuations, thereby reducing false alarms; The detector captures the difference between online inference accuracy and accuracy without data drift by comparing the distribution model of online inference accuracy and accuracy without data drift within the adaptive detection window, thus determining the occurrence of data drift; This method can resist transient or irrelevant accuracy fluctuations caused by input changes, improving the accuracy of data drift detection;

[0008] The retraining planner selectively retrains models by analyzing the inference and retraining correlations between different models in a multi-model system. The core of the retraining planner is that if there is a correlation between two models, retraining one model may simultaneously improve the accuracy of the other model. Therefore, the retraining planner reduces unnecessary model retraining costs, saves resources, and improves the accuracy of the overall application.

[0009] The resource harvesting scheduler fully utilizes idle resources during the keep-alive period of the serverless function, collecting and using these resources for model retraining, thereby improving resource utilization efficiency. Furthermore, the resource harvesting scheduler can cope with dynamically changing resource demands, especially when facing constantly changing model retraining loads and keep-alive periods. By constructing idle resources during the keep-alive period of the serverless function into a directed acyclic graph, dynamic and intermittent resources can be continuously and effectively utilized, while also sharing model weights.

[0010] As an improvement of the present invention, the sliding range of the detection window size in the data drift detector is:

[0011]

[0012] in, Indicates the first window size, Indicates the maximum detection window size. Indicates that Mapped between 0 and 1 Indicates the average precision within the window. This is a hyperparameter that indicates the rate at which the window size decreases.

[0013] As another improvement of the present invention, the data drift detector determines whether data drift has occurred by including the following steps:

[0014] S101: Calculate at time point Detection window maximum normalized precision gradient on:

[0015]

[0016] in, Indicates in the detection window The maximum range of internal inference accuracy, Indicates at a point in time and The detection window inside, Indicates at a point in time and The sliding timestamp within, Indicates the first The first detection window, the first Inference precision under each timestamp Indicates the first The first detection window, the first Inference precision under a timestamp;

[0017] S102: Generate a gradient distribution model for maximum normalized precision values The model is a detection window. At the point of time At the appointed time Distribution of maximum normalized precision gradients for different timestamps during the period:

[0018]

[0019] Among them, time points , , These represent the detection windows. The start time, sliding timestamp, and end time within;

[0020] S103: The above formula models the distribution of the maximum accuracy amplitude without data drift offline under different window sizes. When performing online inference, this detection window can be obtained. Maximum inference accuracy range Furthermore, by combining it with the precision distribution model... contrast Maximum inference accuracy range Exceeding precision distribution model When the fluctuation range is within a certain range, data drift occurs in the detection window;

[0021] S104: When data drift occurs, the system triggers the retraining process for the model and adds the models that need retraining to the retraining set in sequence. In the middle, when When it is an empty set, it represents the detection window. No data drift occurred.

[0022] As another improvement of the present invention, the selection of the optimal preceding model in the retraining planner includes the following steps:

[0023] S201: Calculate the Kullback-Leibler (KL) divergence of the inference accuracy distribution among multiple models to identify model pairs with strong inference correlation;

[0024]

[0025] in, Display window Internal Model Correlation of inference accuracy between them and Representing the model respectively and In the Within each window, timestamps The changing distribution of reasoning accuracy Indicates the total time of the detection window;

[0026] S202: Select the most relevant preceding model obtained in step S201;

[0027]

[0028] The preceding model represents the one with the greatest accuracy gain.

[0029] As another improvement of the present invention, in the retraining planner, maximizing the retraining gain by determining the retraining strategy includes the following steps:

[0030] First, in the budget constraints submitted by the user Precision constraints Next, the preceding model selected in step S202 is used with a strategy Perform selective retraining to maximize retraining gain.

[0031]

[0032] Wherein, constraint (1) indicates that the average accuracy of all detection windows must meet the user's accuracy constraint. , Represents the total number of detection windows; constraint (2) represents the computational cost required to complete the inference and retraining tasks to meet the user's budget constraint. , Indicates the consumption during retraining. Price per second This represents the total amount of resources consumed during the retraining process. This indicates the total execution time after rounding to the nearest billing unit during retraining;

[0033] In addition, retraining gain Indicates the first Within each window, from the retraining set Select the optimal set of retrained models And by determining the number of retraining samples and number of training sessions Retrain:

[0034]

[0035] in, Function representation system with retraining strategy The process of retraining.

[0036] As another improvement of the present invention, a heuristic retraining strategy selection algorithm is designed in the retraining planner to iteratively update... Determine the retraining strategy To further solve the above optimization problem:

[0037] To determine the optimal set of retrained models Within each detection window, from the retraining set The preceding model with the highest accuracy gain is selected to determine... In addition to satisfying the two constraints (1) and (2), when the accuracy gain ratio of two adjacent detection windows is When, take it as a certainty One of the conditions for achieving maximum accuracy gain; in order to determine The method uses Euclidean distance to identify the retraining samples with the greatest differences in adjacent detection windows.

[0038] As a further improvement of the present invention, the method for sharing model weights in the resource harvesting scheduler is as follows: a virtual memory file system is created using tmpfs and mounted to each retraining task to realize the updating of model weights and isolated access to multiple models.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] (1) In the design of the continuous learning system, this invention proposes a new system that combines drift detection, selective retraining and efficient resource scheduling. By integrating a data drift detector with robust fluctuations, a retraining planner based on multi-model correlation analysis and a resource harvesting scheduler centered on resource collection, higher inference accuracy and lower resource consumption are achieved.

[0041] (2) The system of the present invention includes a fluctuation-robust data drift detector. The detector accurately identifies data drift by dynamically adjusting the detection window and combining it with a prediction model, thereby avoiding misjudgment caused by input fluctuation in traditional detection methods and significantly reducing unnecessary retraining tasks.

[0042] (3) The system of the present invention is different from the existing system that directly retrains all data drift models. The present invention is based on the inference correlation between multiple models and includes a retraining planner based on multi-model correlation analysis. By analyzing the mutual influence between models, only the key models are retrained, which effectively reduces resource overhead and ensures the improvement of overall accuracy.

[0043] (4) The system of the present invention includes a resource harvesting scheduler centered on resource collection, which makes full use of the keep-alive resources in the server's non-aware function, dynamically schedules idle resources to support retraining tasks by constructing a directed acyclic graph of available idle resources, and optimizes the resource sharing and security of tasks by using a memory file system.

[0044] (5) By making efficient use of the server’s non-perceptual function resources, this invention optimizes the triggering and execution of retraining tasks in continuous learning, achieving a significant breakthrough in the balance between accuracy and efficiency, and providing a new technical path for continuous learning in video analysis scenarios. Attached Figure Description

[0045] Figure 1 This is a schematic diagram of the server-insensitive continuous learning system for multi-model video analysis according to the present invention. Detailed Implementation

[0046] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.

[0047] Example 1

[0048] A server-insensitive continuous learning system for multi-model video analytics, such as... Figure 1 As shown, users submit video frames to the system, and the data distribution of these video frames changes over time, i.e., data drift. Developers pre-deploy a multi-model video analytics application and service constraints on the system. The multi-model video analytics application is built in the form of a directed acyclic graph and runs through a server-insensitive function, where the... The reasoning process of each model is represented by symbols. The training process is represented by symbols. Indicates. For example... Figure 1 As shown, the multi-model video analytics application includes six models: #1, #2, #3, #4, #5, and #6. Service constraints include accuracy constraints and cost constraints. During online inference, the system collects resources from the server's non-aware functions during the keep-alive period and uses these resources according to the priority of retraining tasks. The system includes at least a data drift detector, a retraining planner, and a resource harvesting scheduler to address the trade-off between accuracy loss and performance degradation caused by data drift.

[0049] Data Drift Detector: This detector employs a robust dynamic detection mechanism to avoid misjudgments caused by input fluctuations in traditional methods. Its aim is to accurately identify accuracy degradation caused by data drift and prevent misjudgments due to input fluctuations. The detector takes as input data the changes in inference accuracy and historical accuracy amplitudes before data drift, and outputs the determination of whether drift has occurred. First, the detector dynamically adjusts the detection range using an adaptive window, based on the average accuracy of the previous window and... The function calculates the current window size to balance detection accuracy and response speed. Secondly, by capturing the distribution of historical inference accuracy changes, the detector constructs a precision distribution model without data drift, which is then compared with the precision changes in future windows. Specifically, within the non-drift window, the detector calculates the magnitude of the inference accuracy change using the maximum normalized gradient and incorporates its distribution into the precision distribution model. If the fluctuation range of online inference accuracy exceeds the normal fluctuation range of this precision distribution model, it is determined that data drift has occurred. Through this process, the drift detector can significantly improve detection accuracy while reducing false positives and unnecessary resource consumption.

[0050] First use continuous windows Content division Each window The content inside is represented as , Indicates the first Inference accuracy of each window, Indicates the first Window size. Set the sliding range for detecting window size. ,in Indicates the maximum detection window size:

[0051]

[0052] in, Indicates the first The average precision within a window, when the difference The smaller the size, according to Function properties Smaller; hyperparameters This indicates the rate at which the window size decreases.

[0053] In a fluctuation-robust data drift detector, the difference between inference accuracy and normal accuracy fluctuation is compared by using an accuracy distribution model that has not experienced data drift, thereby determining whether data drift has occurred.

[0054] S101: Calculate at a certain point in time Detection window Maximum normalized precision gradient on:

[0055]

[0056] in, Indicates in the detection window The range of internal inference accuracy is normalized to reduce the impact of different dimensions of multiple task models within an application.

[0057] S102: Generate a representative detection window from arrive Maximum normalized precision gradient distribution at different time points:

[0058] ;

[0059] S103: The above formula models the distribution of the maximum precision amplitude without data drift offline under different window sizes. This detection window can be obtained during online inference. Maximum inference accuracy range Furthermore, by comparing it with the accuracy distribution model... contrast When the maximum inference accuracy amplitude Exceeding precision distribution model When the fluctuation range is within a certain range, data drift occurs in the detection window.

[0060] S104: When data drift occurs, the system triggers the retraining process for the model and adds the models that need retraining to the retraining set in sequence. In the middle. When When empty, it indicates a detection window. No data drift occurred.

[0061] The retraining planner analyzes the inference correlations between models to select the optimal set of retrained models, thereby reducing resource overhead and improving accuracy. The input is the directed acyclic graph structure of the multiple models and their inference accuracy distribution; the output is a list of retrained models and their priorities. Redundant models are removed while ensuring accuracy recovery, and key models are prioritized for retraining. This method effectively reduces resource consumption while avoiding resource contention between inference and retraining.

[0062] First, by calculating the Kullback-Leibler (KL) divergence of the inference accuracy distribution among models, model pairs with strong inference correlation are identified. Next, the preceding model with the strongest correlation is selected for retraining, as prioritizing the training of this preceding model provides the greatest accuracy gain. By prioritizing multiple models, the most important models are gradually selected for retraining, ensuring that resources are allocated to key models first. Finally, considering budget and time constraints, a retraining sampling strategy and stopping criteria are set to ensure a balance between resources and accuracy, achieving resource-efficient selective retraining and avoiding the resource waste of full-model retraining.

[0063] In the retraining planner, the correlation between multiple models is determined by analyzing the changes in inference accuracy between models. Since multi-model video analytics applications are constructed from multiple models in the form of a directed acyclic graph, retraining the preceding model has a corresponding accuracy gain for the subsequent model; then, the preceding model with the largest accuracy gain is selected for priority training to improve retraining efficiency.

[0064] S201, Inference Relevance: Calculate the similarity of inference accuracy distributions among multiple models using Kullback-Leibler (KL) divergence.

[0065]

[0066] in, Display window Internal Model Correlation of inference accuracy between them and Representing the model respectively and In the Within each window, timestamps The changing distribution of reasoning accuracy;

[0067] S202. Selecting the preceding model with the largest accuracy gain: Since multi-model video analysis applications are constructed from multiple models in the form of a directed acyclic graph, retraining the preceding model provides a corresponding accuracy gain for subsequent models. Therefore, this system prioritizes training the preceding model with the largest accuracy gain. The following formula characterizes this process, addressing the uncertainty of subsequent models;

[0068]

[0069] in, The preceding model represents the one with the greatest accuracy gain. Therefore, retraining it first can not only improve its own accuracy, but also propagate the improvement in accuracy to its related subsequent models, thereby reducing unnecessary retraining.

[0070] In the retraining planner, the preceding model selected in step S202 is retrained by combining the accuracy and budget constraints submitted by the user, setting the retraining sampling strategy and stopping criteria.

[0071] Model retraining under constraints: The correlation between models can be used to plan a continuous learning scheme and avoid resource contention. Therefore, the goal of a server-aware continuous learning system is to retrain within the user-submitted budget constraints. Precision constraints Below, with strategy Perform selective retraining to maximize retraining gain. :

[0072]

[0073] Wherein, constraint (1) indicates that the average accuracy of all detection windows needs to meet the user's accuracy constraint. Constraint (2) indicates that the computational cost required to complete the inference and retraining tasks needs to meet the user's budget constraint.

[0074] This system follows the pricing calculation standard of AWS Lambda (Amazon Server Invisible Platform), and under constraint (2), Is every Price per second It refers to the amount of resources consumed based on memory allocation. It is the execution time after rounding to the nearest billing unit.

[0075] Determine the retraining strategy: retraining gain Indicates the first Within each window, from the retraining set Select the optimal set of retrained models And by determining the number of retraining samples and number of training sessions Retrain:

[0076]

[0077] In the retraining planner, iterative updates Determine the retraining strategy Then, we can further solve the optimization problem mentioned above.

[0078] In order to determine Using step S202, within each round of detection window, the retrained set is... Select the preceding model with the highest accuracy gain; in order to determine In addition to satisfying the constraints (1) and (2) above, when the accuracy gain ratio of two adjacent detection windows is When, take it as a certainty One of the conditions for achieving maximum accuracy gain; in order to determine The method uses Euclidean distance to identify the retraining samples with the greatest differences in adjacent detection windows.

[0079] Resource Harvesting Scheduler: Centered on resource collection, it fully utilizes the keep-alive resources of server-agnostic functions. Under dynamic request loads, it achieves unified scheduling of idle resources by constructing a directed acyclic graph of available idle resources; simultaneously, it employs an in-memory file system to efficiently update model weights, improving scheduling efficiency and ensuring model security. The inputs to the resource harvesting scheduler are the idle resources of server-agnostic functions, the list of retrained models, and task priorities; the outputs are the task scheduling plan and execution results.

[0080] In the resource harvesting scheduler, a directed acyclic graph of available and idle resources is established, utilizing known conditions to ensure the continuity and availability of dynamic and intermittent resources. These conditions include the keep-alive time obtained through analysis of historical workloads, and the data obtained from offline analysis for each model. The retraining time is also a factor. Furthermore, if requests continue to arrive and the idle resources during the keep-alive period cannot meet the developer's expected budget, a new server-agnostic function will be initiated for retraining.

[0081] Secondly, to facilitate sharing model weights across different functions and maintain continuity during retraining, this invention uses the TMPFS method to create a virtual memory file system and mounts it to each retraining task. In this way, the memory file system enables rapid updates of model weights, and the TMPFS method also supports isolated access to multiple models.

[0082] Therefore, the resource harvesting scheduler analyzes historical load data and combines it with the function keep-alive period to construct a directed acyclic graph of available idle resources among multiple model instances, which is used to characterize the temporal sequence and distribution of available idle resources. During scheduling, resources are allocated according to the importance of tasks by combining the priority list generated by the retraining planning method, and the shared model weights are utilized using the memory file system to reduce data transfer overhead. The scheduling strategy ensures the continuity of retraining tasks under dynamic request loads, and the task execution order can be flexibly adjusted even when the keep-alive period is interrupted. In addition, when keep-alive period resources are insufficient, the system can start a new server-agnostic function for supplementary training, ensuring the stability of task completion. Through efficient resource utilization, the overall resource overhead is significantly reduced, while the execution efficiency of retraining tasks is improved.

[0083] In summary, this invention provides a server-insensitive continuous learning system for multi-model video analysis, including a data drift detector that is robust to fluctuations, accurately identifying accuracy degradation caused by data drift and avoiding misjudgments due to input fluctuations; a retraining planner that analyzes the correlation between multiple models in a directed acyclic graph, removes redundant models, and prioritizes models for retraining based on their importance to accuracy recovery; furthermore, to fully utilize keep-alive resources for retraining, a resource harvesting scheduler is included to effectively arrange dynamic and intermittent retraining tasks without affecting normal analysis tasks. The synergistic effect of these three structural and technical solutions resolves the contradiction between accuracy loss and performance consumption caused by data drift, achieving a high-precision and high-resource-efficiency continuous learning system, providing an innovative solution for video analysis scenarios.

[0084] It should be noted that the above content merely illustrates the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. For those skilled in the art, various improvements and modifications can be made without departing from the principle of the present invention, and all such improvements and modifications fall within the scope of protection of the claims of the present invention.

Claims

1. A server-insensitive continuous learning system for multi-model video analysis, characterized in that: It includes at least a data drift detector, a retraining planner, and a resource harvesting scheduler. The data drift detector uses an adaptive drift detection window to dynamically adjust the size of the detection window and constructs a precision distribution model based on historical data without data drift. The precision distribution model determines whether data drift exists by comparing the magnitude of precision change during online inference. The specific determination method is as follows: within each non-data drift window, the magnitude of inference precision change within each window is calculated using the maximum normalized precision gradient, and its distribution is incorporated into the precision distribution model. If the fluctuation magnitude of online inference precision exceeds the fluctuation magnitude of the precision distribution model, it is determined that data drift has occurred. The retraining planner: When the system determines that data drift has occurred in the window through the data drift detector, it selects the optimal set of retraining models for model retraining by analyzing the inference correlation between different models in the multi-model system; The selection method is as follows: the inference correlation between different models in the multi-model system is judged by the change in inference accuracy between models, and the preceding model with the largest accuracy gain is selected for optimized training. The resource harvesting scheduler: After the system determines the training priority of multiple models through the retraining planner, it dynamically schedules retraining tasks by utilizing the idle resources of the server's non-aware function. During the scheduling process, a directed acyclic graph of available idle resources among multiple model instances is constructed based on historical load data and function keep-alive period, and model weights are shared.

2. The server-insensitive continuous learning system for multi-model video analysis as described in claim 1, characterized in that: The sliding range of the detection window size in the data drift detector is: ； in, Indicates the first window size, Indicates the maximum detection window size. Indicates the first -1 average precision within a window Indicates will Mapped between 0 and 1 This is a hyperparameter that indicates the rate at which the detection window size decreases.

3. The server-insensitive continuous learning system for multi-model video analysis as described in claim 2, characterized in that: The data drift detector determines whether data drift has occurred by the following steps: S101: Calculate at time point Detection window Maximum normalized precision gradient on: ； in, Indicates in the detection window The maximum range of internal inference accuracy, Indicates at a point in time and The detection window inside, Indicates at a point in time and The sliding timestamp within, Indicates the first The first detection window, the first Inference precision under each timestamp Indicates the first The first detection window, the first Inference precision under a timestamp; S102: Using the maximum normalized precision gradient obtained in step S101, construct a precision distribution model under the no-data-drift window. : ； Among them, time points , , These represent the detection windows. The start time, sliding timestamp, and end time within the timeframe. Indicates the maximum range of inference precision; S103: Obtain the detection window based on the accuracy distribution model obtained in step S102. Maximum inference accuracy range Compare it with the accuracy distribution model In comparison, when the maximum inference accuracy range Exceeding precision distribution model When the fluctuation range is within a certain range, data drift occurs in the detection window; S104: When data drift occurs, the system triggers the retraining process and adds the models that need retraining to the retraining set in sequence. When the retraining set is empty, it indicates that the detection window is empty. No data drift occurred.

4. The server-insensitive continuous learning system for multi-model video analysis as described in claim 1, characterized in that: The selection of the optimal preceding model in the retraining planner includes the following steps: S201: Calculate the Kullback-Leibler divergence of the inference accuracy distribution among multiple models to identify highly correlated model pairs; ； in, Display window Internal Model Correlation of inference accuracy between them and Representing the model respectively and In the Within each window, timestamps The changing distribution of reasoning accuracy Indicates the total time of the detection window; S202: Select the most relevant preceding model obtained in step S201; ； The preceding model represents the one with the greatest accuracy gain.

5. The server-insensitive continuous learning system for multi-model video analysis as described in claim 4, characterized in that: In the retrained planner, under the budget constraints submitted by the user... Precision constraints Next, the preceding model selected in step S202 is used with a strategy Perform selective retraining to maximize retraining gain. ； Wherein, constraint (1) means that the average accuracy of all detection windows must meet the user's accuracy constraint. , Represents the total number of detection windows; constraint (2) represents the computational cost required to complete the inference and retraining tasks to meet the user's budget constraint. , Indicates the consumption during retraining. Price per second This represents the total amount of resources consumed during the retraining process. This indicates the total execution time after rounding to the nearest billing unit during retraining; Retraining Gain Indicates the first Within each window, from the retraining set Select the optimal set of retrained models And by determining the number of retraining samples and number of training sessions Retrain: ； in, Function representation system with retraining strategy The process of retraining.

6. The server-insensitive continuous learning system for multi-model video analysis as described in claim 5, characterized in that: In the retraining planner, the number of training iterations The specific determination conditions are: satisfying both constraints (1) and (2), and the accuracy gain ratio of two adjacent detection windows is... ; Number of retraining samples The specific method for determining this is as follows: based on Euclidean distance, identify the retraining sample with the greatest difference between adjacent detection windows.

7. The server-insensitive continuous learning system for multi-model video analysis as described in claim 1, characterized in that: In the resource harvesting scheduler, the method for sharing model weights is as follows: a virtual memory file system is created using tmpfs and mounted to each retraining task to achieve model weight updates and isolated access to multiple models.

Images