AI video large model training method based on 5G network slicing

Abstract

According to the AI video large model training method based on 5G network slicing, the current training stage of an AI video large model is determined, and the current resource demand of the AI video large model in the current training stage is determined; adjusting the slice bandwidth, slice delay or slice reliability parameter of the 5G network slice according to the current resource demand; performing priority scheduling and service quality setting by using the adjusted 5G network slices, so as to reduce delay and packet loss rate of the AI video large model; carrying out load balancing and data stream scheduling by using the adjusted 5G network slice to maximize the data volume transmitted by the AI video large model in unit time; and according to the network state of the 5G network slice, adjusting training parameters of the AI video large model. Therefore, the training and reasoning performance of the AI video large model can be effectively improved, and the challenge faced by the traditional network communication technology is solved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a method for training large AI video models based on 5G network slicing. Background Art

[0002] With the rapid development of artificial intelligence and multimedia technology, AI (Artificial Intelligence) video models have been widely used in video analysis, video generation and other fields. However, the training and reasoning process of AI video models places extremely high demands on computing resources and data transmission. Traditional network communication technologies are difficult to meet these requirements, often leading to problems such as slow transmission, excessive load and high latency, which affect the training effect and reasoning performance of large models. Therefore, how to optimize the training and reasoning process of AI video models and improve the training speed and reasoning performance has become a key issue that needs to be solved urgently. Summary of the invention

[0003] The purpose of this application is to solve at least one of the above-mentioned technical defects, especially the technical defects in the prior art of how to optimize the training and reasoning process of large AI video models and improve the training speed and reasoning performance.

[0004] In a first aspect, the present application provides an AI video large model training method based on 5G network slicing, the method comprising:

[0005] Determine the current training stage of the AI ​​video big model, and determine the current resource requirements of the AI ​​video big model in the current training stage;

[0006] Adjust the slice bandwidth, slice delay or slice reliability parameters of the 5G network slice according to the current resource requirements;

[0007] Leverage adjusted 5G network slices for priority scheduling and quality of service settings to reduce latency and packet loss in large AI video models;

[0008] Use the adjusted 5G network slices for load balancing and data flow scheduling to maximize the amount of data transmitted per unit time for large AI video models;

[0009] Adjust the training parameters of the AI ​​video large model according to the network status of the 5G network slice.

[0010] In one embodiment, the step of determining the current resource requirements of the AI ​​video large model in the current training stage includes:

[0011] If the current training phase is the data preprocessing phase, the current resource requirements of the AI ​​video large model include storage resources;

[0012] If the current training stage is the early stage of training, the current resource requirements of the AI ​​video large model include computing resources;

[0013] If the current training stage is the late training stage, the current resource requirements of the AI ​​video large model include data resources.

[0014] In one embodiment, the step of adjusting a slice bandwidth, a slice delay or a slice reliability parameter of a 5G network slice according to current resource requirements includes:

[0015] If the current resource requirements include storage resources, increase the slice bandwidth of the 5G network slice, and improve the slice reliability parameters of the 5G network slice.

[0016] In one embodiment, the step of adjusting a slice bandwidth, a slice delay or a slice reliability parameter of a 5G network slice according to current resource requirements includes:

[0017] If the current resource requirements include computing resources or data resources, the slice bandwidth of the 5G network slice is increased, and the slice latency of the 5G network slice is reduced.

[0018] In one embodiment, the step of adjusting the training parameters of the AI ​​video big model according to the network status of the 5G network slice includes:

[0019] If it is determined that the network condition of the 5G network slice is poor, the number of samples used in each iteration training of the AI ​​video large model will be reduced.

[0020] In one embodiment, the method further comprises:

[0021] Large AI video models are protected through data encryption algorithms, access control policies, and detection and prevention of malicious activities.

[0022] In one embodiment, the method further comprises:

[0023] Based on preset policies, trigger the creation, maintenance, deletion and update of 5G network slices.

[0024] In a second aspect, the present application provides an AI video large model training device based on 5G network slicing, the device comprising:

[0025] A current resource demand determination module, used to determine the current training stage of the AI ​​video big model, and determine the current resource demand of the AI ​​video big model in the current training stage;

[0026] A network slice parameter adjustment module, used to adjust the slice bandwidth, slice delay or slice reliability parameters of the 5G network slice according to the current resource requirements;

[0027] Scheduling and service quality setting module, used to use the adjusted 5G network slices for priority scheduling and service quality setting, so that the AI ​​video large model can reduce latency and packet loss rate;

[0028] The load balancing and data flow scheduling module is used to use the adjusted 5G network slice to perform load balancing and data flow scheduling to maximize the amount of data transmitted per unit time for the AI ​​video large model;

[0029] The training parameter adjustment module is used to adjust the training parameters of the AI ​​video large model according to the network status of the 5G network slice.

[0030] In a third aspect, the present application provides a storage medium: the storage medium stores computer-readable instructions, and when the computer-readable instructions are executed by one or more processors, the one or more processors execute the steps of the AI ​​video large model training method based on 5G network slicing in any of the above embodiments.

[0031] In a fourth aspect, the present application provides a computer device, comprising: one or more processors, and a memory;

[0032] Computer-readable instructions are stored in the memory. When the computer-readable instructions are executed by one or more processors, the steps of the AI ​​video large model training method based on 5G network slicing in any of the above embodiments are performed.

[0033] It can be seen from the above technical solutions that the embodiments of the present application have the following advantages:

[0034] In the AI ​​video large model training method based on 5G network slicing provided in this application, by dynamically adjusting the bandwidth, delay and reliability of the 5G network slice, the network conditions can be optimized according to the current resource requirements of the AI ​​video large model, which helps to reduce the delay in the training process and improve the data transmission speed, thereby speeding up the training process of the model; in the inference stage, the adjusted 5G network slice is used for priority scheduling and service quality setting, which can significantly reduce the delay and packet loss rate in the inference process and improve the response speed and accuracy of the model; through load balancing and data flow scheduling, network resources can be effectively allocated to maximize the amount of data transmitted per unit time, thereby improving the overall utilization efficiency of the 5G network and avoiding resource waste; according to the real-time network status of the 5G network slice, the training parameters of the AI ​​video large model are adjusted so that the model can run adaptively in different network environments, and the dynamic adjustment mechanism helps to ensure the stability and reliability of the training and inference process. In this way, the training and inference performance of the AI ​​video large model can be effectively improved, and the challenges faced by traditional network communication technology can be solved. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0036] Figure 1 A flowchart of a method for training a large AI video model based on 5G network slicing provided in an embodiment of the present application;

[0037] Figure 2 A schematic diagram of the structure of an AI video large model training device based on 5G network slicing provided in an embodiment of the present application;

[0038] Figure 3 A schematic diagram of the internal structure of a computer device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0039] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0040] The present application provides an AI video large model training method based on 5G network slicing. The following embodiments are described by taking the method applied to a computer device as an example. It can be understood that the computer device can be any device with data processing functions, including but not limited to a single server, a server cluster, a personal laptop, a desktop computer, etc. Figure 1 As shown, the method includes:

[0041] S101: Determine the current training stage of the AI ​​video big model, and determine the current resource requirements of the AI ​​video big model in the current training stage.

[0042] Among them, AI video big model refers to the deep learning model used for video analysis or generation, which is usually composed of a large number of parameters and can process complex video data. The training phase refers to the different stages of the model in the training process, usually including data preparation, model building, training, verification and testing. Resource requirements refer to the computing resources required by the model at a specific training stage, including computing power, memory, bandwidth, etc., to ensure the smooth progress of the training process.

[0043] In this step, you can monitor the indicators during the training process, such as loss function, accuracy, etc., to determine which stage you are currently in. For example, when the training loss of the model drops to a certain threshold, it can be considered that the data preparation stage has been completed and the formal training stage has begun. In addition, you can also use a state machine or event-driven approach to automatically switch the training stage according to the preset training steps and record the current status.

[0044] In the current training phase, you can use performance monitoring tools, such as CPU and GPU utilization monitoring, memory usage, etc., to collect resource usage data, which can be used to evaluate the computing power required in the current phase. The algorithm model predicts the resources required in different training phases, and then calculates the current resource requirements to ensure the reasonable allocation of resources.

[0045] It can be understood that determining the training stage can accurately manage the training process and ensure that the model obtains the appropriate training strategy at different stages. For example, in the model initialization stage, more relaxed hyperparameters can be selected, while more stringent parameter settings can be adopted in the formal training stage. In this way, unnecessary computing overhead can be reduced, the efficiency of model training can be improved, and resource consumption can be avoided at unnecessary stages. Accurately assessing current resource requirements helps to reasonably allocate computing resources and prevent resource bottlenecks. When the current resource requirements are known, task scheduling can be optimized to ensure high-intensity training when computing resources are sufficient, and lightweight operations when resources are limited. Through real-time monitoring and dynamic adjustment, resource utilization can be maximized, training costs can be reduced, while model training and reasoning can be accelerated, ultimately improving the performance of large AI video models.

[0046] S102: Adjust slice bandwidth, slice delay or slice reliability parameters of the 5G network slice according to current resource requirements.

[0047] Among them, 5G network slicing is a network virtualization technology that divides the physical network into multiple independent virtual networks to meet the needs of different applications. Each slice can be optimized according to specific service quality requirements. Slice bandwidth refers to the data transmission rate allocated to a specific application in a network slice, which determines the speed of data transmission in the network. Slice latency refers to the time required for data to reach the destination from the source. The smaller the latency, the better the real-time performance. Slice reliability parameters refer to the ability of network slices to maintain data integrity and validity during data transmission, including packet loss rate, retransmission mechanism, etc.

[0048] In this step, for the adjustment of slice bandwidth, the current network traffic and resource usage can be monitored in real time, and the bandwidth can be dynamically allocated using the network management software. Through interface calls, the bandwidth limit of the slice can be adjusted to match the current resource demand. For example, when a large amount of data transmission is required during model training, the bandwidth can be increased, while during low-load periods, the bandwidth can be reduced to optimize resource usage. For the adjustment of slice delay, the slice delay can be optimized by changing the routing strategy or selecting a low-latency transmission channel. The network monitoring tool can be used to analyze the current network delay and adjust the network path based on real-time data. It is also possible to configure QoS (Quality of Service) policies to prioritize low-latency transmission of key data streams to ensure the real-time nature of training and reasoning. For the adjustment of slice reliability parameters, the reliability of the slice can be improved by setting a retransmission mechanism and monitoring the packet loss rate. The number of retransmissions and the time interval are automatically adjusted when a high packet loss rate is detected. By dynamically adjusting these parameters, the reliability of data transmission can be optimized according to the current network conditions to ensure data integrity.

[0049] It can be understood that dynamically adjusting bandwidth can ensure that network resources are fully utilized at high-demand moments, avoiding training delays or failures caused by insufficient bandwidth. Improving bandwidth utilization efficiency can reduce overall latency, enhance the stability of data transmission, and help improve the effectiveness of model training and reasoning. Low latency is critical for real-time applications, especially in AI video analysis tasks that require instant feedback. By adjusting the latency, data can be ensured to be transmitted quickly in the network. Optimizing latency can also help improve the responsiveness of the model, making it more sensitive in real-time scenarios. By optimizing reliability parameters, data loss can be reduced, ensuring the integrity and accuracy of data during training and reasoning. Improving reliability can enhance the robustness of the model in unstable network environments, ensuring that tasks can be completed smoothly even in the event of network fluctuations.

[0050] S103: Use the adjusted 5G network slice to perform priority scheduling and service quality setting to enable the AI ​​video large model to reduce latency and packet loss rate.

[0051] Priority scheduling refers to the dynamic allocation of network resources and transmission priorities in network management according to the urgency and importance of data flows to ensure that critical data is processed in a timely manner. Quality of Service (QoS) settings refer to a series of measures taken to ensure that network services meet specific performance requirements, including the configuration of parameters such as bandwidth, latency, packet loss rate, and reliability. Latency refers to the time required for data to be sent and received. The smaller the latency, the better the real-time performance. Packet loss rate refers to the proportion of data packets that fail to successfully reach the destination during data transmission. The lower the packet loss rate, the higher the reliability of data transmission.

[0052] In this step, for priority scheduling, different types of data streams can be assigned priorities through packet marking technology. Important data streams, such as training and inference data, are marked as high priority, and packet transmission is scheduled according to these markings. By using a flow control algorithm, the priority of each data stream can be dynamically adjusted according to the current network status to ensure that important data streams are given priority when the network is congested. For quality of service settings, target values ​​for bandwidth, latency, and packet loss rate for different data streams can be set by configuring QoS policies. Using network management tools, network performance can be monitored in real time and these parameters can be automatically adjusted as needed. For example, during periods of high traffic, mission-critical data streams can be prioritized and bandwidth allocation for non-critical streams can be reduced to ensure the smooth progress of important operations.

[0053] It can be understood that by utilizing the high speed and low latency characteristics of 5G network slicing, data transmission can be optimized and latency and packet loss can be reduced through priority scheduling, quality of service (QoS) settings, and other methods. Specifically, priority scheduling can ensure that critical data streams are transmitted in the network in a timely manner, thereby reducing latency, which is especially important in AI video tasks that require real-time feedback. By prioritizing important data, the response speed and accuracy of the model can be improved. An effective scheduling mechanism can also reduce resource competition under high load conditions and ensure the stability and reliability of the model. By setting QoS, it is possible to ensure that network performance meets the requirements of specific applications and reduce latency and packet loss. This is especially important for AI video analysis tasks that require high reliability and low latency.

[0054] S104: Use the adjusted 5G network slice to perform load balancing and data flow scheduling to maximize the amount of data transmitted per unit time for the large AI video model.

[0055] Among them, load balancing is a technology for allocating computing and network resources, aiming to optimize resource usage and avoid performance degradation caused by overloading a resource. Data flow scheduling refers to the management and arrangement of data transmission in the network to ensure that data packets are efficiently transmitted according to priority and bandwidth requirements. The amount of data transmitted per unit time refers to the total amount of data that can be successfully transmitted by the network in a specific time period, usually expressed in bits per second (bps) or bytes per second (Bps).

[0056] In this step, for load balancing, you can use load balancing algorithms such as polling and least connections to dynamically monitor the traffic and load of each network slice, distribute data flows to slices with lower current loads, and ensure that no single slice carries too much traffic. Through network management tools, you can obtain slice performance data in real time, and automatically adjust traffic distribution based on this data to ensure efficient use of resources. For data flow scheduling, you can use scheduling algorithms such as priority queues and FIFO to arrange the transmission order of data in the network according to the type and priority of the data packet. High-priority data packets will be sent first to reduce latency.

[0057] It can be understood that the use of load balancing and data flow scheduling between 5G network slices can ensure that data can be transmitted on the most appropriate path, reduce congestion, and improve transmission efficiency. Specifically, load balancing can effectively disperse network traffic, avoid performance bottlenecks caused by overloading a certain slice, and ensure that large AI video models achieve stable performance during training and reasoning, especially in the case of high data volume transmission. Effective load balancing improves resource utilization, allowing the model to remain efficient when facing different loads, thereby improving the reliability and speed of data transmission. Data flow scheduling can optimize the order of data transmission according to priority, ensure that key data streams for model training and reasoning are transmitted in a timely manner, reduce delays, and through reasonable scheduling, effectively reduce packet loss rate, improve the integrity and effectiveness of data transmission, thereby ensuring the efficient performance of the model.

[0058] S105: Adjust the training parameters of the AI ​​video large model according to the network status of the 5G network slice.

[0059] Among them, the network status of 5G network slices refers to the performance indicators of 5G network slices at a specific moment, including information such as bandwidth utilization, latency, packet loss rate, network load, etc., reflecting the current health status of the network and resource availability. The training parameters of the AI ​​video big model refer to the various configurations used to train the AI ​​model. The training parameters of the AI ​​video big model include accuracy, precision, recall rate and F1 value. The accuracy of the model refers to the proportion of correct predictions of the model in all predictions; precision refers to the proportion of samples predicted as positive by the model that are actually positive; recall rate refers to the proportion of samples predicted as positive by the model in all samples that are actually positive; F1 value is the harmonic mean of accuracy and recall rate, which can comprehensively evaluate the performance of the model.

[0060] In this step, you can use network monitoring tools to collect real-time status data of 5G network slices, including metrics such as bandwidth, latency, and packet loss rate. For example, this can be achieved through SNMP (Simple Network Management Protocol) or network performance monitoring software. Based on the network status data, you can set rules or algorithms to dynamically adjust training parameters. For example, when the network latency is high, you can reduce the learning rate to slow down the training process and ensure that the model gradually converges in an unstable network environment. In addition, if the network condition is good, the system can increase the batch size to improve training efficiency and make full use of the available bandwidth.

[0061] It is understandable that monitoring the network status can provide real-time feedback on the current network conditions, ensuring that the training process adapts to environmental changes. Keeping abreast of network conditions helps optimize the model training process and ensure that the model can still be effectively trained under undesirable network conditions. Dynamic adjustment of training parameters can improve the stability and learning efficiency of the model under different network conditions. For example, lowering the learning rate when the network is congested can reduce the uncertainty of training and ensure that the model adjusts parameters incrementally rather than causing large parameter updates due to network fluctuations. By optimizing training parameters according to network status, network resources can be maximized, data transmission efficiency per unit time can be improved, model training effects can be enhanced, and ultimately the performance and reliability of large AI video models can be improved.

[0062] In the above embodiments, by dynamically adjusting the bandwidth, delay and reliability of the 5G network slice, the network conditions can be optimized according to the current resource requirements of the AI ​​video large model, which helps to reduce the delay in the training process and improve the data transmission speed, thereby speeding up the training process of the model; in the inference stage, the adjusted 5G network slice is used for priority scheduling and service quality setting, which can significantly reduce the delay and packet loss rate in the inference process and improve the response speed and accuracy of the model; through load balancing and data flow scheduling, network resources can be effectively allocated to maximize the amount of data transmitted per unit time, thereby improving the overall utilization efficiency of the 5G network and avoiding resource waste; according to the real-time network status of the 5G network slice, the training parameters of the AI ​​video large model are adjusted so that the model can run adaptively in different network environments, and the dynamic adjustment mechanism helps to ensure the stability and reliability of the training and inference process. In this way, the training and inference performance of the AI ​​video large model can be effectively improved, and the challenges faced by traditional network communication technology can be solved.

[0063] In one embodiment, the step of determining the current resource requirements of the AI ​​video large model in the current training phase includes:

[0064] If the current training phase is the data preprocessing phase, the current resource requirements of the AI ​​video large model include storage resources;

[0065] If the current training stage is the early stage of training, the current resource requirements of the AI ​​video large model include computing resources;

[0066] If the current training stage is the late training stage, the current resource requirements of the AI ​​video large model include data resources.

[0067] Among them, the data preprocessing stage refers to the operations such as cleaning, formatting and feature extraction of raw data before model training, so as to prepare high-quality data sets for subsequent training. Storage resources refer to computing resources used to store data, including hard disk space, memory, etc. In the data preprocessing stage, storage resources are used to store raw data, processed data and temporary data. The early training stage refers to the early stage of model training, when the model parameters have not yet fully learned the data characteristics and mainly rely on computing resources for training. Computing resources refer to hardware resources used for numerical calculations, including CPU, GPU, TPU, etc. These resources are used to perform forward and backward propagation of the model during training. The late training stage refers to the final stage of model training, when the model is close to convergence, and the focus is on adjusting the details of the model and verifying the performance of the model. Data resources refer to the additional data required in the later stage of training, usually including validation sets and test sets, which are used to evaluate the generalization ability of the model.

[0068] Specifically, the current training stage can be determined through a state machine or logic. For example, a flag or configuration file can be used to indicate which stage the model is currently in, including data preprocessing, early training, or late training. Combined with monitoring indicators such as loss function, training time, etc., the training stage is automatically identified and the corresponding resource requirements are output. For the data preprocessing stage, storage requirements can be evaluated, including analyzing the size of the original data and the size of the preprocessed data to determine the required storage resources. For the early stage of training, computing resource requirements can be evaluated, and the utilization of the CPU and GPU can be monitored to ensure that the computing load of model training can be met. For the late stage of training, the size of the validation and test sets can be analyzed to determine the required data resources and ensure their availability.

[0069] In this embodiment, determining the current training stage helps to clarify resource requirements and ensure that the training process will not be interrupted or fail due to insufficient resources. Different stages have different requirements for resources, and timely identification can avoid waste of resources. Through dynamic monitoring, resource allocation can be adjusted in real time to ensure that necessary computing and storage support are provided at each stage, thereby improving training efficiency. For example, sufficient storage space is required in the data preprocessing stage to efficiently process data, while powerful computing power is required in the early stages of training to accelerate model learning. By rationally allocating resources, the overall training process can be optimized, training time and resource waste can be reduced, and ultimately the performance and effect of the model can be improved.

[0070] In one embodiment, the step of adjusting the slice bandwidth, slice delay or slice reliability parameters of the 5G network slice according to the current resource demand includes:

[0071] If the current resource requirements include storage resources, increase the slice bandwidth of the 5G network slice, and improve the slice reliability parameters of the 5G network slice.

[0072] Specifically, for slice bandwidth, the bandwidth usage of the current slice can be monitored in real time through the network management platform or API interface, and the slice bandwidth can be dynamically adjusted when the storage demand increases. You can also set policies to automatically allocate additional bandwidth to ensure that data can be transmitted quickly when storage resource requirements are high, such as when large-scale data transmission or backup is performed. For slice reliability parameters, you can reduce the packet loss rate and increase the number of retransmissions by configuring the QoS policy of the network device. For example, when storage requirements increase, the retransmission mechanism of the network slice can be improved to ensure that data is not lost during transmission. Monitor the network traffic status and adjust the reliability parameters in real time, for example, increase the redundant transmission of the network slice to ensure the reliability of critical data.

[0073] In this embodiment, increasing the slice bandwidth can effectively increase the data transmission speed, especially when the storage resource demand for processing a large amount of data increases, it can ensure that the data is quickly written and read, thereby improving the overall processing efficiency. Improving the reliability parameter can ensure the integrity and validity of the data during the data transmission process, especially when the network load is high, which is crucial to ensure that the stored data is not lost or damaged. By increasing the slice bandwidth of the 5G network slice and improving the slice reliability parameter, efficient data transmission can be effectively supported when the storage resource demand increases, and the performance and reliability of the model can be improved.

[0074] In one embodiment, the step of adjusting the slice bandwidth, slice delay or slice reliability parameters of the 5G network slice according to the current resource demand includes:

[0075] If the current resource requirements include computing resources or data resources, the slice bandwidth of the 5G network slice is increased, and the slice latency of the 5G network slice is reduced.

[0076] Specifically, for slice bandwidth, network management tools can be used to monitor the bandwidth utilization of the current slice and dynamically allocate additional bandwidth when increased demand for computing or data resources is detected. Automatically adjust the bandwidth allocation of the slice based on real-time traffic conditions to ensure that computing tasks and data transmission can proceed smoothly. For slice latency, low-latency data streams can be prioritized by configuring QoS policies. For example, traffic priority marking technology can be used to ensure that high-priority data streams are transmitted first, thereby reducing latency. When network conditions permit, routing policies can be adjusted to select network paths with lower latency to further reduce data transmission time.

[0077] In this embodiment, increasing the slice bandwidth can support higher data transmission rates, especially when the demand for computing resources or data resources increases, ensuring that data and computing tasks can interact quickly, thereby improving overall performance. Adequate bandwidth can reduce task execution time and improve the efficiency of model training and reasoning, especially in scenarios where large amounts of data are required for calculation. Reducing slice latency can significantly improve the response speed of real-time tasks and ensure timeliness between computing and data transmission. Lower latency can reduce congestion in data transmission, thereby ensuring that the model can still maintain good performance and reliability under high load conditions. In this way, by increasing the slice bandwidth of 5G network slices and reducing slice latency, efficient data transmission can be effectively supported when the demand for computing resources or data resources increases, and the performance and responsiveness of the model can be improved.

[0078] In one embodiment, the step of adjusting the training parameters of the AI ​​video big model according to the network status of the 5G network slice includes:

[0079] If it is determined that the network condition of the 5G network slice is poor, the number of samples used in each iteration training of the AI ​​video large model will be reduced.

[0080] Iterative training refers to the process of optimizing model parameters through multiple back propagation and weight updates during model training. Each iteration uses a portion of the training data for update. The number of samples refers to the amount of training data used in each iterative training. The size of the sample size directly affects the speed and effect of the training process.

[0081] It is understandable that reducing the number of samples per iteration can reduce the demand for network bandwidth, avoid data transmission delays and packet loss when the network is not in good condition, and ensure the stability of the training process. Reducing the number of samples can also speed up each iteration, thereby improving the real-time nature of training and ensuring that the model can be updated in an undesirable network environment.

[0082] Specifically, the status data of 5G network slices, including indicators such as bandwidth, latency, and packet loss rate, can be obtained in real time through network monitoring tools. When the network condition is poor, the corresponding adjustment mechanism is triggered. During the training process, the number of samples for each iteration can be dynamically modified according to the current network condition. A threshold can be set. For example, when the delay exceeds a certain value, the number of samples used for each iteration is automatically reduced. The number of samples can be adjusted by configuring training parameters or using adaptation algorithms, such as learning rate decay, to ensure effective training in an unstable network environment.

[0083] In this embodiment, by dynamically adjusting the number of samples, the training efficiency of the model in an unstable network environment can be optimized to avoid training interruptions or reduced results due to network problems. This flexible adjustment mechanism can also ensure that the normal training process can be quickly restored after the network conditions improve, thereby improving the overall training effect and model performance.

[0084] In one embodiment, the method further comprises:

[0085] Large AI video models are protected through data encryption algorithms, access control policies, and detection and prevention of malicious activities.

[0086] Among them, data encryption algorithm refers to the mathematical algorithm used to convert raw data into ciphertext to protect the confidentiality of the data. Only users with the correct key can decrypt and access the original data. Access control policy refers to the rules that define who can access system resources and their permissions. These policies usually limit or allow access based on user identity, role or access requirements. Detecting and preventing malicious activities refers to ensuring the security of the system by monitoring, analyzing and responding to potential security threats and attack methods. This process includes detecting abnormal activities in real time and taking measures to prevent these activities from affecting the system.

[0087] Specifically, encryption algorithms such as AES, RSA, etc. can be used to encrypt the training data and model parameters stored in the database. During the data transmission process, protocols such as SSL / TLS can also be used to encrypt the transmission data. The encryption key can also be updated regularly to ensure that even if the key is leaked, the attacker cannot access the encrypted data for a long time. For access control, role-based access control policies can be defined and implemented to ensure that only authorized users can access the relevant resources of the AI ​​video large model. User identity can also be verified through authentication mechanisms such as OAuth, JWT, etc. For malicious activities, intrusion detection systems and intrusion prevention systems (IPS) can be deployed to monitor abnormal activities in the system in real time and generate alerts. Machine learning algorithms can also be combined to identify potential malicious activities. Implement an automated response mechanism to take appropriate measures for identified malicious activities, such as isolating infected systems and blocking suspicious access.

[0088] In this embodiment, data encryption can effectively protect the confidentiality of sensitive data and prevent unauthorized users from accessing training data and model parameters. After encryption, even if the data is stolen, attackers cannot directly use the data, thereby reducing potential losses. Access control can ensure that only authorized personnel can access resources, thereby reducing the risk of internal and external attacks. Timely detection and prevention of malicious activities can significantly reduce the impact of security incidents on the model and ensure the stability and availability of the AI ​​video large model. In this way, through data encryption algorithms, access control strategies, and comprehensive security protection measures for detecting and preventing malicious activities, the security of the AI ​​video large model can be effectively improved and protected from potential security threats.

[0089] In one embodiment, the method further comprises:

[0090] Based on preset policies, trigger the creation, maintenance, deletion and update of 5G network slices.

[0091] The preset policy refers to a set of predefined rules and conditions that guide the triggering and execution of certain operations. Creation, maintenance, deletion, and update refer to the lifecycle management of network slices, including the generation of new slices, performance monitoring and optimization of existing slices, termination of slices, and modification or optimization of slice attributes.

[0092] Specifically, according to the preset policies, network load, user demand and other indicators can be monitored to automatically trigger the creation of new 5G network slices. For example, the slice creation function of the network management platform can be called through the API to specify the required resources and service quality parameters. After that, the new slice profile can be associated with the network resources to ensure that the new slice has the required bandwidth and latency performance. For slice maintenance, the performance indicators of existing slices, such as bandwidth utilization, latency and packet loss rate, can be monitored regularly, and resource allocation can be adjusted according to the preset policies. Dynamic optimization of slices can be achieved through automated scripts or scheduling tools. When the slice load increases, more resources can be automatically allocated to maintain service quality. When the slice is no longer needed, the deletion request of the slice is triggered based on the monitored resource utilization and the preset policy. By calling the API of the network management platform, the resource configuration related to the slice can be safely released. During the deletion process, ensure the cleanup of data and the release of resources to avoid resource waste. The update operation of the existing slice can be triggered according to the changes in network status or user needs. This may include adjusting parameters such as bandwidth and latency. Modify the slice configuration through the network management tool to optimize its performance. Updates can be automated, with real-time data and preset thresholds used to determine when to update.

[0093] In this embodiment, the creation of slices can flexibly respond to changing network needs and ensure that appropriate resources and service quality can be provided. The creation of new slices helps to improve the overall efficiency and responsiveness of the network. Maintaining existing slices can ensure the continuous stability of network performance, adjust resources in time to cope with changes, and ensure that service quality is not affected. Deleting slices that are no longer needed can free up system resources and avoid waste of resources, which helps to improve the overall performance of the network and ensure that other slices can obtain more resources. Updating slices can ensure that they always meet the current network conditions and model requirements and maintain the optimal performance of the slices. In this way, by triggering the creation, maintenance, deletion, and update of 5G network slices based on preset policies, the flexibility and resource utilization of the network can be effectively improved, ensuring the provision of high-quality services under different conditions.

[0094] The following describes the AI ​​video large model training device based on 5G network slicing provided in the embodiment of the present application. The AI ​​video large model training device based on 5G network slicing described below and the AI ​​video large model training method based on 5G network slicing described above can refer to each other. Figure 2 As shown, the present application provides an AI video large model training device based on 5G network slicing, the device comprising:

[0095] A current resource requirement determination module 201 is used to determine the current training stage of the AI ​​video big model and determine the current resource requirement of the AI ​​video big model in the current training stage;

[0096] A network slice parameter adjustment module 202, configured to adjust the slice bandwidth, slice delay or slice reliability parameters of the 5G network slice according to current resource requirements;

[0097] The scheduling and service quality setting module 203 is used to use the adjusted 5G network slice to perform priority scheduling and service quality setting, so that the AI ​​video large model can reduce delay and packet loss rate;

[0098] A load balancing and data flow scheduling module 204 is used to use the adjusted 5G network slice to perform load balancing and data flow scheduling to maximize the amount of data transmitted per unit time of the AI ​​video large model;

[0099] The training parameter adjustment module 205 is used to adjust the training parameters of the AI ​​video large model according to the network status of the 5G network slice.

[0100] In one embodiment, the current resource demand determination module 201 includes:

[0101] A first current resource demand determination unit, configured to determine, if the current training stage is a data preprocessing stage, the current resource demand of the AI ​​video large model includes storage resources;

[0102] A second current resource demand determination unit, configured to determine, if the current training stage is an early training stage, the current resource demand of the AI ​​video large model includes computing resources;

[0103] The third current resource demand determination unit is used to determine that if the current training stage is the late training stage, the current resource demand of the AI ​​video large model includes data resources.

[0104] In one embodiment, the network slice parameter adjustment module 202 includes:

[0105] The first network slice parameter adjustment unit is used to increase the slice bandwidth of the 5G network slice and improve the slice reliability parameters of the 5G network slice if the current resource demand includes storage resources.

[0106] In one embodiment, the network slice parameter adjustment module 202 includes:

[0107] The second network slice parameter adjustment unit is used to increase the slice bandwidth of the 5G network slice and reduce the slice delay of the 5G network slice if the current resource demand includes computing resources or data resources.

[0108] In one embodiment, the training parameter adjustment module 205 includes:

[0109] The training parameter adjustment unit is used to reduce the number of samples used in each iterative training of the AI ​​video large model if it is determined that the network condition of the 5G network slice is poor.

[0110] In one embodiment, the apparatus further comprises:

[0111] The model security protection module is used to provide security protection for large AI video models through data encryption algorithms, access control policies, and detection and prevention of malicious activities.

[0112] In one embodiment, the apparatus further comprises:

[0113] The network slice management module is used to trigger the creation, maintenance, deletion and update of 5G network slices based on preset policies.

[0114] In one embodiment, the present application also provides a storage medium storing computer-readable instructions. When the computer-readable instructions are executed by one or more processors, the one or more processors execute the steps of the AI ​​video large model training method based on 5G network slicing as described in any of the above embodiments.

[0115] In one embodiment, the present application also provides a computer device, which stores computer-readable instructions. When the computer-readable instructions are executed by one or more processors, the one or more processors execute the steps of the AI ​​video large model training method based on 5G network slicing as described in any of the above embodiments.

[0116] Indicatively, Figure 3 As shown, Figure 3 This is a schematic diagram of the internal structure of a computer device provided in an embodiment of the present application. The computer device 300 may be provided as a server. Figure 3 , the computer device 300 includes a processing component 302, which further includes one or more processors, and a memory resource represented by a memory 301, for storing instructions executable by the processing component 302, such as an application. The application stored in the memory 301 may include one or more modules, each corresponding to a set of instructions. In addition, the processing component 302 is configured to execute instructions to perform the AI ​​video large model training method based on 5G network slicing in any of the above embodiments.

[0117] The computer device 300 may further include a power supply component 303 configured to perform power management of the computer device 300, a wired or wireless network interface 304 configured to connect the computer device 300 to a network, and an input / output (I / O) interface 305. The computer device 300 may operate based on an operating system stored in the memory 301, such as Windows Server TM, Mac OS X TM, Unix TM, Linux TM, Free BSD TM, or the like.

[0118] Those skilled in the art will understand that Figure 3 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0119] Finally, it should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply that there is any such actual relationship or order between these entities or operations. Moreover, the term "include", "comprise" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements not only includes those elements, but also includes other elements not clearly listed, or also includes elements inherent to such process, method, article or equipment. In the absence of more restrictions, the elements defined by the sentence "including one..." do not exclude the existence of other identical elements in the process, method, article or equipment including the elements. Herein, "one", "one", "said", "the" and "it" may also include plural forms, unless the context clearly indicates another way. A plurality refers to at least two cases, such as 2, 3, 5 or 8, etc. "And / or" includes any and all combinations of the relevant listed items.

[0120] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The various embodiments can be combined as needed, and the same or similar parts can refer to each other.

[0121] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for training a large AI video model based on 5G network slicing, characterized in that: The method comprises: Determining a current training stage of the AI ​​video big model, and determining a current resource requirement of the AI ​​video big model in the current training stage; Adjusting slice bandwidth, slice delay or slice reliability parameters of the 5G network slice according to current resource requirements; Utilize the adjusted 5G network slice to perform priority scheduling and service quality settings so that the AI ​​video big model can reduce latency and packet loss rate; Using the adjusted 5G network slice to perform load balancing and data flow scheduling to maximize the amount of data transmitted per unit time by the AI ​​video large model; Adjust the training parameters of the AI ​​video big model according to the network status of the 5G network slice.

2. The AI ​​video large model training method based on 5G network slicing according to claim 1 is characterized in that: The step of determining the current resource requirements of the AI ​​video large model in the current training stage includes: If the current training stage is a data preprocessing stage, the current resource requirements of the AI ​​video large model include storage resources; If the current training stage is the initial training stage, the current resource requirements of the AI ​​video large model include computing resources; If the current training stage is the late training stage, the current resource requirements of the AI ​​video large model include data resources.

3. The AI ​​video large model training method based on 5G network slicing according to claim 2 is characterized in that: The step of adjusting the slice bandwidth, slice delay or slice reliability parameter of the 5G network slice according to the current resource demand includes: If the current resource demand includes storage resources, the slice bandwidth of the 5G network slice is increased, and the slice reliability parameters of the 5G network slice are improved.

4. The AI ​​video large model training method based on 5G network slicing according to claim 2 is characterized in that: The step of adjusting the slice bandwidth, slice delay or slice reliability parameter of the 5G network slice according to the current resource demand includes: If the current resource demand includes computing resources or data resources, the slice bandwidth of the 5G network slice is increased, and the slice delay of the 5G network slice is reduced.

5. The AI ​​video large model training method based on 5G network slicing according to claim 1 is characterized in that: The step of adjusting the training parameters of the AI ​​video large model according to the network status of the 5G network slice includes: If it is determined that the network condition of the 5G network slice is poor, the number of samples used in each iterative training of the AI ​​video large model is reduced.

6. The AI ​​video large model training method based on 5G network slicing according to any one of claims 1 to 5, characterized in that: The method further comprises: The AI ​​video large model is protected through data encryption algorithms, access control strategies, and detection and prevention of malicious activities.

7. The AI ​​video large model training method based on 5G network slicing according to any one of claims 1 to 5, characterized in that: The method further comprises: Based on preset policies, the creation, maintenance, deletion and update of the 5G network slice are triggered.

8. An AI video large model training device based on 5G network slicing, characterized in that: The device comprises: A current resource requirement determination module, used to determine the current training stage of the AI ​​video big model, and determine the current resource requirement of the AI ​​video big model in the current training stage; A network slice parameter adjustment module, used to adjust the slice bandwidth, slice delay or slice reliability parameters of the 5G network slice according to current resource requirements; A scheduling and service quality setting module, used to use the adjusted 5G network slice to perform priority scheduling and service quality setting, so that the AI ​​video large model can reduce delay and packet loss rate; A load balancing and data flow scheduling module, used to use the adjusted 5G network slice to perform load balancing and data flow scheduling to maximize the amount of data transmitted per unit time by the AI ​​video large model; A training parameter adjustment module is used to adjust the training parameters of the AI ​​video large model according to the network status of the 5G network slice.

9. A storage medium, characterized in that: The storage medium stores computer-readable instructions, and when the computer-readable instructions are executed by one or more processors, the one or more processors execute the steps of the AI ​​video large model training method based on 5G network slicing as described in any one of claims 1 to 7.

10. A computer device, characterized in that: include: one or more processors, and memory; The memory stores computer-readable instructions, and when the computer-readable instructions are executed by the one or more processors, the steps of the AI ​​video large model training method based on 5G network slicing are performed as described in any one of claims 1 to 7.

Images