Method and device for computing resource allocation, equipment and medium

Abstract

The invention relates to the technical field of computing resource scheduling, and discloses a computing resource allocation method, device and equipment and a medium, and the method comprises the steps: constructing a global data set in a long-tail scene; a high-dimensional sparse matrix is constructed based on the global data set, the high-dimensional sparse matrix comprises a plurality of data dimensions, feature dimension reduction is performed based on the data dimensions to obtain a feature matrix after dimension reduction, and a feature sub-matrix is generated based on the feature matrix after dimension reduction; evaluating the feature sub-matrix, and determining the priority of the feature sub-matrix in the current scene based on an evaluation result; allocation of computing resources is adjusted based on the priority. According to the method, the vehicle-mounted system can preferentially process high-priority long-tail scenes under the condition that resources are limited, the processing efficiency is high, resource waste and processing delay are avoided, and the efficiency of the whole data processing chain can be optimized, so that the vehicle-mounted system has certain flexibility and coping capacity when facing complex and rare long-tail scenes.

Description

Technical Field

[0001] The present invention relates to the technical field of computing resource scheduling, and in particular to a method, device, equipment and medium for computing resource allocation. Background Art

[0002] Ordinary high-speed driving, lane changes, or waiting at traffic lights can be trained based on a large number of scene driving scenarios. However, in the actual driving environment, there are still many uncommon marginal situations, such as blizzards or heavy fog in extreme weather, or emergency accidents, or construction and sudden obstacles in complex road conditions. The probability of these uncommon scenarios is low, but they are crucial to the safety and reliability of the autonomous driving system. These rare, complex and unpredictable special scenarios and extreme situations that occur during the autonomous driving process refer to the long-tail problem in autonomous driving. Related technologies still have many shortcomings in dealing with long-tail problems in autonomous driving systems.

[0003] In the actual driving process, dealing with long-tail problems requires more computing resources and real-time response capabilities. How to effectively manage and allocate resources to ensure that there are enough resources to deal with long-tail problems at critical moments is an important problem point in the autonomous driving system. Summary of the invention

[0004] In view of this, the present invention provides a method, apparatus, device and medium for computing resource allocation to solve the problem that the rarity of long-tail scenes leads to data scarcity, and the autonomous driving system is unable to perform sufficient training and fitting, resulting in low efficiency in processing long-tail scenes.

[0005] In a first aspect, the present invention provides a method for allocating computing resources, the method comprising: constructing a global data set in a long-tail scenario; based on the global data set, constructing a high-dimensional sparse matrix, the high-dimensional sparse matrix comprising multiple data dimensions, performing feature dimensionality reduction based on the data dimensions to obtain a reduced feature matrix, and generating a feature sub-matrix based on the reduced feature matrix; evaluating the feature sub-matrix, and determining the priority of the feature sub-matrix in the current scenario based on the evaluation result; and adjusting the allocation of computing resources based on the priority.

[0006] In an optional embodiment, the construction of a global data set for a long-tail scenario includes: collecting sensor data to construct a sensor historical data set and a sensor real-time data set; constructing a computational load historical data set and a computational load real-time data set based on the second central processing unit utilization, the second graphics processing unit utilization, and the second memory utilization of the vehicle-mounted system in processing the sensor data; determining a long-tail scene recognition record based on the scene type, the score corresponding to the scene type, the response time of the vehicle-mounted system for processing the scene type, and the resources consumed by the vehicle-mounted system for processing the scene type, and constructing a long-tail scene historical data set and a long-tail scene real-time data set based on the long-tail scene recognition record; integrating the sensor historical data set, the computational load historical data set, and the long-tail scene historical data set to construct a historical data set; integrating the sensor real-time data set, the computational load real-time data set, and the long-tail scene real-time data set to construct a real-time data set.

[0007] In an optional embodiment, the collecting of sensor data includes: adjusting the sensor data collection frequency based on scene complexity and vehicle system load; determining the scene complexity based on the frequency of the current scene appearing in the scene history records, the current environment complexity and the current traffic complexity; determining the vehicle system load based on the first central processing unit utilization, the first graphics processing unit utilization and the first memory utilization in the vehicle system; and collecting the sensor data based on the sensor data collection frequency.

[0008] In an optional implementation, after obtaining the reduced-dimensional feature matrix, the method further includes: completing a low-rank matrix based on the reduced-dimensional feature matrix to obtain a completed matrix; and generating the feature submatrix based on the completed matrix.

[0009] In an optional implementation, generating a feature sub-matrix based on the feature matrix after dimensionality reduction includes: performing a convolution operation based on the feature matrix after dimensionality reduction; and during the convolution operation, adjusting the weight of the convolution kernel to generate the feature sub-matrix.

[0010] In an optional embodiment, the characteristic submatrix is ​​evaluated and the priority of the characteristic submatrix in the current scene is determined based on the evaluation results, including: generating nodes in a graph based on the characteristic submatrix; constructing a multi-layer graph based on time series and the nodes; determining the similarity between the nodes based on the temporal similarity and spatial similarity between the multi-layer graphs; and determining the priority of the characteristic submatrix in the current scene based on the similarity.

[0011] In an optional embodiment, the adjustment of the allocation of computing resources based on the priority includes: determining the resource allocation ratio based on the priority and the complexity of the scenario associated with the characteristic sub-matrix; and determining the computing resources allocated to the characteristic sub-matrix based on the resource allocation ratio and the total amount of computing resources currently allocatable.

[0012] In a second aspect, the present invention provides a device for allocating computing resources, the device comprising: a data set construction module, used to construct a global data set in a long-tail scenario; a feature function extraction module, used to construct a high-dimensional sparse matrix based on the global data set, the high-dimensional sparse matrix comprising multiple data dimensions, based on the data dimensions, feature dimensionality reduction is performed to obtain a reduced feature matrix, based on the reduced feature matrix, multiple feature sub-matrices are generated; a priority evaluation module, used to evaluate the feature sub-matrix, and determine the priority of the feature sub-matrix in the current scenario based on the evaluation result; a computing resource allocation module, used to adjust the allocation of computing resources based on the priority.

[0013] In a third aspect, the present invention provides a computer device, comprising: a memory and a processor, the memory and the processor being communicatively connected to each other, the memory storing computer instructions, and the processor executing the method for computing resource allocation of the above-mentioned first aspect or any corresponding embodiment thereof by executing the computer instructions.

[0014] In a fourth aspect, the present invention provides a computer-readable storage medium having computer instructions stored thereon, the computer instructions being used to enable a computer to execute the method for computing resource allocation of the above-mentioned first aspect or any corresponding embodiment thereof.

[0015] In a fifth aspect, the present invention provides a computer program product, comprising computer instructions for causing a computer to execute the method for computing resource allocation according to the first aspect or any corresponding embodiment thereof.

[0016] The method, apparatus, device and medium for calculating resource allocation provided by the embodiments of the present invention form a complete end-to-end dynamic resource scheduling process from global data set construction, feature dimensionality reduction, priority evaluation to resource allocation; by evaluating the priority of the feature sub-matrix, resource allocation is dynamically adjusted, so that the vehicle-mounted system can flexibly respond to complex long-tail scenarios, and the priority-based dynamic adaptive scheduling process ensures that when resources are limited, the vehicle-mounted system can give priority to high-priority long-tail scenarios, with high processing efficiency, avoiding waste of resources and processing delays, and optimizing the efficiency of the entire data processing chain. From data collection to final resource allocation, the entire process realizes adaptive optimization, so that the vehicle-mounted system has certain flexibility and response capabilities when facing complex and rare long-tail scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the related technologies, the drawings required for use in the specific embodiments or the related technical descriptions will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0018] Figure 1 A schematic diagram of a process for calculating resource allocation provided by an embodiment of the present invention is shown;

[0019] Figure 2 A schematic diagram of the process of constructing a global data set in a long-tail scenario is shown;

[0020] Figure 3 A schematic diagram showing the structure of a device for calculating resource allocation provided by an embodiment of the present invention is shown;

[0021] Figure 4 A schematic diagram of the hardware structure of a computer device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0022] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0023] In related technologies, resource scheduling in long-tail scenarios is performed through static resource allocation, and computing resource allocation is mostly static pre-configuration, that is, the vehicle system allocates fixed central processing unit (CPU) or graphics processing unit (GPU) resources in advance according to the average load. Although this method is simple, it is often not flexible enough when dealing with complex scenarios in dynamic environments, especially when dealing with long-tail scenarios, which may lead to over-allocation or under-allocation of computing resources.

[0024] Related technologies also use a combination of edge computing and cloud computing to schedule resources in long-tail scenarios. The introduction of edge computing provides local data processing capabilities for autonomous driving systems and reduces data transmission delays. Combined with the advantages of cloud computing, this technology can allocate some tasks to the cloud for processing, thereby alleviating the burden on onboard computing resources. However, the implementation of edge computing is often limited by hardware resources and network bandwidth, especially when dealing with emergency or complex long-tail scenarios, and real-time performance is still somewhat insufficient.

[0025] Related technologies also use load balancing algorithms, such as round-robin scheduling or dynamic priority scheduling, to ensure that each computing task can obtain reasonable resource allocation. Such methods are generally applicable to standard task load scenarios, but when dealing with long-tail scenarios, the identification of task priorities is not accurate enough, resulting in resource allocation that may not match the complexity and urgency of the task, affecting the real-time response of the system.

[0026] Related technologies also use machine learning to drive dynamic scheduling, and use machine learning to predict the load of the vehicle system and optimize resource scheduling. Based on historical data, the vehicle system can learn and predict future resource requirements and dynamically adjust computing resources. However, this method often relies on large-scale training data and computing power, and its application in real-time complex scenarios is still limited.

[0027] According to an embodiment of the present invention, a method embodiment for computing resource allocation is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0028] In this embodiment, a method for calculating resource allocation is provided, which can be used in a vehicle-mounted terminal. Figure 1 A schematic diagram of a method for computing resource allocation provided by an embodiment of the present invention is shown. Figure 1 As shown, the process includes the following steps:

[0029] Step S101, constructing a global data set in a long-tail scenario.

[0030] In this step, the data in the long-tail scenario can come from multiple sources. The global data set in the long-tail scenario can include a sensor data set, a scene complexity data set, a computational load data set, and a historical long-tail scene recognition data set.

[0031] Step S102, based on the global data set, construct a high-dimensional sparse matrix, the high-dimensional sparse matrix includes multiple data dimensions, based on the data dimensions, perform feature dimensionality reduction to obtain a reduced feature matrix, and based on the reduced feature matrix, generate a feature sub-matrix.

[0032] In this step, a high-dimensional sparse matrix is ​​constructed based on the global data set. The high-dimensional sparse matrix includes multiple data dimensions, including: the high-dimensional sparse matrix M(t) is characterized by the following method:

[0033] M(t)={S(t),L(t),C(t),T(t)}

[0034] Among them, S(t) represents the sensor data matrix, which may include at least one of the real-time data collected by lidar, camera or millimeter wave radar. L(t) represents the scene complexity matrix, which may include at least one of the environmental characteristics such as weather or traffic. C(t) represents the computational load matrix, which can be used to record at least one of the real-time utilization of system resources such as CPU and GPU. T(t) represents the historical long-tail scene recognition data matrix, which is used to represent the rare scene labels recorded in the historical data set. T(t) is used to describe rare scene information. The construction of T(t) is based on historical data to identify rare and complex scenes at each time point. T(t) is independent of other matrices and is not a system load matrix or other matrix related to system resource utilization. T(t) aims to capture and mark special scene information that has occurred in history for identification and reference in real-time applications. Therefore, T(t) is mainly used to identify and record rare scenes in historical data, rather than being directly equal to or associated with current or past system loads.

[0035] Based on the data dimension, feature dimensionality reduction is performed to obtain the feature matrix after dimensionality reduction, which can be achieved through singular value decomposition, principal component analysis or linear discriminant analysis. In this step, different feature dimensions can be processed successively through multi-scale singular value decomposition as shown below to ensure the retention and accurate extraction of information.

[0036]

[0037] Among them, M i (t) is the i-th submatrix, representing the i-th layer features in the sparse matrix, for example, M1(t) can represent sensor data, M2(t) can represent scene complexity data, and so on. Each submatrix is ​​sparsely connected to reduce the interference between features, and the non-diagonal part is set to zero.

[0038] For each submatrix M i (t), the formula for singular value decomposition is:

[0039] M i (t) = U i ·∑ i ·V i T

[0040]

[0041] Among them, U i is a left singular vector matrix, representing the linear combination of the input features, ∑ i is a singular value matrix, representing the matrix sub-M i The importance of the features in (t), V i T is a right singular vector matrix representing the linear combination of the output features, w i (t) is the weight of the i-th layer sub-matrix, which is dynamically adjusted to adapt to the changes in scene complexity (S) and system load (t), α s is the initial weight, ensuring that the weight increases with the increase of scene complexity and system load, so that sparse data in long-tail scenes can get more attention from the overall computing network. β2 is the second weight value, which is used to adjust the effect of scene complexity (S) on weight w i (t) is the coefficient used to balance the effect of scene complexity on weights during matrix decomposition and feature extraction. γ4 is the fourth weight coefficient used to adjust the system load on weight w i (t) impact.

[0042] Through multi-scale singular value decomposition, important features in the matrix are retained, while irrelevant or sparse features are downgraded, and the most representative features can be extracted. By first decomposing the overall information matrix of autonomous driving and then performing multi-scale weight optimization, the vehicle system can effectively reduce the dimension without losing key features.

[0043] Step S103 : evaluating the characteristic sub-matrix, and determining the priority of the characteristic sub-matrix in the current scene based on the evaluation result.

[0044] In this step, based on the complexity, rarity or risk level of the scene, the importance of the long-tail scene is quantified through an improved kernel function evaluation algorithm. This can provide a clear basis for the resource allocation of the vehicle system and ensure that key scenes are given priority when computing resources are tight.

[0045] Step S104: adjusting the allocation of computing resources based on the priority.

[0046] In this step, a computing resource allocation scheme can be obtained based on the priority of the feature submatrix in the current scene and the scene complexity through an improved Bayesian optimization algorithm, which can ensure that the feature submatrix with high priority obtains computing resources first.

[0047] The method for calculating resource allocation provided in this embodiment forms a complete end-to-end dynamic resource scheduling process from global data set construction, feature dimensionality reduction, priority evaluation to resource allocation; by evaluating the priority of the feature sub-matrix, resource allocation is dynamically adjusted, so that the vehicle-mounted system can flexibly respond to complex long-tail scenarios. The priority-based dynamic adaptive scheduling process ensures that when resources are limited, the vehicle-mounted system can give priority to high-priority long-tail scenarios, with high processing efficiency, avoiding waste of resources and processing delays, and optimizing the efficiency of the entire data processing chain. From data collection to final resource allocation, the entire process realizes adaptive optimization, so that the vehicle-mounted system has certain flexibility and response capabilities when facing complex and rare long-tail scenarios.

[0048] In some optional embodiments, collecting sensor data includes: adjusting the sensor data collection frequency based on scene complexity and vehicle system load; determining scene complexity based on the frequency of the current scene appearing in scene history records, current environment complexity, and current traffic complexity; determining the vehicle system load based on a first central processing unit utilization, a first graphics processing unit utilization, and a first memory utilization in the vehicle system; and collecting sensor data based on the sensor data collection frequency.

[0049] In this implementation, the scene complexity can be quantified by the complexity of the current scene through the scene complexity function Complexity(S) to determine the sensor data collection frequency.

[0050]

[0051] Among them, T S Indicates the frequency of the current scene in the scene history, T S The lower the value of F, the rarer the current scene is, and the higher the complexity of the current scene is. env Indicates the complexity of the current environment, which can be a visibility factor of the environment, such as light intensity or fog density, F env The lower the value, the more complex the environment. traffic Indicates the current traffic complexity, which can be estimated based on traffic density or vehicle movement patterns. γ1 is the first weight coefficient, γ2 is the second weight coefficient, and γ3 is the third weight coefficient, which can be dynamically adjusted according to the importance of different scenes.

[0052] The vehicle system load can be quantified by the system load function Load(t) to measure the current usage of the vehicle system computing resources and further used to determine the frequency of sensor data collection.

[0053]

[0054] Among them, CPUutil(t) represents the utilization rate of the first central processing unit in the vehicle system at the current time t, CPUutil(t) represents the utilization rate of the first graphics processing unit in the vehicle system at the current time t, Memory util (t) represents the first memory utilization rate in the vehicle system at the current time t. The higher the value of the load function Load(t), the higher the load of the vehicle system. The sensor data collection frequency is reduced based on the current computing resource usage. This can avoid computing resource overload.

[0055] When the value of the load function Load(t) is high, it means that the utilization of the CPU, GPU, and memory is high, and the vehicle system is close to or has reached the upper limit of its computing resources. Therefore, in order to avoid further increasing the burden on the vehicle system and causing performance bottlenecks, the frequency of sensor data collection needs to be appropriately reduced to reduce the transmission and processing requirements of sensor data, thereby optimizing the use of vehicle system resources.

[0056] When the value of the load function Load(t) reaches the preset threshold, the vehicle system reduces the frequency of sensor data collection, thereby effectively reducing the computational burden brought by sensor data collection. Through the aforementioned adjustment mechanism, it can be ensured that the vehicle system can operate stably when the load is high. At the same time, it can maintain a high collection frequency when the load is low, capture richer scene data, and realize dynamic optimization management of resources.

[0057] The sensor data collection frequency f can be adjusted based on the scene complexity function Complexity(S) and the system load function Load(t) by the following formula: 采集 :

[0058] f 采集 =min(f max ,α1·Complexity(S)+β1·Load(t))

[0059] Among them, α1 is the first weight parameter, β1 is the first weight value, and f max The maximum sensor data collection frequency set for the vehicle system.

[0060] Based on the sensor data acquisition frequency f 采集 , according to the complexity of the scene and the risk level, the sensor collection frequency is automatically adjusted to capture more key information. In simple scenarios, the sensor data collection frequency is automatically reduced to reduce the transmission and processing of redundant information.

[0061] In this way, it can be ensured that when the vehicle system load is light, the sensor data collection frequency in complex scenarios will automatically increase, and when the vehicle load is high, the sensor data collection frequency in simple scenarios will decrease, thereby optimizing the use of vehicle system resources.

[0062] Long-tail scenarios occur very rarely in daily driving, which makes it difficult for the autonomous driving system to obtain sufficient long-tail scenario data during the training process, making it difficult to ensure the reliability and accuracy of the autonomous driving system in long-tail scenarios.

[0063] In some optional embodiments, Figure 2 A schematic diagram of the process of constructing a global data set in a long-tail scenario is shown, Figure 2 As shown, including:

[0064] Step S201 , collecting sensor data, and constructing a sensor historical data set and a sensor real-time data set.

[0065] In this step, the sensor data acquisition frequency f 采集 , to construct the autonomous driving sensor data set. The sensor data in the autonomous driving system is generated by multiple sensors at time t, and the sensor data matrix S(t) is represented as follows:

[0066] S(t)={S L i DAR (t), S Camera (t), S Radar (t), S Ultrasonic (t)}

[0067] Among them, S LiDAR (t) is the point cloud data of the lidar at time t, S Camera (t) is the image data captured by the vehicle camera at time t, S Radar (t) is the target information obtained by the millimeter-wave radar at time t, S Ultrasonic (t) is the close-range obstacle data collected by the ultrasonic sensor at time t.

[0068] The sensor data matrix S(t) contains multi-dimensional information, including spatial position information, object recognition data, and speed data. To further construct the data set, the data is aligned in time series based on the moment of data acquisition and stored.

[0069] Among them, sensor historical data set It can be expressed as:

[0070]

[0071] Real-time sensor data set It only contains data at the current time:

[0072]

[0073] Step S202 , constructing a computing load historical data set and a computing load real-time data set based on the second central processing unit utilization, the second graphics processing unit utilization, and the second memory utilization of the vehicle-mounted system in processing sensor data.

[0074] In this step, the load matrix C(t) is calculated, which includes the utilization of the CPU, GPU, and memory of the vehicle system when processing sensor data.

[0075] C(t)={CPU usage (t),GPU usage (t),MEM usage (t),BW usage (t)}

[0076] Among them, CPU usage (t) is the second CPU utilization at time t, GPU usage (t) is the second GPU utilization at time t, MEM usage (t) is the second memory utilization at time t, BW usage (t) is the network bandwidth utilization at time t. Similar to the sensor data matrix S(t) mentioned above, the load matrix C(t) is also stored in time series to calculate the load history data set. It can be expressed as:

[0077]

[0078] Compute Load Real-time Dataset It can be expressed as:

[0079]

[0080] Step S203, based on the scene type, the score corresponding to the scene type, the response time of the vehicle system for processing the scene type, and the resources consumed by the vehicle system for processing the scene type, determine the long-tail scene recognition record, and based on the long-tail scene recognition record, construct a long-tail scene historical data set and a long-tail scene real-time data set.

[0081] In this step, the long-tail scene recognition matrix L(t) includes the rare and complex scene information recognized at a specific time t.

[0082] L(t)={Scenario type (t), Severity(t),

[0083] Response time(t), Resource_consumption(t)}

[0084] Among them, Scenario type (t) is the scene type identified at time t, such as traffic congestion or extreme weather, Severity(t) is the complexity or rarity score of the scene, which can be set to the score corresponding to the scene type, such as 1-10 points, etc. Response time(t) t) is the response time of the vehicle system to the scenario, Resource_consumption(t) is the resources consumed by the vehicle system when processing the scenario, and the long-tail scenario historical data set It can be expressed as:

[0085]

[0086] Long-tail scenario real-time dataset It can be expressed as:

[0087]

[0088] Step S204, integrating the sensor historical data set, the computing load historical data set and the long-tail scenario historical data set to construct a historical data set; integrating the sensor real-time data set, the computing load real-time data set and the long-tail scenario real-time data set to construct a real-time data set.

[0089] In this step, after constructing sensor data, computing load information, and long-tail scene recognition records, further integration is performed to construct historical data sets and real-time data sets. history Contains sensor data, computing load information, and long-tail scene recognition records for past time periods:

[0090]

[0091] Real-time dataset D realtime The system processing decision for the current moment is expressed as:

[0092]

[0093] As the running time of the vehicle system increases, the data in the real-time dataset is updated regularly and added to the historical dataset:

[0094] D history ←D history ∪D realtime

[0095] Real-time datasets are updated with the latest data:

[0096] D realtime={S(t+1),C(t+1),L(t+1)}

[0097] Based on this feedback mechanism, the gradual expansion of the historical data set is further ensured, and the processing results of each real-time data will be reflected in the historical data in a timely manner.

[0098] In this way, based on the real-time long-tail scene acquisition framework in the embodiment of the present invention, various sensor data in the autonomous driving system can be integrated, and based on the sensor data acquisition frequency f 采集 , automatically improves the collection rate in complex scenarios, can capture more key data, and the data scarcity problem caused by the rarity of long-tail scenarios in daily driving is alleviated through dynamic sampling and historical data accumulation mechanisms. It further ensures that as the system runs, the historical data set continues to expand, thereby providing more data support for long-tail scenarios in training and actual applications, and improving the reliability and accuracy of the system in these scenarios.

[0099] In some optional implementations, after obtaining the reduced-dimensional feature matrix, the method further includes: completing a low-rank matrix based on the reduced-dimensional feature matrix to obtain a completed matrix; and generating a feature sub-matrix based on the completed matrix.

[0100] In this embodiment, due to the characteristics of the long-tail scenario, some important information may be lost in the process of matrix dimensionality reduction. In order to restore these features to a certain extent, the following low-rank matrix completion method can be used to repair the lost information on the sub-matrix:

[0101]

[0102] Among them, ∥M i (t)∥ * is the nuclear norm of the matrix, indicating the low-rank property of the submatrix, λ is the regularization coefficient, ensuring that the matrix can restore the low-rank structure of the original data after compression, M reduced,i (t) is the sub-matrix after dimensionality reduction, and M i (t) is the matrix completed by the repair mechanism.

[0103] This repair mechanism can effectively restore key sparse features after dimensionality reduction, avoid information loss affecting subsequent priority evaluation, and ensure information integrity in long-tail scenarios. Through sparse matrix decomposition and feature repair mechanisms, the system can effectively compress high-dimensional sensor data, reduce computational complexity, and ensure that important features are retained. This is actually more important when dealing with long-tail scenarios, which can improve the accuracy of feature recognition of the vehicle system and enhance the accuracy of subsequent evaluation and decision-making.

[0104] Through the aforementioned improved sub-matrix feature repair mechanism, the key information in the dimensionality reduction process is supplemented, and the obtained feature matrix after dimensionality reduction is still a high-dimensional and complex matrix to a certain extent. The embodiment of the present invention proposes a sparse weighted convolutional neural network (Sparse Weighted Convolutional Neural Network, referred to as SW-CNN) extraction layer to further extract the intrinsic correlation between these features and automatically identify which features are more important for the current scene.

[0105] In some optional implementations, generating a feature sub-matrix based on the feature matrix after dimensionality reduction includes: performing a convolution operation based on the feature matrix after dimensionality reduction; and adjusting the weight of the convolution kernel during the convolution operation to generate a feature sub-matrix.

[0106] In this embodiment, a convolution operation is performed on each feature matrix after dimension reduction, that is, a submatrix, and the i-th submatrix M sub,i Convolution operation is performed in the matrix to extract the relationship between the internal features of the sub-matrix.

[0107]

[0108] in, is the feature activation value in the submatrix, is the weight of the convolution kernel in the matrix, is the eigenvalue in the submatrix, K (sub) is the convolution kernel in the sub-matrix, and σ is the activation function.

[0109] Within each sub-matrix, the weights of the convolution kernel are still dynamically adjusted according to the sparse features.

[0110]

[0111] η (sub) (t)=α2·Complexity(S)+β3·R(S)+γ5·f(t)

[0112] Among them, η (sub) (t) is the sub-matrix adjustment factor, Complexity(S) is the scene complexity, representing the rarity or difficulty of the scene, R(S) is the risk level of the scene, such as the sudden danger or emergency situation that the system may face, f(t) is a time-related factor, indicating that the higher the system's real-time requirements, the more aggressive the adjustment of the convolution weight, α2 is the second weight coefficient, which is used to adjust the impact of the scene complexity in the sub-matrix adjustment factor. β3 is the third weight value, which is used to adjust the impact of the scene's risk level in the sub-matrix adjustment factor. γ5 is the fifth weight coefficient, which is used to adjust the impact of time-related factors in the sub-matrix adjustment factor.

[0113] Based on the aforementioned convolution operation, key information for scene recognition is generated.

[0114] In this way, through SW-CNN feature extraction, the features in the sub-matrix can be further integrated, so that the vehicle system can quickly identify and process important features in emergencies. The repaired multi-dimensional feature matrix is ​​fused with feature regions. Since the data in long-tail scenes are sparse and unevenly important, the weights of the convolution kernels are dynamically adjusted to enable the neural network to prioritize the features that are critical to scene recognition.

[0115] Different from the fusion in related technologies, which is processed across objects and across sensor devices, the embodiment of the present invention performs fusion processing between sub-matrices. Each sub-matrix only contains one type of highly correlated features. By performing fusion within each sub-matrix, the computational burden of complex fusion between multi-dimensional data can be reduced. It is suitable for sparse data in long-tail scenarios, because sparse features are often concentrated in certain specific dimensions, rather than being evenly distributed in all data dimensions. In addition, in long-tail scenarios, the data source may contain noise or irrelevant features. If global fusion is performed in all dimensions, noise may interfere with the overall result. Limiting the fusion to the sub-matrix can effectively reduce the impact of noise and make the fusion of each sub-matrix more accurate.

[0116] In some optional embodiments, the feature submatrix is ​​evaluated, and the priority of the feature submatrix in the current scene is determined based on the evaluation results, including: generating nodes in a graph based on the feature submatrix; constructing a multi-layer graph based on time series and nodes; determining the similarity between nodes based on the temporal similarity and spatial similarity between the multi-layer graphs; and determining the priority of the feature submatrix in the current scene based on the similarity.

[0117] In this embodiment, the characteristic submatrices are set as nodes in the graph, and a multi-layer graph is constructed based on the time series and nodes in the following manner:

[0118] G sub (t)=(V sub (t),E sub (t))

[0119] Among them, V sub (t) is a node of the graph, representing the state of different sub-matrices at a certain time step t, E sub (t) is the edge of the graph, which represents the relationship and interaction between different sub-matrices at a certain time step t. In order to handle multi-dimensional scenarios, the embodiment of the present invention not only focuses on the sub-matrix features at the current moment, but also considers the changes of these features over time, so it is based on a time-series multi-layer graph structure. Each time step forms a sub-graph G sub(t), multiple time steps constitute a dynamic multi-layer graph G sub (t1), (t2), …, G sub (t n ).

[0120] Each node not only represents a single submatrix, but also includes its historical and real-time feature interaction information. The constructed graph not only contains the feature submatrix of a single time slice, but also introduces the time evolution characteristics of each feature submatrix to obtain a multi-layer graph.

[0121] Based on the graph kernel technology, the feature sub-matrix is ​​evaluated to clarify the priority of the feature sub-matrix in the current scene. Based on the hybrid graph kernel of space and time, the complex interactive relationship of the feature sub-matrix in the spatial and temporal dimensions is captured. That is, the criterion for judging the priority level is mainly: the similarity between the sub-matrices. If two processed sub-matrices, that is, the sets of important scene features, are very similar in space or time, it indicates that the two scenes are likely to capture similar events or dynamics. For example, in one sub-matrix, the sensor data indicates that an obstacle suddenly appears on the road ahead. If another sub-matrix also shows a similar sensor pattern, they are identified as scenes with the same risk and their processing priority is increased.

[0122] An improved kernel evaluation method is used to calculate the priority of each feature submatrix in the current scene. Based on the combination of the Weisfeiler-Lehman kernel and the temporal kernel, the feature changes in the feature submatrix in space and time are comprehensively evaluated:

[0123]

[0124] Among them, K(G i ,G j ) Characterization G i and G j The similarity between the two submatrix graphs, H is the number of layers of the Weissfeller-Lyman graph kernel, which is used to calculate the spatial similarity of the nodes, k h is the h-th layer similarity calculation, are the node labels of the h-th layer submatrix graph, where α3 is the weight of spatial similarity, β4 is the weight of temporal similarity, and K time It is the time graph kernel, which measures the similarity of sub-matrices at different time steps and can be set as the basic Gaussian kernel function.

[0125] Through the above hybrid graph kernel formula, the spatial correlation and temporal correlation between feature sub-matrices are evaluated to ensure that feature sub-matrices with high similarity can obtain higher priorities.

[0126] Based on the similarity, the priority of the feature submatrix in the current scene is determined, including: priority evaluation through the following formula:

[0127] Priority(V i )=λ1·Sim space (V i ,V j )

[0128] +λ2·Sim time (V i (t1),V j (t2))+λ3·R i

[0129] Among them, Sim space is the spatial similarity of the feature submatrices at the same time, which is used to evaluate the feature correlation between different feature submatrices. time is the temporal similarity of the feature submatrix between different time steps, which is used to evaluate the change of the feature submatrix features over time. i is the complexity, sparsity or risk level of the scene, which is used to ensure that the feature submatrices with high risk and strong sparsity are processed first. λ1 is the weight of spatial similarity, which is used to adjust the priority contribution of spatial similarity. λ2 is the weight of temporal similarity, which is used to adjust the priority contribution of temporal similarity. λ3 is the weight of scene importance, which is used to adjust the priority contribution of scene importance.

[0130] In this way, the constructed temporal multi-layer graph kernel evaluation method combines information in spatial and temporal dimensions to ensure that the on-board system can identify complex behaviors and environmental changes in long-tail scenarios, and ensure the system's efficient response and processing in diverse and complex scenarios.

[0131] In the priority evaluation stage, a hybrid kernel model is constructed by combining the Weisfeiler-Lehman kernel and the time kernel to measure the spatial and temporal correlations between feature sub-matrices. The vehicle system not only relies on the data at the current moment, but also considers the changing trend in the time dimension, making the priority evaluation more comprehensive and accurate. The hybrid kernel model can accurately capture the dynamic similarities between feature sub-matrices, especially in rare long-tail scenarios. It can effectively evaluate the importance of complex scenarios and ensure that the vehicle system prioritizes key scenarios when resources are tight. It not only improves the resource utilization efficiency of the vehicle system, but also enhances the resilience of the vehicle system in dealing with sudden scenarios.

[0132] By real-time monitoring and evaluation of scenario complexity, feature sub-matrix priority, and vehicle system load, resource allocation is dynamically adjusted, enabling the vehicle system to flexibly respond to complex long-tail scenarios.

[0133] The dynamic and adaptive scheduling process ensures that the system can prioritize high-priority long-tail scenarios when resources are limited, avoiding resource waste and processing delays. The efficiency of the entire data processing chain can be optimized, and adaptive optimization is achieved throughout the entire process from data collection to final resource allocation, giving it a certain degree of flexibility and coping capabilities when facing complex and rare long-tail scenarios.

[0134] In some optional embodiments, the allocation of computing resources is adjusted based on priority, including: determining the resource allocation ratio based on the priority and the complexity of the scenario associated with the feature submatrix; and determining the computing resources allocated to the feature submatrix based on the resource allocation ratio and the total amount of computing resources currently allocatable.

[0135] In this embodiment, the dynamic allocation of computing resources can be performed by an improved Bayesian optimization algorithm. The goal of the improved Bayesian optimization is to maximize the processing efficiency of the vehicle system under high-priority tasks. An objective function f(V) is proposed, which is the processing efficiency corresponding to the resources obtained by the characteristic submatrix. The optimization goal is to maximize the processing efficiency of the high-priority submatrix:

[0136]

[0137] Among them, Priority(V u ) is the characteristic submatrix V u The priority of Resource(V u ) is assigned to the feature submatrix V u The computing resources, T comp (V u ) is the characteristic submatrix V u The expected calculation time can reflect the processing speed of the computing resources for the characteristic submatrix.

[0138] By searching in multidimensional space, the high priority submatrix V u Resource allocation Resource (V u ) makes the objective function f(V u ) reaches its maximum value.

[0139] Further according to the results of Bayesian optimization, the following resource allocation formula is used to make the sub-matrices with high priority and in complex scenarios obtain more resource allocation:

[0140]

[0141] Among them, Resource (V u ) is assigned to the feature submatrix V u The computing resources of Priority(V i ) is the characteristic submatrix Vi The priority of C(S i ) is the characteristic submatrix V i The complexity of the associated scenario can be calculated based on the sparsity, complexity and risk level of the scenario. Total Resource is the total amount of computing resources currently available for allocation, and n is the total number of feature sub-matrices.

[0142] The resource allocation of each feature submatrix depends not only on its priority but also on the complexity of the scene. Submatrices with high priority and in complex scenes will receive more resource allocation.

[0143] When faced with multiple feature submatrix tasks, each feature submatrix represents a different scenario, task or sensor data, and the vehicle system needs to determine which tasks should be prioritized. Through priority evaluation, the vehicle system can clearly understand the importance of each submatrix. Through the improved Bayesian optimization algorithm, a resource allocation plan can be obtained based on priority and scenario complexity to ensure that high-priority submatrices get computing resources first.

[0144] Resources can be dynamically allocated to feature submatrix tasks based on the total amount of resources in the current system, such as the computing power of CPU, GPU, etc., and the priority of the task. Feature submatrices with high priority get more resources, while feature submatrices with low priority are allocated fewer resources based on their actual needs. Through this mechanism, the system avoids waste of resources and ensures that the system processing speed can be maintained when processing highly complex scenarios.

[0145] In this way, after completing the feature submatrix priority evaluation, the vehicle system dynamically adjusts the allocation of computing resources on the vehicle or cloud according to the priority of each feature submatrix, scene complexity, and system load. This ensures that high-priority long-tail scenes can be processed first when resources are limited, so as to maintain the efficient operation of the system in complex scenarios.

[0146] In some optional implementations, after the above resource allocation, the tasks of each feature submatrix will be executed in the vehicle system, and the high-priority feature submatrix will have priority in obtaining the computing power of the vehicle system to ensure the completion time and response speed of key tasks, while low-priority tasks can remain uninterrupted. After resource allocation, the vehicle system does not stop the dynamic monitoring and adjustment of resources, and continues to detect certain tasks. If their priority changes, such as the danger of a certain scene increases, the vehicle system can re-evaluate the priority of the feature submatrix and perform secondary allocation of resources to cope with the newly emerged high-priority scenes. Ultimately, the ability to process long-tail scenes is greatly improved.

[0147] Long-tail scenarios may involve a wide variety of factors, including complex behaviors of traffic participants, such as sudden intrusion of pedestrians or vehicles, and environmental changes, such as terrain or weather. The on-board system needs to be able to identify and handle these emergencies.

[0148] In the embodiment of the present invention, multi-scale singular value decomposition and sparse matrix dimensionality reduction can be used to extract and retain important feature information in long-tail scenarios. The proposed SW-CNN feature extraction method can further integrate the features in the feature submatrix, so that the vehicle system can quickly identify and process important features in emergency situations.

[0149] In addition, the time-series multi-layer graph kernel evaluation method constructed in the embodiment of the present invention can combine information in spatial and temporal dimensions to ensure that the vehicle-mounted system can identify complex behaviors and environmental changes in long-tail scenarios, and dynamically allocate computing resources to give priority to high-risk and complex scenarios, thereby ensuring the efficient response and processing of the vehicle-mounted system in diverse and complex scenarios.

[0150] The embodiments of the present invention can effectively solve the problems of data scarcity, diversity and complexity in long-tail scenarios through innovative sensor data collection and data processing technologies, feature extraction mechanisms and resource allocation strategies.

[0151] In the present embodiment, a device for computing resource allocation is also provided, and the device is used to implement the above-mentioned embodiment and preferred implementation mode, and the descriptions that have been made are not repeated. As used below, the term "module" can implement a combination of software and / or hardware of a predetermined function. Although the device described in the following embodiments is preferably implemented in software, the implementation of hardware, or a combination of software and hardware is also possible and conceived.

[0152] This embodiment provides a device for calculating resource allocation. Figure 3 A schematic diagram of the structure of a device for calculating resource allocation provided by an embodiment of the present invention is shown, Figure 3 As shown, including:

[0153] The data set construction module 301 is used to construct a global data set in a long-tail scenario.

[0154] The feature function extraction module 302 is used to construct a high-dimensional sparse matrix based on the global data set. The high-dimensional sparse matrix includes multiple data dimensions. Based on the data dimensions, feature dimensionality reduction is performed to obtain a reduced feature matrix. Based on the reduced feature matrix, multiple feature sub-matrices are generated.

[0155] The priority evaluation module 303 is used to evaluate the feature sub-matrix and determine the priority of the feature sub-matrix in the current scene based on the evaluation result.

[0156] The computing resource allocation module 304 is used to adjust the allocation of computing resources based on the priority.

[0157] In some optional implementations, the data set construction module 301 includes:

[0158] The first data set construction unit is used to collect sensor data and construct a sensor historical data set and a sensor real-time data set.

[0159] The second data set construction unit is used to construct a computing load historical data set and a computing load real-time data set based on the second central processing unit utilization, the second graphics processing unit utilization and the second memory utilization of the vehicle-mounted system in processing sensor data.

[0160] The third data set construction unit is used to determine the long-tail scene recognition records based on the scene type, the score corresponding to the scene type, the response time of the vehicle system for processing the scene type, and the resources consumed by the vehicle system for processing the scene type, and to construct a long-tail scene historical data set and a long-tail scene real-time data set based on the long-tail scene recognition records.

[0161] The fourth data set construction unit is used to integrate the sensor historical data set, the computing load historical data set and the long-tail scenario historical data set to construct a historical data set; integrate the sensor real-time data set, the computing load real-time data set and the long-tail scenario real-time data set to construct a real-time data set.

[0162] In some optional implementations, the first data set construction unit includes:

[0163] The first data set construction subunit is used to adjust the sensor data collection frequency based on the scene complexity and the vehicle system load; determine the scene complexity based on the frequency of the current scene appearing in the scene history record, the current environment complexity and the current traffic complexity; determine the vehicle system load based on the first central processing unit utilization, the first graphics processing unit utilization and the first memory utilization in the vehicle system; and collect sensor data based on the sensor data collection frequency.

[0164] In some optional implementations, the feature function extraction module 302 further includes:

[0165] The first feature extraction unit is used to complete the low-rank matrix based on the feature matrix after dimension reduction to obtain a completed matrix; and to generate a feature sub-matrix based on the completed matrix.

[0166] In some optional implementations, the feature function extraction module 302 further includes:

[0167] The second feature extraction unit is used to perform a convolution operation based on the feature matrix after dimensionality reduction; during the convolution operation, the weight of the convolution kernel is adjusted to generate a feature sub-matrix.

[0168] In some optional implementations, the feature function extraction module 302 further includes:

[0169] The third feature extraction unit is used to generate nodes in the graph based on the feature sub-matrix; construct a multi-layer graph based on the time series and the nodes; determine the similarity between the nodes based on the time similarity and spatial similarity between the multi-layer graphs; and determine the priority of the feature sub-matrix in the current scene based on the similarity.

[0170] In some optional implementations, the computing resource allocation module 304 includes:

[0171] The first computing resource allocation unit is used to determine the resource allocation ratio based on the priority and the complexity of the characteristic sub-matrix associated scenario; based on the resource allocation ratio and the total amount of computing resources currently allocatable, determine the computing resources allocated to the characteristic sub-matrix.

[0172] The further functional description of each of the above modules and units is the same as that of the above corresponding embodiments and will not be repeated here.

[0173] The device for computing resource allocation in this embodiment is presented in the form of a functional unit, where the unit refers to an application specific integrated circuit (ASIC) circuit, a processor and memory that executes one or more software or fixed programs, and / or other devices that can provide the above functions.

[0174] The embodiment of the present invention also provides a computer device having the above Figure 3 An apparatus for computing resource allocation is shown.

[0175] See also Figure 4 , Figure 4 is a schematic diagram of the structure of a computer device provided by an optional embodiment of the present invention, such as Figure 4As shown, the computer device includes: one or more processors 10, a memory 20, and interfaces for connecting various components, including high-speed interfaces and low-speed interfaces. Various components are connected to each other using different buses for communication, and can be installed on a common mainboard or installed in other ways as needed. The processor can process instructions executed in the computer device, including instructions stored in or on the memory to display graphical information of a graphical user interface on an external input / output device (such as a display device coupled to an interface). In some optional embodiments, if necessary, multiple processors and / or multiple buses can be used together with multiple memories and multiple memories. Similarly, multiple computer devices can be connected, and each device provides some necessary operations (for example, as a server array, a group of blade servers, or a multi-processor system). Figure 4 A processor 10 is taken as an example.

[0176] The processor 10 may be a central processing unit, a network processor or a combination thereof. The processor 10 may further include a hardware chip. The hardware chip may be a dedicated integrated circuit, a programmable logic device or a combination thereof. The programmable logic device may be a complex programmable logic device, a field programmable gate array, a general purpose array logic or any combination thereof.

[0177] The aforementioned memory 20 stores instructions executable by at least one processor 10, so that the aforementioned at least one processor 10 executes the method shown in the above embodiment.

[0178] The memory 20 may include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function; the data storage area may store data created according to the use of the computer device, etc. In addition, the memory 20 may include a high-speed random access memory, and may also include a non-transient memory, such as at least one disk storage device, a flash memory device, or other non-transient solid-state storage device. In some optional embodiments, the memory 20 may optionally include a memory remotely arranged relative to the processor 10, and these remote memories may be connected to the computer device via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0179] The memory 20 may include a volatile memory, such as a random access memory; the memory may also include a non-volatile memory, such as a flash memory, a hard disk or a solid state drive; the memory 20 may also include a combination of the above types of memory.

[0180] The computer device also includes an input device 30 and an output device 40. The processor 10, the memory 20, the input device 30 and the output device 40 may be connected via a bus or other means. Figure 4 The example of connecting through bus is taken in the following.

[0181] The input device 30 can receive input digital or character information, and generate key signal input related to the user settings and function control of the computer device, such as a touch screen, a keypad, a mouse, a track pad, a touch pad, an indicator rod, one or more mouse buttons, a trackball, a joystick, etc. The output device 40 may include a display device, an auxiliary lighting device (such as a light emitting diode) and a tactile feedback device (such as a vibration motor), etc. The above-mentioned display device includes but is not limited to a liquid crystal display, a light emitting diode, a display and a plasma display. In some optional embodiments, the display device can be a touch screen.

[0182] The embodiment of the present invention also provides a computer-readable storage medium. The method according to the embodiment of the present invention can be implemented in hardware, firmware, or can be implemented as a computer code that can be recorded in a storage medium, or can be implemented as a computer code that is originally stored in a remote storage medium or a non-temporary machine-readable storage medium and will be stored in a local storage medium through a network download, so that the method described herein can be stored in such software processing on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. Among them, the storage medium can be a magnetic disk, an optical disk, a read-only storage memory, a random access memory, a flash memory, a hard disk or a solid-state hard disk, etc.; further, the storage medium can also include a combination of the above types of memories. It can be understood that a computer, a processor, a microprocessor controller, or programmable hardware includes a storage component that can store or receive software or computer code. When the software or computer code is accessed and executed by a computer, a processor, or hardware, the method shown in the above embodiment is implemented.

[0183] A part of the present invention may be applied as a computer program product, such as a computer program instruction, which, when executed by a computer, can call or provide the method and / or technical solution according to the present invention through the operation of the computer. Those skilled in the art should understand that the existence of the computer program instruction in a computer-readable medium includes, but is not limited to, a source file, an executable file, an installation package file, etc., and accordingly, the way in which the computer program instruction is executed by the computer includes, but is not limited to: the computer directly executes the instruction, or the computer compiles the instruction and then executes the corresponding compiled program, or the computer reads and executes the instruction, or the computer reads and installs the instruction and then executes the corresponding installed program. Here, the computer-readable medium may be any available computer-readable storage medium or communication medium accessible to the computer.

[0184] Although the embodiments of the present invention have been described in conjunction with the accompanying drawings, those skilled in the art may make various modifications and variations without departing from the spirit and scope of the present invention, and such modifications and variations are all within the scope defined by the appended claims.

Claims

1. A method for computing resource allocation, characterized in that The method comprises: Build a global dataset for long-tail scenarios; Based on the global data set, a high-dimensional sparse matrix is ​​constructed, wherein the high-dimensional sparse matrix includes multiple data dimensions, feature dimension reduction is performed based on the data dimensions to obtain a feature matrix after dimension reduction, and a feature submatrix is ​​generated based on the feature matrix after dimension reduction; Evaluate the feature sub-matrix, and determine the priority of the feature sub-matrix in the current scene based on the evaluation result; Allocation of computing resources is adjusted based on the priorities.

2. The method according to claim 1, characterized in that The construction of a global data set in a long-tail scenario includes: Collect sensor data and build sensor historical data sets and sensor real-time data sets; Based on the utilization rate of the second central processing unit, the utilization rate of the second graphics processing unit, and the second memory of the vehicle system in processing the sensor data, a computing load historical data set and a computing load real-time data set are constructed; Determine a long-tail scene recognition record based on the scene type, the score corresponding to the scene type, the response time of the vehicle system for processing the scene type, and the resources consumed by the vehicle system for processing the scene type, and construct a long-tail scene historical data set and a long-tail scene real-time data set based on the long-tail scene recognition record; Integrate the sensor historical data set, the computing load historical data set, and the long-tail scenario historical data set to construct a historical data set; The sensor real-time data set, the computing load real-time data set and the long-tail scenario real-time data set are integrated to construct a real-time data set.

3. The method according to claim 2, characterized in that The collecting of sensor data includes: Adjust the frequency of sensor data collection based on scene complexity and vehicle system load; Determining the scene complexity based on the frequency of the current scene appearing in the scene history records, the current environment complexity, and the current traffic complexity; Determining the vehicle system load based on a first central processing unit utilization rate, a first graphics processing unit utilization rate, and a first memory utilization rate in the vehicle system; The sensor data is collected based on the sensor data collection frequency.

4. The method according to claim 1, characterized in that: After obtaining the feature matrix after dimension reduction, the method further includes: Based on the feature matrix after dimension reduction, low-rank matrix completion is performed to obtain a completed matrix; Based on the completed matrix, the characteristic sub-matrix is ​​generated.

5. The method according to claim 1, characterized in that The generating a feature sub-matrix based on the feature matrix after dimension reduction includes: Performing a convolution operation based on the feature matrix after dimension reduction; During the convolution operation, the weight of the convolution kernel is adjusted to generate the feature sub-matrix.

6. The method according to claim 1, characterized in that The step of evaluating the characteristic sub-matrix and determining the priority of the characteristic sub-matrix in the current scene based on the evaluation result includes: Based on the characteristic submatrix, generating nodes in the graph; Based on the time series and the nodes, a multi-layer graph is constructed; Determining the similarity between the nodes based on the temporal similarity and the spatial similarity between the multi-layer graphs; Based on the similarity, the priority of the feature sub-matrix in the current scene is determined.

7. The method according to claim 1, characterized in that The adjusting the allocation of computing resources based on the priority comprises: Determining a resource allocation ratio based on the priority and the complexity of the scenario associated with the feature submatrix; The computing resources allocated to the characteristic sub-matrix are determined based on the resource allocation ratio and the total amount of computing resources currently available for allocation.

8. An apparatus for computing resource allocation, characterized in that The device comprises: The dataset construction module is used to construct a global dataset in long-tail scenarios; A feature function extraction module is used to construct a high-dimensional sparse matrix based on the global data set, wherein the high-dimensional sparse matrix includes multiple data dimensions, perform feature dimension reduction based on the data dimensions to obtain a feature matrix after dimension reduction, and generate multiple feature sub-matrices based on the feature matrix after dimension reduction; A priority evaluation module, used to evaluate the feature sub-matrix and determine the priority of the feature sub-matrix in the current scene based on the evaluation result; A computing resource allocation module is used to adjust the allocation of computing resources based on the priority.

9. A computer device, characterized in that: include: A memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the method for computing resource allocation according to any one of claims 1 to 7 by executing the computer instructions.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer instructions, and the computer instructions are used to enable a computer to execute the method for computing resource allocation according to any one of claims 1 to 7.

Images