133results about How to "Reduce memory usage" patented technology

The invention relates to the technical field of network acceleration, and discloses an edge-end large language model reasoning acceleration method and accelerator, and the method comprises the following steps: reconstructing a calculation process of a decoding stage, and carrying out the deep fusion of a multi-head attention mechanism and the calculation operation of a feedforward network; the weight and key value data are stored in HBM, and the coefficient and the accumulated attention score are stored in DDR; for linear matrix calculation, a unified matrix calculation unit is used for executing multi-precision matrix operation; for nonlinear function calculation, a mathematical transformation and linear fitting method is adopted, a Softmax function is converted into operation with 2 as the bottom through a bottom conversion formula, and truncation and third-order linear fitting are conducted on a Sigmoid function; and constructing a key value screening algorithm based on the accumulated attention score, dynamically adjusting a key value storage position, maintaining a recent key value cache region and an important key value cache region in a limited cache space, and realizing key value efficient cache in long text reasoning.

The invention relates to an agent selection method and device, a task processing method and device, equipment, a medium and a product. The method comprises the steps of obtaining a target task and preference information for the target task; decomposing the target task to obtain sub-tasks of the target task; obtaining capability labels of the candidate intelligent agents; the capability label comprises quantized values of a plurality of capability dimensions; according to a weight vector in the preference information, carrying out weighted calculation on the quantized numerical value of each capability dimension of the candidate intelligent agent to obtain a comprehensive score of the candidate intelligent agent; and based on the comprehensive score, selecting a target agent for executing the subtask from the candidate agents. By adopting the method, the resource utilization efficiency can be improved.

The invention discloses an aero-engine vibration signal generation method under extreme imbalance, and belongs to the technical field of aero-engine fault signal generation. The method comprises the following steps: acquiring an original vibration signal at a target position of the aero-engine, and performing preprocessing operation; taking the preprocessed real vibration signal as a training sample, constructing a forward noise adding process, and training the improved one-dimensional diffusion generation network to learn a reverse denoising mapping relation; and inputting random Gaussian noise into the one-dimensional diffusion generation network after training convergence, and generating a target vibration signal matched with a real vibration signal feature through a reverse diffusion denoising process. Through targeted improvement of the diffusion generation network, the problem of data imbalance is effectively relieved, the model deployment cost is reduced, the signal generation capability of the model for minority types of faults is improved, and the method is suitable for fault signal generation scenes of aero-engines under complex working conditions.

The invention discloses a file system event processing method and device based on an intelligent filtering rule, and the method comprises the steps: obtaining a file system original event of a monitoring directory, carrying out the cross-platform standardization processing, and extracting event feature vectors of four dimensions: time, a file, a process and a path; the method comprises the following steps: constructing a rule base containing software operation modes such as an editor, a compiler and a version control tool, performing mode matching on an event sequence, calculating confidence, determining a filtering strategy type based on a matching result, executing corresponding processing, and dynamically optimizing system parameters through statistical analysis. The technical problems that an existing file monitoring system is serious in event redundancy and frequent in false triggering are solved, effective file operation and temporary intermediate operation can be intelligently distinguished, invalid synchronous triggering is greatly reduced, the accuracy is improved to 95% or above from 60-70% of a traditional scheme, and invalid event processing is reduced by 90% or above.

The present application relates to a kind of WebGL-based pipe network digital twin model light weight rendering method, method includes: resource preconfiguration: the grid data of different three-dimensional model is generated corresponding instance template;Through instance template generation instance is stored in instance pool;In scene graph, create include call data and dynamic data lightweight placeholder, call data is used to extract instance, and dynamic data is used to adjust the dynamic display of instance;Lightweight rendering: set the visibility screening strategy of placeholder, whether current placeholder is visible is judged in turn through the screening condition of level, distance and perspective, view cone intersection;Set instance management strategy, for the placeholder screened as visible, corresponding instance is called from instance pool and is bound, and the binding relationship is recorded, for the placeholder screened as invisible, unbind, and the instance after unbinding is marked as idle;The placeholder of bound instance is rendered and output.Effectively reduce the memory occupation when rendering.

This invention discloses a real-time rendering method and system for physically simulated volumetric clouds. The method includes the following steps: constructing a dynamic noise mixing model, and constructing complex cloud layers by mixing multiple noises based on the dynamic noise mixing model; constructing a multi-scattering illumination model, and using the multi-scattering illumination model to perform single-scattering calculations and multi-scattering approximations on the complex cloud layers to simulate illumination; based on the completed illumination simulation, dividing the cloud density field of the complex cloud layers into several voxel blocks, and performing adaptive light travel processing to complete the real-time rendering of volumetric clouds. This invention, without sacrificing image quality and rendering efficiency, integrates advanced cloud modeling theory with efficient GPU acceleration technology, and uses a mixed noise model to generate diverse cloud structures, accurately reproducing various typical cloud types such as high-altitude cirrus clouds, cumulus clouds, and stratus clouds.

This invention discloses a hybrid inverse kinematics method for rotating biprisms that deeply integrates deterministic physics prior models with reinforcement learning. It solves the long-standing problem of balancing computational efficiency and pointing accuracy in existing rotating biprism beam pointing inverse kinematics techniques. This invention employs a reinforcement learning compensation method based on the paraxial approximation method, preserving its computational efficiency. The paraxial approximation method solves the problem through basic arithmetic operations, a single inverse trigonometric function, and simple conditional judgments; the number of operations remains constant regardless of the input parameters, resulting in a time complexity of O(n log n). Spatially, the paraxial approximation method requires only a small number of scalar variables, resulting in a space complexity of O(n log n), lower memory usage, and better numerical stability.

The present application provides a kind of topography probability distribution intelligent identification method based on lightweight convolutional neural network, constructs special terrain identification network by fusing the advantages of MobileNetV3 and ShuffleNetV2, improves model accuracy by combining channel attention mechanism and knowledge distillation technology, realizes the real-time inference of embedded platform using structured pruning and INT8 quantization.At the same time, this method not only outputs the probability distribution of five types of terrain and the confidence evaluation based on information entropy, but also realizes the unsupervised ROI automatic positioning through the improved Grad-CAM++ algorithm, without additional labeling of boundary box data.The present application can realize high-precision, high-robustness terrain identification with very low hardware cost and computing overhead, and provide reliable environment perception basis for exoskeleton proactive adaptive control.

The application discloses a Vision Transformer model quantization method based on attention head importance guidance, which adopts a PTQ method based on weight disturbance and output reconstruction. Firstly, a full-precision pre-training model is prepared, and then the weights of the full-precision output part of each attention head are saved, and the Fisher information of each attention head is calculated. Then, the full-precision output is obtained as the benchmark output by inputting the calibration data set into the full-precision model, and the quantization parameters are initialized. According to the Fisher information score of each attention head, the loss before and after quantization is weighted. The AdaRound strategy is adopted to optimize the rounding strategy by minimizing the quantization loss. The application effectively utilizes the characteristics of the multi-head attention mechanism, preferentially optimizes the important attention head in the quantization process of the Vision Transformer model, and retains more effective information, and further improves the model performance in the case of low bits, so that the model has more possibilities in the deployment of edge devices.

The application relates to the technical field of storage computing, and discloses a system, a method and an electronic device for directly searching data in a storage medium, the system comprising a large-scale vector database, a nearest neighbor search operation unit and a host computer, the large-scale vector database is arranged in the storage medium and is used for storing a vector set; the nearest neighbor search operation unit is arranged in the storage medium; the host computer is provided with a CPU and a cache unit; when the nearest neighbor search operation unit receives an NNS_CSD instruction sent by the CPU, the vector of the database is read, similarity measurement is carried out, and the search result meeting the condition is sent to the cache unit, the application directly searches data in the storage medium, the NNS_CSD instruction is used to start the calculation in the storage medium, the nearest neighbor search operation unit in the storage medium is used for vector reading and similarity measurement, compared with a traditional architecture, the migration of data is greatly reduced, and the overall performance is improved.

The embodiment of the invention provides a subprogram instance creating method and device, equipment, a medium and a program product. The method comprises the following steps: in response to a calling request of a target subprogram in a session between a client and a database, searching a target shared instance from a shared memory; multiplexing the public data in the searched target shared instance, and creating a private instance structure; and generating a private instance stored in the private memory of the session according to the private instance structure and the private data. According to the method, when the session calls the subprogram, the shared instance is searched from the shared memory, and the private instance for the session is created by multiplexing the public data in the shared instance, so that multiple sessions do not need to repeatedly compile and store the same public data, and the memory overhead is reduced, and therefore, in a high-concurrency scene, the session can be quickly and efficiently called. The risk of insufficient memory is avoided, and the stability of database operation is improved.

The invention provides a tree structure data storage and query method and system and a medium, relates to the technical field of data storage, and solves the problems of performance bottleneck and hierarchical information loss of tree structure data storage and query in the prior art. The method comprises the following steps: generating and storing node data including a node identifier, a father node identifier, a hierarchy depth and a path identifier; the hierarchy depth of the node is obtained by adding 1 to the hierarchy depth of a father node of the node, and the path identifier is formed by splicing a path identifier of the father node of the node, a node identifier of the node and a predefined separator in sequence; in cross-hierarchy query, taking a path identifier of a target node as a left fuzzy matching condition, and performing condition screening in combination with hierarchy depth; and in the whole subtree query, performing reconstruction according to a query result. According to the method, retrieval based on the table structure in the database does not need to generate extra memory overhead, so that the query efficiency can be greatly improved, and the performance overhead can be greatly reduced.

The application provides a multi-agent collaborative decision-making and disambiguation method and system. The method comprises the following steps: registering the dependency relationship between the task context and the event trigger condition concerned by the agent to the coordination center, updating the dynamic dependency view through the coordination center, analyzing the knowledge update time through the coordination center, and generating a task context identifier set of the agent; constructing a fact set based on the task context identifier set, inputting the fact set into a rule reasoning engine, executing the rules in the collaborative rule base to generate a strategy, and obtaining a collaborative strategy; and distributing the collaborative strategy to the corresponding agent and executing a cleaning process at the end of the task generation period. The application solves the efficiency bottleneck of event routing in a distributed intelligent system from the system architecture level, greatly improves the real-time response capability and scalability of the system in a high-concurrency scenario while ensuring decision consistency.

The image decoding method includes: determining a context for use in a current block, from among a plurality of contexts (S204); and performing arithmetic decoding on a bit sequence corresponding to the current block, using the determined context (S210), wherein in the determining (S204): the context is determined under a condition that control parameters of neighboring blocks of the current block are used, when the signal type is a first type, the neighboring blocks being a left block and an upper block of the current block (S206); and the context is determined under a condition that the control parameter of the upper block is not used, when the signal type is a second type (S207).

The invention provides a dynamic header rendering method and system based on EasyExcel, and belongs to the technical field of data export and report generation. During export, an export request containing a service type and a request parameter is received, a callback address, a request mode and header field configuration are read from a database according to the service type, and an export report is generated; the field configuration at least comprises a field key, a field name, a data type and sorting information, and can comprise a header hierarchy and a style merging rule; the system constructs a header list required by EasyExcel, obtains service data according to callback configuration and converts the service data into Listt; mapt, Mapt; the method comprises the following steps of: (1) carrying out String, Objectgt Gt; and the form and the header field key are dynamically matched and written. An Excel file is generated in a paging writing mode so as to reduce memory occupation, and finally the file is stored in an object storage and returned to a downloading address. The method supports quick adjustment of the multi-task type header, reduces code modification and repeated model class maintenance, and improves the export stability of large data volume.

This invention relates to the fields of computer data processing and database technology, specifically to an intelligent search system for business objects based on metadata. The system includes: a context request parsing module: receiving abstract search requests and extracting environment parameters and permission features to generate a dynamic context snapshot; a logical structure construction module: combining snapshot execution rule injection operations to generate a logical abstract syntax tree containing implicit filtering conditions; a multi-dialect compilation and execution module: translating the syntax tree into physical execution statements matching the target storage engine dialect according to physical mapping rules; and an adaptive optimization feedback module: monitoring execution latency data and dynamically correcting the index strategy definition. This invention solves the maintenance bottleneck caused by the tight coupling between business logic and physical storage in heterogeneous data source environments, achieving zero-code modification and hot deployment of business logic.

The application discloses an extensible laser mapping method based on prior topological information and belongs to the technical field of multi-sensor fusion of unmanned vehicles. The method comprises the following steps: S1, multi-sensor data acquisition; S2, data preprocessing; S3, radar odometry; S3.1, IMU pre-integration; S3.2, radar point cloud matching; S3.3, pose fusion; S3.4, map construction; S4, map updating and expansion based on topological information; S4.1, dynamic loading module; S4.2, map association; S4.3, map updating evaluation and map expansion and fusion. The method realizes an incremental expansion, globally consistent and resource-efficient laser mapping system by tightly coupling multi-sensor data, embedding prior topological constraints and implementing dynamic resource management, and provides key mapping technical support for long-term autonomous navigation in a dynamic environment when the operation area is expanded.

The invention relates to an unmanned aerial vehicle intelligent path planning method and system based on visual signals, and belongs to the field of path planning. The method comprises the following steps: optimizing a neural network through progressive pruning; fuzzy coordinates and scene types are provided for the unmanned aerial vehicle, the unmanned aerial vehicle flies to a fuzzy coordinate area, and when the unmanned aerial vehicle is close to the fuzzy coordinate area, the unmanned aerial vehicle is switched to a pure vision mode to perform environment perception through a neural network, and a three-dimensional semantic map is constructed; performing real-time path planning based on the three-dimensional semantic map to obtain a flight path; the method comprises the following steps: performing autonomous flight along a flight path, when approaching an operation target, performing autonomous operation to obtain real-time monitoring data, storing the real-time monitoring data in airborne storage equipment, and returning the real-time monitoring data after returning. And an unmanned aerial vehicle intelligent path planning system with high robustness, high autonomy and high task adaptability is constructed.

The application discloses a front-end data compression method, device, equipment, medium and product. The method comprises the following steps: a main thread receives service data, and when a preset compression triggering condition is met, the main thread dynamically converts a pre-embedded script code string into an executable sub-thread, and the script code string comprises compression logic of a lightweight compression engine; the main thread encodes the service data and stores the service data in a binary buffer, and transfers the memory ownership of the binary buffer to the sub-thread through a message communication interface; the sub-thread creates a typed array view based on the binary buffer, calls the lightweight compression engine to compress the typed array view, generates a compression result, and transfers the memory ownership of the compression result to the main thread; and the main thread reports the compression result or locally stores the compression result. The method can realize efficient data compression in the front end, ensure smooth running of a front-end page, and reduce memory occupation and transmission delay.

The invention discloses a processing method and system for OFD document conversion, and the method comprises the steps: triggering an OFD conversion operation when a PDF document is exported, so as to extract document structure information and document resource data; performing format conversion on the document structure information and the document resource data, and then adding the document structure information and the document resource data to an OFD resource library; performing text recognition on the text resource data to obtain text segment content information; performing position identification on the text segment content information and storing the text segment content information in the OFD resource library; according to the OFD document conversion method, the image resource data are subjected to image recognition to obtain image feature information, so that the image resource data are identified in the OFD resource library, the problems of format and style loss, incomplete resource processing, low conversion efficiency, insufficient integration level, poor cross-platform compatibility and the like in the existing OFD document conversion technology are solved, and low-cost, high-safety and multifunctional OFD document conversion is realized.

The invention belongs to the technical field of infrared imaging system testing, and discloses an MRTD intelligent testing method for an infrared imaging system, and the method comprises the steps: collecting the original four-rod target image data of thermal infrared imagers of the same model at different temperatures under the spatial frequency under the condition that the spatial frequency is not changed; cutting the collected original image data, and only obtaining an important area of a four-bar target image; carrying out expansion processing on the image; annotating the images, and dividing the images into a training set, a test set and a verification set; an MRTD intelligent detection model is constructed; training and accuracy testing are carried out on the MRTD intelligent detection model, label value prediction is carried out on the image data of the verification set, and the temperature difference value corresponding to the predicted label value N is the MRTD value of the thermal infrared imager under the spatial frequency. According to the method, the real-time performance of detection is ensured, and the interference of repetitive work and subjective factors in a traditional test method is reduced.

This invention belongs to the field of terrain elevation modeling and simulation technology, specifically disclosing an ELM-based terrain elevation data service method, including: creating a dataset containing the latitude, longitude, and elevation values ​​corresponding to all rasters; designing an ELM structure; training the ELM structure using the dataset, and using the ELM structure, the parameters of each layer, the correspondence between elevation values ​​and latitude and longitude, and normalized parameters as the ELM fitting model and saving it; obtaining the longitude and latitude range of the currently needed terrain, setting the parameters of the ELM fitting model within the current longitude and latitude range to obtain an instantiated ELM fitting model; receiving elevation data requests from entity nodes within the current longitude and latitude range through the RP interface, calculating and sending the required elevation values ​​to the entity nodes; having the following advantages: solving the problem of traditional terrain data storage requiring the establishment of a terrain database; and solving the problem of large amounts of terrain elevation data occupying memory during simulation runtime.

The application provides a processing method, a terminal device and a storage medium. The processing method comprises the following steps: in response to a preset condition being met, loading target data in a multilingual language model that matches a target language and / or a target country. In the technical solution of the application, when the multilingual language model is used, only the target data in the language model that matches the target language and / or the target country is loaded, so that the memory occupation can be reduced.

This invention discloses a high-performance computing-based system and method for evaluating colorectal cancer screening strategies, belonging to the field of medical information processing technology. It includes: a dual-pathway disease model module supporting both adenoma-carcinoma pathways and serrated adenoma pathways, modeled using age-dependent normal distribution progression time; an individualized risk scoring module integrating risk factors such as family history, inflammatory bowel disease, obesity, and diabetes; a screening strategy simulation module supporting combined screening strategies using different tools, employing compliance parameters specific to the Chinese urban population; a high-performance computing module using Numba JIT compilation technology for acceleration, multi-process parallel processing for further acceleration, and data structure optimization such as integer age to reduce memory usage; and a health economics evaluation module calculating indicators such as QALY, LYG, and ICER. This invention achieves significant performance improvements while maintaining model accuracy, greatly reducing simulation time for large-scale populations.

The invention relates to the technical field of artificial intelligence, can be applied to the fields of financial science and technology and medical science and technology, and discloses a lexical element cutting method, device, equipment and medium for a multi-modal visual language model.The method comprises the steps that a question image and a text are received and converted into a visual lexical element set and a target lexical element set; on the basis of the visual lexical element set and the target lexical element set, a contribution matrix is calculated with the goal of minimizing the optimal transmission cost from the visual lexical elements to the target lexical elements by fusing a visual perception cost function of feature similarity, a spatial relationship and a center distance, so that the contribution degree of each visual lexical element to the generation of the target lexical element is globally quantified; and according to the contribution value sorting and in combination with a preset threshold value, performing one-time cutting on the visual lexical element set, reserving the key Token, and selectively recovering part of the cut Token through uniform sampling so as to adapt to a fine task requirement. According to the method, accurate and efficient Token screening is realized, the calculation and memory overhead is remarkably reduced, an additional calculation step does not need to be introduced, and the reasoning efficiency is improved.

This invention provides a Chinese text-to-speech method based on an offline unit library constructed using a multimodal large model, relating to the field of text-to-speech technology. The method includes the following steps: receiving the target text to be synthesized; parsing the target text for scene tags, sentiment tags, and stress positions using a lightweight semantic parsing model; querying a pre-constructed offline unit library based on the parsing results, and retrieving corresponding candidate acoustic prosodic units for each Chinese character or word; constructing a state network with the candidate acoustic prosodic units as nodes; calculating the optimal path based on the inter-unit transition cost and semantic matching degree in the offline unit library using the Viterbi algorithm, and selecting the globally optimal acoustic prosodic unit sequence; and smoothly concatenating the candidate acoustic prosodic units in the optimal acoustic prosodic unit sequence to generate and output the target speech.

This invention discloses a method, system, equipment, and medium for the synchronous distribution and analysis of aerospace simulation training scenarios. The method includes: the director notifying both sides of the adversarial training to acquire scenario information; both sides processing the acquired scenario information to generate scenario model data and satellite model data respectively; both sides loading the scenario model data and satellite model data and conducting adversarial exercises; and the maneuver information of one side being sent to the other side after a delay by the director. This invention employs a unified distribution of basic simulation scenario information and satellite, ground station, orbit, and sensor data by the director; both sides acquiring data and constructing dynamic model mapping relationships through modular interfaces, and synchronously initializing using an orbit predictor; real-time acquisition of one's own satellite maneuver parameters and delayed maneuver information during the adversarial process; and improved information interaction efficiency and security through a lightweight data transmission mechanism.

The invention relates to the technical field of mobile terminal artificial intelligence computing, in particular to a mobile terminal lightweight reasoning optimization method and system based on PyTorch Android Lite, and the method comprises the following steps: model preparation, model conversion and optimization, environment configuration during operation, and model deployment and execution. The method has the beneficial effects that through training-reasoning separation and use of the Lite Interpreter, the memory occupation during operation is reduced from more than 500MB to less than 50MB, so that a complex model can stably operate on middle and low-end mobile phones. Vulkan is used as an abstraction layer, calling of GPUs / NPUs of different manufacturers is effectively unified, the problem that PyTorch Mobile has insufficient support for novel hardware is solved, and better performance transportability is achieved.

A method, apparatus, and device for liveness detection based on large model screening and distillation are disclosed, relating to the field of data processing. In this method, liveness image data is input into a preset teacher processing model to generate a first liveness confidence score and a teacher intermediate layer feature map, and a first spatial saliency map is calculated. Liveness image data is then input into a preset student processing model to generate a second liveness confidence score and a student intermediate layer feature map. The feature differences between the teacher and student intermediate layer feature maps are calculated, and these differences are weighted spatially to generate a weighted feature difference. The spatially weighted feature imitation loss and confidence imitation loss are calculated. The internal parameters of the preset student processing model are adjusted and optimized to obtain a target student processing model, and liveness images to be detected are then detected based on this target student processing model. Implementing the technical solution provided in this application improves the accuracy of liveness detection.