227 results about "Model transformation" patented technology

The invention relates to the technical field of artificial intelligence chips, and discloses a model performance estimation method and device and computer equipment, and the method comprises the steps: determining model configuration data and candidate distributed strategies of a target model; converting the original calculation graph corresponding to the single-card deployment state based on the model configuration data and the strategy configuration data of the candidate distributed strategies to obtain a distributed overhead calculation graph corresponding to the multi-card deployment state; and performing performance estimation on the candidate distributed strategy according to the basic overhead and the additional overhead in the distributed overhead calculation graph to obtain strategy performance data of the target model under the candidate distributed strategy, thereby realizing conversion of a single-card model into a multi-card model. And based on the multi-card model, simulation calculation of the multi-card interconnection mode is realized on the premise of limited hardware resources, so that the influence of a communication operator and a topological structure corresponding to the multi-card interconnection mode in the overall operation of the model is reflected, and the upper limit of the model performance can be accurately evaluated in a simulator verification stage before silicon is applied.

The invention discloses a rotor reliability constrained rolling bearing assembly parameter robust design method. The method comprises the following steps: constructing a dynamic model for an actual rotor-bearing system; constructing an uncertainty parameter vector and a design variable vector; a target function based on robustness and a constraint function based on reliability are constructed, so that an uncertainty optimization model is obtained; constructing an augmented input variable, and establishing a candidate orthogonal polynomial basis function set; on the basis, constructing and evaluating polynomial chaos-Kriging models for the target function and the constraint function respectively, and screening out an optimal polynomial chaos-Kriging model; calculating the expectation and the standard deviation of the target function and the failure probability of the constraint function under each design variable vector; and converting the uncertainty optimization model into an unconstrained single-target optimization model, randomly generating population individuals of a heuristic optimization algorithm in a feasible region of design variables, and iteratively searching an optimal solution of the unconstrained single-target optimization model as a rolling bearing assembly scheme.

The invention relates to the technical field of deep learning, in particular to a PyTorch framework-based compatibility expansion system, which comprises a cross-framework model converter, a heterogeneous hardware abstraction layer, a hybrid computational graph execution engine, an intelligent distributed trainer and a self-adaptive optimizer, the cross-frame model converter converts input model formats of different frames into PyTorch executable formats, the heterogeneous hardware abstraction layer supports rear ends of various hardware and automatically selects an optimal calculation path, and the hybrid calculation graph execution engine fuses dynamic graph flexibility and static subgraph optimization and supports dynamic control function execution flow. An intelligent distributed trainer automatically selects a parallel scheme and is compatible with multi-protocol communication, a self-adaptive optimizer performs dynamic optimization based on hardware characteristics and a PyTorch tensor, a cross-frame model converter supports model conversion of multiple deep learning frames, the cost of migration among different frames by a user is reduced, and the universality of the PyTorch frame is improved.

A processor-implemented method for converting a trained language model into a language model in an architecture of mixture of experts (MoE), and a computing device using the same is provided. The method for converting a trained language model into a language model in an architecture of mixture of experts using a computing device according to an embodiment of the disclosure may include dividing a plurality of layers included in a target language model and extracting a feed-forward network (FFN) included in each of the plurality of layers, generating an MoE block of the MoE language model, which corresponds to the feed-forward network, generating an input tensor, comparing output tensors between the feed-forward network and the MoE block for the input tensor to obtain a first loss, and updating a weight of the MoE block, based on the first loss.

The invention discloses a vehicle collision accident severity prediction method and device, equipment and a storage medium, and relates to the technical field of artificial intelligence, and the method comprises the steps: collecting the multi-modal data of a vehicle collision accident, and carrying out the preprocessing of the multi-modal data; converting the pre-processed structured numerical data into a first feature map through a Grubrum angle field method; converting the preprocessed unstructured text data into a text semantic vector through a pre-training language model, and converting the text semantic vector into a second feature map; performing size alignment on the first feature map and the second feature map, and performing splicing on a channel dimension to obtain a multi-channel fusion image; and inputting the multi-channel fusion image into a first deep learning model, and outputting an accident severity prediction result. According to the method, the recognition sensitivity and the prediction recall rate of serious injury accidents are effectively improved, meanwhile, complex artificial feature engineering is avoided, and the generalization ability and the interpretability of the model are enhanced.

Described are a system, method, and computer program product for distributed execution of a machine-learning model on a server cluster. The method includes initiating retrieval of a machine-learning model from a data repository and converting the machine-learning model to an executable format. The method includes transmitting the converted machine-learning model to each node of the server cluster and executing the converted machine-learning model on each node. The method includes generating an initial performance metric based on execution of the converted machine-learning model on each node. The method includes transmitting the plurality of initial performance metrics from each node to an external processor and combining the plurality of initial performance metrics to produce a combined performance metric. The method includes modifying a model hyperparameter of the machine-learning model based on the combined performance metric and executing the modified machine-learning model in a computer system to evaluate real-time event data.

The invention discloses an end-to-end zero code AI model automatic training and deployment system and method, and belongs to the technical field of artificial intelligence and machine learning engineering, and the system comprises a user interaction layer which is used for providing a graphical user interface and receiving a model training instruction configured by a user; the business logic layer responds to the instruction, encapsulates the instruction into a structured work order, automatically executes data management, model training and model conversion tasks based on the work order, and generates a model deployment package capable of directly running on at least one target hardware platform; and the resource management layer is used for monitoring and scheduling bottom computing resources to support task execution of the business logic layer. According to the method, a Web-based graphical interface is provided, all operations are completed through clicking, dragging and pull-down selection, dependence on programming and command line operations is thoroughly eliminated, and complete AI workflow zero-code operation from data to deployment in a single system is achieved.

The invention belongs to the technical field of electric power system dispatching optimization and electric power and electric quantity balance analysis, and relates to a time sequence random production simulation method considering a flexibility demand in an electric power market environment, which comprises the following steps: 1, constructing an electric power system time sequence-random production simulation framework considering the flexibility demand in the market environment; 2, the upper layer model obtains a start-stop scheduling scheme by taking the minimization of the total operation cost of the system as a target, calculates the time sequence flexibility supply capability of each type of unit, and provides an input basis for the lower layer; 3, constructing a multi-state flexibility supply model of the unit under different time scales in a lower layer model, carrying out multi-time scale system flexibility supply and demand probabilistic matching by adopting an equivalent electric quantity function method, and realizing flexibility index evaluation under a short time scale; 4, converting and solving the model; according to the invention, support is provided for coordinating the operation economy and flexibility of the system under security constraints.

The invention discloses a model discretization method and device, electronic equipment and a computer readable storage medium. Comprising the steps that a to-be-discretized CAD model is obtained, the to-be-discretized CAD model is analyzed, geometric features of the to-be-discretized CAD model are obtained, and the geometric features at least comprise faces and edges; for any adjacent edge A and edge B, discretizing the edge A and the edge B according to the relationship between the edge A and the edge B, and discretizing the edge A and the edge B respectively to obtain a discrete point column of the edge A and a discrete point column of the edge B; and for any adjacent surface C and surface D, discretizing the surface C and the surface D according to the relationship between the surface C and the surface D, and discretizing the surface C and the surface D respectively to obtain a discrete grid of the surface C and a discrete grid of the surface D. According to the method, when the model is converted into the net-shaped structure, it is guaranteed that selfing of the grids does not occur.

The invention provides an AI model cross-platform deployment system and method based on a unified interface and an intermediate presentation layer, and is used for solving the technical problem that the deployment adaptation of an AI model among different cloud platforms is difficult. The system adopts a hierarchical architecture design and comprises a unified interface layer, a middle presentation layer and a platform adaptation layer. Wherein the unified interface layer provides a standardized model deployment interface specification; the middle presentation layer converts the AI model into a platform-independent middle format and manages model dependence and configuration information; and the platform adaptation layer is responsible for converting the intermediate format into a deployment format of a target platform and calling a corresponding platform API to complete deployment. Through the innovative design of the unified interface and the middle presentation layer, the development cost of cross-platform deployment is remarkably reduced, the portability of the model is improved, and the technical problem in the cross-platform deployment process of the model is effectively solved.

According to the model conversion method provided by the invention, the height map of the terrain and the weight distribution maps of the plurality of material layers are acquired, the target three-dimensional grid model is generated according to the height map, the weight distribution maps of the plurality of material layers are combined to generate the weight map, and based on the position of the target three-dimensional grid model in the world coordinate system, the target three-dimensional grid model is converted into the target three-dimensional grid model. Calculating texture mapping coordinates; and generating a material instance corresponding to the target three-dimensional grid model according to a preset material template, the weight map and the texture mapping coordinate. Through effective integration and conversion processing of terrain height information and multilayer material weight information, automatic conversion from two-dimensional terrain data to a three-dimensional grid model with complete material information is realized, and it is ensured that the finally generated three-dimensional model can accurately express a complex material mixing effect. And the visual authenticity and rendering efficiency of the terrain model are obviously improved.

The invention relates to the technical field of spiking neural networks, in particular to a spiking neural network conversion method and system based on differentiable adaptive optimization, and the method comprises the steps: constructing and pre-training a target ANN model; in a pre-trained ANN model, performing simulation processing on the activation value of each layer by using a global differential value representation operator; on the basis of a loss function, an original weight parameter in the target ANN model and a trainable excitation threshold in a global differentiable value representation operator are placed in the same optimization framework for joint fine tuning, and an optimal weight and an optimal threshold learned by each layer are obtained; and an equivalent SNN model is constructed. According to the method, precise correspondence between continuous activation and pulse sequences can be optimized, the method is suitable for multiple pulse coding modes, quantization errors are reduced to the maximum extent, the deployment requirements of neuromorphic hardware and an edge intelligent computing platform can be deeply adapted, and an efficient and universal model conversion solution is provided for the neuromorphic computing industry.

The invention discloses a building structure explosion multistage early warning and damage assessment integrated analysis method, which relates to the field of building structure safety, and comprises a BIM-FEM model conversion module, an FEM explosion response and damage assessment module, an Internet of Things gas sensor acquisition module, an early warning analysis and grading early warning module and a data storage module. And an efficient automatic model conversion interface between the BIM and the FEM is developed. The interface realizes bidirectional data conversion between the BIM model and the finite element analysis model; a gas explosion load is applied to finite element analysis software, dynamic response and damage of each component under the gas explosion load are studied through numerical simulation, and anti-explosion damage evaluation attributes of each component are formed. The problem of low modeling efficiency of a complex structure in finite element analysis software is effectively solved, and automatic data extraction, model mapping and result write-back between the BIM model and the FEM model are realized.

The invention discloses a predefined space-time pitch angle control method of a variable-speed wind generating set, and belongs to the technical field of variable-speed wind generating set control. The method is characterized by comprising the following steps of 1, obtaining a dynamic model of the variable-speed wind turbine according to the Betz theory; step 2, obtaining a predefined space-time reaching law with a buffer area according to the improved barrier function; 3, constructing a sliding mode variable, and converting a non-affine model of the variable speed wind turbine into an affine model based on an unknown smooth nonlinear function; and 4, compensating the unknown smooth nonlinear function by using a neural network, and obtaining a pitch angle controller according to the neural network. According to the predefined space-time pitch angle control method of the variable-speed wind generating set, the state space is divided into the multiple parts including the buffer areas, dependence on model parameters is reduced, the buffeting phenomenon is weakened through the improved obstacle function, and then the power generation efficiency is improved, and the operation cost is reduced.

The invention relates to a segmented fusion SPICE to IBIS model conversion method and system, and the method comprises the steps: obtaining data, precisely extracting voltage / current curve data from a simulation result of SPICE simulation, carrying out the transient analysis, guaranteeing that a finally generated IBIS model can accurately reflect the electrical behavior of the SPICE model, increasing the voltage / current sampling point density of DC scanning in the SPICE simulation, and obtaining the electrical behavior of the SPICE model. Particularly, in a non-linear severe change region, a segmentation interval is adaptively adjusted according to curvature change, so that the accuracy of curve data extraction is ensured; a compensation item is added in a [Model] section of the IBIS model to compensate the packaging parasitic effect, the accuracy of model conversion is improved, meanwhile, simulation data under the multi-PVT condition is compressed through PCA, redundant information is reduced, the model generation efficiency is improved, and high-precision and high-efficiency conversion from the SPICE model to the IBIS model is achieved.

Embodiments described herein systems and methods for dynamically guiding a generative artificial intelligence model through structured, goal-oriented interactions using machine- generated rubrics for instructing operations of the generative Al model. The system selects, for each conversational turn, a first rubric defining response structure and a second rubric defining evaluation criteria. A generative model formulates prompts based on these rubrics and receives participant responses. An assessment module evaluates the responses against the second rubric to update the interaction context. A rubric orchestration module uses the updated context to select new rubrics for subsequent turns. This iterative process transforms a stateless generative model into a persistent, adaptive agent capable of conducting multi-turn interactions aligned with defined objectives.

A method and a system for performing domain adaptations of social media text by using lexical data transformations are provided. The method includes: receiving a first data set that is usable for training a machine learning (ML) model that is designed to perform natural language processing tasks; training the ML model by using the first data set; receiving a second data set that relates to a social media platform; transforming a subset of the first data set into a third data set that is suitable for the social media platform; and retraining the ML model by using a combination of the first data set, the second data set, and the third data set. The transformations may include injecting emojis, emoticons, user mention indicators, hashtags, retransmission indicators, URLs, and / or inverse lexical normalizations that are often used in social media posts.

The invention discloses an intelligent self-adaptive time sequence diagram analysis and structuring method. The method comprises the following steps: preprocessing an input time sequence diagram; identifying and extracting basic structure elements in the time sequence diagram; extracting text information in the time sequence diagram by using an optical character recognition technology; the recognized signal lines are analyzed, and signal state changes and key event points are extracted through a neural network; detecting arrows representing time sequence constraints, and performing arrow detection and classification by using a target detection network; establishing an association relationship between the graphic elements and the text; the association relationship and the identified elements are verified and corrected by using grammar rules and global layout information of the time sequence diagram, and unreasonable results are filtered; converting the time sequence relation model into a standardized structured data format; and the overall recognition performance is continuously optimized based on interactive correction and feedback of confidence. According to the method, the workload of manually converting the time sequence diagrams by engineers is greatly reduced, the accuracy of time sequence constraint extraction is improved, and various time sequence diagram formats and styles are supported.

In some implementations, a method for custom fitting and manufacturing parametric products comprises: generating a parametric model of a custom product based on past learned data pertaining to the custom product and user-specific data of a user; translating the parametric model of the custom product to a plurality of parameters of the custom product; based on the plurality of parameters, generating a first graphical representation of the custom product; generating a first user interface and displaying the first user interface; receiving a plurality of adjustment parameters for improving a fit of the custom product in relation to the user-specific data; based on the plurality of adjustment parameters and the plurality of parameters of the custom product, generating, for the custom product, a parametric fit model having a plurality of fit model parameters; transmitting the parametric fit model to a manufacturer to cause the manufacturer to produce a parametric physical product.

The invention relates to an agricultural water right distribution method based on dual-stage stochastic programming and a neural network. The method comprises the following steps: collecting historical data to construct a data set; the irrigation water demand and the available water supply are predicted by inputting the data set into the long-short-term memory neural network; establishing a two-stage stochastic programming model, introducing a decision variable to convert the model into a deterministic sub-model, and solving an optimal water distribution target and configuration water quantity through an interactive algorithm; and according to the optimal water distribution target and the configured water quantity, generating water right intelligent distribution schemes under different water inflow scenes, and dynamically adjusting a water right distribution threshold value between the regions. Through deep fusion of double-stage stochastic programming and the neural network, a data-driven and model-driven water right intelligent distribution framework is constructed, a long-short-term memory neural network is introduced to predict a future irrigation water demand and an available water supply interval value, the adaptability of the model to dynamic climate is improved, and the water right distribution efficiency is improved. And the optimal distribution threshold values under different inflow water levels are quantified, so that the water demand of a high marginal benefit area is ensured.

The invention relates to a large-model heterogeneous reasoning engine method, device and equipment and a storage medium. The method comprises the following steps: acquiring an end side model generated by performing model conversion on a neural network model by a host end; constructing a calculation graph at the equipment end according to the end side model; a model reasoning task request is received, a task type is obtained, the computational graph is analyzed, differentiated task distribution is conducted on all nodes of the analyzed computational graph based on the task type, and all the nodes are bound with heterogeneous computing units executing different tasks; and according to a task allocation result, executing a model reasoning task through the heterogeneous computing unit bound with the allocated node. According to the method, the nodes are bound with the heterogeneous computing units executing different tasks, differentiated task distribution is performed on the nodes in the computing graph according to the task types, the computing power utilization rate of the heterogeneous computing units can be maximized when the end-side large model reasoning task is executed, the high throughput and low time delay requirements during reasoning task execution are ensured, and the reasoning efficiency of the end-side large model reasoning task is improved. The execution efficiency is improved.

In variants, the method can include generating mapping model training data, determining the mapping model, and predicting a program based on a transformer. The method can optionally include evaluating the mapping model, running analyses on the program, and / or utilizing the program and / or generated program analyses. The method functions to convert transformer models into programs that can be characterized and / or analyzed using program analysis techniques.

The invention relates to a method for reducing the network dimension of a base model, comprising the following steps:-providing (S1) a base model having a pre-trained weight matrix, which is trained for solving a target task; -transforming (S2) the base model into a one-time model with a matrix of weights; adding (S3) at least one, in particular network dimension-specific, low-rank matrix to each weight matrix of the one-time model; -performing (S4) a neural architecture search (NAS) on the one-time model to extract at least one sub-model of the one-time model having a reduced network dimension based on the low-rank matrix and a weight matrix of the one-time model until a termination criterion is reached; and-providing (S5) at least one sub-model having a reduced network dimension, in particular for implementation on an embedded system.

Embodiments of the present disclosure provide a model conversion method, related device and medium. The method deploys a model conversion image to a node where a target chip is located, converts an original model into a target model corresponding to the target chip through the model conversion image. Then, the original model is checked through an original model checking image and the target model is checked through a target model checking image, a conversion deviation value capable of judging the model conversion effect is obtained according to a first checking result and a second checking result, and automatic checking of the converted target model is realized. Based on the checking result (conversion deviation value), a target model meeting the requirements is constructed into an image and stored in an image warehouse for subsequent calling. Embodiments of the present disclosure aim to ensure the usability of the converted model. Embodiments of the present disclosure can be applied to computer vision, natural language processing, speech recognition, autonomous driving, edge computing and the like.

Aspects of the subject disclosure may include, for example, an intent-based network (IBN) management system that effects changes in a communication network based on high-level intents. High-level intents are translated into operator-level intents by a large language model (LLM). Conflicts between operator-level intents are resolved, and the operator-level intents are mapped to intent functions. The intent functions are then mapped to policies that may effect changes in the network in accordance with the high-level intents. Other embodiments are disclosed.

The invention discloses a virtual sensor modeling and deployment method and system based on spatio-temporal feature fusion and a gradient boosting strategy, and relates to the technical field of industrial process monitoring and control, embedded intelligent sensing and data-driven modeling. The method comprises the following steps: collecting multi-source sensor data under a multi-environment condition and preprocessing the multi-source sensor data; constructing a spatial-temporal characteristic system which simultaneously represents historical memory, dynamic change and a multivariable coupling relationship, and performing characteristic screening and weight reduction; under a gradient lifting framework, adopting automatic hyper-parameter optimization to obtain a virtual sensor model with compromise between performance and complexity; the trained model is converted into a unified reasoning format irrelevant to a platform, and real-time reasoning is achieved on a resource-limited embedded control unit. Compared with the prior art, the method has higher prediction precision and generalization ability under the complex dynamic working condition, the model size and the single reasoning delay are controllable, and the method is suitable for being deployed and applied in a vehicle-mounted ECU, an industrial controller and an edge node.

The invention discloses a video benchmark model reasoning framework deployed at an embedded terminal. The video benchmark model reasoning framework comprises a model lightweight module, a cross-format conversion module, a data preprocessing module, a reasoning acceleration module and a result verification module. The model lightweight module is used for performing pruning and quantification processing on the original visual language reference model, and reducing the model parameter scale and the calculation complexity; the cross-format conversion module is used for sequentially converting the lightweight PyTorch model (. Pth) into an ONNX format, and then converting the ONNX format into an OM format suitable for a domestic AI processor; the data preprocessing module is used for carrying out standardization processing on input visual data (images) and text data, wherein the standardization processing comprises size adjustment, channel conversion and normalization; the reasoning acceleration module is used for optimizing a model reasoning path and supporting multi-thread parallel computing; and the result verification module is used for comparing the reasoning results of the converted model and the original model to ensure that the precision loss is within a preset threshold value.

This invention provides an automatic modeling method and system for low-voltage power grid data based on visual data acquisition, relating to the field of low-voltage power grid technology. The method includes: dividing the low-voltage power grid area into multiple sub-regions and constructing multiple local coordinate systems; analyzing the power grid equipment to be acquired and determining at least one matching local coordinate system; configuring a visual data acquisition task template, performing data acquisition to obtain at least one set of equipment modeling data; constructing a multi-region coordinate system transformation model, performing coordinate system transformation to obtain at least one set of equipment modeling transformation data corresponding to the equipment modeling data, and generating visual modeling simulation results for the data acquisition task. This invention solves the technical problem in existing technologies where, in the process of low-voltage power grid data modeling, a large, unified global modeling coordinate system is typically established for data acquisition or modeling of all power grid equipment. While this achieves a unified effect, it is not accurate enough, affecting the true reflection of power grid data and its application effectiveness.

The application relates to an unmanned vehicle-mounted camera target detection method based on an improved FCOS algorithm, and relates to the field of computer vision. An image is collected by an unmanned vehicle, the image is pretreated, and then is put into an improved FCOS network model for training. In the training process, the model performs feature extraction, prediction, loss calculation and parameter updating on the image. After multiple iterations, a trained detection model file can be obtained. After model conversion, the model can be applied and deployed on terminal equipment such as an unmanned vehicle. The application has stronger feature extraction capability, and the two-stage model constructed has better detection effect on small targets, effectively improves the recognition accuracy of the model, and improves the missed detection and false detection of the model.

The invention discloses a vehicle dynamics modeling method based on a logic tree, and relates to the technical field of automobile electronic control, a model to be inspected and tested is selected from an ECU end and a software development platform, the position / working path of the model is determined, and the relative position of the position of the model in the whole model is determined; the model is converted into a logic tree, after a tree structure is generated, execution logic with missing coverage is sorted based on the tree structure, an automatic test framework based on the tree structure is generated, and errors are rapidly positioned and maintained; finally, the model is copied to the vehicle-mounted ECU or the software development platform again, and updating of the test and inspection model is completed. According to the method, the logic and the content of the model are simply described through a graphical tree structure, the node can be directly described according to the model, the execution logic of the model can also be viewed from a code tree, the system is modularized, and a clear interface is displayed, so that the complexity is reduced and the maintainability is improved.