340 results about "Training time" patented technology

The invention discloses a cloud data server mixed training method and system, which are applied to a cloud data server cluster comprising a plurality of computing resources. According to the method, a model structure, a data set, a time delay constraint and an isolation constraint of a training task are analyzed to generate a task portrait, resource configuration, an operation state and an isolation capability of a calculation node are collected to generate a resource portrait, and a co-located interference degree index table of a task and resource combination is constructed based on historical monitoring data. During scheduling, a heterogeneous computing power utilization rate, an estimated training time delay deviation and a co-located interference degree are taken as indexes, a candidate resource allocation scheme is subjected to weighted evaluation to generate a mixed training scheduling strategy, and resource isolation is implemented through container and accelerator multi-instance division. In the operation process, the training time delay and the actual co-located interference degree are continuously monitored, the scheduling weight and the co-located interference degree index are dynamically adjusted according to the deviation, closed-loop optimization of multi-task mixed training is achieved, the heterogeneous resource utilization rate is increased, and time delay default and co-located interference are reduced.

The embodiment of the invention provides a model training method, an information recommendation method, equipment, a storage medium and a program product. According to the embodiment of the invention, a multi-encoder-multi-sub-decoder-total decoder hybrid model architecture is provided, sample data of different information modes correspond to different encoders-sub-decoders, and a mode of processing all training sample data by a single model is converted into a divide-and-conquer mode. The internal complexity of each encoder-sub-decoder is relatively low, the complexity of model training can be reduced, resource consumption can be saved, and different encoder-sub-decoders can be trained in parallel, so that the model training time can be shortened; and furthermore, by utilizing a dual decoding mechanism of the sub-decoder and the global decoder, parameters of the encoder can be continuously adjusted through local optimization and global optimization, the performance of the model is optimized, the accuracy of a reasoning result is improved, the convergence speed of the model is accelerated, the model training efficiency is further improved, and the model training time is saved.

The invention discloses a photovoltaic power prediction method based on improved empirical mode decomposition and an optimized long short-term memory network, and the method comprises the steps: firstly carrying out the preprocessing of abnormal value elimination, missing value filling, normalization and the like of photovoltaic power and related meteorological data, and improving the data quality; then, an improved empirical mode decomposition (EE-ANEMD) algorithm is adopted to decompose the preprocessed power sequence into a multi-scale intrinsic mode function component and a residual term, and high-frequency noise, intermediate-frequency fluctuation and a low-frequency trend are effectively separated; global optimization is carried out on the hidden layer unit number, the initial learning rate and the maximum number of training times of the LSTM network through an improved sparrow search algorithm (ISSA), finally, the optimized LSTM is utilized to carry out training prediction on each component, and results are fused and subjected to reverse normalization to obtain a final value. Experiments show that the test set RMSE of the method is reduced compared with that of a single LSTM, the mid-term prediction precision is remarkably improved, and reliable technical support is provided for power system dispatching, new energy consumption planning and photovoltaic power station operation and maintenance.

The invention discloses a load identification method and system based on multi-view learning and a deep wide residual network. The method comprises the following steps: acquiring voltage and current data at an outlet of a current collector; extracting two types of load feature views from the collected data, and adding data labels to form a labeled data set; generating two training subsets from the labeled data set, wherein the initial training subset comprises a part of labeled samples and unlabeled samples; performing cross training on the two depth wide residual classification networks by using the training subset, and training to obtain a depth wide residual classification network model based on multi-view learning; and extracting two feature views from unknown electric appliance data collected in real time, inputting the two feature views as label-free data into the depth wide residual classification network model based on multi-view learning, updating the model, and obtaining a load identification result. The method has the advantages of improving the model recognition accuracy and generalization performance, enhancing the robustness of the model, improving the model recognition accuracy and saving the model training time.

The invention relates to a federal multi-language machine translation method based on efficient fine tuning, and belongs to the technical field of natural language processing. Aiming at the problems of high communication cost and long training time in a federated learning-based multi-language machine translation method, the invention provides a federated multi-language machine translation method based on efficient fine tuning, which comprises the following steps of: efficiently fine-tuning a multi-language translation model of a client; performing gradient similarity clustering on the fine-tuned multi-language translation model; carrying out average aggregation based on the clustered multi-language translation model; and deploying a federal multi-language machine translation device based on efficient fine tuning. According to the method, the calculation and communication overhead is greatly reduced while the translation performance is kept, and the method is suitable for distributed translation tasks in a multi-language scene.

How to reduce the error of visible light positioning is the technical problem focused in the current positioning technology, and reduction of the positioning error by adopting the neural network technology is an important scheme for solving the problem. The invention provides a deep learning fusion system for visible light positioning and a method thereof, and aims to solve the problems of unstable model training effect, insufficient positioning precision and poor generalization ability caused by insufficient data acquisition amount in the prior art, or the problems of long training time caused by overlarge data scale and the like. According to the method, the corresponding small model neural network or large model neural network is selected for training and prediction according to the scale of the collected data volume in different operation stages of the system, and a dynamic judgment and periodical comparison mechanism is combined, so that quick and stable positioning output is realized when the data is limited; when the data is sufficient, a high-precision and high-robustness positioning effect is realized, so that the adaptability and reliability of the system in practical application are improved.

An adaptive model partitioning method and system applied to distributed training, which method and system belong to the technical field of deep learning, and aim at solving the technical problem of how to implement, in respect of distributed training, distributed model training by means of combining deep reinforcement learning and a Bayesian optimization algorithm. The method comprises the following steps: constructing a Q network on the basis of a deep neural network, and defining state information, actions and a reward function, wherein the state information comprises feature vectors of partitioned models, and training times, resource utilization rates and inter-node communication overheads of when the partitioned models are subjected to distributed training by means of distributed computing nodes, each action is a partitioning strategy used by an agent under the current state information, the reward function R is used for evaluating the effect of the current partitioning strategy, and the Q network uses the state information as input to predict and output a Q value of each action that the agent may take; and performing multiple iterative training on a deep reinforcement adaptive model, so as to obtain a final partitioning strategy and parameters of the Q network.

The invention discloses an enhanced 3D (three-dimensional) surface reconstruction method based on 3D (three-dimensional) Gaussian Splitting. Global consistent depth priori is obtained through virtual stereo pair rendering; constructing a factor graph and introducing a cross-view geometry / luminosity consistency constraint to form local beam adjustment loss; the prior is used as a learnable parameter to be combined with 3DGS to be optimized, and meanwhile, the Pull loss is assisted to pull low-credibility pixels; and finally, multi-loss function end-to-end training is adopted. The method comprises the following steps of: in Tanksamp; the method has the advantages that F1 is equal to 0.58 and DTU Chamfer is equal to 0.48 mm on a Temples data set, the training time is only 20 min, compared with the prior art, geometric accuracy SOTA and speed magnitude improvement are achieved at the same time, and the method is suitable for VR / AR, robot and industrial measurement scenes.

The invention discloses an AI model parameter initialization method based on model parameter and structure multi-modal fusion, and belongs to the technical field of artificial intelligence. The method comprises the following steps: firstly, collecting historical pre-training model data of a cross-model architecture and a cross-data set, processing model parameters into a token sequence, training a Transform codec, converting a model structure into a graph, and training GAT to extract structural features; a multi-modal feature data set is constructed after a related network is frozen, structural features serve as core conditions, a conditional diffusion model DDPM is trained through AdaLN modulation and residual module coupling, and multi-modal feature fusion of'parameter feature-structural feature 'is established. In the reasoning stage, the structural features of the unknown model are extracted, the random parameters of the unknown model are combined with the submerged space shape determined by the encoder, sampling is conducted through the conditional diffusion model, anti-token processing is conducted through the decoder, and adaptive initialized model parameters are generated for the unknown model. According to the method, cross-model structure high-quality parameter initialization is realized, the model training time is shortened, and the large model training requirement is met.

The invention discloses a crystal grain detection and analysis method and related equipment, effective crystal grains are obtained through corrosion expansion pretreatment denoising, OfficientNet and bidirectional feature pyramid network detection, contour obtaining through region segmentation and three-step filtering, qualification is judged in combination with a threshold value, a report is generated, and the existing technical problems are solved in a targeted mode. On the aspect of reducing calculation complexity, preprocessing reduces model data volume and interference, the model abandons complex modules, parameters are few, calculation cost is low, hardware occupation can be reduced, and detection and training time can be shortened; on the aspect of improving generalization ability, features are preprocessed and purified, interference of imaging conditions and material types is reduced, the number of layers of the model is simplified, multi-scale fusion ability is achieved, and detection in different fields can be adapted without optimization of a specific data set; on the aspect of reducing data dependence, invalid and pseudo crystal grains are accurately removed through three-step filtering, defects of a small-scale or low-quality data set are cooperatively made up, dependence on a high-quality large-scale data set is greatly reduced, and robustness is improved.

The application provides a multimodal large model security protection method, device and equipment, which comprises the following steps: obtaining a sample image and a sample text, and performing feature fusion on the sample image and the sample text to obtain a multimodal feature; performing a first perturbation operation on the multimodal feature to obtain a first perturbation feature, inputting the first perturbation feature into an initial multimodal large model to obtain a first prediction label; determining a first loss value and a second loss value based on the first prediction label; adjusting the initial multimodal large model based on the first loss value to obtain an intermediate multimodal large model; performing a second perturbation operation on the multimodal feature to obtain a second perturbation feature, inputting the second perturbation feature into the intermediate multimodal large model to obtain a second prediction label, and determining a third loss value based on the second prediction label; and adjusting the initial multimodal large model based on the second loss value and the third loss value to obtain a target multimodal large model. Through the application scheme, the calculation resources are saved, and the training time is reduced.

The invention provides a big language model training method and device, a data processing method and device, equipment, a storage medium and a program product. The method comprises the following steps: acquiring a sequence data sample, wherein the sequence data sample comprises a plurality of data units; performing attention calculation on the sequence data sample through a to-be-trained model to obtain a first attention intensity matrix; determining a time scheduling factor corresponding to the training time step, and determining a space scaling factor corresponding to each data unit; determining a mask matrix based on the time scheduling factor and the spatial scaling factor; adjusting the first attention intensity matrix based on the mask matrix to obtain a second attention intensity matrix, and determining an attention feature sample of the sequence data sample based on the second attention intensity matrix; and determining a loss value based on the attention feature sample, and updating parameters of the to-be-trained model based on the loss value to obtain a model obtained by training of the training time step. Through the method, the performance of the large language model can be improved.

The invention relates to an underwater marshalling path planning method and system, and the method comprises the steps: carrying out the rasterization of a target region in a water area search region based on the environment data of the water area search region, and generating a detection capability matrix; based on the detection capability matrix, generating a minimum open coverage center point set of the detection range of the marshalling serving as a whole for the target area; taking the minimum open coverage center point set as a traversal object, generating an initial candidate path sequence population through a greedy algorithm, and executing iterative optimization on paths in the initial candidate path sequence population by utilizing a multi-objective optimization algorithm to generate a global optimal path; decomposing the global optimal path into an independent path of each platform in the marshalling; and smoothing all the independent paths based on a preset constraint condition to obtain an actual application path of each platform in the marshalling, and completing underwater marshalling path planning. According to the method, real-time calculation is carried out by utilizing environment information, so that the training time, the hardware cost and the trouble of data annotation are saved.

The invention discloses a dynamic nerve radiation field real-time three-dimensional reconstruction method and device based on window attention and gradient equilibrium, and the method comprises the steps: carrying out the stratified sampling of rays, and obtaining sampling points; projecting the sampling points to six groups of two-dimensional feature planes; querying corresponding feature vectors from the six groups of feature planes through bilinear interpolation, and generating spatio-temporal features through Hadamard product fusion; inputting the fused spatial-temporal characteristics into a decoding network of a Swin Transform, and outputting the color and the volume density of a corresponding sampling point; and carrying out volume rendering integration along the ray of the camera to generate a composite image, and optimizing and updating the model according to a multi-target loss function between the composite image and a real image. The invention aims to solve the problems of large memory consumption, poor rendering quality, long training time and poor geometric consistency in the prior art, and realizes real-time three-dimensional reconstruction of a dynamic scene by using six-plane grid representation, Swin Transform feature generation and a gradient equilibrium mechanism.

The embodiment of the invention discloses a health state diagnosis method and device of a tobacco flake charging machine and electronic equipment. The method comprises the steps of constructing an original charging machine health state diagnosis model corresponding to a historical data set through a width learning algorithm, determining a target charging working condition corresponding to a collected target composite feature vector, and updating the original charging machine health state diagnosis model according to the target charging working condition, and finally, a target health state diagnosis result corresponding to the target composite feature vector is determined based on the target charging machine health state diagnosis model, so that the purpose of automatically and accurately diagnosing the health state of the tobacco flake charging machine in real time according to the composite feature vector of the tobacco flake charging machine is achieved. And moreover, the incremental diagnosis requirements of different charging working conditions are met, the training time and resources of a subsequent model are saved, and the production efficiency is further improved.

The application discloses a continuous acquisition and data model construction system and method of an intelligent sensor. In order to overcome the problems that the training data acquisition of the chemical sensor in the prior art is inconsistent with the actual application, the data acquisition efficiency is low, and effective data is insufficient, in the training process, the sensor unit is placed in the test unit, and the concentration control unit is used to adjust the concentration of different samples, so that the data is continuously acquired. In the prediction process, the sensor unit is placed in the sample input unit, is input into the concentration prediction model after correction by the state calibration unit, and the sample concentration output is obtained. The continuous data sampling method is used, the recovery process is not performed, and the balance state is reached in a relatively short time, so that the training time is greatly reduced; the response and recovery stage data are continuously acquired, so that the training data is consistent with the signal and concentration change process in the actual application; in the prediction process, the state of the sensor is dynamically detected and different models are called, so that the model accuracy is improved.

The invention discloses a cross-scene self-adaptive flexible load dynamic interval prediction method. The method comprises the following steps of multi-modal scene identification, dynamic interval prediction and cross-scene self-adaptive migration. According to the cross-scene adaptive flexible load dynamic interval prediction method provided by the invention, the cross-scene adaptive capability is realized, automatic identification of industrial, commercial and resident scenes is realized, cross-scene prediction MAPE errors are reduced to be within 3%, which is better than that of a traditional model, the interval reliability is improved, quantile non-cross constraint enables PICP to be stabilized to be more than 95%, PINAW is reduced to be 5%-8%, and the interval reliability is improved. According to the method, the requirement of power grid dispatching for a narrow and reliable interval is met, knowledge migration is efficient, through field adaptive migration learning, the new scene model training time is shortened by 60%, the data demand quantity is reduced by 50%, and the engineering application cost is reduced.

The invention belongs to the technical field of image and text cross-modal processing, and particularly relates to an unsupervised contrast cross-modal Hash retrieval method based on lifelong learning, which comprises the following steps: respectively encoding image and text data through pre-trained DINOv2-Small and BGEv1.5-Small, inputting encoded features into a contrast learning module, and carrying out Hash retrieval on the image and text data through a Hash retrieval module; and designing a self-adaptive false negative sample cancelling mechanism based on a Hash memory bank to identify potential false negative sample pairs, and converting the potential false negative sample pairs into positive semantic information, so that the model can correctly close the positive sample pairs and push away the negative sample pairs, and the discriminative representation learning is enhanced. Besides, in order to avoid the situation that the model needs to be retrained when new category data arrives, a prompt-based frozen pre-training model continuous learning strategy is designed, the plasticity and stability of the model can be dynamically balanced in a non-label stream data scene, and catastrophic forgetting of Hash learning is effectively relieved. Dynamic distribution of the streaming data is considered, training time and computing resources are saved, and continuous and efficient retrieval of the label-free streaming data is achieved.

The present application belongs to the technical field of medical data recognition, and particularly relates to a depression recognition method and system based on differential evolution feature selection. Original electroencephalogram data is acquired, and the acquired original electroencephalogram data is preprocessed; based on the preprocessed electroencephalogram data, a differential evolution-based feature selection method is used to select an optimal feature subset; a self-adaptive hierarchical fusion network model based on Transformer is constructed, and training is performed based on the optimal feature subset to obtain a depression recognition reference model and a reference recognition result. The purpose of feature selection on electroencephalogram data is to remove redundant features and irrelevant features, thereby obtaining an optimal feature subset. The screened feature subset is put into training, which not only saves training time but also obtains a higher recognition rate than original data training.

This invention relates to an intelligent database performance evaluation method, belonging to the field of database performance evaluation. It addresses the problems of low accuracy, long training times, and inefficient model construction in existing database performance evaluation methods. In the early stages of model training, this invention categorizes multiple variables into frequent and infrequent variables. A linear regression model is used to learn the impact of infrequent variables on database performance. The system parameters of the linear regression model are then input into a second-stage hierarchical machine learning model. Frequent variables are directly input into this model to learn their impact on database performance. The trained hierarchical machine learning model is then used as the database performance evaluation model. This invention is primarily used for evaluating database performance.

The application discloses a model training process-oriented computing power chip hardware Trojan design method, and relates to the field of integrated circuit security.The method is designed in view of the computing architecture of a GPU, and contains a flip-flop, a finite state machine and a load circuit.The flip-flop identifies the typical characteristics of deep learning training by monitoring the arithmetic instruction intensity and the Warp concurrent activity of the GPU runtime; when the characteristics meet a preset threshold, the finite state machine is triggered, so that the hardware Trojan enters an activated state.The load circuit has two types of designs: the first type is a gradient flipping circuit which interferes with the gradient update direction by modifying the sign bit output by a floating point calculation unit, so that the model cannot converge; the second type is a redundant calculation circuit which consumes computing resources by inserting invalid floating point operations, thereby prolonging the training time.The application designs and triggers the hardware Trojan by utilizing the architecture characteristics of the GPU, has high concealment, and can effectively destroy the accuracy and efficiency of a deep learning training task.

This invention provides an energy efficiency optimization method for the large-scale deployment of ultra-reliable low-latency (URLLC) devices, comprising: Step 1: Calculating the short packet collision probability and short packet reception failure probability of a single device, and calculating the packet loss rate of the device based on the calculated short packet collision probability and short packet reception failure probability; Step 2: Calculating the packet loss rate of each device, and grouping devices with similar communication environments into one category based on the calculated packet loss rate; Step 3: Optimizing the energy efficiency of all devices categorized in Step 2 using a deep reinforcement learning method. The beneficial effects of this invention are: 1. The method can effectively reduce the transmission power of sensing devices sending short packets and the number of times the same short packets are repeatedly sent, significantly improving the communication energy efficiency of URLLC devices and extending their battery life; 2. The method of this invention uses a classification method, which significantly shortens the training time of deep reinforcement learning.

The invention discloses a multi-tower section stress rapid prediction method and system for a mixed tower drum, and the method comprises the steps: building a tower drum modal physical model of a target mixed tower drum, and carrying out the modal analysis to obtain a modal parameter set of the target tower drum; establishing a tower section stress prediction hybrid neural network model of the target mixed tower drum based on the modal parameter set of the target tower drum; inputting a pre-constructed training data set into the tower section stress prediction hybrid neural network model for training, wherein the training data set comprises a plurality of seismic oscillation training time histories and a high-precision stress time history of each target tower section corresponding to each seismic oscillation training time history; and inputting the target seismic oscillation time history into the trained tower section stress prediction hybrid neural network model to obtain a target prediction stress time history of each target tower section. According to the method, the calculation efficiency, the prediction precision and the physical interpretability of the prediction result can be considered.

The present application relates to the technical field of distributed model training, and provides a model distribution parallel training process automatic optimization method, system and device. The method comprises the following steps: defining a task plane according to the communication and computing power of a neural network model, dividing the task plane to obtain a divided task plane; constructing a training time consumption model according to the divided task plane to obtain total time consumption; using a tournament algorithm to design a genetic algorithm to solve the training time consumption model according to the total time consumption to obtain a training optimization result; and embedding the training optimization result into the neural network model for optimization and updating. The present application models the training process in detail under the scene of a mesh pipeline and a heterogeneous network and computing power equipment. According to the established training process-training time consumption model, a meta-heuristic algorithm is applied to search for an optimal solution to realize automatic optimization of the distributed parallel training process.

The application provides a communication method, device and equipment and a storage medium. The method comprises the following steps: a network device sends a training time delay constraint instruction to a terminal device participating in federated learning; the terminal device receives the training time delay constraint instruction from the network device; and the training time delay constraint instruction is used to indicate the reporting deadline of a model training result. Since the network device can send the training time delay constraint instruction to the terminal device participating in the federated learning, the training time delay constraint instruction is used to indicate the reporting deadline of the model training result, so that in the case that there are two or more terminal devices participating in the federated learning, different terminal devices can upload the model training result according to the training time delay constraint instruction, so as to reduce the time difference of the model training result uploaded by different terminal devices participating in the federated learning, thereby ensuring the overall training time delay of the federated learning, reducing the possibility of late reporting of the terminal device, and helping to reduce the waste of network resources.

The invention provides a self-adaptive network security situation awareness method and device combined with online learning, and the method comprises the steps: collecting real-time network state data and system load index data of a preset data source, carrying out the preprocessing of the data to form a multi-modal time sequence segment, inputting a pre-training time sequence data prediction model activated by employing Monte Carlo Dropout through a sliding window, and carrying out the prediction of the real-time network state data and system load index data. The method comprises the following steps of: calculating a prediction value of a next time period, outputting a prediction value and an uncertainty quantity of the next time period, calculating a threat probability through historical residual probability distribution fitting, realizing double-index risk assessment based on a preset threshold interval system, dividing into three types of states, and finally, respectively triggering online learning, configuration maintenance or intervention disposal flow for different states. According to the method, by introducing uncertainty quantized double-index evaluation and state-driven online learning closed loop, crossing of network security situation awareness from static detection to dynamic self-adaption is achieved, and the two core problems of insufficient real-time performance and concept drift in an edge computing scene are effectively solved.

The application discloses a cloud data server hybrid training method and system, which is applied to a cloud data server cluster containing various computing resources. The method analyzes the model structure, dataset, time delay constraint and isolation constraint of the training task to generate a task image, collects the computing node resource configuration, running state and isolation capability to generate a resource image, and constructs a co-location interference degree index table of the task and resource combination based on historical monitoring data. During scheduling, the heterogeneous computing power utilization rate, estimated training time delay deviation and co-location interference degree are taken as indexes to perform weighted evaluation on the candidate resource allocation scheme to generate a hybrid training scheduling strategy, and resource isolation is implemented through container and accelerator multi-instance division. During running, the training time delay and actual co-location interference degree are continuously monitored, the scheduling weight and co-location interference degree index are dynamically adjusted according to the deviation, closed-loop optimization of multi-task hybrid training is realized, the heterogeneous resource utilization rate is improved, and time delay violation and co-location interference are reduced.

The application discloses a digital twin upper limb rehabilitation training robot and method considering patient emotion, which comprises: a teleoperation master robot for obtaining accurate action information of a trainer; at least one wearable exoskeleton mechanical arm for obtaining force, heart rate, electromyographic signal and facial expression of the patient respectively, restoring the action of the trainer and making the same movement with the patient; a computer system carrying a digital twin, comprising a digital twin model established according to the connecting rod relationship of the mechanical arm, and a man-machine interaction module and a voice broadcast module; according to different training stages, the action of the trainer is learned and output to guide the exoskeleton mechanical arm to replicate the action. In the training process, the trainer can participate and help the patient to restore the action of the trainer through the teleoperation teaching mode, and at the same time, the patient emotion is considered to adjust the training progress and intensity, so that better training effect is achieved and the training time is saved.