1089 results about "Training time" patented technology

The invention belongs to the technical field of speech recognition. The invention provides a low-resource language self-adaptive speech recognition method based on AdLoRA Plus. The low-resource language self-adaptive speech recognition method comprises the following steps of: obtaining a speech According to the embodiment of the invention, the AdLoRA Plus fine tuning technology is introduced, so that the speech recognition precision of the Whisper model is remarkably improved in a low-resource language environment. Through a low-rank adapter layer and dynamic learning rate adjustment, efficient parameter updating is realized under limited training data. According to the method, a strategy of freezing basic parameters is adopted, only an AdLoRA Plus adapter layer is trained, calculation overhead and training time are remarkably reduced, pre-training knowledge is kept, and overfitting is prevented. The method has wide adaptability and expansibility, and is ensured to be outstanding in practical application. The voice recognition precision, the reasoning efficiency and the training efficiency are effectively improved, multiple challenges in application in small languages and specific fields in the prior art are solved, and the method has important practical application value.

The invention discloses an industrial equipment fault prediction system based on federated learning and a privacy protection method, and relates to the field of industrial equipment fault prediction. The data acquisition preprocessing module extracts fault features through compressed sensing downsampling, screens and uploads the fault features; the federal learning training module adopts a layered architecture and a dynamic algorithm to schedule a learning rate; the fault prediction and diagnosis module constructs a space-time diagram neural network and fuses a physical model to improve generalization; the privacy protection security communication module performs homomorphic encryption storage and zero-knowledge proof verification update; the knowledge graph construction reasoning module constructs a dynamic graph, locates a fault root cause through causal reasoning, and supports cross-device knowledge migration. By adopting the quantum and federated learning technology, the industrial equipment fault diagnosis accuracy is high, the attack resistance is high, the encryption efficiency is greatly improved, the model training time is shortened, cross-equipment knowledge migration is realized, the operation and maintenance cost is reduced, and the intelligent operation and maintenance development of the industrial equipment is promoted.

The invention relates to a multi-vehicle intelligent driving cooperative training method based on a block chain in the Internet of Vehicles, and belongs to the technical field of mobile communication. Aiming at the problems of data privacy leakage, too high synchronous federated learning time delay, inaccurate node contribution degree evaluation, insufficient single-chain block chain expandability and the like existing in intelligent driving cooperative training in the Internet of Vehicles, an asynchronous federated learning framework with fusion of a cloud-edge-end three-layer architecture and a double-layer block chain is constructed. A local training strategy is optimized through multi-agent reinforcement learning to minimize the total time delay of a system, a dynamic reputation evaluation mechanism based on training interaction timeliness, confirmed site occupancy and model quality contribution is designed to screen high-reputation nodes, and asynchronous model verification and PBFT main chain global aggregation are realized by adopting a DAG block chain. According to the method, on the premise of guaranteeing data privacy, the model sharing efficiency is improved by more than 30%, the training time delay is reduced by 40%, the convergence stability of a global model is effectively enhanced, and the expansibility bottleneck of a traditional architecture is solved.

An improved YOLOv8-based industrial pipeline defect detection method and system are provided. The method includes: acquiring a pipeline surface image; and recognizing a defect position and a defect type in the pipeline surface image by using a pipeline defect detection model. The model is based on an improved YOLOv8 network, which replaces the original means with WIoU loss and a Sophia optimizer during training, and a final model can quickly and accurately recognize the defect position and the defect type in the pipeline surface image. Compared with a conventional YOLOv8 algorithm, training stability, convergence speed, and recognition accuracy are improved by replacing a CIoU loss function with a WIoU loss function. An official AdamW optimizer is replaced with a Sophia optimizer, training time of the model can be greatly shortened, a lot of computing resources can be saved, and less memory is occupied.

The invention provides a VR-based dynamic visual function training method and device, a medium, a program product and a terminal. An initial training task type and configuration parameters are generated according to a visual evaluation result of a patient, a sighting mark and a stimulation signal for visual function training are constructed, and training is executed through a VR unit. In the training process, feedback data of a patient is collected in real time, a real-time evaluation result is analyzed and generated, task types, configuration parameters and sighting marks are dynamically adjusted, and new stimulation signals are generated to continue training, so that dynamic optimization of training content is achieved. Through real-time evaluation and dynamic adjustment, the problem that the training content is not matched with the requirements of the patient is solved, the defects of a traditional method in the aspects of individuation and adaptability are overcome, the training effect is remarkably improved, the requirements of the patient are better met, the training time is shortened, and the efficiency is improved. The device is especially suitable for rehabilitation training of various visual function defects, and has remarkable advantages in the aspect of dynamic visual function recovery.

The invention relates to the technical field of computer vision image enhancement, and provides a zero sample learning defogging image enhancement method and device based on image decomposition, and the method comprises the steps: introducing three initial image layer generation networks: an initial clean image layer generation network, an initial atmospheric light image layer generation network, and an initial transmissivity image layer generation network; a target atmospheric light layer generation network and a target transmissivity layer generation network are obtained through joint training and are respectively used for determining a target atmospheric light layer and a target transmissivity layer of a to-be-defogged image, and a defogged image corresponding to the to-be-defogged image is solved by combining an atmospheric scattering model. In the joint training process of the three initial layer generation networks, only the to-be-defogged image is needed, a large number of training samples do not need to be introduced, the problem that in the prior art, the training samples of an image defogging model are difficult to obtain can be avoided, the calculated amount in the training process can be reduced, the training duration of the network can be shortened, and the training efficiency of the image defogging model is improved. And the training time cost of the network is reduced.

According to an embodiment, a performance benchmark database is obtained, where the performance benchmark database at least includes structural data of one or more deep neural network models, time performance data and computing resource consumption data of a plurality of deep learning applications based on the one or more deep neural network models; a training dataset is extracted based on the performance benchmark database, where the training dataset has a plurality of parameter dimensions, the plurality of parameter dimensions including: structures of deep neural network models of the plurality of deep learning applications, resource configuration of the plurality of deep learning applications, and training time of the plurality of deep learning applications; and correspondence among the parameter dimensions of the training dataset is created so as to create an estimation model for estimating resources utilized by deep learning applications.

System and techniques to facilitate the translation of a sign language into another language are described herein. A modular architecture may be used in which the output of different classifiers may be used to produce intermediate representations, or final translations, of the sign language. These classifiers may be trained on different types of signs to enhance accuracy while reduce training time and complexity.

The invention relates to the technical field of medical treatment, in particular to a child fracture rehabilitation management system based on virtual reality, which comprises a multi-modal sensing and data acquisition module, an intelligent rehabilitation training module, a multi-dimensional evaluation and feedback module and a data intelligent analysis and prediction module, compared with the defects that an existing rehabilitation training technology mainly depends on subjective experience of doctors to formulate a static scheme and lacks multi-source data dynamic feedback, the method comprises the steps that firstly, a comprehensive and accurate patient ability portrait is constructed; intelligent dynamic adjustment of training parameters is realized by utilizing an LSTM gating mechanism, a forgetting gate filters invalid historical data, an input gate fuses real-time action quality evaluation, an output gate generates a personalized difficulty gradient, and a patient ability threshold is continuously memorized in combination with a cell state, so that a training scheme has an adaptive evolution ability, and the training efficiency is improved. The parameter adjustment period of a traditional rehabilitation scheme is compressed from a week level to a real-time level, the accuracy rate of risk prediction of muscle strain and the like is improved, and the utilization rate of effective training time is increased.

The invention discloses a Gaussian splash model training method and device, a medium and a program product, and relates to the technical field of computer graphics, and the method comprises the steps: carrying out the initialization of an original Gaussian splash model based on a sparse input sample image; according to the sparse input sample image and a preset Gaussian densification strategy, multiple rounds of iterative training are carried out on the original Gaussian splash model to obtain a progressive Gaussian splash model, and the Gaussian densification strategy comprises the following steps: when the number of times of iterative training of the original Gaussian splash model meets a preset number range, the number of times of iterative training of the original Gaussian splash model meets the preset number range; the densification threshold value used by the original Gaussian splash model in the iterative training process is inversely proportional to the number of iterative training times. According to the method, a relatively high densification threshold value is used at the initial stage of training, so that an over-fitting risk caused by over-high densification degree at the initial stage of training under a sparse input view angle is reduced.

The invention discloses a training method of a voice recognition model based on noise deconstruction, a voice recognition method, a device, equipment and a medium. According to the method, a staged training strategy that a noise unwrapping module is firstly isolated and trained and then a Conformer-Transducer architecture is finely adjusted and trained is adopted, so that high calculation complexity and training difficulty caused by simultaneous training of a plurality of complex modules are avoided. In the isolation training stage, the performance of the noise unwrapping module can be quickly optimized; in the fine tuning training stage, the trained noise unwrapping module is utilized to concentrate on optimizing the Conformer-Transducer architecture, so that the training efficiency is improved, and the training time and the consumption of computing resources are reduced. In the isolation training and fine tuning training process, the parameters of part of modules are frozen, the number of parameters needing to be optimized is reduced, and therefore the calculation complexity is reduced. Noise and pure voice in a voice signal are deconstructed through the noise unwrapping module, and accurate semantic understanding is carried out in combination with a Conformer-Transducer architecture, so that the whole voice recognition model has higher robustness to noise.

The invention discloses an unmanned aerial vehicle double-layer conflict resolution method for an urban low-altitude dense traffic flow airspace, and the method comprises the steps: constructing a mixed integer nonlinear programming model based on a conflict resolution strategy before an intelligent unmanned aerial vehicle takes off, and optimizing the mixed integer nonlinear programming model through an improved random fractal search algorithm; after the intelligent unmanned aerial vehicle takes off, a dual depth Q network is constructed and network training is carried out on the dual depth Q network, and the trained dual depth Q network is utilized to enable the intelligent unmanned aerial vehicle to output conflict-free maneuvering actions at each time step. According to the method, a double-layer conflict resolution framework is provided in combination with deep reinforcement learning, and conflict resolution before flight and real-time conflict resolution are combined, so that the occurrence of air collision is greatly reduced compared with that of any method which is independently used, and the intelligent unmanned aerial vehicle can avoid the current obstacle in an ultra-low-altitude urban dense traffic flow environment, thereby improving the traffic flow resolution efficiency. And meanwhile, secondary conflicts are reduced, and the training time of the real-time conflict solving method based on reinforcement learning is also remarkably shortened.

The invention discloses a one-stage severe weather fruit detection method based on dynamic coding, and the method comprises the steps: collecting fruit image data on site in an agricultural scene, carrying out the data preprocessing of the collected image data, setting the hyper-parameters of a to-be-constructed one-stage severe weather fruit detection model based on dynamic coding, and carrying out the calculation of the hyper-parameters of the to-be-constructed one-stage severe weather fruit detection model. Constructing and training a first-stage severe weather fruit detection model based on dynamic coding; detecting fruits in the test set image, outputting a detection result mask, and comprehensively evaluating the model; the method solves the problems of poor generalization and applicability, tedious steps of a monitoring method and low real-time performance in the prior art, and has the characteristics that accurate detection can be carried out without pre-processing the image, and the training time and tedious steps of image pre-processing are reduced.

The invention relates to a video generation model acquisition method and device, computer equipment, a storage medium and a computer program product. The method comprises the following steps: acquiring a training sample set for a target face action; executing the following training process by using the training sample set to train a time sequence feature extraction network: inputting a plurality of preorder video frames corresponding to the video samples into the time sequence feature extraction network to obtain time sequence features corresponding to the plurality of preorder video frames; inputting the time sequence features corresponding to the plurality of preorder video frames into a pre-trained image generator to obtain a predicted video frame corresponding to the target video frame; training a time sequence feature extraction network according to the difference between the target video frame and the predicted video frame; and obtaining a video generation model according to the trained time sequence feature extraction network and a pre-trained image generator. By adopting the method, a video with relatively high quality can be generated.

The invention discloses a structural XCT fiber bundle identification and segmentation method based on improved U-Net, which is used for identifying fiber bundles in XCT slices of a woven CMC structural member, and the network is applied to the CMC field for the first time. According to the method, the down-sampling operation of the original U-Net network is greatly modified, the characteristic extraction capability of the model on the fiber bundle is improved, the parameter quantity of the model is reduced, and the network model can be trained more stably, so that the efficiency of segmenting and identifying the fiber bundle by the model is improved. Wherein an Inception structure is adopted, so that the visual field of the model is widened, and the recognition and segmentation capability of the model on fiber bundles with different sizes and boundaries thereof in CT is enhanced. And residual connection is added in an original model by adopting a ResNet structure, so that gradient propagation in training is ensured, the depth of the improved U-Net is deepened, and the accuracy of segmentation recognition is improved. And the bottleneck structure is adopted, so that the parameter quantity of the whole model is greatly reduced, the training time is shortened, and the recognition and segmentation speed is increased.

An unmanned aerial vehicle (UAV) identification method based on blind source separation and deep learning is disclosed, in the method, firstly, the one-dimensional radar cross section millimeter wave data set of the UAV is acquired, and the mixed signal is obtained by mixing, and the improved FastICA algorithm is used to separate it. Secondly, the separated signal is converted into a two-dimensional image by data transformation, and the two-dimensional image is augmented, the obtained data set is divided into training set, validation set and test set. Thirdly, establish a UAV classification model based on Improved ResNet18 and train this model on the training set to achieve UAV classification. In the present invention, the training time of the network is not greatly increased while the network identification accuracy is improved, so that it can well complete the UAV type identification, and the design is more reasonable and effective.

The invention provides a traceable federal incremental learning method based on grouping feature aggregation, and relates to the technical field of federal learning, and the method comprises the following steps: S1, carrying out end node traceable task learning; carrying out local pruning and fine tuning based on the initial sharing model, and judging whether tasks are repeated or not according to label distribution similarity; s2, constructing a cloud-end side collaborative grouping model; sending the global shared knowledge generated by the server model and the aggregation weight corresponding to the global shared knowledge to the end node; s3, at the end node, the client updates own model parameters based on the local cross entropy loss, the global shared knowledge generated by the server model of each group and the aggregation weight corresponding to the global shared knowledge, and outputs the updated model; and S4, replacing the new sub-model in the S1 with the updated model or updating the sub-model, repeating the steps S1 to S3, and stopping training until specified training times are reached. The method aims to effectively cope with challenges caused by storage limitation and task repeatability.

Disclosed is a method and apparatus for predicting a training time. The method includes generating a first execution graph representing a distributed learning process of a neural network model, wherein the first execution graph includes a computation node representing a repeated computation of the neural network model based on a parallelism technique; generating a second execution graph based on the first execution graph and a target operator corresponding to the computation node, wherein the second execution graph includes an execution time of a hardware kernel performing the target operator; and predicting a training time of the neural network model based on the second execution graph.

The invention provides a large model parameter optimization and self-adaptive adjustment method based on reinforcement learning, and relates to the technical field of optimization and self-adaptive adjustment of large model parameters based on reinforcement learning in machine learning, and the method comprises an overall architecture fusing a large model main body, a reinforcement learning module, an environment perception module and a reward feedback mechanism. The method is used for realizing dynamic optimization and adaptive adjustment of large model parameters, and can rapidly increase the learning rate according to environment feedback in the initial stage of large model training, so that the model parameters rapidly approach to the direction of an optimal solution, and the early-stage exploration time of training is greatly shortened. In the later stage of training, the learning rate can be accurately reduced, model oscillation is avoided, and it is ensured that the model is stably converged to a globally optimal solution. According to the dynamic adjustment mechanism, the number of iterations of training is effectively reduced, the training efficiency is greatly improved, a large number of computing resources and time cost can be saved, and for example, in large-scale image classification model training, the training time can be shortened by more than 30%.

The invention discloses a decentralized federated learning communication topological structure optimization method and device and a medium, and relates to the field of distributed machine learning and optimization algorithms, and the method comprises the steps: optimizing a communication topological structure of a terminal node through employing an ant colony optimization algorithm when a decentralized federated learning system is subjected to iterative training; a fitness function of the ant colony optimization algorithm takes minimum communication load and communication delay as a target function; summarizing and averaging the model parameters of each terminal node model in the optimal communication topological structure of the terminal nodes to obtain an average model parameter; updating the model parameters of the terminal node model of each terminal node in the optimal communication topological structure by adopting the average model parameters to obtain the updated terminal node model of each terminal node; and when the number of iterative training times reaches the maximum number of training times, completing training of the decentered federated learning system, and outputting an optimization result. According to the invention, the communication efficiency in the decentralized federated learning system is improved.

Systems and methods disclose herein procedures for accelerated tree learning. In one class, the acceleration is based on self-adapting learning rates, while in another class, the acceleration is based on a plurality of learning rates, wherein each learning rate varies over the training; each learning rate increases linearly as a respective pseudo residual maintains a direction across sequential training iterations; and each learning rate decreases exponentially as the respective pseudo residual changes direction across sequential training iterations. The latter can be incorporated with other methodologies, such as momentum-augmented gradient boosting and Nesterov Accelerated Gradient Boosting. These systems and methods for accelerated tree learning exhibit a marked reduction in training time and resources required for gradient boosted trees.

The invention discloses a semi-centralized edge federated segmentation learning method of a federated learning system facing Internet of Things terminal equipment under a wireless network. The method comprises the following steps: establishing a to-be-trained model; splitting the to-be-trained model into a server-side first model and a client-side first model; distributing the client first model to each client in the to-be-trained client group; training each client first model to obtain a client second model and a first gradient corresponding to each client; according to the plurality of first gradients, updating the server-side first model to obtain a server-side second model and server-side second model features; according to the server-side second model features, updating each client-side second model to obtain a client-side third model corresponding to each client-side in the to-be-trained device group; repeating the execution; selecting the models meeting the first condition from the plurality of client third models for aggregation to obtain an aggregated client third model; and combining the aggregated third client model and the aggregated second server model to obtain a trained model. The method has the characteristics of low computing power consumption and short training time delay.

The invention discloses an image classification method and system based on pulse Transform of dynamic gradient adjustment. The method comprises the following steps: acquiring an image classification data set and randomly dividing the image classification data set into a training set, a test set and a verification set; the method comprises the following steps: designing MALIF neurons to reconstruct LIF neurons in an SNN network, introducing a dynamic time step mechanism module to construct a pulse Transform model for image classification, and training by using a training set; wherein the MALIF neurons can learn a membrane potential time constant and a voltage threshold, time and space information can be better utilized, and the MALIF neurons can efficiently transmit information; and obtaining a classification result of the to-be-tested image sample by using the trained pulse Transform model. The method is suitable for the field of image classification, reduces the training time of the model, has better performance, and improves the prediction precision and accuracy.

The method comprises the following steps: obtaining training data of a voice data construction model, preprocessing the obtained training data to obtain original waveform features and multi-scale fundamental frequency features, inputting the features into the model to carry out model training, and obtaining an attention enhancement model; in the training process of the detection model, hyper-parameters of the detection model are set, a loss function is continuously reduced, the trained detection model is obtained when the set training frequency is reached, test data of the voice data construction model are obtained, and preprocessing is carried out to obtain original waveform features and multi-scale fundamental frequency features; the original waveform features and the multi-scale fundamental frequency features are input into a trained model for model testing, in the testing process, a voice authenticity classification result is obtained, the performance of the model is evaluated, and the performance of the detection model can be effectively improved.

The invention discloses a search intention recognition method and device, equipment and a storage medium, and relates to the technical field of reinforcement learning, and the method comprises the following steps: generating training data based on an original corpus; performing supervision fine tuning processing on the lightweight language model based on the training data to obtain a fine tuning trained lightweight language model; performing strategy optimization on the lightweight language model subjected to fine tuning training through an improved GRPO reinforcement learning algorithm to obtain a lightweight language model subjected to reinforcement training; constructing a double-model hybrid architecture based on a short thinking chain model and a thinking chain-free model in the lightweight language model subjected to enhanced training; and outputting a final sequence of the related documents of the search terms of the user through the double-model hybrid architecture. According to the technical scheme of lightweight language model fine tuning, reinforcement learning optimization and double-model hybrid reasoning, the target search scene recognition accuracy is improved, the real-time response speed is high, and training time consumption and reasoning resource requirements are reduced.

The invention provides a parallel policy determination method, electronic equipment, a storage medium and a program product. The method comprises the following steps: determining cluster communication information, model calculation information and model resource occupation information according to information of a cluster used when a large model is trained and a structure of the large model; according to the cluster communication information and the model calculation information, modeling training time required by large model training to obtain a training time model; according to the model resource occupation information, carrying out modeling on resources occupied by large model training to obtain a resource occupation model; and according to the training time model and the resource occupation model, determining a target parallel strategy from multiple groups of parallel strategies as a parallel strategy used during large model training. Thus, by constructing the training time model and the resource occupancy model, the optimal parallel strategy can be automatically screened out from the multiple groups of parallel strategies, manual intervention is not needed, and therefore efficiency can be improved, and complexity can be reduced.

The invention provides a model training method. The method comprises the steps of collecting information of a first model; and making a re-calculation strategy according to the communication time of the first model in the communication stage and the execution time required by the plurality of operators in the first model. The information of the first model comprises execution time respectively required by a plurality of operators in the first model, the recalculation strategy comprises at least one recalculation operator in the plurality of operators and opportunity for executing a recalculation process by the at least one recalculation operator, the recalculation operator is an operator used for executing the recalculation process in the plurality of operators, and the opportunity for executing the recalculation process by the at least one recalculation operator in the plurality of operators is the opportunity for executing the recalculation process by the at least one recalculation operator. The opportunity at which the at least one recalculation operator performs the recalculation process includes performing the recalculation process in parallel with the communication phase of the first model. The recalculation process is parallel to the communication process of the model, so that the model training time is shortened, and the throughput of model training is improved.

The invention discloses an evolutionary multitask intrusion detection feature selection method based on double-view dimension reduction, and belongs to the crossing field of network security and machine learning. According to the method, simplified and complementary tasks are generated through an improved double-view-angle dimension reduction method based on a filtering method and a grouping method, and rapid recognition of foreground areas in a high-dimensional search space is promoted; and through a multi-task optimization mechanism based on double archives, feature subsets with the same performance are maintained, convergence guidance is provided, and the balance between convergence and diversity among tasks is realized, so that a plurality of equivalent feature subsets with high intrusion detection accuracy and fewer feature numbers are searched and obtained, and the accuracy of intrusion detection is improved. The accuracy and the training efficiency are improved, and meanwhile, diversified and more interpretable decision support is provided for the model. According to the method, the Pareto optimal feature subsets with the same performance but different features can be quickly searched in the high-dimensional feature space, and more flexible and diversified selections are provided for decision makers, so that the accuracy of intrusion detection is improved, and the model training time is shortened.

The present invention provides a method and system 100 for classifying defects in a wafer using wafer defect images based on a deep learning network. Embodiments herein use the synergy between multiple modalities of wafer defect images to make classification decisions. Additionally, by adding a mixture of modalities, information can be obtained from different sources, such as color images, ICI, black and white images, etc., to classify defect images. In addition to the mixed modalities, reference images can be used for each modality. The reference image of each modality image is provided to the deep learning model to focus on the defect itself rather than the underlying lithography associated with the defect image. Further, the reference image can be provided to the training process of the deep learning model, which can significantly reduce the number of labeled images and the training time for the deep learning model to converge.

The invention discloses a structure optimization method based on a mechanical embedded heterogeneous graph neural network. Comprising the following steps: pre-training a heterogeneous graph neural network, and carrying out transfer learning on the pre-trained heterogeneous graph neural network to obtain a pre-training model, and embedding a preset mechanical loss function in the pre-training model to obtain a mechanical embedded heterogeneous graph neural network, and inputting the real-time load condition of the engineering structure into the mechanical embedded heterogeneous graph neural network to output a second predicted value. According to the method and the device, the calculation efficiency and the calculation precision are considered in a mode of combining the pre-training model and neural network re-parameterization, the training time of the pre-training model is greatly shortened, the prediction precision is also improved, and based on the mechanical constraint and the adaptive weight coefficient in the improved loss function, the iteration process is more robust, and the prediction precision is improved. Therefore, the optimization of the engineering structure is realized.