23results about How to "Shorten training time" patented technology

The embodiment of the invention relates to the technical field of video detection, in particular to a video behavior recognition method, device and equipment based on multi-mode large model fine tuning. A specific embodiment of the method comprises the steps of obtaining a video behavior recognition data set and an initial video behavior recognition model; for each piece of video behavior identification data, executing the following steps: determining a key frame image corresponding to a video frame sequence included in the video behavior identification data; storing the key frame image to a preset image storage position; based on the label text template, the key frame image, the image storage path and a behavior label included in the video behavior identification data, a training sample is generated, and the training sample comprises a sample image and sample label data; training the initial video behavior recognition model, and determining the initial video behavior recognition model as a video behavior recognition model; and inputting the target video data into the video behavior recognition model to obtain video behavior text information. According to the embodiment, the training duration of the video behavior recognition model can be prolonged.

The invention discloses a microwave-based polyethylene pipe hot-melt welding defect detection system, and relates to the technical field of polyethylene pipe hot-melt welding joint defect detection, and the method comprises the following steps: firstly, collecting microwave reflection signals of normal and defect-containing hot-melt welding joints, and carrying out normalization; the processed signals are sliced, sliced fragments are labeled, time domain and frequency domain features are extracted, and a defect sample data set is constructed; carrying out feature selection on the extracted features, and selecting 10 features with the highest contribution to the recognition accuracy; constructing a support vector classifier model, performing training by using the data set after feature selection, and performing hyper-parameter optimization on the model to obtain an optimal classification model; designing and realizing a polyethylene pipe hot-melt welding defect detection system, and detecting whether a welding position contains defects and defect positions or not; according to the invention, accurate detection of polyethylene pipe hot melting welding defects can be realized, defects and positions can be rapidly identified, and technical support is provided for quality safety detection of polyethylene pipes.

The embodiment of the invention provides a malicious domain name generation method, a domain name generation model training method and related equipment, and belongs to the technical field of network security. The malicious domain name generation method comprises the following steps: extracting sequence features of an obtained sample domain name character sequence through an initial domain name generation model, and determining a sample generation domain name based on the sequence features; determining an initial parameter set of the current initial domain name generation model, and performing Bayesian optimization on model parameters of the initial domain name generation model based on the sample generation domain name and the initial parameter set to obtain a trained domain name generation model; obtaining a plurality of malicious domain names to be processed, and generating a domain name character sequence based on the plurality of malicious domain names; and inputting the domain name character sequence into the trained domain name generation model, and predicting to obtain a target generated domain name with the same feature as the malicious domain name sequence. The generation efficiency of the malicious domain name can be improved, and then the detection efficiency of the malicious domain name is improved.

A lightweight image super-resolution method based on dynamic reparameterization includes: S1 inputting a preprocessed low-resolution original image into a CNN-based lightweight image super-resolution backbone network and image reconstruction module to obtain a super-resolution image; S2 calculating a loss function based on the super-resolution image and the corresponding high-resolution original image after preprocessing, and updating the network weights through backpropagation; S3 expanding the convolutional layer with the largest contribution score in the iteration cycle into an enhanced edge-oriented reparameterization branch module according to the iteration cycle and initialization conditions, and pruning all reparameterization branches with the smallest contribution score that have been expanded into reparameterization branch modules in the iteration cycle; repeating steps S1-S3 until the backbone network converges. This method improves reconstruction performance without changing the lightweight network parameters of the backbone network and maintaining fast inference speed. Compared to other reparameterization methods, it can significantly reduce training time and memory consumption.

The application provides a wall-penetrating target behavior recognition method based on channel state information, which comprises the following steps: preprocessing the collected CSI signals; performing PCA data dimension reduction on the preprocessed CSI signals, removing redundant signals and irrelevant information, and extracting optimal subcarriers; performing first-order difference processing on the signals extracted after PCA dimension reduction, and then using a method based on a buffer sliding window to segment effective feature signal segments; converting the effective feature signal segments into feature images with time-frequency domain features through STFT, and inputting the feature images into a pre-trained SE-ResNet18 convolutional neural network for behavior recognition classification. The method can penetrate the wall to realize the behavior recognition of the human target behind the wall. Compared with the traditional deep learning network, the method adopts a small sample transfer learning method combined with a pre-trained model, and the recognition accuracy can reach 91.67% under the conditions of fewer training times, fewer iteration times and shorter training time, so that the behavior recognition classification task can be effectively completed.

The present application relates to the field of signal detection, and particularly to a non-cooperative unmanned aerial vehicle burst frequency point detection method and device based on a Seg-Yolo model. The Seg-Yolo model proposed by the present application does not require prior sequences of non-cooperative unmanned aerial vehicle signals at all, and does not require prediction of specific sequences, and can directly realize continuous, discontinuous, overlapping and non-overlapping signal frequency point detection in a time-frequency graph. Through preprocessing, image noise reduction and the Seg-Yolo model, the present application reduces the image volume that needs to be processed and simplifies the size of the deep learning network, can accurately lock the center frequency of each frequency point on the basis of reducing the training time and improving the training efficiency, and improves the accuracy and reliability of signal extraction based on complex frequency hopping signals.

The embodiment of the invention provides a voice data processing method, electronic equipment, a vehicle and a storage medium, and the method comprises the steps: obtaining multiple segments of voice data, and generating text data corresponding to the voice data; cutting each segment of voice data based on a preset duration range to obtain at least one segment of sub-voice data corresponding to each segment of voice data and text data corresponding to each segment of sub-voice data; extracting acoustic features and semantic features of each segment of sub-voice data, and extracting a text vector of text data corresponding to each segment of sub-voice data; and storing the acoustic features, the semantic features and the text vectors of each segment of sub-voice data into a feature database. According to the embodiment of the invention, the voice data can be cut based on the preset duration range, the duration difference of different voice data is reduced, the time consumption of the voice feature extraction process is avoided, the effective voice features can be extracted, the quality of the voice data is effectively improved, and the training efficiency of the large voice model is improved.

The application belongs to the technical field of computers and discloses a parallel segmentation learning method and device based on clusters, equipment and a storage medium. The method comprises the following steps: obtaining a plurality of to-be-learned user terminals and user communication information of each to-be-learned user terminal; performing cluster division on each to-be-learned user terminal according to each to-be-learned user terminal and the user communication information of each to-be-learned user terminal, and determining a plurality of target user clusters; and performing cluster serial segmentation learning according to an aggregated user terminal model and each target user cluster to obtain a target user terminal model, wherein the aggregated user terminal model is obtained by performing parallel segmentation learning on each target user cluster according to the target spectrum resources of each to-be-learned user terminal. Through the above method, the overall training time delay in the segmentation learning process is effectively reduced, the efficiency of segmentation learning is improved, the negative effects caused by network heterogeneity and dynamics are inhibited, and the convergence and accuracy of existing segmentation learning technology are ensured.

The invention provides an improved MobileNetV3 lightweight electric tower identification algorithm, and belongs to the technical field of electric tower identification algorithms, and the algorithm comprises the steps: firstly, introducing an expansion convolution method into MobileNetV3 to expand the receptive field range of a convolution kernel, so as to reduce the loss of electric tower characteristics; then, a channel attention mechanism SENet in the network is replaced with a two-dimensional attention mechanism SimAM introduced without parameters, and the parameters of the network are reduced; and finally, in normalization operation, batch data normalization of the backbone network is replaced with group data normalization of which the accuracy is stable under various batch sizes, so that errors caused by small-batch data processing are reduced. Experiments on a self-made electric tower database prove that compared with a baseline algorithm, the network keeps a high recognition rate of 95.1%, the parameter quantity of the network is reduced, and rapid convergence and accurate recognition of the network on an electric tower image are realized.

The invention discloses an optical receiver construction method, device and equipment, a communication system and a storage medium, and relates to the field of network communication, and the method comprises the steps: constructing a first neural network equalizer of a first communication link based on a neural network; configuring a model parameter of a second neural network equalizer as an initial model parameter of the first neural network equalizer; based on the initial model parameter, performing fine tuning training on the first neural network equalizer by using training data of the first communication link to obtain a target neural network equalizer; the target neural network equalizer is suitable for the optical receiver of the first communication link; the first communication link is different from the second communication link. Through migration and multiplexing of knowledge, the construction process of the optical receiver is converted from repetitive low-efficiency labor to efficient adaptive learning, so that the deployment efficiency and expandability of an optical fiber communication system are greatly improved.

The application discloses a kind of based on deep learning's extremely weak light environment under moving target scattering imaging method and device, it is related to extremely weak light environment under moving target imaging field, for the speckle image obtained by detector, establish neural network to obtain position information and original image recovery information.The accurate position information of target is obtained using prior information and recovered speckle to carry out trajectory recovery.N fixed positions in the detection range are placed target, and the sampling graph of fixed position for training is obtained.N fixed positions of sample speckle are used to train neural network, and after training, fixed parameter becomes the neural network for low-light imaging position classification.Each fixed position of sample speckle is used to train neural network respectively, and after training, fixed parameter becomes the neural network for low-light imaging reconstruction in fixed position.The application uses multiple position neural networks to recover speckle together, obtains more environmental information, and is conducive to generating the accuracy of detection imaging under the complex motion state of target.

The invention discloses a working system S3-S5 motor temperature rise test method and device capable of shortening the interruption period, and relates to the technical field of motor test. A long-term problem is split into a plurality of independent short-term problems by the PPINN model, and the short-term problems are supervised by a rapid coarse-grained (CG) solver, so that the high calculation efficiency of training the neural network by using small data sets can be utilized to the maximum extent, the training time can be shortened, a big data set can be decomposed into a plurality of small data sets for training by the PPINN model, and the training efficiency is improved. Therefore, under the condition of not enough GPU resources, the PINN can still be trained by using a big data set on the CPU resources, data acquired by the equivalent temperature rise test is used for prediction model training and prediction of equivalent motor temperature rise test time and a temperature rise value, only one time of measurement is needed for one time of temperature rise test, and compared with a traditional method, the experiment time can be greatly shortened, and the test efficiency is improved. The test efficiency is improved, and the labor intensity and cost are reduced.

The application relates to a robot autonomous navigation method based on local information, which comprises the following steps: step 1, a virtual navigation platform is built, an autonomous navigation model is designed, and training data of a virtual navigation process are collected; step 2, based on the training data, a deep reinforcement learning algorithm is used to train the autonomous navigation model; step 3, the trained autonomous navigation model is derived and encapsulated; and step 4, the encapsulated autonomous navigation model is applied to an actual navigation task. Compared with the prior art, the application has the advantages of high scene applicability and low calculation complexity.

The invention provides a data distribution method, device and equipment and a computer readable storage medium. The method comprises the following steps: acquiring a plurality of resource clusters and the unit computing power of an arithmetic unit included in each resource cluster; according to the unit computing power of each computing unit, combining each computing unit to obtain a plurality of combined clusters, and determining the cluster computing power of each combined cluster; obtaining multiple pieces of training data; and based on the cluster computing power of each combined cluster, performing grouping processing on the plurality of pieces of training data to obtain a training data set corresponding to each combined cluster, and distributing the training data set to the corresponding combined cluster so as to train a model by using each combined cluster through the corresponding training data set. According to the method and the device, the computing power of different computing units in a plurality of resource clusters can be activated to support the training of the model, so that the training time can be shortened, and the training efficiency can be improved.

The present disclosure relates to a method, device, readable medium and electronic device for processing a computational graph. The method for processing a computational graph obtains a target computational graph corresponding to a machine learning model to be trained, performs node fusion processing on the target computational graph to obtain a fused computational graph including multiple subgraphs; according to multiple preset segmentation methods of each operator node, effectively segments the operator nodes in the fused computational graph to obtain multiple segmentation data flow graphs corresponding to the fused computational graph, and the segmentation data flow graph is a target computational graph with an effective segmentation method; then, a parallel configuration strategy corresponding to the target computational graph is searched according to the segmentation data flow graph, so that a model training strategy that is a hybrid of pipeline parallelism, data parallelism and model parallelism can be automatically and efficiently generated, which can effectively improve the model training efficiency of the machine learning model and shorten the model training time.

The invention discloses a multi-feature charging load prediction method and system, and the method comprises the steps: obtaining a historical load sample data set, and carrying out the preprocessing of the historical load sample data set, and obtaining a complete data set; optimizing the complete data set by using a first optimization algorithm and a second optimization algorithm to obtain training parameters; and constructing a prediction model by using the training parameters, predicting the load according to the prediction model to obtain a prediction result, and performing data analysis on the prediction result, thereby improving the prediction accuracy and stability, and providing technical support for accurate prediction of the charging load.

The invention provides a point cloud data sparse convolution training method and system. The method comprises the steps of obtaining to-be-trained point cloud data; performing preprocessing operation on the point cloud data to obtain a target convolutional layer; disassembling the target convolutional layer into a convolutional layer group; the convolutional layer group comprises a plurality of convolutional layers of which the size is smaller than that of the target convolutional layer; extracting features of the convolutional layer group; performing normalization processing on the features to obtain numerical values and threshold values corresponding to the features; screening the numerical values by using the threshold value to obtain a target numerical value; and obtaining a target feature based on the target numerical value, so as to solve the problems that when feature extraction is performed on the point cloud data at present, sparsity of the point cloud data disappears, so that the calculation amount of the point cloud data is very huge, the training duration and size of a model are greatly increased, and the risk of overfitting is improved.

The invention provides a microwave device large-range modeling method and system based on an adjoint neural network, and the method comprises the steps: obtaining a range of a plurality of device parameters of a microwave device, dividing the range of the device parameters into a plurality of sub-intervals with overlapping intervals, building a sub-model based on the plurality of sub-intervals of the plurality of device parameters, and carrying out the modeling of the plurality of device parameters. Each sub-model is correspondingly provided with an adjoint neural network; pre-training each sub-model based on the training data, and performing auxiliary training through the adjoint neural network corresponding to each sub-model; constructing a sigmoid function corresponding to each overlapping interval for the upper boundary and the lower boundary of each overlapping interval; determining a correction function of each sub-model based on the sub-interval of each device parameter corresponding to each sub-model; and constructing all the sub-models provided with the correction functions into a large-range model, outputting electromagnetic frequency domain characteristic values corresponding to the to-be-calculated device parameter group based on the large-range model, and outputting derivative information corresponding to the device parameter group through an adjoint neural network.

The application discloses a three-dimensional model light ray tracing method based on meta learning, and belongs to the technical field of three-dimensional model image processing, and comprises the following steps: step one, constructing a general model for light ray tracing; step two, training a previous model by using a stereo matching convolutional neural network; step three, jointly training the previous model to obtain a new model; and step four, finding customized parameters and a customized model of the new model meeting a specific light ray tracing effect.The application trains an old model by using a stereo matching convolutional neural network based on the construction of a general model, trains general parameters meeting a specific light tracing effect, obtains new potential general parameters through the convergence of a loss function, adapts to a specific light tracing effect, effectively trains the old model by removing 90% of previous data and 98% of previous data, reduces the training data capacity, improves the accuracy of new and old data, and thus solves the problems of long training time, large training data capacity and low light ray tracing effect accuracy of the light ray tracing technology.

The invention relates to the technical field of data sharing, in particular to a federated learning-oriented multi-party authorization system, which comprises a model aggregation module for processing a local model of a participant to output a federated model; the verification module is connected with the model aggregation module and is used for verifying the use permission of the user according to the authorization certificate of the participant; the control module is respectively connected with the model aggregation module and the verification module, and is used for executing the verification module when judging that the operation safety of the multi-party authorization system does not meet the requirement according to the byte quantity proportion of the authorization certificate of the user; and adjusting the transmission rate of the authorization certificate or determining the running memory proportion of the secret share generator according to the average generation number of the secret shares, and adjusting the training period of the federal model for the first time according to the byte difference quantity before and after decoding the authorization certificate. According to the invention, the operation security of the multi-party authorization system is improved.

The invention relates to a data training optimization method and system, and the method comprises the steps: obtaining the to-be-trained data size of a to-be-trained task and the node processing amount of each idle node through a data management layer, calculating the total amount of to-be-loaded data at the time according to the node processing amount, and carrying out the calculation of the total amount of to-be-loaded data; the data to be trained is loaded and distributed in batches according to the total amount of the current data to be loaded, idle nodes store training results generated by data training in a data management layer, and when the data management layer distributes the total amount of the current data to be loaded next time, if it is judged that idle nodes needing training results exist, the data to be loaded is distributed to the data management layer. If yes, the training results are distributed together when the total amount of the data to be loaded is distributed. Therefore, the to-be-trained tasks are uniformly managed through the data management layer, the to-be-trained data volume is recorded and distributed in batches, repeated data loading and distribution are avoided, the loading and distribution efficiency is improved, training results needed by idle nodes are distributed in advance, the transmission time is shortened, and the training efficiency is improved.

The application discloses a panoramic camera auxiliary-based GNSS non-line-of-sight signal identification and application method, constructs a light MN-DLV3-CRF model to optimize the boundary information, obtains a clear segmentation boundary, and constructs a sky map with shielding information; the sky map is divided into at least a non-shielding sky area, a building shielding area and a tree shielding area according to the segmentation boundary; when a satellite signal falls into the building shielding area and the tree shielding area, the satellite signal is determined as an NLOS signal; if the satellite signal falls into the tree shielding area, the NLOS weight of the tree is calculated by considering the NLOS weight model of the tree. The method generates a shielding sky map by semantic segmentation of a sky image, assists a GNSS receiver in identifying NLOS signals shielded by buildings, trees and the like, improves positioning precision in complex scenes such as urban canyons and shaded roads, and is suitable for high-precision positioning scenes such as vehicle navigation, unmanned aerial vehicle positioning and intelligent transportation systems.

Embodiments of the present application provide a metasurface unit spectrum prediction method, device and equipment and a storage medium, belonging to the technical field of metasurfaces. The method comprises: obtaining at least two target metasurface unit structure image sets; wherein the number of sampling points of images of different image sets is different; training a preset original spectrum recognition model based on each image set in turn according to the order of the number of sampling points from small to large to obtain a target spectrum recognition model; obtaining a to-be-predicted metasurface unit structure image, inputting the to-be-predicted metasurface unit structure image into the target spectrum recognition model for prediction, and outputting a target transmission coefficient; and performing inversion calculation on the target transmission coefficient to obtain an amplitude spectrum and a phase spectrum of the to-be-predicted metasurface unit structure image. By using the progressive training mode, the number of training images is reduced, the training time is shortened, and the target spectrum recognition model trained has a short prediction time and a high prediction result accuracy.