53 results about "Batch training" patented technology

This invention provides a method and system for predicting the state of power equipment based on online testing adaptation, comprising: collecting power equipment state data in real time through sensors and forming test samples; filtering a set of adapted historical samples from a historical sample memory bank that meet preset conditions in terms of similarity to the test samples in the latent space through a transferable historical sample selection module, wherein the historical sample memory bank stores historical power equipment state data; performing time-frequency domain hybrid data augmentation on the test samples and the adapted historical sample set through a transferable online augmentation module to generate an augmented sample set; inputting the augmented sample set into a pre-trained power equipment state prediction model for batch training, dynamically adjusting the model parameters to adapt to the distribution shift; and fusing the output of the dual-stream predictor of the power equipment state prediction model to generate the power equipment state prediction result for the next time period. This invention can perform power equipment state prediction.

The invention discloses a semi-supervised domain self-adaptive deep forgery detection method, relates to the technical field of forgery detection, and solves the technical problem that a model cannot fully adapt to data distribution of a new domain due to the fact that a small amount of annotated data and a large amount of unannotated data in a target domain are generally difficult to use at the same time in an existing method. The method comprises the following steps: performing multi-batch training on a deep neural network model by adopting samples, wherein each batch of samples comprise a source domain labeled sample, a target domain labeled sample and a target domain unlabeled sample; in the training process, a joint loss function is optimized by adopting stochastic gradient descent, network parameters are iteratively trained, any input image is detected by using a trained Xception model, and the probability that the image is real or forged is output; according to the method, a semi-supervised field adaptive framework is introduced, so that the feature distribution difference between a source domain and a target domain is effectively reduced, and the cross-domain generalization ability is remarkably enhanced.

The application discloses a model training data construction method and device, which can be applied to various scenes such as cloud technology, artificial intelligence, intelligent transportation and Internet of Vehicles. The method comprises the following steps: obtaining a set of original object images and a set of object wearing images corresponding to a virtual wearing object in a virtual object display platform; extracting object key point information corresponding to each object wearing image in the set of object wearing images and a wearing object category corresponding to each object wearing image; determining an object wearing image with a matching result representing a successful matching result in the set of object wearing images as a preliminary screening wearing image; screening an image meeting a preset condition from the preliminary screening object wearing image to obtain a target wearing image; and constructing a training data set of the virtual wearing object based on the set of original object images and the target wearing image. The application realizes rapid and accurate construction of a batch training data set of the virtual wearing object.

This paper presents a graph neural network node classification method that integrates meta-learning and mini-batch training, belonging to the field of information technology. First, the method utilizes the METIS algorithm to divide the original large-scale graph data into multiple non-overlapping connected subgraphs. Then, by constructing a hybrid selection mechanism based on label coverage and label entropy, subgraphs with high information content and strong representativeness are selected from the subgraph pool as the meta-learning task. Subsequently, iterative training is performed on the selected subgraphs using the meta-learning framework to capture the general prior features of the graph structure, thereby obtaining a set of initial parameters for the model with rapid adaptability. Finally, these optimized initial parameters are transferred to the mini-batch training stage on the full dataset, guiding the model to achieve rapid convergence through high-quality initialization. On large-scale benchmark datasets, this method significantly reduces the number of iterations required by the model while maintaining the same classification accuracy as current mainstream graph neural network models, thus greatly shortening the overall training time.

The invention provides a model training method and device, electronic equipment and a storage medium, and relates to the technical field of artificial intelligence system performance optimization and deep learning framework scheduling. The method comprises the following steps: in a current batch training process of a target model, acquiring execution information of a corresponding bottom kernel when a processor executes a plurality of computational operators of the target model; according to the execution information, determining a global key path, related to the total calculation duration, of the target model in the current batch training process, and identifying target operators forming training iteration delay in calculation operators contained in the global key path; and generating a delay optimization instruction for the target operator, and in response to the delay optimization instruction, adjusting training configuration parameters corresponding to the target operator in an upper-layer training framework of the processor so as to apply the adjusted training configuration parameters in a subsequent batch training process of the target model. According to the scheme, the training performance of model training can be improved.

Systems, methods, and computer-readable media provide a graph processing system that incorporates a graph neural network (GNN) based recommender system (RS), as well as a method for training a GNN based RS to address feature leakage that leads to overfitting of the trained GNN based RS. A message correction algorithm is used to modify a user node embedding and a positive item node embedding generated by the graph neural network when generating mini batches of training triples used to train the GNN based RS. The GNN message passing operations are performed on one graph only, in contrast to existing approaches which typically run GNN message passing operations on multiple adjusted input graphs constructed for multiple training triples.

The invention discloses a long-tail radiation source individual identification method and device based on field generalization and a storage medium. The core of the method is that a plurality of types of source domain data sets distributed in a long-tail mode are obtained; sampling to obtain batch training data; after extracting feature representation by using a feature extractor, inputting the feature representation into a classifier and a domain discriminator at the same time; based on a classification prediction result, combining a field prediction result of a gradient inversion mechanism and class balance loss to construct an integrated overall objective function; all model parameters are jointly optimized by minimizing the target function, and an optimized recognition model is obtained and is finally used for performing high-precision individual recognition on target domain radiation source signals. Through collaborative optimization of classification precision, category balance and domain generalization ability, the problem of insufficient recognition performance of the model on an unknown target domain in a complex scene with training data long tail distribution and domain offset is effectively solved.

This invention relates to the field of fluid mechanics and transient process analysis of hydraulic systems, and discloses a method for determining the occurrence boundary and risk level of liquid column separation. The method utilizes a visual test bench to conduct multiple experiments under different combinations of test parameters, collecting data and simultaneously recording the liquid column separation state. Based on Buckingham's π theorem, dimensional analysis is used to construct multiple independent dimensionless parameter sets from the test parameter combinations. These are used as input features, with the separation state as the target variable. A machine learning algorithm is introduced for multiple batch training to calculate importance scores, selecting the dominant dimensionless parameter with the highest and most stable score. Subsequently, regression analysis is used to fit the data boundary, solving for the intersection point and outputting the critical equilibrium point as a quantitative benchmark for the occurrence boundary. After confirming separation, high-scoring dimensionless parameters are selected to construct composite judgment parameters, which are compared with preset thresholds to classify the risk level. This invention can accurately define the critical equilibrium point of cavitation occurrence and quantitatively assess its risk severity.

The invention relates to a multi-teacher mixed distillation method and system and computing equipment, and the method comprises the following steps: evaluating and screening teacher models, and obtaining pre-training models with cross-domain migration potential to form a teacher set; selecting a lightweight neural network as a student model, and configuring an independent MLP mapping layer for each teacher model in the teacher set to adapt to an output dimension; sampling is carried out in a pre-training data set of each teacher model, and mixed batch training data containing multi-field samples is constructed; and inputting the mixed batch training data into the student model, outputting a prediction result through an MLP mapping layer, and constructing a loss function in combination with teacher model output to complete distillation training. The student model obtained by the invention is no longer limited to a certain specific scientific field or a certain data feature, but becomes a relatively general scientific time sequence signal analysis model, and shows excellent performance in multiple fields.

The invention provides a virtual training method, system and device, electronic equipment and a storage medium. According to the method, the acquired training instruction is responded, the virtual scene corresponding to the training instruction is loaded, the operation instruction is acquired through the input device, the input device can be a mouse or a keyboard or other devices, VR equipment or other high-cost devices are not needed in the training process, and batch training can be achieved; obtaining an operation instruction for any second entity virtual object; wherein the operation instruction comprises an operation position relation between the second entity virtual object and the first entity virtual object; and generating a training report according to the position relationship corresponding to each second entity virtual object and the sequence of obtaining the operation instruction of each second entity virtual object. According to the virtual training party disclosed by the invention, not only is the result monitored, but also the assembly process is finely monitored, so that the training efficiency is improved to a certain extent.

The application discloses a differential privacy model training method based on a buffer mechanism, a medium and a system, wherein the method comprises the following steps: sampling a training data set to obtain small-batch training data; calculating a gradient value and processing the gradient value to generate a private gradient, which is applied to a current model parameter to obtain a preselected weight; sampling a verification data set to obtain small-batch verification data; calculating a loss change value, and performing clipping and noise adding processing on the loss change value to generate a noisy loss change value; judging whether the noisy loss change value is less than a preset rejection threshold; if yes, adding the candidate weight to the buffer; when the number of candidate weights is equal to a preset number threshold, determining the optimal weight according to the relative noisy loss change value, and updating the current model parameter based on the optimal weight; the method can effectively protect privacy, improve the updating quality of the model in the training process, and improve the accuracy of the final prediction result.

The application discloses a general domain self-adaptive image classification model implementation method and system, wherein the method comprises the following steps: setting a source domain image dataset with classification labels and a target domain image dataset different from the source domain data distribution and without classification labels as image data for classification model training; training closed set and open set classifiers in the source domain classification model, outputting classification probabilities belonging to each category of the source domain through the closed set classifier, outputting in-class and out-of-class classification probabilities and maximum open set entropy through the open set classifier; determining the most confused categories of the target domain from the maximum open set entropy and setting a transition zone, separating simple samples outside the transition zone with a self-confidence higher than a preset first threshold value and difficult samples inside the transition zone with a self-confidence lower than a preset second threshold value; and performing batch training on the classification model by combining a cross-entropy loss, a selection optimization strategy loss and a neighbor aggregation strategy loss function, so as to obtain a final general domain self-adaptive image classification model.

The invention relates to the technical field of artificial intelligence, and particularly provides a forklift tray tracking method, system and device based on a lightweight model and a medium, and the method comprises the steps: constructing a rotating target data set of a standard tray, and carrying out the targeted data enhancement and preprocessing; secondly, performing three-point improvement on a YOLOv12 algorithm: introducing a brand new StarNet backbone network, constructing a high-dimensional implicit feature space through star operation, and improving feature expression capability while reducing parameter quantity; a dynamic hybrid convolution module is designed in the neck network, multi-scale features are adaptively extracted by using multi-branch deep convolution and a dynamic weight fusion mechanism, and the flexibility of the model is enhanced; and a lightweight rotation detection head is provided, and the rotation angle of the tray is efficiently predicted while the parameter quantity is reduced and the small-batch training stability is improved through application group normalization and a shared convolution structure. And finally, combining the improved detection model with a ByteTrack tracking algorithm of the optimized adaptive rotating frame to form a complete identification tracking system.

The application provides a pipeline parallel distributed training method, device and system for a deep neural network. The method comprises: sequentially transmitting each group of small batch training data constituting a current batch of training data to an edge server via an optical network, so that the edge server and the cloud server cooperatively perform asynchronous parallel cooperative training on different sub-task models constituting the deep neural network in a communication and training decoupling manner, and the edge server sequentially outputs the gradients corresponding to each group of small batch training data; and sequentially receiving the gradients of each group of small batch training data. The application can ensure correct transmission of data during model training, reduce communication overhead during training, achieve load balancing between devices, improve model training efficiency, effectiveness and resource utilization of devices participating in training.

The invention relates to the technical field of financial risk management, and discloses a financial risk intelligent management method based on deep learning. The method comprises the following steps: acquiring multi-dimensional financial data such as market transaction, user credit, asset mobility and macroeconomic indicators, and optimizing a data acquisition process; extracting features by using a deep belief network to obtain optimized feature data; and dividing batches according to the optimized feature data, analyzing a risk trend through small-batch training of a deep neural network, and obtaining a risk assessment index. And based on the index, adjusting parameters and weights of the graph neural network by using a Bayesian optimization algorithm, and generating an optimization model. Receiving a real-time financial data flow and an optimization model, searching and deducing a risk path through a Monte Carlo tree, and outputting a risk control decision signal; a near-end gradient method and a branch and bound method are combined to optimize asset configuration to realize risk hedging; and finally, tracking the state after risk control by using particle filtering, and feeding back to a relevant link adjustment strategy.

The invention discloses a reinforcement learning training method and device of a model, electronic equipment and a storage medium, and relates to the technical field of artificial intelligence such as machine learning. According to the specific implementation scheme, a batch of training data sets are obtained, wherein the training data sets comprise multiple pieces of training data; based on a pre-configured hyper-parameter, segmenting the training data set to obtain a plurality of micro-batch training data groups; carrying out asynchronous parallel processing on the generation tasks and the reward calculation tasks of the plurality of training data groups; and performing parameter adjustment on the model based on task processing results of the plurality of training data groups.

Systems and methods are disclosed that perform efficient training of a graph neural network (GNN) using graph structure-aware randomized mini-batching. For example, nodes from a graph may be obtained. Subsequently, the nodes of the graph may be grouped into communities and then the order of the communities as well as the nodes within each of the communities may be shuffled. Based on shuffling the order of the communities and the nodes within the communities, mini-batches for training the GNN may be determined. Following, based on a sampling bias, a sub-graph may be constructed for each of the mini-batches to obtain a plurality of sub-graphs. The sampling bias may indicate a bias for sampling intra-connections instead of inter-connections. After, the GNN may be trained based on the constructed sub-graphs.

The application discloses a kind of multimodal data distributed training method and system, it is related to robot field, including: each configuration data pool is equally divided into N shards, and the shard with same serial number in different configuration data pool is distributed to the same computing node;Each computing node loads corresponding shard data subset according to the shard serial number distributed;Each computing node independently executes modal decision;According to modal decision result, corresponding data loader is selected to load homogeneous batch training data of corresponding mode from the data subset of the node;Each computing node sequentially executes forward propagation and back propagation to homogeneous batch training data, and obtains the gradient of each node;Global gradient fusion is carried out to the gradient of all computing nodes, and global gradient is obtained;All computing nodes update model parameters using global gradient.The application is physically isolated from the forward interference and Loss scale conflict between different modal data by shard scheduling and isolated training.

The application relates to the technical fields of machine learning, artificial intelligence and medical health, and proposes a pre-training classification model training method, device, equipment and medium, which comprises the following steps: acquiring a batch training sample set; determining a target label category contained in the batch training sample set; expanding the batch training sample set according to the target label category and an initial label embedding layer corresponding to the target label category, to obtain an expanded batch training sample set; and training a pre-training classification model and the label embedding layer by using a contrast learning loss function according to the expanded batch training sample set, to obtain a trained target classification model and a target label embedding layer. Through the technical scheme, the model training process is simplified, overfitting is avoided, and the generalization effect is improved.

The invention relates to a color difference recovery model training method and device. The method comprises the steps of performing multi-batch training on a color difference recovery model, and training the color difference recovery model through joint training data in the training process of each batch to obtain a model output image; obtaining an updated color difference recovery model corresponding to the current training batch based on the joint loss function, the model output image and the joint training data; under the condition that a preset training completion condition is not met currently, processing the joint loss function based on a loss function weight corresponding to the current training batch to obtain a processed joint loss function; and a trained color difference recovery model is obtained. Through adoption of the method, training is performed under the condition of considering physical priori knowledge, and a light-weight chromatic aberration recovery model with relatively high generalization ability, relatively high chromatic aberration recovery accuracy and relatively good image quality can be obtained.

The invention discloses a training method of a knock signal defect recognition model, relates to the technical field of intelligent nondestructive testing and signal mode recognition, and can at least partially solve the problem of weak minority class recognition capability caused by strong subjectivity of manual interpretation, insufficient robustness of manual features and unbalanced class samples in the prior art. The method comprises the steps that knocking signal samples are acquired and subjected to category labeling, and a sample set is constructed; event detection, peak value alignment, fixed-length segmentation and preprocessing are carried out on the knocking signal sample; dividing a training verification test set and constructing small-batch training data streams; inputting the preprocessed sample into a one-dimensional deep convolutional neural network for multi-layer feature extraction to obtain a feature vector; inputting the feature vectors into a classification head and outputting probability distribution through a Softmax classifier; a cross entropy loss function is combined with an optimization algorithm for training until convergence; and outputting the deployable model. According to the invention, end-to-end automatic feature learning and multi-class defect discrimination are realized.

The application discloses a kind of training methods of knock signal defect identification model, it is related to intelligent nondestructive testing and signal mode identification field, it can at least partially solve the problems that the subjectivity of artificial interpretation is strong, manual feature robustness is insufficient in prior art, class sample imbalance leads to the weak recognition ability of minority class.The present application comprises: obtaining knock signal sample and carrying out class labeling, construct sample set;Knock signal sample is detected, peak alignment, fixed-length segmentation and pre-processing are carried out;Divide training verification test set and construct small batch training data stream;The pre-processed sample is input into one-dimensional deep convolutional neural network to extract multi-layer features, and a feature vector is obtained;The feature vector is input into the classification head and the probability distribution is output by the Softmax classifier;Cross-entropy loss function is used in combination with optimization algorithm for training until convergence;Output deployable model.The present application realizes end-to-end automatic feature learning and multi-class defect discrimination.

A deep reinforcement learning optimization compensation method for nonlinear intermittent processes, which expands a three-dimensional input data matrix into a two-dimensional matrix; performs standardization processing; constructs a JY-KPLS model; constructs a JY-KPLS model optimization problem; solves the optimization problem; calculates the similarity of historical and query data; according to the similarity, m old data and n new data are selected from the old and new process data sets respectively, and the difference is calculated with the current query data; the deviation sample is used as the data set to establish a JITL-JYKPLS local model to solve the mismatch problem; the compensated model and the optimization system are interacted and trial-and-error trained; if the total reward value of the current batch training exceeds the total reward value of the previous batch training, the optimized system of the current training is used for batch-to-batch optimization; otherwise, the optimization system of the previous batch is used for batch-to-batch optimization; the final product quality is output. The method can significantly improve the quality of the final product.

The invention discloses a preference reinforcement learning robustness improvement method based on a dynamic denoising classifier, and relates to the technical field of artificial intelligence and machine learning. The method comprises the following steps: constructing a dynamic denoising reward model, wherein the dynamic denoising reward model comprises a basic reward predictor and a dynamic denoising classifier; sampling training data from the preference data set, and inputting the training data into the basic reward predictor and the dynamic denoising classifier for cooperative training; and taking a basic reward predictor in the trained dynamic denoising reward model as a reward function, and integrating the reward function into a reinforcement learning algorithm to optimize a strategy network of the intelligent agent. According to the method, a double-branch network model containing a basic reward predictor and a dynamic denoising classifier is innovatively constructed, dynamic recognition and suppression of noise labels are realized through a cooperative training mechanism, and a multi-stage dynamic threshold adjustment strategy is adopted, and adaptive learning rate scheduling and progressive batch training are combined, so that the dynamic recognition and suppression of the noise labels are realized. And the immunity of the model to noise data is effectively improved.

The invention discloses a human body posture estimation method based on selective boundary aggregation-star convolution-grouping normalization (SBA-StarC-GN). According to the method, on the basis of a Hyper-YOLO-Pose architecture, a selective boundary aggregation module (SBA), a MANet-StarC module and a GN-Pose Head based on grouping normalization are integrated, so that the problems of complex background, shielding and inaccurate small-scale joint point detection are solved. The SBA module adaptively fuses high and low-level semantics and spatial features through a bidirectional attention mechanism; the MANet-StarC module is combined with a star-shaped convolution structure and context anchor point attention (CAA) to enhance long-range dependence modeling and suppress background interference; according to the GN-Pose Head, grouping normalization is adopted to replace batch normalization, the influence of small batch training on precision is reduced, and the calculation overhead is reduced. According to the method, the mAP50 reaches 85.3% and the mAP50: 95 reaches 47.7% on an MPII human body posture data set, so that the method is obviously superior to an existing YOLO series model, and is suitable for a real-time and lightweight human body posture estimation application scene.

This invention proposes an optimization method for parallel deep convolutional neural networks based on Winograd convolution, comprising: S1, the model batch training stage, employing the feature filtering strategy FF-CSNMI based on cosine similarity and normalized mutual information, which eliminates redundant feature computation by filtering and then fusing, thus solving the problem of excessive redundant feature computation; S2, the parallel parameter update stage, employing the parallel Winograd convolution strategy MR-PWC, which reduces the computational cost of convolution in big data environments by using parallelized Winograd convolution, thereby improving the performance of convolution operations and solving the problem of insufficient convolution operation performance in big data environments; S3, the parameter combination stage, employing the task migration-based load balancing strategy LB-TM, which reduces the average response time of each node in the parallel system by balancing the load among nodes, improving the efficiency of parallel parameter merging, thus solving the problem of low efficiency in parallel parameter merging. This invention significantly improves both parallel efficiency and classification performance.

The invention discloses a CNN-BIGRU photovoltaic power generation prediction method and device based on clustering-decomposition. The method comprises the following steps: collecting original data; preprocessing the original data; inputting a density peak clustering algorithm, and automatically identifying a clustering center by calculating a local density rho and a relative distance delta to obtain a plurality of representative weather scene clusters; for each scene cluster, extracting a corresponding power sequence; generating a high-frequency sub-sequence and a low-frequency sub-sequence; a shared CNN-BiGRU network architecture is constructed; training is carried out by adopting a small-batch training strategy of different scenes and different scales; rolling prediction is completed, and prediction results of the high-frequency subsequences and the low-frequency subsequences are reconstructed and synthesized into a final prediction value; and an output result is subjected to reverse normalization processing and then is used as a photovoltaic power prediction value. The technical problems that an existing photovoltaic power prediction method is low in prediction precision and poor in model robustness under the changeable weather condition are solved, and particularly the problem that errors are suddenly increased under the weather sudden change scenes such as cloudy and sunny-to-rainy weather is solved.

The application discloses a large model-based SysML state machine diagram formal requirement verification method, and belongs to the technical field of computer software development.The method is as follows: collecting SysML state machine diagram data sets, tracing corresponding requirement texts, and then translating and verifying the SysML state machine diagram and the requirement texts; processing the SysML state machine diagram and the requirement texts; setting a prompt template for a large model, and batch training two large-scale data sets; obtaining translation results of the SysML state machine diagram and the requirement texts, and representing the translation results in a language recognizable by NuSMV; making corresponding modifications on the obtained code; and inputting the obtained target code into NuSMV for formal verification.The application improves the verification efficiency, makes the verification method more universal, reduces cumbersome work when the formal verification method is applied in different fields, and can adapt to various requirement verification scenes.

The invention provides a user value prediction method based on multi-view feature expression, which comprises the following steps: acquiring user out-domain payment behavior data and app in-domain behavior data, and constructing a multi-cycle batch training sample; discretizing user data, inputting the discretized user data into a specific Embedding layer, and outputting dense representation of each feature; inputting the dense representation of the features into a multi-gate hybrid expert network module, and learning and outputting multi-view feature representation adapted to each period; mapping the feature representation of each period to output the customer lifelong value, and designing and adopting a zero-expansion logarithmic loss optimization task with cutting; and finally, designing sequence auxiliary loss combined with multi-cycle auxiliary optimization by utilizing the predicted lifetime value of the customer in each cycle. According to the method, the technical problems of insufficient embedding quality and low prediction precision when a traditional customer lifelong value CLTV prediction method is used for processing complex multi-distribution data are solved, the accuracy and robustness of customer lifelong value prediction are improved, and effective technical support is provided for customer value management in the fields of finance, e-commerce and the like.

The invention discloses a semantic-guided discriminant and robust unsupervised cross-modal hashing method, which comprises the following steps of: acquiring a multi-modal training data set of texts and images to obtain a pre-processed batch training data set; constructing and initializing a multi-modal feature extractor and a soft symbol hash layer; training to obtain a cross-modal Hash model based on the multi-modal feature extractor and the soft symbol Hash layer; and taking a text or an image as input based on the cross-modal hash model, and retrieving in an image or text pool to obtain a semantic matching result. According to the invention, based on native semantic information, image and text multi-modal feature extraction is realized by using a pre-training encoder and an attention mechanism, and the model is endowed with comprehensive extraction of image and text fine-grained features; through guidance and supervision of a self-adaptive robust prototype supervised learning mechanism, a semantic center is used as a category center of the Hash codes to carry out unsupervised classification, so that classification discrimination of the Hash codes and robust Hash representation learning under a pseudo tag are realized, and cross-modal efficient retrieval is realized.