48results about How to "Efficient training process" patented technology

The invention provides a deep learning large model training method and system for heterogeneous equipment, belongs to the technical field of model training, and aims to solve the problem that a heterogeneous GPU cluster cannot be effectively utilized when a traditional scheme is used for training a deep learning large model. Different network layers of a deep learning large model to be trained are divided into a plurality of stages, forward propagation and back propagation calculation of all network layers in each stage is executed by independent virtual equipment, and meanwhile, utilization of GPU resources of different structures is coordinated in combination with a proposed hybrid parallel training strategy, so that efficient model training is realized.

The invention provides a big language model training method and device, electronic equipment and a storage medium, and relates to the technical field of artificial intelligence. The training method of the large language model comprises the following steps: acquiring an instruction response data pair; wherein the instruction response data pair is composed of an instruction and a response generated by utilizing a large language model according to the instruction, and the response comprises a text of a preset field; extracting constraint information contained in the response from the response contained in the instruction response data pair; generating a training data set according to the response, constraint information contained in the response and an instruction corresponding to the response; and using the training data set to continuously train the pre-trained large language model to obtain a target large language model, and using the target large language model to generate a response according to the instruction. According to the method, the large language model capable of accurately executing the instruction with the complex constraint can be efficiently trained, so that the utilization rate of computing resources is effectively improved.

The invention relates to the technical field of medical image analysis, in particular to a training method and system for a cataract recognition model and a storage medium. The method comprises the following steps: acquiring cataract eye image samples and non-cataract eye image samples in different stages; lesion region segmentation and image synchronous simulation enhancement are carried out on the cataract eye image samples and the non-cataract eye image samples in different stages to obtain cataract lesion enhanced image samples and non-cataract enhanced image samples; extracting local detail features of the lesion and extracting structural features of a mask pair, and constructing a multi-task convolutional neural network learning framework to train a cataract recognition model and optimize focus weighted loss; and obtaining a cataract related key feature set and carrying out online incremental learning updating to generate a cataract recognition online incremental updating model. According to the invention, the accuracy and robustness of cataract recognition model training can be improved.

The invention relates to the technical field of radar weak target detection, in particular to a radar weak target detection method and device based on deep learning. According to the technical scheme, the method comprises the following steps: data preprocessing and enhancement: receiving an original echo signal of a radar, generating three-dimensional time frequency-wavelet coefficient spatial data by combining time frequency analysis with wavelet transform, and expanding a data set in multiple modes; and feature extraction network construction: constructing a 3D-CNN and LSTM fused network, wherein the 3D-CNN has a convolution operation with a variable convolution kernel size. According to the method, the data are richer and more comprehensive through time-frequency analysis in combination with wavelet transform and data enhancement, and multi-dimensional features can be accurately captured through 3D-CNN and LSTM fusion and an adaptive convolution kernel; and the training effect is ensured through the adaptive learning rate and the early stop method, and the accuracy and adaptability of weak target detection are jointly improved.

The invention relates to a continuous identity authentication method based on a multi-level memory enhanced auto-encoder, which belongs to the technical field of continuous authentication and comprises the following steps: S1, collecting user behavior data; s2, performing normalization processing on the user behavior data; s3, constructing a VAE-based data enhancement model, and performing training by using normalized data; s4, constructing a data reconstruction model based on MulMemAE, and performing training by using normalized data and VAE enhanced data; s5, inputting the user behavior data collected and normalized in real time into the trained data reconstruction model to reconstruct the user behavior data; s6, comparing the reconstructed user behavior data with the input user behavior data, and calculating a reconstruction error; s7, the reconstruction error is compared with a preset threshold value, and if the error is lower than the preset threshold value, the user is a legal user; on the contrary, if the user is a famous person, the user is required to log in again

The invention provides a multi-source software-defined science and education resource recommendation method and system based on a knowledge graph, and the method comprises the steps: constructing an artificial intelligence knowledge graph based on the relation between artificial intelligence knowledge entities; obtaining science and education resource data of a science and education institution, extracting knowledge entities in the science and education resource data, and connecting the knowledge entities to the artificial intelligence knowledge graph to obtain a science and education resource knowledge graph; extracting local resource features / user features based on a graph neural network to obtain local resource features and user features of each client; constructing a local knowledge graph based on the science and education resource knowledge graph, and performing model training by taking the local resource features of each client as initial features until the model converges; and inputting the users and the resources into the graph neural network for prediction, and recommending the resources with high scores to each user. According to the scheme, the relationship between knowledge entities in the field of artificial intelligence is effectively extracted, a structured knowledge graph is formed, and the recommendation precision is improved.

The invention discloses a large model training resource evaluation and strategy recommendation method, which comprises the following steps of: 1) constructing a large model database, and storing model state data of different types and sizes; the model database supports expansion and is used for synchronously updating the latest open source large model; 2) constructing a calculation overhead evaluation database of different distributed strategies; defining the video memory efficiency as an index for measuring the degree of storing redundant information on each video card; the calculation efficiency is defined as the proportion of the time for measuring the effective calculation of the graphics card, namely the time for updating the model state and the time for communication; 3) scanning existing computing power resources, traversing the strategy priority queue, and selecting a first hit strategy as a recommendation strategy; aiming at the current available computing resources, recommending a proper distributed training strategy, and giving a super parameter of training recommendation; and according to the selected large model type, estimating the training overhead, scanning node computing resources, giving a training strategy, automatically constructing a configuration file of a training framework, and cooperating with a training platform to realize efficient training of the large model.

The invention discloses a rehabilitation training data processing method and device, and relates to the technical field of medical information data processing, and the rehabilitation training data processing method comprises the following steps: 1, data acquisition and preprocessing: collecting and preprocessing physiological, log and video data of a to-be-rehabilitated user, and integrating the physiological, log and video data into a multi-source data set; 2, high-performance model construction and knowledge distillation: selecting and training a teacher network model, constructing a student network model, and obtaining a model with relatively good performance and high calculation efficiency through knowledge distillation; step 3, differential evolution algorithm optimization: defining differential evolution algorithm parameters, and obtaining an optimized student network model through iterative optimization; and step 4, deep reinforcement learning parameter adjustment: utilizing the optimized student network model to generate feature data, and adjusting rehabilitation training parameters through reinforcement learning. The data processing efficiency and accuracy are remarkably improved, and the complexity is reduced.

The invention provides an unsupervised learning method applied to a millimeter-wave radar sensing task, electronic equipment and a storage medium, and can be applied to the technical field of millimeter-wave radars. The method comprises the following steps: acquiring a radar echo signal by using a target antenna array combination, and performing signal processing on the radar echo signal to obtain a positive sample pair; performing feature extraction on the positive sample pair by using an encoder and a momentum encoder to obtain a positive sample feature pair, and projecting the positive sample feature pair to obtain a positive sample projection pair; processing the positive sample projection pair and the negative sample projection pair stored in the queue memory bank by using a predefined comparison loss function to obtain a comparison loss value; performing parameter updating on the encoder based on the comparison loss value, performing parameter updating on the momentum encoder based on the encoder after parameter updating, and storing the positive sample projection pair into a queue memory bank; and iteratively carrying out the above operations in an unsupervised manner until a preset training condition is satisfied, and obtaining a pre-trained encoder.

The invention relates to a tongue feature analysis method and device, electronic equipment and a storage medium. A trained tongue detection model and a trained tongue segmentation model are fused to obtain a target segmentation model, an image data set is segmented through the target segmentation model to obtain a tongue segmentation image data set, and the tongue segmentation image data set is labeled to obtain a tongue feature classification data set. Training a tongue feature classification model based on the tongue feature classification data set to obtain a trained tongue feature classification model, obtaining to-be-detected target image data, inputting the target image data into a target segmentation model to obtain tongue segmentation image data of the to-be-detected target image data, and obtaining the tongue segmentation image data of the to-be-detected target image data. According to the method, the trained tongue feature classification model is used for performing feature analysis on the tongue segmentation image data to obtain a tongue feature analysis result, so that the training efficiency of the model is improved, and the classification stability and reliability are also remarkably improved.

The invention belongs to the technical field of building cooling load prediction, and discloses an airport terminal cooling load prediction method and a related device, and the method comprises the steps: collecting main influence index real-time data affecting the terminal cooling load prediction; inputting the collected main influence index real-time data influencing the airport terminal cooling load prediction into the ISSA-GRNN-GRU hybrid prediction model, and outputting to obtain an airport terminal cooling load prediction result; wherein the construction process of the ISSA-GRNN-GRU hybrid prediction model comprises the following steps: acquiring influence constituent element data of terminal building cold load prediction, and training and combining a pre-constructed generalized regression neural network and a pre-constructed gating circulation unit neural network by adopting an improved sparrow search algorithm to obtain the ISSA-GRNN-GRU hybrid prediction model; the overall prediction error can be reduced, and the reliability of the prediction result is improved.

This invention discloses a D2D user resource allocation method and storage medium based on a deep reinforcement learning algorithm, relating to the field of wireless communication technology. The method includes: constructing a wireless network model and discretizing the D2D transmit power; constructing a user signal-to-noise ratio calculation model with maximizing the communication system throughput as the optimization objective; setting a prediction policy network π, a prediction value network Q, a target policy network π′, and a target value network Q'; modeling the D2D communication environment as a Markov decision process, treating the D2D transmitter as an agent, iteratively loading the parameters of the target policy network to generate a policy that interacts with the environment, determining the state space, action space, and reward function; using the MAAC algorithm to optimize the policy for each D2D user; using a soft update method to iteratively update the parameters of the target policy network and the target value network until training is complete; and having the D2D user download the trained parameters of the target policy network and improve their policy.

A robustness measurement method for LeNet-5 networks based on adversarial boundary constraints is presented, relating to the field of deep learning model testing. The main steps include: for each training sample, dynamically generating adversarial examples based on the model's current state during training iterations; constructing a composite loss function based on standard cross-entropy loss and dynamic boundary constraint loss; performing end-to-end training on all parameters of the LeNet-5 network; and using the overall approximate robustness boundary as a measure of model robustness after training. This method effectively improves the resistance of the LeNet-5 model to fast gradient sign-based adversarial attacks without altering the basic structure of the LeNet-5 network by designing a new loss function.

The invention relates to a method for AOI. The method comprises the steps that an image of a detected part of a first domain is obtained; inputting the image of the detected part into a detection model, wherein the detection model comprises a general feature extraction sub-model and at least one task detection sub-model related to a specific AOI task of the first domain; generating, by the general feature extraction sub-model, a first feature representation of the image based on the image of the detected component; at least one detection result related to a particular AOI task of the first domain is generated by the at least one task detection sub-model based on the first feature representation of the image.

This invention relates to the field of space mission planning technology, and provides a space multi-objective mission planning method that integrates reinforcement learning and curriculum learning. The method includes: constructing a space multi-objective mission simulation environment comprising an orbital dynamics model, an objective characteristic model, and a spacecraft constraint model; designing a progressive task sequence for agent training based on a curriculum learning method; and training the agent in the simulation environment using the progressive task sequence based on a deep reinforcement learning framework, updating network parameters by collecting experience data, and finally obtaining the optimal sequence decision-making strategy. This invention solves the technical problems of low decision quality, poor training efficiency, and weak policy adaptability caused by complex environments and sparse rewards in space multi-objective mission planning by guiding reinforcement learning through curriculum learning for stable and efficient training.

The application provides a training method, device and equipment of a reinforcement learning model and a robot control method, which are used to improve the walking stability of a robot. The training method of the reinforcement learning model comprises the following steps: in response to receiving a first speed instruction, determining a first body observation; inputting the first speed instruction and the first body observation into a to-be-trained encoder to obtain first auxiliary decision information; the first auxiliary decision information comprises a first estimated linear velocity and a first tensor; inputting the first auxiliary decision information into a to-be-trained policy network to obtain a first action, and updating a first simulation state based on the first action to obtain a second simulation state; inputting the first action and the first auxiliary decision information into a to-be-trained decoder to obtain a first predicted state; and updating first parameters of the to-be-trained encoder and second parameters of the to-be-trained policy network based on a difference between the second simulation state and the first predicted state to obtain a target reinforcement learning model comprising a target encoder and a target policy network.

This invention provides a clustered federated multi-task learning method and apparatus for the Internet of Things (IoT). By clustering IoT terminal devices, the data distribution within the same cluster becomes more approximate. A federated multi-task learning algorithm is executed within each cluster, with global training and personalized training tasks performed on each IoT terminal device. This achieves data sharing within the cluster while fully utilizing local data from each IoT terminal device for training on personalized tasks, thus efficiently utilizing local data and improving training effectiveness. During local training on each IoT terminal device, the number of training rounds is adjusted based on computing power, fully utilizing the computing resources of each IoT terminal device and improving model training efficiency. Regularization constraints applied to personalized training tasks using the global model effectively prevent overfitting, control the degree of personalization, and improve model quality.

According to the method for training the target model based on reinforcement learning, the target model is used for executing tampering detection on an input image, and the method comprises the steps that a training sample is obtained, the training sample comprises a target image and boundary information of a plurality of truth value boxes, and the truth value boxes are borders of a part of images marked in the target image and comprising tampering content; and inputting the target image into the target model, and outputting boundary information of the plurality of prediction frames. And calculating the overall overlap ratio of the plurality of prediction frames and the plurality of true value frames. Determining a first reward score based on a preset piecewise function according to a numerical range of the overall overlap ratio; the function corresponding to each numerical range of the piecewise function is a constant function, and the value of the corresponding constant function is increased along with the improvement of the overall overlap ratio. And based on a reinforcement learning algorithm, according to the first reward score, updating the target model.

An alloy performance prediction method fusing vision-language-process multi-modal data comprises the steps that a multi-modal data set of an alloy material is constructed, and the multi-modal data set comprises three kinds of modal input including process parameters such as temperature and time of solid solution and aging treatment, SEM image visual information and SEM image description text information and corresponding alloy mechanical property true values; training a visual encoder ResNet50 model through comparative learning and training a language encoder BERT model through mask language modeling to obtain an SEM image and vector codes of language description of the SEM image; splicing and fusing process parameters such as temperature and time of solid solution and aging treatment and the vector codes obtained in the second step, and training a random forest regression device according to corresponding alloy mechanical properties; and the random forest regression device obtained through training is used for alloy performance prediction. According to the method, the structured process data, the unstructured SEM image data and the derived text description data are subjected to collaborative fusion and joint modeling for the first time, complementarity among different modal data is fully utilized, and more comprehensive and more three-dimensional digital representation of the alloy state is constructed; the multi-modal fusion framework and the feature learning mechanism are suitable for wide material systems. Meanwhile, the adopted'depth representation + random forest 'hybrid model has the advantage of high training efficiency while ensuring high prediction precision.

The present application relates to the technical field of artificial intelligence and federated machine learning, and particularly relates to an asynchronous federated learning aggregation method and system based on clustering cache. The present application clusters clients by intermediate features of the clients. In global aggregation, intra-cluster aggregation is firstly performed, and then global aggregation is performed based on client clusters. The client clusters are dynamically updated. In intra-cluster aggregation, an active set and a slow set are introduced to realize an asynchronous participation mechanism of intra-cluster aggregation. Weighted aggregation of the active set and the slow set solves the problem of system heterogeneity. The present application solves the defect of insufficient timeliness of existing federated learning methods in a data heterogeneous scene, and guarantees the timeliness and model training accuracy of federated learning in the data heterogeneous scene.

The invention discloses a long text recognition method and device based on a Transform architecture, a medium and a product, and relates to the field of text recognition, the method comprises the following steps: constructing a long text recognition model based on the improved Transform architecture; the improved Transform architecture is characterized in that an attention module is replaced by a linear attention module; the execution process of the linear attention module is as follows: performing linear transformation on the current input of the current position; determining an output vector based on the memory state of the previous location and the current input; determining an input gate and a forgetting gate based on the memory state of the previous position and the current input, and updating the memory state; using the long text data set to train a long text recognition model based on a lexical weighting method; and lexical element recognition is carried out based on the trained long text recognition model. According to the method, the limited storage space can be utilized more efficiently, and the performance on the long text task is more effectively improved.

The invention discloses a large model training and reasoning-oriented containerized operation method and computing equipment, and the method comprises the steps: receiving a training starting request or reasoning starting request for a large model, and determining a training parameter corresponding to the training starting request or a reasoning parameter corresponding to the reasoning starting request; for the training starting request, creating a trainer instance and introducing training parameters, starting a training container through the trainer instance based on the training parameters, and running a training program through the training container so as to execute a large model training task based on the training program; and for the inference starting request, creating an inference device instance and transmitting inference parameters, starting an inference container based on the inference parameters through the inference device instance, and running an inference program through the inference container so as to execute a large model inference task based on the inference program. On the basis, training and reasoning tasks of the large model can be efficiently and conveniently executed in a containerized environment, and an embeddable containerized operation capability can be provided for a model development platform, so that other platforms can conveniently and rapidly integrate the development capability of the large model.

The invention provides a work memory ability training method based on electroencephalogram neural feedback, a storage medium and equipment, and the method comprises the steps: obtaining electroencephalogram signal data of a tested object in a process of executing a training task, and obtaining task execution result data of the tested object, the training task at least comprising a digital memory breadth task and a spatial memory breadth task; extracting physiological features of the electroencephalogram signal data, and calculating a cognitive state index and a global cognitive ability index according to the physiological features and the task execution result data; the training task is adjusted according to the global cognitive ability index, and the tested object is trained again according to the adjusted training task; and when it is judged that at least one of the training times, the global cognitive ability index and the cognitive state index meets a preset training condition, generating a training report at least including the cognitive state index of the tested object during each training. The computing resource consumption, the training delay and the deployment cost in the multi-dimensional training of the working memory ability are reduced.

The present invention discloses a continuum robot motion control method based on deep reinforcement learning, which applies the deep reinforcement learning DDPG algorithm to the continuum robot motion control, and trains the control strategy by making the robot continuously interact with the environment. The present invention realizes the autonomous motion control of the continuum robot by constructing a virtual training environment for the continuum robot, establishing a DDPG algorithm model, constructing a reward function calculation model, accumulating an experience replay pool, training a deep reinforcement learning network, and deploying the trained model to the actual robot system. The DDPG algorithm in the present invention adopts a deterministic strategy, has good network convergence, reduces excessive Q estimation values ​​by using a target network, and adds random action noise, which is conducive to exploring the optimal strategy; priority experience replay is used during training, which can speed up the training speed and improve the training effect; and through the domain randomization method, the deviation between the training environment and the real environment is reduced, and the success rate of model deployment is improved.

The invention discloses a cigarette carton missing detection method for a cigarette carton box. The method comprises the following steps: S1, collecting image data of the cigarette carton through a plurality of cameras when the carton is opened; s2, performing illumination adjustment on the acquired image data, and applying image enhancement processing to generate an enhanced image; s3, inputting the enhanced image into a target detection model, detecting whether the cigarette bar is missing, outputting a cigarette bar bounding box and confidence coefficient by the target detection model through extracting multi-scale features of the image and applying attention weight calculation, and judging a bar missing state according to a confidence coefficient threshold value; and S4, when it is detected that the cigarette bar is missing, triggering a data recording operation, recording related data of a missing event, including a timestamp, a box body number and an image fragment, and packaging, transmitting and storing the data. Through the steps of collecting images by multiple cameras, adjusting illumination, integrating a target detection model of an attention mechanism and automatically recording and transmitting data, the accuracy of detecting the carton lacking of the cigarette box of the cigarette piece is remarkably improved.

The invention relates to generating audio using autoregressive generative neural networks. Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for generating a prediction of an audio signal. One of the methods includes: receiving a request to generate an audio signal; obtaining a semantic representation of the audio signal; generating an acoustic representation of the audio signal conditioned by at least the semantic representation using one or more generative neural networks; and processing at least the acoustic representation using a decoder neural network to generate a prediction of the audio signal.

This invention discloses a method for rapid prediction of data center temperature fields based on convolutional neural networks, comprising: acquiring the geometric information of the data center structure; performing geometric modeling and mesh generation on the data center structure to obtain the mesh structure of the data center; acquiring the operating parameters of the equipment within the data center; acquiring the temperature field data of the data center and dividing the temperature field data according to a height plane to generate the output dataset for training the convolutional neural network model; converting the mesh structure of the data center at different heights into a spatial feature map structure and recording the corresponding height information and operating parameter information to generate the input dataset for training the convolutional neural network model; and using the trained convolutional neural network model for temperature field prediction. This method can achieve accurate and reliable rapid prediction of the temperature field of data center rooms.

The invention discloses a data expansion and exoskeleton joint end-to-end torque estimation method based on a diffusion model, and the method comprises the steps: carrying out the normalization processing of time series data from a multi-source sensor, carrying out the fragmentation according to a fixed length, and constructing a training sample with a motion class label; then, training a classifier-free conditional diffusion model by adopting a sample, and simultaneously learning conditional and unconditional denoising mapping relationships in a manner of randomly inactivating category conditions; in the generation stage, based on a classifier-free condition guidance mechanism, multi-modal time series data with specified motion category features are gradually generated from random noise; and finally, fusing the generated synthetic data with real acquired data to train a joint torque end-to-end prediction network, thereby realizing joint torque estimation of input sensor time sequence data. The method improves the prediction precision, generalization ability and stability of the end-to-end torque estimation model in a multi-action and few-sample scene, and has a good engineering application value.

A general aspect of the present disclosure relates to a method for training a machine learning module for validating or validating a closed loop test platform for a test system. The method includes providing a plurality of data sets from two or more closed-loop test platforms, wherein the closed-loop test platforms include at least one reference test platform. Each data set contains input data and associated output data of a system under test in a respective closed-loop test platform. The method further includes determining a similarity of a plurality of segments of the input data for a pair of the plurality of data sets. Each pair of test data sets includes a data set of a reference test platform. The method further includes determining a similarity of segments of the output data associated with the segments of the input data; and estimating input and output data of a data set of a closed-loop test platform to be validated or validated and a data set of a plurality of data sets of the at least one reference test platform by using a specific similarity in view of a function of the input and / or output data as the data set of the closed-loop test platform to be validated or validated The similarity aspects train a machine learning module to generate a machine learning module for validating or validating a closed loop test platform.

The invention provides a distributed edge machine room computing power integration method and system based on a DPU, and relates to the field of processors, and the method comprises the steps: transmitting parameter copy management data and consistency control function data to a DPU layer corresponding to a DPU card deployed in an edge machine room, and obtaining a target DPU layer; constructing a hardware hierarchical communication topology combining aggregation in the machine room and cross-machine room controlled transmission in the target DPU layer to obtain a target DPU, and executing bandwidth sensing, QoS scheduling, cross-machine room parameter coordination and traffic shaping control through the target DPU; the gradient processing and parameter synchronous control logic data corresponding to the edge machine room are transmitted to a target DPU, and hardware-level unloading and cross-machine-room intelligent scheduling are carried out through the target DPU; gPU resource virtualization and network virtualization control are carried out through the target DPU, so that the multiple edge machine rooms form a unified training resource pool logically, and a distributed edge machine room computing power integration result is obtained.