216 results about "Linear layer" patented technology

The invention relates to the technical field of model compression, and provides a large model compression method and device, a task processing method and equipment and a storage medium, and the large model compression method comprises the steps: carrying out the layer-by-layer quantification of a linear layer of a to-be-compressed initial large model, and obtaining a first large model; the initial large model is a pre-trained large language model constructed based on an expert hybrid architecture; performing route calibration on each expert sub-model in the first large model to obtain a second large model; in the reasoning process of the second large model, based on the task type of a to-be-executed target task, the importance of each expert sub-model in the task type is evaluated; and performing dynamic pruning on each expert sub-model based on importance so as to compress the second big model. Through a compression mode of combining static quantification and dynamic pruning, on the basis of ensuring the model performance, the memory and calculation overhead required by large model reasoning can be reduced, and efficient operation of the large model on light-weight equipment with limited video memory resources is facilitated.

The invention provides a model training method and device, equipment and a storage medium, and relates to the technical field of computers, in particular to the technical field of neural network models and model training. The specific implementation scheme is as follows: a calculation unit executes quantization matrix multiplication based on Hadamard pre-transformation on an activation tensor and a weight tensor of a target model stored in a memory so as to generate an output tensor of a linear layer based on a low-precision tensor with smaller data bit width; using the output tensor and a subsequent network layer of the target model to complete forward propagation so as to obtain a loss value; and according to the loss value, updating model parameters of the target model stored in a memory through a back propagation algorithm. By means of the technical scheme, on the premise that the model training precision is guaranteed, memory resource occupation and the calculation amount in the calculation process can be remarkably reduced, and the training cost is reduced.

The invention discloses a cpu + gpu-based tensor parallel computing algorithm for a large model, which comprises the following steps of: S1, combining row parallelism and column parallelism for a linear layer in the model by adopting a mixed-dimension tensor segmentation mode, and dynamically adjusting a segmentation proportion according to a model structure and hardware resources; s2, a CPU and GPU cooperative computing mechanism is constructed, part of tasks which have large video memory requirements and are relatively simple in computation are allocated to the CPU, the GPU is responsible for computing intensive tasks, communication between CPU-GPU is optimized, and overlapping of CPU-GPU communication and GPU computation is achieved; and S3, implementing a dynamic resource allocation and load balancing strategy, monitoring load conditions of the CPU and the GPU in real time, dynamically adjusting task allocation according to calculation requirements of different layers of the model and use conditions of hardware resources, and adopting a self-adaptive batch processing size adjustment strategy. The invention obviously reduces the occupation of the video memory, reduces the communication cost and improves the utilization rate of computing resources.

The invention discloses a multi-wave-potential pulse radar target detection method and device based on deep learning and a medium, and the method comprises the steps: obtaining radar original echo data in different detection scenes, and forming a radar original echo data set; training the target detection network by using the radar original echo data set to obtain a target detection model; in an input preprocessing layer of the target detection network, a real part and an imaginary part of radar original echo data are separated, linear layer networks with different weight matrixes are used for carrying out feature mapping on the real part and the imaginary part of different wave position data, the features of the real part and the imaginary part are calculated, then the phase spectrum features are calculated, and a multi-dimensional feature map is formed; training a target detection network by using the echo data set; and obtaining echo data to be detected, and inputting the echo data to the target detection model for detection. According to the method, high-precision and low-calculation-overhead target detection can be realized in a complex scene, and meanwhile, the detection performance and robustness are remarkably improved.

The invention relates to a model optimization method and device for a low-resource scene, electronic equipment and a medium, and the method comprises the steps: obtaining a pre-training model, the pre-training model being a Transform model, for any linear layer in the Transform model, connecting a pair of low-rank matrixes in parallel in the linear layer to form a LoRA increment path, and freezing a backbone network in the pre-training model; for any linear layer in the Transform model, a learnable gate is added on a LoRA increment path, and an improved LoRA module is formed; and obtaining training data, training the pre-training model according to the training data to obtain a target model, and performing semantic analysis on the to-be-processed text according to the target model. According to the method provided by the invention, through collaborative design of structural innovation and an optimization mechanism, the performance and stability of the efficient parameter fine tuning method in a low-resource scene are remarkably improved, and the method has a good industrial application prospect.

The invention relates to the technical field of satellite altimetry, in particular to an altimetry method based on a residual network multi-mode deep learning model, which can remarkably improve the robustness of SSH inversion under sea condition noise interference, and is characterized in that a fusion self-attention mechanism and a residual network multi-mode deep learning model ViTResNetMDL is established, and the model ViTResNetMDL is applied to improve the robustness of the SSH inversion under sea condition noise interference. A satellite-borne GNSS-R sea surface height inversion geometric physical model is established, a ViTResNetMDL model is composed of ResNet50, ViT and MHL-NN, the first part adopts a ResNet50 convolution structure to extract effective features from an original DDM, the second part adopts a ViT layer which captures global features in an effective scattering area DDM based on a self-attention mechanism in a Transform module, the third part is composed of linear layers, and the fourth part is composed of linear layers. And using the MHL-NN to retrieve the sea surface height.

The embodiment of the invention discloses a quantitative model degradation correction method and device, electronic equipment and a storage medium, relates to the technical field of data processing, and is convenient for reducing errors generated by a quantitative model. The method comprises the following steps: acquiring a key operator of an original model; implementing the key operator by adopting a model reasoning framework to obtain a model reasoning operator of a quantized version; registering the model reasoning operator into a deep learning training framework, and replacing a corresponding operator in the deep learning training framework with the model reasoning operator to obtain a quantitative model; injecting a fine-tuning low-rank weight into each linear layer of the quantitative model to obtain an injected quantitative model; performing fine tuning on the post-injection quantification model to obtain a fine tuning parameter result; and loading the fine tuning parameter result to the quantitative model. The method is used for degeneration correction of the quantitative model.

The invention discloses a model training method and device, equipment, a storage medium and a program product, and the method comprises the steps: carrying out the type conversion of an input tensor of a second data type based on a zooming parameter corresponding to a linear layer in a forward process, and obtaining an input tensor of a first data type; after the calculation of the linear layer is completed, converting the output tensor of the first data type into an output tensor of a second data type; in a reverse process, performing type conversion on the model parameter of the first data type based on a scaling parameter corresponding to a linear layer to obtain a model parameter of a second data type; determining a gradient of the second data type based on the model parameter of the second data type and the loss value of the second data type; and updating model parameters of the model by using the gradient of the second data type to obtain a trained model. Therefore, under the condition that the model precision is guaranteed, the model training efficiency is improved, and the requirements for computing power and storage during model training are reduced.

The invention discloses a dynamic scene re-operation mirror video generation method and system based on a diffusion model, and belongs to the technical field of computer vision and video generation. A diffusion generation architecture with a control branch is adopted, and the core is composed of an embedded layer, a main branch and a control branch. In the control branch, the output of each sub-block is added with the output of the corresponding block of the main branch after being processed by the zero initial linear layer, and the sum is input into the next block of the main branch. During training, generating a rendered video by using the target video and the reference video in the same scene; and inputting the target video latent variable after noise addition into the control branch, inputting the splicing result of the target video, the reference video and the rendering video latent variable into the main branch, and simultaneously providing the text latent variable of the reference video for the two branches as a condition. During generation, the model finally generates a target video latent variable through step-by-step denoising and decodes the target video latent variable into a target track video, and it is ensured that the motion of a moving object in a scene of a generated video and a reference video is consistent at the same time.

The invention belongs to a point cloud reconstruction method, and provides a power equipment three-dimensional point cloud reconstruction method and a related device for solving the technical problems that the accuracy of reconstructed and restored object form is low and key features are difficult to capture during detail restoration in two common curved surface reconstruction methods at present. And inputting into a three-dimensional point cloud reconstruction model of the power equipment to obtain a reconstruction result. Wherein the calculation method in the three-dimensional point cloud reconstruction model of the power equipment comprises local feature weighting, global feature fusion and occupancy prediction, the local feature weighting is based on the similarity of query point and neighborhood point features, the global feature fusion is based on the global feature weighting, and the occupancy prediction is realized through linear layer decoding. The detail reconstruction precision is effectively improved, and the accuracy of overall reconstruction is guaranteed.

The invention relates to an NDVI product generation method and device based on a diffusion model, equipment and a medium, and relates to the technical field of remote sensing data processing, and the method comprises the steps: obtaining a plurality of NDVI images corresponding to a target time period; each NDVI image is input into a pre-constructed diffusion generation model, and an NDVI product corresponding to the NDVI image is generated; wherein the diffusion generation model which is constructed in advance comprises a Stable Diffusion model and a LoRA fine tuning module; the Stable Diffusion model comprises a VAE module and a diffusion module constructed on the basis of a U-Net framework; and the LoRA fine tuning module is embedded in a Query linear layer, a Key linear layer and a Value linear layer in the attention structure. The method has the technical effect of effectively generating the high-time-resolution NDVI product by using the diffusion model.

The invention discloses a classification fault detection method based on a graph neural network, and the method comprises the steps: inputting time series data, projecting the input data to a set dimension through a linear layer, employing a two-layer one-dimensional convolutional neural network to extract time features, dynamically generating an adjacent matrix through a radius according to the Euclidean distance, and based on the adjacent matrix, in each layer of graph convolution, carrying out the classification fault detection of each layer of graph convolution. The method comprises the following steps of: performing spatial aggregation on adjacent node features, learning the weight of important nodes by using an attention mechanism, repeating three-layer GATv2 convolution, adding residual connection and normalization processing into output of each layer, finally obtaining node feature representation of spatial aggregation, and finally converting extracted spatial and temporal features into classification results through a decoder to detect faults; the method solves the problem that the prior art cannot be competent for multi-source data and diversified features in a modern industrial system, combines a memory one-dimensional convolutional neural network and a GNN deep learning method, can extract spatial and temporal features at the same time, and is significantly superior to a traditional statistical method in fault detection accuracy.

The invention discloses a lane line detection method based on SDAMamba and FFCDConv, and relates to the field of lane line detection. The method comprises the following steps: preprocessing a training set lane line image to obtain an updated training set; improving the DAMamba model to obtain an SDAMamba model for extracting multi-scale features of the updated training set image; extracting depth multi-scale features by using a feature pyramid neck network; mapping the depth multi-scale feature to a Hough parameter space by using DHT and selecting a Hough feature corresponding to a lane; jointly inputting the Huffcharacteristics and the depth multi-scale characteristics into an FFCDConv module, predicting the characteristics of different lane instances, and obtaining a predicted lane position map through a lane decoder; and inputting the predicted lane position map into a linear layer to determine a final lane position, and obtaining a lane line detection image. According to the method, the problem of breakage easily occurring in a lane line detection result can be repaired in a targeted manner, and the confusion rate of adjacent lanes is remarkably reduced.

The invention relates to a model training method and device, a storage medium and a program product. The method comprises the following steps: in a forward propagation stage, for any check point module, storing input and output of the check point module in a video memory, and releasing an intermediate activation value of the check point module in the video memory; the input of the check point module is used for recalculating an intermediate activation value, and the output of the check point module is used for forward calculation of subsequent modules of the check point module; and in a back propagation stage, for any check point module, responding to the condition that the last layer of the check point module is a linear layer, skipping forward calculation of the last layer, calculating a gradient according to a gradient formula corresponding to the last layer, and completing back propagation of the check point module according to the gradient of each layer in the check point module. According to the method and the device, the calculation cost can be remarkably reduced while the calculation precision and the video memory saving amount are the same as those of a standard re-calculation scheme.

The invention discloses a robot action sequence generation method based on a hierarchical attention mechanism, and the method comprises the steps: 1, inputting a plurality of robot operation images into a CNN layer, and extracting an image feature sequence; projecting the multiple pieces of joint position information into a joint state vector sequence through a linear layer of the CNN layer; taking a to-be-executed action sequence as a prediction target; taking the joint state vector sequence, the image feature sequence and the prediction target as input data of an encoder; 2, mapping the joint state vector sequence into a style variable by adopting an encoder, and adding randomly initialized learnable vector marks at the front ends of the image feature sequence, the joint state vector sequence and the action sequence to be executed; and 3, based on the sequence sparse attention, the trained decoder is adopted to predict a future action sequence of the mechanical arm of the robot. The invention provides a new solution for improving the fine operation capability of the robot.

The invention provides a Bayesian nerve radiation field modeling method and system based on uncertainty perception and dynamic importance sampling. The method comprises the following steps: replacing a full connection layer in a multi-layer perceptron with a Bayesian linear layer to obtain a Bayesian neural radiation field BN-NeRF model; acquiring a data set containing house source photos of different viewing angles and corresponding camera positions, and training a BN-NeRF model by using the data set; performing preliminary coarse sampling on each light passing through the house source scene to obtain a coarse sampling point set, and performing uncertainty evaluation on each sampling point in the coarse sampling point set by adopting a trained BN-NeRF model; according to the uncertainty evaluation result corresponding to the preliminary coarse sampling, performing secondary sampling on each light passing through the housing resource scene to obtain a fine sampling point set; and integrating the coarse sampling point set and the fine sampling point set to generate a final sampling point set, calculating the color and volume density of each sampling point in the final sampling point set by adopting a trained BN-NeRF model so as to carry out volume rendering, and generating a final house viewing picture.

The invention discloses a method and system for compressing a potential safety hazard recognition model for urban rail transit engineering, and the method comprises the steps: selecting a pre-trained Swin Transform as a teacher model, and carrying out the parameter adjustment of an urban rail potential hazard image. And embedding gating vectors in a W-MSA / SW-MSA attention head and an MLP linear layer, dynamically scaling channel output, strengthening sparsity through L1 punishment, and selecting a threshold value according to verification set performance to complete adaptive pruning. And after pruning, adjusting the bit width based on a weight L1 norm and probability distribution. Knowledge distillation is carried out on the student model and the teacher model after pruning quantification, an attention contrast loss function containing feature and probability distribution consistency loss is constructed, and temperature coefficient dynamic adjustment is introduced to determine final loss. And collaborative optimization of model lightweight and performance maintenance is realized.

The invention discloses an elastic Bragg breakwater structure response prediction method, and belongs to the technical field of ocean engineering structure dynamic response prediction, and the method comprises the steps: obtaining the front and rear wave surface elevation and motion response data of a breakwater; a double-flow attention mechanism based on physical prior guidance is constructed, and wave-structure interaction is decoupled into two parallel attention flows of structural dynamic evolution and wave-structure coupling feedback; a self-attention mechanism is combined with a time decay physical prior coding structure motion feature, and a bidirectional cross attention mechanism based on a resonance characteristic is combined with a wave-structure phase relation physical prior coding coupling feature; the double-flow features are adaptively integrated through a global context fusion module, and a multi-step prediction sequence is output through linear layer decoding; and a time-frequency domain joint loss function optimization model is adopted, so that the prediction precision is improved. According to the method, physical prior is explicitly embedded into an attention mechanism and is improved by more than 16% compared with an optimal baseline model, and the generalization ability is enhanced by more than 13% under the working condition that waves do not appear.

The invention discloses a method and a device for decoupling a slow-thinking area and a quick-response area in a large language model. The method comprises the following steps: constructing a slow thinking prompt and a fast response prompt, respectively inputting the slow thinking prompt and the fast response prompt into a large language model, and recording a neuron activation value of each linear layer of the large language model; calculating an activation value difference vector based on the neuron activation values corresponding to the slow thinking prompt and the fast response prompt, and obtaining a prompt difference vector based on the activation value difference vector; iteratively processing a plurality of reasoning problems, obtaining positive value position intersection of prompt difference vectors in all reasoning problems, and identifying a key neuron set related to slow thinking; and constructing a gradient mask matrix based on the position information of the key neurons in the key neuron set, and performing fine tuning on the key neurons by adopting a mask mode based on the gradient mask matrix to obtain an optimized large language model. The technical problem that an existing model is poor in reasoning ability is solved.

The present application belongs to the technical field of grain informatization processing, and particularly relates to a kind of imperfect grain kernel identification method and system.Visible light image and reflectance spectrum data of grain kernel are input into identification model;From the features extracted from the visible light image and reflectance spectrum data of grain kernel respectively by the feature extraction module, the input features are obtained;The spectral Mamba branch in the identification model sequentially performs linear layer and convolution processing on the normalized input features, and then inputs the state space model to obtain the output;The spatial Mamba branch performs dimension transposition on the normalized input features, and then sequentially performs linear layer and convolution processing, and then inputs the state space model to obtain the output;The guide branch sequentially performs linear layer processing on the normalized input features, and then inputs the ConvNeXt classifier, and then adjusts the output of the ConvNeXt classifier combined with the SiLU activation function to obtain the output;According to the fusion result of the output of the guide branch, the spectral Mamba branch, the spatial Mamba branch and the input features, the identification result is obtained.

The invention provides a numerical feature embedding method based on a scale perception radial basis function and a situation awareness method for the health degree of power equipment, and the method comprises the steps: obtaining a logarithm of to-be-coded numerical data x, and obtaining a base number D and an index L; performing RBF (Radial Basis Function) expansion on the base number D to obtain the expression of the base number D; carrying out soft sub-bucket distribution on the index L to obtain a distance from the index L to each soft sub-bucket, taking the distance as a coefficient of the soft sub-bucket, and weighting each soft sub-bucket according to the coefficient of the soft sub-bucket to obtain representation of the index L; and converting the representation of the index L into two numbers by adopting a gated linear layer network, and scaling the representation of the base number D to obtain the coded representation of the numerical data x.

The invention discloses a time sequence basic model construction method based on representation enhancement and an electromagnetic recognition system. The method comprises the following steps: constructing a time sequence basic model; dividing an input time sequence into a plurality of blocks, and performing linear mapping and one-dimensional convolution operation on each block; key point selection and nonlinear enhancement are carried out on initial block embedding, and initial enhanced block embedding is obtained; the method comprises the following steps of: inputting data into an encoder, extracting key points by applying Fourier transform and inverse Fourier transform to block embedding, carrying out cross attention operation on the block embedding and key point embedding, carrying out self-attention operation on the block embedding, and carrying out operation of a general transformer module; and inputting the time sequence characteristics extracted by the multi-layer transformer module into a linear layer, and identifying the type of the electromagnetic signal. According to the method, fine-grained time information is effectively reserved, the effectiveness of time sequence feature extraction by the time sequence basic model is remarkably enhanced, and meanwhile, the accuracy of electromagnetic signal type recognition is greatly improved.

The invention belongs to the technical field of data mining, and discloses an aero-engine missing data filling method based on a neural network with an exogenous variable graph. Aiming at the detection parameter data and the operation parameter data, a sliding window is used for dividing according to the time dimension to obtain an intermediate matrix; the method comprises the following steps of: performing sliding window division on high-dimensional representation of input data to construct a time sequence-attribute graph; the attention is used for calculating the feature representation of the action intermediate matrix and the action time sequence-attribute graph, and a final weight matrix of the time sequence-attribute graph is obtained through combination. Convolution operation is carried out on the time sequence-attribute graph, so that the node can dynamically aggregate influences from multiple historical moments and multi-dimensional attributes, and meanwhile, the directional effect of exogenous variables on endogenous variables is effectively fused. And splicing all the convoluted feature representations, fusing time sequence information of different time points through a pooling layer, and finally calculating by adopting two linear layers and an activation function to obtain a final filling result.

The invention discloses a direction consistency constraint-based adaptive image classification method during testing. The method comprises the following steps of: 1, constructing an image classification training data set and preprocessing the image classification training data set; 2, constructing a classification model and completing pre-training; 3, decomposing the weight of each linear layer in the pre-training model to obtain a modulus length matrix and a unit direction matrix, and introducing an orthogonal matrix to construct an effective weight structure for updating in a test stage; and 4, in a test stage, calculating the difference between the model output features and the source domain features in terms of statistics, constructing a loss function with statistics alignment as a target, and on the premise of keeping the unit direction matrix unchanged, updating the modulus length matrix and the orthogonal matrix to realize adaptive adjustment of the image classification model during test. According to the method, the problem of performance degradation caused by input distribution change is effectively relieved, and the classification precision of an image classification model under a complex disturbance condition is remarkably improved.

The invention provides an ARM architecture-oriented symmetric cipher encryption and decryption method and system, and relates to the technical field of encryption and decryption, the symmetric cipher adopts a block cipher, the block cipher adopts a round function to perform confusion and diffusion on data blocks in each round of encryption and decryption, and in a linear layer component for realizing diffusion, the data blocks are subjected to encryption and decryption in a byte copying and filling manner. The method comprises the following steps of: filling 8-bit or 16-bit original variables to be calculated to bits required by an ARM (Advanced RISC Machines) architecture, carrying out combined operation of cyclic shift, XOR and bit permutation on the filled variables, and carrying out low-bit extraction on an operation result according to the original bits of the variables to obtain an operation result of the original variables; according to the method, the linear layer efficiency of the small-size symmetric cryptography can be remarkably improved, a new thought and method are provided for lightweight symmetric cryptography algorithm design, and the method is of great significance to cryptography algorithm performance improvement in the Internet of Things and intelligent terminals.

A deep learning model based on attention using an embedding scheme for continuous variables of tabular data. A method of constructing the deep learning model based on attention includes converting tabular data of structured data having a mixture of categorical variables and continuous variables into embedding values and training a network model including a transformer block, a linear layer block, and a sharing function for the sharing of an attention between the transformer block and the linear layer block by using the embedding values.

The application provides a matrix operation accelerator combining wavelength division multiplexing and MZI cascade network, relates to the field of optical neural networks, and comprises an input signal layer, a weight signal layer, a summation layer and a nonlinear layer.The input signal layer is used for realizing matrix operation of optical signals through a Mach-Zehnder interferometer array; the weight signal layer is used for applying an electrical signal to a micro-ring modulator array to adjust a weight signal; the summation layer is used for separating the results of the action of the optical signals of different wavelengths through the weight signal; and the nonlinear layer is used for converting the optical signals into electrical signals through a photodetector array to realize a nonlinear activation function in the electrical domain.The application introduces N different wavelengths in the network formed by MZI cascade, so that the number of times of executing matrix operation is increased by N times each time, high-speed convolution operation is facilitated, and the size of the micro-ring modulator is relatively small, so that the energy efficiency and area ratio of the MZI cascade network calculation can be effectively increased.

The application discloses a kind of visual cognitive drive small sample image classification method, system and medium, the method includes pre-training stage and meta-learning stage;The pre-training stage includes: obtaining initial training set, data enhancement is carried out to the initial training set, obtains training dataset;The training dataset is embedded coding, mapping is carried out to embedding by linear layer, pre-trains the classification task of upstream;The meta-learning stage includes: the training dataset is divided into support set and query set;Image of the support set and the query set is mapped to embedding space, obtains corresponding embedding vector representation;The similarity of embedding in high-dimensional space is evaluated by relationship network, and the classification result of small sample learning output is obtained: the application can solve the problem that embedding space cannot be correctly perceived and the relationship between support set prototype and query set embedding is not accurate in the prior art.

Tuning weights of a neural network of a model for processing input of the model representing information about a technical system and outputting an output of the model for operating a technical system. The model includes a linear layer for mapping a multidimensional input of the layer depending on the weights to a multidimensional output of the layer. The model is configured to determine the input of the layer depending on the input of the model, and to determine the output of the model depending on the output of the layer. A method includes providing training data include the input of the model and a ground truth for the output of the model corresponding to the input of the model in the training data, providing a set of tuning methods for tuning the weights, determining the principal components decomposition of a weight matrix including the weights.

The invention discloses a traffic flow long-time-sequence prediction method based on a graph convolutional network, and aims to solve the problems of poor precision and large calculation overhead and memory overhead of an existing traffic flow prediction method under long-time-sequence prediction requirements. Firstly, a training sample is constructed for preprocessed data through a sliding window strategy; secondly, a traffic flow long-time-sequence prediction model based on the graph convolution network is constructed, the model is formed by stacking a plurality of TimeModule layers, after passing through a fast Fourier transform and periodic graph convolution module, weighted merging is carried out according to amplitude, output processed by the plurality of TimeModule layers passes through a linear layer, and predicted traffic state data is output; carrying out model training by taking a mean absolute error as a loss function; and finally, inputting the traffic state data to be analyzed into the model to obtain predicted traffic state data. Through verification, in a long-time-sequence (60 time steps) traffic flow prediction task, the model has the performance of high prediction structure precision, small parameter quantity, short training time and the like, and has good generalization ability.