23 results about "Matrix compression" patented technology

The invention discloses a tensor autoregression model coefficient estimation optimization system and method based on a sketch matrix, and relates to the technical field of multi-dimensional time series data processing. The system comprises a multi-modal data acquisition module, a self-adaptive sketch matrix construction module, a tensor autoregression model construction module, a multi-stage coefficient estimation optimization module and the like. The method comprises the following steps: preprocessing multi-source time sequence data to construct a tensor, generating sparse sketch matrix compressed data based on PARAFAC decomposition and differential sampling, constructing a tensor autoregression model containing a lag term, optimizing a coefficient by combining an L1-L2 mixed regular term and an ADMM algorithm, dynamically updating the model, and deploying the model. According to the method, the calculation complexity is reduced through the sketch matrix, the comprehensive efficiency is improved by more than or equal to 50%, data key information retention is guaranteed, the reconstruction error is less than or equal to 5%, the prediction precision is reduced by less than or equal to 10%, edge equipment is adapted, the method can be applied to scenes such as Internet of Things monitoring and financial analysis, the contradiction between efficiency and precision in high-dimensional data processing is solved, and the operation and maintenance cost is reduced.

Disclosed are systems and techniques for performing matrix multiply operations on an expressive sparse matrix representation with limited metadata. The techniques include receiving a sparse matrix, metadata corresponding to the sparse matrix, and a matrix operand. The sparse matrix contains a first number (N) of elements to retain from a dense matrix which comprises at least a second number (M) of elements. The metadata corresponding to the sparse matrix is based on a third number (P) of positions and a format determined during compression of the dense matrix. The techniques include selecting, by one or more selection circuits, a subset of elements of the matrix operand based on the metadata corresponding to the sparse matrix and performing one or more matrix multiply operations on the sparse matrix and the subset of elements of the matrix operand.

This invention discloses a parallelizable adaptive lossless matrix compression method. It involves dividing a large first matrix into smaller second matrices, assigning a splitting process to each second matrix, and having these splitting processes divide all second matrices into multiple bit matrices in parallel. Compression threads are then assigned to each bit matrix, and these compression threads adaptively select the optimal compression strategy from a set of compression strategies. Finally, the compression threads compress all bit matrices in parallel according to the determined compression strategy, forming a compressed matrix and a flag array as the final compressed data of the first matrix. By hierarchically and parallelizing the compression process, the memory overhead of the compression process is effectively reduced and the compression speed is improved. Adaptive compression is achieved by selecting the optimal compression strategy through trial compression, effectively improving the data compression ratio. Lossless data compression is achieved by constructing a flag array.

The present application relates to the technical field of fault diagnosis, in particular to a power grid fault analysis method and system, which comprises constructing a fusion tensor and performing dimension reduction processing on the fusion tensor to generate a fusion feature matrix; constructing a convolutional autoencoder model to detect abnormal fusion feature matrix; classifying power grid faults according to abnormal data of the abnormal feature matrix, and calculating power grid equipment fault area based on the result of power grid fault classification. The present application has the beneficial effect of ensuring the integrity and accuracy of data by using a high-precision low-rank tensor completion algorithm to complete power grid operation data. Key features are extracted from multi-dimensional data collected by multiple source sensors and a fusion feature matrix is constructed using Tucker decomposition and matrix compression technology, effectively improving the efficiency of data processing and the ability of feature expression. The application of the convolutional autoencoder model enhances the detection accuracy of abnormal features, and the improved weighted summation and hierarchical clustering-based method improves the accuracy of fault classification and positioning.

A system and method for performing multiplication for a neural network, e.g. for data of one or more layers in a neural network, may include loading a portion of a compressed version of a sparse input matrix into a cache memory; uncompressing a subset of the data in the portion of the compressed version of the sparse input matrix; and multiplying a sparse kernel matrix by the subset of the data using a set of instructions which are themselves created based on the sparse kernel matrix.

The invention discloses a fully expressive sparse matrix representation with finite element data. Systems and techniques for compressing a dense matrix into a fully expressive sparse matrix representation with finite element data are disclosed. The techniques include generating a sparse matrix with corresponding metadata based on the dense matrix. Generating a sparse matrix with respective metadata includes: identifying a first number (M) of elements to be compressed, a second number (N) of elements to be retained, and a third number (B) of bits used by each metadata value; determining a metadata value of each of the N elements of the dense matrix; packing the first metadata value having more than B bits into a second metadata value having B bits; and generating a sparse matrix containing the N elements of the dense matrix. The techniques include storing a sparse matrix and corresponding metadata, wherein the corresponding metadata includes a second metadata value.

The invention relates to a method for testing interface shear strength of an optical fiber embedded composite material based on an annular pressure head, and belongs to the technical field of interface performance testing of optical fibers and composite materials. Comprising the following steps: providing a group A sample and a group B sample of which the cross sectional areas are smaller than the area of a test pressure head; the A group of samples are composite material samples embedded with optical fibers, the end surfaces of the optical fibers and the surface of the matrix are coplanar, and the B group of samples are composite material samples not embedded with optical fibers; acquiring annular pressure data; obtaining matrix data; calculating shear strength; correcting the matrix compression load based on the ratio of the effective compression area of the A group of samples to the effective compression area of the B group of samples to obtain a matrix equivalent compression load; and calculating the difference between the total load and the equivalent compression load of the matrix to obtain an interface shear force, and further calculating the interface shear strength. The interference of the plastic deformation of the matrix on a test result is eliminated, and the accuracy and the reliability of a calculation result are remarkably improved.

The invention provides a high-order sparse matrix LDL efficient decomposition calculation acceleration method, which can be used for a high-order sparse matrix linear equation set solving calculation scene. The accelerator is constructed based on an FPGA platform, and efficient decomposition of a high-order sparse matrix is achieved through a software and hardware cooperative computing mechanism. The software level realizes symbolic analysis of a sparse positive definite matrix, calculation sequence dependency modeling and LDL decomposition hardware customization instruction generation; and the hardware acceleration unit realizes streamlined execution of decomposition tasks by performing instruction fetching, decoding and execution of customized instructions, and completes high-order matrix numerical calculation. The system comprises an instruction generation unit, an instruction analysis unit, a calculation module, a data access module and a control scheduling module, and all the modules work cooperatively to improve the degree of parallelism and the storage access efficiency. Matrix data types support fp16, fp32 and fp64, matrix compression is realized by adopting a sparse column (CSC) format and a matrix rearrangement method, and storage overhead and memory access delay are remarkably reduced while high calculation precision is ensured. Experiments show that compared with traditional CPU and GPU platforms, the sparse matrix solution performance and energy efficiency of the method are remarkably improved by one order of magnitude, and the method is suitable for high-precision engineering calculation and edge real-time solution scenes such as engineering simulation, structural mechanics, electromagnetic analysis and attitude calculation.

The invention provides an efficient solution calculation acceleration method for a high-order sparse matrix linear equation set. The high-order sparse matrix linear equation set is widely applied to the fields of scientific calculation, engineering simulation and the like. A traditional FPGA-based solving method is low in hardware resource utilization rate and complex in data flow scheduling, so that the calculation efficiency is limited, and the efficient solving requirement of a large-scale sparse matrix is difficult to meet. In order to solve the problem, the invention provides a high-order sparse matrix-oriented linear equation set efficient solution calculation acceleration method, a sparse matrix compression storage format is combined, a linear equation set solution special circuit and an acceleration solution algorithm based on LDL decomposition are designed, and the calculation efficiency is effectively improved.

The utility model relates to the technical field of compression molding device, and disclose a kind of matrix compression molding device, and the anti-skid bottom plate is fixedly sleeved in the position around the surface of fixed bottom plate, first cross plate is movably connected in the position of both sides of fixed bottom plate top, second cross plate is movably connected in the position of both sides of first cross plate on fixed bottom plate top, second screw rod is movably connected in the position of the bottom of first cross plate in fixed bottom plate interior, first screw rod is movably connected in the position of the bottom of second cross plate in fixed bottom plate interior, servo motor is drivingly connected in the position of both ends of second screw rod and first screw rod, the utility model is equipped with servo motor, first screw rod, second screw rod, first cross plate, second cross plate, is favorable to after matrix compression molding, by servo motor drive first screw rod, second screw rod is rotated, and then make first cross plate, second cross plate separate to around, and then quickly the matrix of compression molding is demoulded, increase the production efficiency of matrix.

Embodiments herein relate to a neural network compression technique in which a weight matrix within a neural network is transformed via matrix multiplication with an orthogonal matrix. The orthogonal matrix is derived from a calibration data set, which is generally selected to widely represent expected runtime input data, and the transformation causes the resulting modified weight matrix to have components ordered by relative importance. The modified weight matrix is incorporated into a compressed neural network with fewer weights. By removing one or more components of lower importance, the size of the neural network (and thus its storage and execution overhead) is reduced while still maintaining acceptable performance levels.

Systems and methods for decompressing a meaningful sparse matrix representation with limited metadata are disclosed. The methods include receiving a sparse matrix and metadata corresponding to the sparse matrix. The sparse matrix is ​​a compressed representation of a dense matrix. The sparse matrix contains a first number N of elements to be retained from the dense matrix, which comprises at least a second number M of elements. The metadata corresponding to the sparse matrix is ​​based on a third number P of positions and a format determined during the compression of the dense matrix. The methods include generating an uncompressed matrix based on the sparse matrix and the metadata corresponding to the sparse matrix.

Systems and methods for performing matrix multiplication operations with a meaningful sparse matrix representation containing limited metadata are disclosed. The methods include receiving a sparse matrix, metadata corresponding to the sparse matrix, and a matrix operand. The sparse matrix contains a first number N of elements, which are to be retained from a dense matrix comprising at least a second number M of elements. The metadata corresponding to the sparse matrix is ​​based on a third number P of positions and a format determined during compression of the dense matrix.The procedures involve selecting a subset of elements of the matrix operand by one or more selection circuits based on the metadata corresponding to the sparse matrix, and performing one or more matrix multiplication operations on the sparse matrix and the subset of elements of the matrix operand.

Systems and methods for compressing a dense matrix into a meaningful sparse matrix representation with limited metadata are disclosed. The methods include generating a sparse matrix with corresponding metadata based on a dense matrix. Generating the sparse matrix with corresponding metadata involves identifying a first number M of elements to be compressed, a second number N of elements to be retained, a third number P of positions, and a format; determining a metadata value for each of the N elements of the dense matrix based on the identified P and the identified format, wherein the dense matrix has at least M elements; and generating the sparse matrix containing the N elements of the dense matrix.The methods involve storing the sparse matrix and the corresponding metadata, where the corresponding metadata includes the metadata value for each of the N elements of the dense matrix.

Disclosed are systems and techniques for compressing a dense matrix into an expressive sparse matrix representation with limited metadata. The techniques include generating a sparse matrix with corresponding metadata based on a dense matrix. Generating the sparse matrix with corresponding metadata includes identifying a first number (M) of elements to compress, a second number (N) of elements to retain, a third number (P) of positions, and a format; determining a metadata value for each of N elements of the dense matrix based on the identified P and the identified format, wherein the dense matrix includes at least M elements; and generating the sparse matrix containing the N elements of the dense matrix. The techniques include storing the sparse matrix and the corresponding metadata, wherein the corresponding metadata comprises the metadata value for each of the N elements of the dense matrix.

This invention discloses a hierarchical TCAM message parsing and field extraction system based on parse graph optimization, relating to the field of communication technology. It includes: a software calculation and configuration module for performing graph topology optimization on the protocol parsing graph to generate multiple independent parsing subgraphs; generating a hardware resource configuration table corresponding to each parsing subgraph based on the key bit set corresponding to each parsing subgraph, and distributing the hardware resource configuration table to the key bit extraction unit, mapping table matching unit, flow classification module, and field extraction module on the hardware side, thereby configuring the logic structure of the target circuit according to the hardware configuration information and realizing dynamic reconfiguration of the protocol parsing function. This invention decomposes the protocol parsing graph into parsing subgraphs and adopts a multi-level unit fusion architecture at the hardware level to achieve pipelined protocol parsing. Combined with key bit matrix compression technology, it can significantly reduce TCAM storage requirements and effectively solve the timing and storage bottlenecks in large-scale message header parsing.

The application discloses a matrix compression method, a matrix processing method and related devices, and belongs to the field of data processing. The method comprises the following steps: dividing a first matrix stored in a memory to obtain a plurality of first sub-matrices; compressing non-zero element groups in each first sub-matrix in the plurality of first sub-matrices according to a first direction to obtain a plurality of second sub-matrices corresponding to the first matrix; dividing each second sub-matrix in the plurality of second sub-matrices to obtain a plurality of third sub-matrices; and generating a compression result of the first matrix based on the plurality of third sub-matrices. According to the application, only non-zero element groups exist in the first sub-matrices, so that a part of elements with a value of zero in the first matrix is removed, and therefore, the compression result of the first matrix can effectively reduce the number of zero elements that need to be stored, thereby saving data storage space.

Methods and systems for compressing a matrix are described. A matrix having a plurality of rows formed from a respective plurality of vectors is partitioned into a plurality of sub-matrices, each sub-matrix containing a respective set of sub-vectors from one or more consecutive columns of the matrix. For each given sub-matrix, the sub-vectors are clustered into a plurality of clusters. For each given cluster, a centroid and a variance are computed and stored based on the sub-vectors belonging to the given cluster. A mapping of each vector to a respective cluster in each sub-matrix is stored. The stored centroids, the stored variances, and the stored mappings form a compressed data set for reconstructing the matrix.

The application provides a GPU parallel implementation method for multi-frame multi-code rate LDPC decoding, and belongs to the technical field of satellite communication channel coding. The method comprises the following steps: obtaining check matrix compression parameters and frame offset parameters by processing a check matrix in a CPU according to input LLR information and code rate information of multi-frame LDPC codes in signaling frame information; allocating memory in the CPU and the GPU, and respectively storing log-likelihood ratio information LLR transmitted by a front end and initializing check parameter transmission nodes C2V, variable parameter transmission nodes V2C, variable node log-likelihood ratio LOV and decoding results Out_Dec; dynamically adjusting the check matrix compression parameters according to the frame offset parameters, and updating nodes in the GPU through kernel functions by using the check parameter transmission nodes C2V, the variable parameter transmission nodes V2C and the variable node log-likelihood ratio LOV; and copying the decoding results back to the Host end. The application eliminates memory access conflicts caused by dynamic code rates by preprocessing input LLR data and analyzing multi-frame check matrix parameters, so that the utilization rate of GPU computing cores is improved.

Disclosed are systems and techniques for compressing a dense matrix into a fully-expressive sparse matrix representation with limited metadata. The techniques include generating a sparse matrix with corresponding metadata based on a dense matrix. Generating the sparse matrix with corresponding metadata includes identifying a first number (M) of elements to compress, a second number (N) of elements to retain, and a third number (B) indicating the number of bits each metadata value uses; determining a metadata value for each of N elements of the dense matrix; packing a first metadata value having more than B bits into a second metadata value having B bits; and generating the sparse matrix containing the N elements of the dense matrix. The techniques include storing the sparse matrix and the corresponding metadata, wherein the corresponding metadata comprises the second metadata value.

The invention discloses a self-adaptive large model fusion method and system based on preference data driving, and solves the technical problem that the performance of a fused multi-task model is far lower than that of a single expert model due to the fact that serious parameter conflicts are extremely easy to generate when a plurality of task vectors are merged by the existing large model fusion method. The method comprises the steps of obtaining a basic model and a plurality of expert models, and generating task vectors corresponding to the expert models by calculating parameter difference values of the basic model and the expert models; preprocessing the task vector, outputting a corresponding compression component and a task component, and constructing a hierarchical coefficient matrix based on the task component; generating a preference data set in combination with no-label prompt of downstream tasks of the expert model and task vectors, and iteratively optimizing the coefficient matrix by means of an optimizer and a direct preference optimization loss function to obtain an optimal fusion coefficient matrix; and finally, updating the basic model according to the matrix, the compression component and the basic model parameters, and determining a target static model.

The application discloses a rail transit station building drawing matrix compression method and system, and the method comprises the following steps: extracting line segment information of building walls and structural columns in a station building drawing, wherein the line segment information comprises straight line segment information and arc line segment information; converting all arc line segments into corresponding straight line segments, and calculating the segment length according to the column radius of the structural column; extracting the head and tail points of all line segments, obtaining the X and Y coordinates of each head and tail point, sorting the X and Y coordinates, and constructing an image matrix, with the X and Y coordinates serving as the horizontal and vertical axes of the image matrix; filling the geometric information and attribute information of building elements in the image matrix, and filling negative numbers in the edge of the image matrix through an edge filling technique, to represent the peripheral contour information of the rail transit station.

The invention discloses an expressive sparse matrix representation with finite element data. Systems and techniques for compressing a dense matrix into an expressive sparse matrix representation with finite element data are disclosed. The techniques include generating a sparse matrix with corresponding metadata based on the dense matrix. Generating a sparse matrix with respective metadata includes: identifying a first number (M) of elements to be compressed, a second number (N) of elements to be retained, a third number (P) of locations, and a format; determining a metadata value for each of N elements of a dense matrix based on the identified P and the identified format, where the dense matrix includes at least M elements; and generating a sparse matrix containing the N elements of the dense matrix. The techniques include storing a sparse matrix and respective metadata, wherein the respective metadata includes metadata values for each of N elements of a dense matrix.