4 results about "Transpose" patented technology

This application discloses a method and related apparatus for determining the output result of a model, belonging to the field of data processing technology. The network model includes an attention mechanism module; the method includes: determining a query matrix, a key matrix, and a value matrix through the attention mechanism module based on the current input information of the network model; processing the transpose of the key matrix to obtain an intermediate matrix, wherein the intermediate matrix has the same number of rows and columns as the transpose matrix, and the average value of each column element in the intermediate matrix is ​​less than the average value of the corresponding column elements in the transpose matrix; determining an output matrix through the attention mechanism module based on the query matrix, the intermediate matrix, and the value matrix; and determining the current output result of the network model based on the output matrix. Therefore, this application reduces the value of elements in the transpose matrix by processing it, thereby preventing numerical overflow and computational instability at the computational source.

The embodiment of the invention discloses a linear system acceleration solving method, a weather prediction method and a related device, and the method comprises the steps: carrying out the iterative calculation on a quantum computer according to a first type of target vector and a second type of target vector of a to-be-solved linear system, and obtaining a three-diagonal matrix, the plurality of first-class target vectors form a first-class target matrix, the plurality of second-class target vectors form a second-class target matrix, and the product of the transpose of the first first-class target vector in the first-class target matrix and the first second-class target vector in the second-class target matrix is 1; according to the unknown vector of the subspace linear system obtained by solving the subspace linear system of the linear system to be solved, the unknown vector of the linear system to be solved is determined, and the coefficient matrix of the subspace linear system is a tridiagonal matrix. By adopting the embodiment of the invention, the solving speed of the linear system can be improved when a large-scale sparse matrix is processed, and memory resources occupied by solving the linear system are saved.

Disclosed is a neural network operation method using an NPU, the method comprising the steps of: inputting, by the NPU of a computing device, a matrix stored in a memory and a transposed matrix of the matrix to a matrix multiplication operator; generating, by the matrix multiplication operator, a first output matrix by performing a matrix multiplication operation on the transposed matrix and the matrix; and storing, by the NPU, main diagonal elements of the first output matrix in the memory as a first reduced array of the matrix. The first reduced array is an array including sums of squares of element values included in each column constituting the matrix.

The present application belongs to the technical field of data processing. Disclosed are a method for determining an output result of a model, and a related apparatus. A network model comprises an attention mechanism module. The method comprises: on the basis of current input information of a network model, determining a query matrix, a key matrix and a value matrix by means of an attention mechanism module; processing a transposed matrix of the key matrix, so as to obtain an intermediate matrix, wherein the intermediate matrix has the same number of rows and columns as those in the transposed matrix, and the average value of elements in each column in the intermediate matrix is less than the average value of elements in a corresponding column in the transposed matrix; on the basis of the query matrix, the intermediate matrix and the value matrix, determining an output matrix by means of the attention mechanism module; and on the basis of the output matrix, determining a current output result of the network model. It can be seen therefrom that in the present application, a transposed matrix of a key matrix is processed, so as to decrease the values of elements in the transposed matrix, thereby preventing a computational source from the problems of numerical overflow and computational instability.