12 results about "Linear operators" patented technology

The application provides a model quantification method, device and equipment and storage medium, the method comprises: using integer algorithm to execute a plurality of operators of transformer model for image processing, the plurality of operators of the transformer model comprises linear operator and nonlinear operator, wherein the linear operator comprises matrix multiplication operator, the matrix multiplication operator adopts symmetric quantization to quantize floating point value to integer value;The nonlinear operator comprises an activation function operator and a layer normalization operator, the activation function operator adopts polynomial fitting to quantize floating point value to integer value, and the layer normalization operator adopts the mean and standard deviation of input data in the channel dimension to quantize floating point value to integer value.The transformer operator is quantified, so that the inference work of the transformer model for natural language is based on integer operation, so that it can be truly deployed on FPAG chip for practical application.

The embodiment of the invention provides a nonlinear operator approximate calculation device and method, a neural network processor and a medium, and belongs to the technical field of neural networks. The device comprises: a floating point number evaluation unit receiving original floating point data of an original neural network operator, and performing compensation interval evaluation on the original floating point data according to a preset floating point value domain range to obtain compensation interval evaluation information; if the compensation interval evaluation information represents that the original floating point data is not in the preset floating point value domain range, the index splitting unit splits the original floating point data into a first floating point number and a second floating point number; the operation compensation unit performs fitting compensation on the first floating-point number to obtain first output data; and the splicing unit splices the first output data and the second floating-point number passing through the original neural network operator to obtain target output data. According to the invention, the computing resources and time of a computer system can be reduced, the full-value-domain compensation of the neural network operator is completed with few hardware resources, the computing precision is improved, and the reasonability of network reasoning is ensured.

The invention discloses a non-linear operator calculation method and system for privacy protection machine learning, and provides efficient and safe basic support for private calculation of operations such as a non-linear activation function in a machine learning model on the premise that multi-party cooperative calculation is performed and a plurality of participants do not expose local data. According to the framework, additive secret sharing and mask secret sharing are combined, and an efficient sharing conversion protocol is constructed and used for supporting conversion operation between different sharing types. Furthermore, the invention provides a series of sub-protocols such as security replacement, security comparison and most significant bit (MSB) extraction, and lays a key foundation for subsequent construction of security calculation protocols of non-linear operators such as ReLU, DReLU, MaxPool and the like.

The invention relates to a stability analysis method and system for a boundary layer of a high-enthalpy ablated wall surface, and is applied to the technical field of aerospace, and the method comprises the steps: taking a Landau-Teller equation and a chemical reaction rate equation as source items, constructing a control equation, and solving the control equation to obtain a steady laminar flow field; analyzing the steady laminar flow field by using LST to obtain a transmission coefficient eigenvalue; if the first modal time scale and the second modal time scale are greater than or equal to 1, correcting the first modal time scale and the second modal time scale based on the roughness of each position of the wall surface of the high-speed aircraft, and constructing a first modal linear operator and a second modal linear operator; further constructing NPSE, and solving to obtain an NPSE result; when an NPSE result is decomposed and linearized according to Fourier subharmonics, a coupling source item of a subharmonic equation is constructed according to a second modal linear operator, secondary instability analysis is carried out, and a transition initial position and the amplitude and frequency of disturbance development are obtained. And the accuracy of stability analysis can be improved.

The present application belongs to the field of integrated circuit manufacturing online measurement, and discloses a snapshot overlay error measurement method and system, which comprises the following steps: measuring a measured object, obtaining two measurement spectra with positive and negative preset deviations respectively, coherently demodulating the measurement spectra, and shifting the specific frequency channel to the zero frequency channel; processing the spectrum data after frequency shifting by using a linear operator to obtain the coefficients corresponding to the positive and negative preset deviations; constructing a characteristic quantity according to the linear combination of the real part and the imaginary part of the coefficients, then calculating the corresponding characteristic quantity according to the coefficients corresponding to the positive and negative preset deviations; and calculating the overlay error according to the characteristic quantity based on the linear relationship between the characteristic quantity and the overlay error. The present application can solve the overlay error with multi-wavelength coupling without using traditional Fourier analysis and truncation operation, has high precision and robustness to noise, and can be used for data processing of snapshot overlay error measurement in multiple scenes.

The application belongs to the technical field of artificial intelligence, and particularly relates to a pure integer quantization method of a visual Transformer model and a related device; the pure integer quantization method of the visual Transformer model comprises the following steps: based on a selected visual Transformer model, linear components and nonlinear components are respectively quantized and integrated to obtain a pure integer quantization model; when the nonlinear components are quantized, integer approximation algorithms are used to respectively reconstruct integer Softmax, GELU and LayerNorm nonlinear operators to obtain reconstructed integer Softmax, GELU and LayerNorm operators; a tensor virtual machine compiler framework is used to perform operator packaging and computation graph optimization on the pure integer quantization model to obtain an optimized pure integer computation graph. The technical scheme disclosed by the application realizes full integer calculation and efficient deployment of the ViT model on the FPGA end side by reconstructing integer operators of the three types of nonlinear operators, namely Softmax, GELU and LayerNorm, and combining computation graph scheduling optimization.

The disclosure provides a computing system, a data processing method, equipment and a medium for high-bandwidth inference, and relates to the technical field of integrated circuits. The computing system is used for performing a decoding stage of a Transformer-based model inference, and comprises a host processor for performing a decoding stage nonlinear operator, an offload subsystem for performing at least a part of a decoding stage linear operator, and a standard high-speed interface module for transmitting an input activation vector and an output result vector between the host processor and the offload subsystem. The offload subsystem comprises a weight lock storage array for storing a weight matrix in a static residence manner, an input vector streaming interface for streamingly receiving the input activation vector, a matrix vector multiplication calculation unit for performing a matrix vector multiplication operation on the input activation vector and the weight matrix, and a result processing module for reducing or arranging the operation result to obtain the output result vector.

The application relates to the technical field of numerical weather prediction, in particular to a layered self-adaptive Gaussian conversion method and equipment for cloud and rain control variables, which can acquire observation data and input the observation data into a numerical prediction model to obtain three-dimensional cloud and rain control variables; a variational assimilation model including a forward transformation operator, an inverse transformation operator, a tangent linear operator and an adjoint operator is constructed; the forward transformation operator can calculate and adjust conversion strength layer by layer, realizing more fine and accurate Gaussian processing; the tangent linear operator and the adjoint operator are used for internal minimum iteration calculation of the assimilation model, and then the inverse transformation operator is used to restore the three-dimensional cloud and rain control variable analysis field to its original physical order of magnitude output, so that the actual three-dimensional cloud and rain control variable analysis field is obtained. According to the technical scheme, the vertical distribution difference of the cloud and rain variables can be accurately adapted, and through the provision of a complete operator chain, the compatibility with an existing assimilation system is ensured, so that the analysis field quality and the prediction ability are improved.

The invention belongs to the technical field of artificial intelligence, and particularly relates to a pure integer quantization method of a visual Transform model and a related device. The pure integer quantization method of the visual Transform model comprises the following steps: based on a selected visual Transform model, respectively quantizing a linear component and a non-linear component, and then integrating the quantized linear component and the quantized non-linear component to obtain a pure integer quantization model; when the non-linear component is quantized, an integer approximation algorithm is adopted to carry out integer reconstruction on three types of non-linear operators, namely Softmax, GELU and LayerNorm, so that a reconstructed Softmax integer operator, a reconstructed GELU integer operator and a reconstructed LayerNorm integer operator are obtained; and adopting a tensor virtual machine compiler frame to carry out operator packaging and computational graph optimization on the pure integer quantization model to obtain an optimized pure integer computational graph. According to the technical scheme disclosed by the invention, integer operator reconstruction is carried out on three types of nonlinear operators, namely Softmax, GELU and LayerNorm, and calculation graph scheduling optimization is combined, so that full integer calculation and efficient deployment of the ViT model on the FPGA end side are realized.

The embodiment of the application provides a conversion method, a storage medium and a program product of an artificial neural network model, relates to the technical field of artificial intelligence, and the method comprises the following steps: after obtaining a pre-trained artificial neural network model, converting each nonlinear operator in the artificial neural network model into a corresponding pulse module, wherein the pulse module comprises a differential expectation compensation module, the differential expectation compensation module is used for calculating an output increment according to cumulative membrane potential, a differential pulse neuron is inserted in each pulse module, the differential pulse neuron updates an encoding activation value when a pulse is emitted, otherwise the encoding activation value remains unchanged, a bias term of a linear operator located in a previous layer of each nonlinear operator is removed, and an initial membrane potential of the differential pulse neuron inserted in the pulse module corresponding to the nonlinear operator is set as the bias term, and the encoding activation value is only updated when a pulse is emitted, so that the loss caused by updating the encoding activation value regardless of whether a pulse is emitted or not is avoided, and energy consumption is significantly reduced.