Efficient convolution implementation method and device based on winograd algorithm and approximate multiplier

Abstract

The present application discloses a high-efficiency convolution implementation method and device based on Winograd algorithm and approximate multiplier. The method includes: dividing the obtained initial feature map into a matrix sequence composed of multiple first matrices, each first matrix Contains matrix elements with 6 rows and 6 columns, and there are matrix elements with 2 rows or 2 columns overlapping between adjacent matrices in the matrix sequence; get each first matrix corresponding to an input matrix and the initial convolution weight of the initial feature map; according to The initial convolution weight is calculated as the second matrix of the Winograd convolution weight; each input matrix is ​​calculated with the Winograd convolution weight using an approximate multiplier to obtain multiple output matrices, wherein each input matrix corresponds to an output matrix; Concatenate multiple output matrices to obtain an output feature map. By adopting the foregoing solution, the number of multipliers required for unit convolution output can be greatly reduced, and the convolution calculation efficiency can be improved.

Description

technical field [0001] The present application relates to the technical field of neural networks, in particular to a method and device for realizing efficient convolution based on Winograd algorithm and approximate multiplier. Background technique [0002] Convolutional neural network is a kind of feed-forward neural network with convolution calculation and deep structure, and it is one of the representative algorithms of deep learning. With the wide application and development of deep learning, convolutional neural networks are used in more and more scenarios, especially in image recognition scenarios, which have achieved breakthrough development. [0003] When the convolutional neural network is deployed on the hardware for calculation, for the traditional convolution calculation method, the input is generally a three-dimensional matrix, which has three dimensions of length, width and channel number, and multiple convolution kernels are convolved with the input matrix ope...

AI Technical Summary

Problems solved by technology

[0005] This application provides a high-efficiency convolution implementation and device based on the Winograd algorithm and an approximate multiplier to solve the problem that the traditional convolution calculation method consumes a large number of multipliers and the calculation efficiency is often relatively low

Images