Convolutional neural network weight compression method and device based on arm architecture fpga hardware system

Abstract

The present application relates to speed-up of neural networks, and discloses a convolutional neural network weight compression method and a device thereof. The method includes obtaining all the convolution kernels of the convolutional neural network; each weight matrix for any convolution kernel is stacked and arranged according to the depth from small to large, and all the arranged multi-layer weight matrices except the first Any non-zero element of : Calculate the depth offset of the current non-zero element relative to the previous non-zero element and calculate the height offset of the current non-zero element relative to the first row and first column element of the current layer weight matrix and width offset, and compress the value of the current non-zero element, the depth offset, the height offset, and the width offset according to the preset compression rules to obtain the corresponding compression result. The embodiment of the present application enables the compressed result to accelerate the subsequent neural network operation on the FPGA.

Description

technical field [0001] The present application relates to the field of neural network speed-up, in particular to convolutional neural network weight compression technology. Background technique [0002] For non-uniformly distributed information, Huffman coding is the theoretically optimal lossless compression scheme. However, Huffman coding is not aimed at huge sparse matrices. For the convolution calculation process of convolutional neural networks with huge sparse matrices, Huffman coding needs to count the number of weights before calculating the Huffman tree to get Long encoding, which undoubtedly increases the burden on the computing system. At present, the process of Huffman encoding and decoding does not take hardware characteristics into consideration, and the implementation effect on FPGA is not good. Contents of the invention [0003] The purpose of this application is to provide a convolutional neural network weight compression method and device thereof, so th...

AI Technical Summary

Problems solved by technology

However, Huffman coding is not for huge sparse matrices. For the convolution calculation process of convolutional neural networks with huge sparse matrices, Huffman coding needs to count the number of weights before calculating the Huffman tree to get Long encoding, which undoubtedly increases the burden on the computing system

At present, the process of Huffman encoding and decoding does not take hardware characteristics into account, and the implementation effect on FPGA is not good.

Images