Implementation method of small neural network oriented to programmable logic device mobile terminal

Abstract

The invention relates to an implementation method of a small neural network oriented to a programmable logic device mobile terminal. The implementation method comprises the following steps of S1, obtaining a feature map after an input image passes through two convolution layers; S2, performing nonlinear mapping on the feature map after each convolution through an activation function; performing maximum pooling operation between different convolution layers by adopting a filter with a step length of 2, selecting main features in the feature map, and reducing the dimension of the feature map toobtain a pooling layer output image; and S3, tiling the pooling layer output image obtained in the step S2 into one dimension, and obtaining an output result after the image passes through the full connection layer. According to the method, the model parameters are reduced by about 4 times, the requirement of the parameters for the storage space is reduced, and the calculation requirement is alsoreduced through the 1-bit input data and the 8-bit weight parameters. When the neural network is deployed at a terminal, a lookup table and pipeline parallel computing method can be adopted to improvethe computing speed.

Description

technical field [0001] The invention relates to the field of mobile terminal operation, in particular to a method for realizing a small-scale neural network oriented to a programmable logic device mobile terminal. Background technique [0002] In recent years, deep neural networks have been widely used in image processing, speech recognition, and natural language processing, and have achieved great success. More and more mobile applications adopt deep neural networks to provide accurate, intelligent and effective services. But deploying neural networks on mobile terminals is a big challenge. Because although these neural networks have strong representation capabilities, their large weight parameters require considerable storage space, memory bandwidth, and computing resources, and mobile terminals with small capacity and limited resources are difficult to meet their needs; The neural network needs a large amount of memory bandwidth to obtain weights and a large number of c...

AI Technical Summary

Problems solved by technology

But deploying neural networks on mobile terminals is a big challenge

Because although these neural networks have strong representation capabilities, their large weight parameters require considerable storage space, memory bandwidth, and computing resources, and mobile terminals with small capacity and limited resources are difficult to meet their needs; The neural network needs a large amount of memory bandwidth to obtain weights and a large number of calculations to complete the dot product, which consumes a considerable amount of energy. For mobile terminals, the constraints of battery power make it difficult to deploy neural networks that consume high power, so Neural networks still have great limitations in practical applications

Images