A neural network reasoning structure optimization method and device

Abstract

The embodiment of the invention provides a neural network reasoning structure optimization method. The method comprises the following steps: when an Mth network layer and an (M + 2) th network layer of a neural network reasoning structure are both normalization layers, the (M + 1) th network layer is a convolutional layer or a full connection layer, and the (M + 1) th network layer is connected with the (M + 2) th network layer and the output of the (M + 1) th network layer is only connected with the (M + 2) th network layer, calling a first preset algorithm to process the (M + 1) th network layer so as to merge the (M + 2) th network layer into the (M + 1) th network layer to obtain a first optimized network layer of the (M + 1) th network layer; And calling a second preset algorithm to process the first optimization network layer of the (M + 1) th network layer so as to merge the Mth network layer into the first optimization network layer of the (M + 1) th network layer. According tothe scheme, the calculated amount and the processing delay in neural network reasoning can be reduced to the maximum extent, so that the purpose of improving the neural network model reasoning efficiency is achieved.

Description

technical field [0001] This application relates to the technical field of machine learning, in particular to a method and device for optimizing a neural network reasoning structure. Background technique [0002] Batch Normalization (BN) is an algorithm born to overcome the difficulty of training due to the deepening of neural network layers. It is a step toward reducing internal covariate shifts, which can reduce the dependence of gradients on the scale of parameters or their initial values, with beneficial effects on the flow of gradients through the network. In addition, in order to speed up the gradient descent convergence speed of the training neural network, the neural network generally adds normalization processing operations to the input data. [0003] Due to the addition of BN or normalization processing, the neural network reasoning structure has increased in terms of processing depth and calculation volume compared with the previous ones, and at the same time incr...

AI Technical Summary

Problems solved by technology

[0003] Due to the addition of BN or normalization processing, the neural network reasoning structure has increased in terms of processing depth and calculation volume compared with the previous one, and at the same time increases the processing delay, which has a negative impact on the efficiency of neural network model reasoning deployment.

Images