Edge calculation-oriented reparametric neural network architecture search method

Abstract

The invention discloses an edge calculation-oriented reparametric neural network architecture search method. The method comprises the following steps of 1, designing a linear operator and multiple branch block structures; 2, constructing a super network by stacking the multiple branch block structures; 3, training the super network through a gradient-based first-stage search algorithm; 4, deleting redundant branches in the super network to construct an optimal sub-network; 5, converting the multi-branch optimal sub-network into a single-branch network; and 6, completing task reasoning by using the single-branch network. The method is used for searching the neural network structure capable of performing re-parameterization, and ensures the reasoning real-time performance and the high efficiency of model operation while ensuring the reasoning precision.

Description

technical field [0001] The invention relates to the technical field of neural network architecture search, in particular to an edge computing-oriented heavy-parameter neural network architecture search method. Background technique [0002] Neural network architecture search is a research hotspot in the field of machine learning in recent years. This technology includes the design of search operators and spaces, and the design of search algorithms. At present, neural network architecture search technology can be used to automatically design neural network models of various sizes, avoiding manual complex parameter adjustment. Among them, one of the most potential applications is designing a lightweight neural network model to improve the application capabilities of neural networks on mobile devices. [0003] In mobile devices, the real-time and accuracy of neural network inference are two major considerations. In the early artificially designed lightweight neural network mod...

AI Technical Summary

Problems solved by technology

[0005] Although the RepVGG series models have greatly improved the actual inference speed of the model, because the structure of the branches is artificially fixed, there is still a lot of room for improvement in the accuracy of the network model

In addition, too many branches will greatly increase the memory space required to train the network model

Images