Super-network construction method and device, super-network use method and device and medium

Abstract

The invention discloses a super-network construction method, a super-network use method, a super-network use device and a medium, and the construction method comprises the steps: setting a linear connection unit in at least one layer of a super-network, the input end of the linear connection unit being connected to the upper layer of an affiliation layer of the linear connection unit, and the output end of the linear connection unit being connected to the lower layer of the affiliation layer; wherein the output and the input of the linear connection unit form a linear relationship, and the linear relationship comprises a linear relationship except that the output is equal to the input. According to the method, the linear connection unit is used in the super network, and compared with a direct connection unit, the characterization capability of the sub-network can be effectively improved, the stability of the network indexes of the sub-network comprising the linear connection unit in the super network is maintained, and the network indexes of the sub-network are prevented from being rapidly reduced in the deep adjustment process.

Description

technical field [0001] This article relates to the technical field of data processing, and in particular relates to a hypernetwork construction method, usage method, device and medium. Background technique [0002] Neural Architecture Search (NAS) is a technology for automatically designing neural networks, which can automatically design high-performance network structures based on sample sets through algorithms. In neural network architecture search, multiple searches are required to generate a separate neural network and obtain network indicators after training, which has the problems of low evaluation efficiency and low search speed. To solve this problem, some NAS methods use a hypernetwork that contains all search networks. Such as figure 1 In the structure of an exemplary supernet (supernet) shown, the supernetwork contains a plurality of layers, and each layer contains a plurality of network units, and a network unit is selected from each layer to connect sequential...

AI Technical Summary

Problems solved by technology

In neural network architecture search, multiple searches are required to generate a separate neural network and obtain network indicators after training, which has the problems of low evaluation efficiency and low search speed

Images