Neural network structure obtaining method and device, and storage medium

Abstract

The invention discloses a neural network structure obtaining method and a device, and a storage medium. The method comprises the steps of acquiring a first parameter of a first neural network structure, and the first parameter at least comprises a first topological relation of all modules in the first neural network structure and a first weight corresponding to each edge in the first topological relation; training a first parameter of the first neural network structure by using the test task data corresponding to the group of test tasks, and obtaining a second parameter of the second neural network structure, the second parameter at least comprising a second topological relationship of each module in the second neural network structure and a second weight corresponding to each edge in thesecond topological relationship, and the second parameter being used for generating a corresponding neural network structure for different tasks. The technical problem that in the prior art, the searched neural network structure is applied to different tasks, and time consumed by multiple times of searching of the different tasks is long is solved.

Description

technical field [0001] The present invention relates to the field of neural networks, in particular to a method and device for obtaining a neural network structure, and a storage medium. Background technique [0002] In recent years, deep neural networks have achieved great success in the fields of artificial intelligence such as computer vision, natural language processing, speech recognition and other tasks. Behind these successes, the neural network structure plays a role that cannot be ignored. For example, the recent ResNet and DenseNet provide strong benchmarks for several computer vision tasks. However, most of the current neural network structures are manually designed through human prior knowledge, which is difficult to deploy automatically. The recently popular neural network search (NAS) can automatically search the network structure for specific tasks, belongs to the category of automatic machine learning, and has achieved world-leading results on multiple data...

AI Technical Summary

Problems solved by technology

However, in these methods, one needs to search different networks for different tasks, which is sometimes repetitive and time-consuming

[0008] For the above problems, no effective solution has been proposed

Images