Network structure processing method of deep neural network and related products

Abstract

An embodiment of the present invention provides a network structure processing method, device and medium of a deep neural network, wherein the processing device includes a neural network coding calculation module and an evolutionary algorithm module; the method includes: the evolutionary algorithm module according to preset target performance requirements Computing candidate structure codes, the candidate structure codes include attributes of multiple neural network layers included in the candidate deep neural network; the evolutionary algorithm module sends the candidate structure codes to the neural network code calculation module; the neural network The code calculation module receives the candidate structure code sent by the evolutionary algorithm module; the neural network code calculation module constructs the network structure of the candidate deep neural network according to the candidate structure code. The embodiment of the present invention can construct a target deep neural network for a specific application scenario.

Description

technical field [0001] The present invention relates to the technical field of neural networks, in particular to the technical field of neural network structure processing, and in particular to a deep neural network network structure processing method, a processing device and a computer storage medium. Background technique [0002] At present, in conventional application fields (such as image recognition, speech recognition, etc.), there are already a lot of better network structures of deep neural networks available for reference. Therefore, these existing deep neural network structures can be called in these conventional application fields. The network structure is used as the initial deep neural network, and then common toolkits (such as TensorFlow (an artificial intelligence programming tool), PyTorch (an artificial intelligence programming tool), etc.) are used to carry its calculation and training, so as to achieve the goal of meeting actual needs Network structure of ...

AI Technical Summary

Problems solved by technology

However, the above-mentioned network structure processing scheme is only applicable to the general application field, which already has a reproducible and reference deep neural network network structure; if an unconventional new field (such as agricultural data analysis and other exploratory scenarios), since there is no suitable deep neural network network structure that can be used directly, the above-mentioned network structure processing scheme is not suitable

Therefore, how to design the network structure of the deep neural network suitable for new scenarios for unconventional application fields is a key issue; generally, engineers will design the deep neural network architecture artificially based on their own personal experience. The resolvability of the deep neural network designed in this way is poor; the training of the deep neural network consumes a lot of computing resources and computing time. If the direction of network construction optimization is not clear, it will lead to a serious waste of time and resources.

Images