Hierarchical network structure search method, device and readable storage medium

Abstract

The invention provides a hierarchical network structure search method, a device and a readable storage medium. The method comprises the steps of S1, constructing a super network; S2, acquiring image data and taking the image data as training data of network parameters and structure parameters of the super network; S3, generating a feature map, calculating a cross entropy loss function of the network parameters, and updating the network parameters of the super network; S4, generating a feature map and a KL divergence loss function, calculating a cross entropy loss function of the structure parameters to obtain a semi-implicit variational discard loss function, training and updating the structure parameters of the super network and obtaining a discard probability; s5, updating the basic unitby using the discard probability, and updating the annealing parameter; s3 to S5 are repeated, and the network parameters and the structure parameters are updated; s6, obtaining a final network. Themethod greatly reduces the search time and reduces the calculation complexity while guaranteeing the higher performance, guarantees the search stability and practicality, can be used in the fields ofimage target detection and classification, and improves the image processing speed.

Description

technical field [0001] The invention relates to the technical fields of artificial intelligence and image processing, in particular to a hierarchical network structure search method, computer equipment and a readable storage medium thereof, and the application of the method in image target detection, semantic segmentation and classification. Background technique [0002] With the development of computing power and deep learning, the development of artificial intelligence is getting faster and faster. In the initial image processing, due to the low quality of the collected images, the requirements for feature extraction are relatively low, so the main technology is manual feature extraction. Later, the image quality continued to improve, and the accuracy requirements for image processing continued to increase. Some statistical pattern recognition methods such as SVM and signal processing methods such as wavelet transform have made some progress in the development of image pro...

AI Technical Summary

Problems solved by technology

Secondly, the existing search method based on the DARTS search space will collapse due to the excessive number of search epochs when searching the network. This collapse phenomenon is reflected in the fact that the final structure becomes a fully connected neural network without parameters. Completely incompetent on image classification tasks

And this collapse phenomenon also makes when the network structure search is directly used for image classification or other downstream tasks, because it is impossible to determine when the network structure converges, resulting in insufficient performance of downstream tasks

Images