Netural Architecture Search Based On One-Shot Model

The neural architecture is crucial to the computing performance of deep learning tasks.Traditional manual design of network architecture is not only time-consuming and laborious,but also has limited design ideas because of previous researchers' design ideas,and the performance of the model is difficult to be guaranteed.Therefore,in recent years,neural network architecture search has become one of the most popular research directions in the field of deep learning.Although there are so many methods for automatically designing neural network structures,most of them still require a lot of computing resources and cannot be widely used.This topic analyzes and summarizes the strengths and weaknesses of the existing neural network architecture search methods.Based on the one-shot model in neural network architecture search,several improved technologies and methods are proposed,and the effectiveness and efficiency of the proposed methods are proved by experiments.And for the image classification task,the model architecture obtained by most existing neural network architecture search methods and manual design is obtained.Firstly,this paper focuses on the training process of the traditional one-shot model,analyzes and summarizes the deficiencies,and proposes a single-path training method in the super-network training stage,which greatly reduces the resource requirements for super-network training.An effective coding method for complex multi-branch search space is proposed to improve the efficiency of network sampling and selection.In the stage of network structure search,a network structure search method based on evolutionary algorithm using network structure coding is proposed to improve the performance of model search.Secondly,because the search network is too large during the training of the super network,the selection of the sub-network model is too broad and not targeted.In this paper,the search space tailoring method is designed and implemented,using the intermediate performance results generated during the training of the supernet Perform masking restrictions to achieve the purpose of cropping the search space.The experimental results show that by cutting the search space during the supernet training,the training effect of the supernet and the search results of the sub-network model can be improved.Finally,aiming at the traditional supernet training methods generally pursuing the balance of network training and thereby ignoring the importance and effective information of the excellent network sub-models that have been trained earlier,this paper proposes a buffer pool-based neural network architecture search algorithm.Successfully use the intermediate information of network training to guide the training and model search process of the subsequent supernet.And this algorithm obtains a very competitive convolutional neural network model for image classification tasks.