Research On Optimization Of Neural Architecture Search Algorithm Based On DARTS

Image classification has always been an important research task in the field of computer vision,and Convolutional Neural Network(CNN for short)plays a key role in image classification tasks.However,the early CNN architectures were all manually designed by experts,which required high professionalism,and finding the correct and suitable architecture was a time-consuming,laborious and error-prone task,which limited the application of CNN.In order to break through the bottleneck of manually designing the neural network architecture,researchers proposed the Neural Architecture Search(Neural Architecture Search,referred to as NAS)algorithm.Among them,The Differentiable Architecture Search(DARTS)model uses a gradient-based differentiable search strategy for architecture search,which effectively solves the problem of high computing and high cost overhead in traditional NAS algorithms,and has become the main object of current NAS algorithm research.However,DARTS uses a stacked network of agents with shared weight parameters in the search phase,which affects the stability of the model and has problems of discontinuous depth and width in the search phase and the evaluation and verification phase.In order to solve the above-mentioned problems in DARTS and improve the performance of the model architecture,this paper carried out the following optimization research:(1)In response to the problems of the current DARTS model,the accuracy of which can be optimized and the lack of long-range information acquisition,two optimizations are made in the search space,one is to propose a new operator to make the model have a larger sensory field and thus can obtain more local information,and the other is to add a residual structure to the macroscopic network architecture to strengthen the information transfer process,the improved model DARTS-VAN proposed by the above two methods reduces the instability of the original model and improves the accuracy rate.(2)To address the problem of the depth gap in the search strategy of DARTS,this paper proposes a dual progressive search strategy,which performs a progressive search in the depth of the model architecture and also performs a progressive search on the channel dimension of the width level,making the model architecture more continuous between the search phase and the evaluation phase,aiming to reduce the architectural performance difference between the two phases.(3)For the problem that the double progressive search strategy occupies large video memory and takes a long time to search the topology of the cell,this paper adopts a partial channel trimming search strategy for acceleration.A part of the channels in the feature matrix is subjected to search operation,and the rest of the channels are directly skipped operation,and finally the two parts are combined to get the new feature matrix,which greatly reduces the search time and computational cost.(4)In this paper,we propose a double progressive and fast differentiable architecture search model DPF-DARTS by combining the dual progressive search strategy and the partial channel pruning search strategy,and introduce an edge regularization method for the weak parameter operation aggregation problem such as jumping connection of the model caused by using the partial channel pruning strategy,where the weight of each edge is jointly determined by the architecture parameters and the edge weight parameters to enhance the overall stability of the model.This paper analyzes the reduction of model search time,the optimization of architecture performance and the effective solution to the defects of DARTS by the above improved method from the theoretical level.Through experiments,this paper verifies the effectiveness of the two models on the CIFAR-10 dataset.In addition,this paper also migrated the model to the ImageNet dataset to verify the migration of the model,and also achieved good and competitive experimental results.