An efficient neural architecture search based on PCDARTS

Deep learning has achieved great achievements in a variety of tasks, such as image classification, speech recognition, and cultural heritage protection. However, manually design deep convolution neural networks for certain tasks requires a long time. Furthermore, it also requires the designer to have certain professional knowledge. Faced with an increasingly complex network architecture, it has become more and more difficult to design only manually. Neural architecture search provides a method to overcome problems in manually designing networks. It mainly includes search space, search strategy, and performance evaluation. In this paper, a new search method, F-PCDARTS, is proposed based on PCDARTS. In the research space, we propose to apply channel shuffle and global context for efficient feature extraction. At the same time, in terms of search strategy, a fixed reduction cell is designed to search normal cells, which solves the difficult and time-consuming problem of search optimization. The experiment results indicate a 9.8% reduction in network search time, requiring only 0.09 GPU days. Furthermore, the validation accuracy of CIFAR10 and CIFAR100 is improved by 0.09% and 0.12%, respectively. The method we propose effectively reduces the time of network search.