1. Introduction
Despite the fast development of neural architecture search (NAS) [52] to aid network design in vision tasks like classification [32], [6], [41], [42], object detection [11], [36], and segmentation [22], there has been an urging demand for faster searching algorithms. Early methods based on the evaluation of a huge number of candidate models [52], [31], [16] require unaffordable costs (typically 2k GPU days). In the light of weight-sharing mechanism introduced in SMASH [2], a variety of low-cost approaches have emerged [1], [27], [24]. DARTS [24] has taken the dominance with a myriad of follow-up works [38], [3], [39], [10], [5], [47]. In this paper, we investigate a single-path based variation of DARTS, typically GDAS [10], for its fast speed and low GPU memory.