Using the code provided by the authors I was unable to get results consistent with those in the paper, with Spearman correlations of only 0.75 (CIfar10), 0.73 (CIfar100), and 0.68 (Imagenet16-120) for the three datasets on NASbench201.
We supplemented the experimental results obtained. We don’t know where the problem lies. All computation on AZ score used the code provided by the authors.
Using the code provided by the authors I was unable to get results consistent with those in the paper, with Spearman correlations of only 0.75 (CIfar10), 0.73 (CIfar100), and 0.68 (Imagenet16-120) for the three datasets on NASbench201.