Notes: RETHINKING ARCHITECTURE SELECTION IN DIFFERENTIABLE NAS
shaw

简介

提出DARTS中的并不能很好的反应operator的strength。因此设计了基于扰动的评价指标来选择operator,在实现上只需要修改DARTS的operator选择阶段(即离散化阶段)。实验证明这个方法不仅能一致提升DARTS类方法的表现,而且能解决DARTS的鲁棒性差的问题。

方法

在darts的supernet训练好之后,进行perturbation-based architecture selection
(PT),包括两点:

  1. 去掉一个operator,用accuracy下降的大小来作为operator选择的依据
  2. 逐步离散化edge,每确定一个edge上的operator就fintune一次
    具体如下
    perturbation-based architecture selection

实验与结论

进行了如下实验:

  1. Comparison with state-of-the-art image classifiers on CIFAR-10. 在现有的DARTS类方法上加上了PT,分数都得到了提升。
  2. Robustness issue of DARTS can be explained by the failure of
    magnitude-based architecture selection. 实验验证了DARTS在某些搜索空间中失效的现象可以通过PT避免
  3. test a baseline by combining progressive tuning with magnitude-based operation selection instead of our selection criterion。证明论文提出的评价指标确实更好体现了operator的strength。
  4. 实验证明了在supernet的训练过程中可以固定为0(即去掉),使用PT后性能没有下降
  • Post title:Notes: RETHINKING ARCHITECTURE SELECTION IN DIFFERENTIABLE NAS
  • Post author:shaw
  • Create time:2021-09-14 21:15:53
  • Post link:https://www.zenwill.top/2021/09/14/Rethinking gradient-based nas/
  • Copyright Notice:All articles in this blog are licensed under BY-NC-SA unless stating additionally.
 Comments