search

JasonWayne / deep-learning-essay

0 stars 0 forks source link

NAS-BENCH-201: EXTENDING THE SCOPE OF REPRODUCIBLE NEURAL ARCHITECTURE SEARCH #23

Open JasonWayne opened 4 years ago

JasonWayne commented 4 years ago

https://www.semanticscholar.org/paper/NAS-Bench-201%3A-Extending-the-Scope-of-Reproducible-Dong-Yang/86d378df2221e16c339bb5a482f20e4ccda4cb5b#related-papers

https://arxiv.org/pdf/2001.00326.pdf

https://arxiv.org/abs/2001.00326

JasonWayne commented 4 years ago
  1. 提供了一个有xx个结构，用同一套超参，训练至收敛，包含推理速度等信息的BenchMark。
  2. 提供了多个NAS算法的实验结果对比。
  3. 和NAS-Bench-101的区别：本文提出的结构，搜索空间没有约束，支持更多种算法，给出了3个数据集，多种metric的评估结果，

https://www.semanticscholar.org/paper/NAS-Bench-201%3A-Extending-the-Scope-of-Reproducible-Dong-Yang/86d378df2221e16c339bb5a482f20e4ccda4cb5b/figure/2

JasonWayne commented 4 years ago

https://d3i71xaburhd42.cloudfront.net/86d378df2221e16c339bb5a482f20e4ccda4cb5b/8-Table5-1.png

文章的实验结果看出，Regularized Evoluation效果最优，REINFORCE次之。