search

changlin31 / DNA

(CVPR 2020) Block-wisely Supervised Neural Architecture Search with Knowledge Distillation
235 stars 35 forks source link

The size of the search space #28

Closed zhengjian2322 closed 2 years ago

zhengjian2322 commented 3 years ago

Hi, I really enjoy your paper! I have a question, why the size of the search space in section 4.1 is 2×10e17?

changlin31 commented 3 years ago

Hi,

Given num_op = 6, search space is calculated by (num_op ^ num_layer). By adding the search space of three different supernets, we have:

block 1 search space: 6^2 + 6^3 + 6^2 = 288 block 2 search space: 6^2 + 6^3 + 6^4 = 1548 block 3 search space: 6^2 + 6^3 + 6^4 = 1548 block 4 search space: 6^3 + 6^4 + 6^4 = 2808 block 5 search space: 6^4+ 6^5 + 6^5 = 16848 block 6 search space: 6

total search space: 288 × 1548 × 1548 × 2808 × 16848 × 6 = 195898498887057408, which is about 2×10^17.

zhengjian2322 commented 2 years ago

Thank you very much! But I have another question why the number of partial models is about 10^4 in each cell?