search

renqianluo / NAO

Neural Architecture Optimization
GNU General Public License v3.0
286 stars 66 forks source link

Is parameter N in train_cifar10_search.sh the same as N in paper? #3

Closed Yuzz1020 closed 5 years ago

Yuzz1020 commented 5 years ago

Hey, I'm still new to this area, but it confuses me that parameter N in train_cifar10_search.sh is multiplied by 3 when generating model. However, in the paper, it seems that N is the total normal convolution cell used to construct final model. So I'm wondering if these two Ns are the same or is N in code equals 3 * N in the paper?

renqianluo commented 5 years ago

@noNameForMe1994 Hi, recent works on NAS commonly follow a pattern of: stem head - convolution cell N - reduction cell - convolution cell N - reduction cell - auxiliary head - convolution cell N - reduction cell (please refer to NASNet, ENAS...). So you see there are totally N3 convolution cells and 2 reduction cells. The N in the script is the same with the N in the paper (we use N=6 as a common practice).