which genotype is best？

Question

which genotype is best？

johsnows opened this issue 5 years ago · comments

after search on the train_search.py, we get many genotype, no restrict skipconnetct genotype and genotypes with number of skip-connect:7,6,5,4,3,2,1,0, which one is best? which one is recommended to evaluate?

Arash Vahdat · Answer 1 · Wed Sep 18 2019 03:05:33 GMT+0800 (China Standard Time)

This can help us too. Will appreciate authors' feedback!

chenxin061 · Answer 2 · Wed Sep 25 2019 14:20:15 GMT+0800 (China Standard Time)

Sorry for the late reply. According to our experiments, keeping 2 skip-connects is recommended. You can refer to our paper for more details.

Arash Vahdat · Answer 3 · Thu Sep 26 2019 00:15:38 GMT+0800 (China Standard Time)

Thanks for the clarification!

Neda Azarmehr · Answer 4 · Mon Jan 27 2020 21:00:58 GMT+0800 (China Standard Time)

@chenxin061 do you recommend keeping 2 skip-connects genotype for any kind of dataset (eg. medical image dataset)?
And the genotype printed at the end is it genotype for the last epoch? As far as I understood, DARTS was giving the genotype after each epoch in the search stage, and we should select the genotype that gives the highest accuracy. I am wondering which genotype should I use for evaluation in PDARTS.

Thank you in advance,

chenxin061 · Answer 5 · Mon Jan 27 2020 21:25:35 GMT+0800 (China Standard Time)

@chenxin061 do you recommend keeping 2 skip-connects genotype for any kind of dataset (eg. medical image dataset)?
And the genotype printed at the end is it genotype for the last epoch? As far as I understood, DARTS was giving the genotype after each epoch in the search stage, and we should select the genotype that gives the highest accuracy. I am wondering which genotype should I use for evaluation in PDARTS.

Thank you in advance,

We found that for image classification (on CIFAR and ImageNet), keeping 2 skip-connects in the genotype forms a good balance between accuracy and model size. For other tasks and datasets, I believe that keeping 2 can also work well but I recommand you to do some ablation study on your own dataset.
The genotype printed at the end is from the last epoch. From my side, I recommand you to use the genotype with two skip-connects or the one fits your dataset best according to the additional ablation study.