search

chenxin061 / pdarts

Codes for our paper "Progressive Differentiable Architecture Search:Bridging the Depth Gap between Search and Evaluation"
Other
360 stars 83 forks source link

Question regarding the paper #35

Open yxchng opened 4 years ago

yxchng commented 4 years ago

Paper:

Screenshot 2020-10-23 at 4 23 28 PM

Code: https://github.com/chenxin061/pdarts/blob/master/train_search.py#L247-L262

Screenshot 2020-10-23 at 4 25 30 PM

I can't quite align the paragraph written in the paper and the corresponding code write in this repo. From what I understand from the code, M is 8 and when there are more than 8 skip connections, the code will generate 8 genotypes by keeping 1-8 skip connections with the largest weights. Is that correct? If so, I am not quite sure what is meant by "this may bring up other skip-connects to the topology, so we repeat this procedure until the desired number is achieved." as this does not seem to happen in the code.

Do you mind explaining more about this? Thank you.

chenxin061 commented 4 years ago

Sorry for the late reply. Your understanding is correct. The while loop executes the repeating procedure. Actually, the repeating procedure is not necessary since zeroing all other sk-probs except the largest sks ones can achieve the same goal. In practice, these two forms do not make a large difference in search cost so I just keep them unchanged.