search

CanyonWind / Single-Path-One-Shot-NAS-MXNet

Single Path One-Shot NAS MXNet implementation with full training and searching pipeline. Support both Block and Channel Selection. Searched models better than the original paper are provided.
151 stars 22 forks source link

supernet training with resource constrain #14

Closed pawopawo closed 4 years ago

pawopawo commented 4 years ago

Thanks for your excellent work!

the supernet trained with this strolling evolution method could be more trustworthy when using it to sample the subnet's performance.

Have you compared the results of searched acrhitecture between strolling evolution and random?

CanyonWind commented 4 years ago

Hi, this is a very good question! Sorry I haven't compared them... yet.

The background of why this was done is that when I tried to search a subnet on a not constrained supernet with fewer FLOPs but more # parameters (the top left corner here), the searched model's real performance is not aligned with what being sampled from the supernet.

So I assumed the reason could be that the top left part has not been trained and adding constraints on supernet might help. But the random constraints obviously also cannot touch that region neither. so I just skipped to try it...

You are very welcomed to give them a try and feedback will be greatly appreciated!

CanyonWind commented 4 years ago

Close the issue for no further response. Please feel free to reopen if necessary.