Skip to content
This repository has been archived by the owner on Jul 3, 2024. It is now read-only.

Handling isomorphic architectures in benchmarked algorithms ? #6

Open
mikelzc1990 opened this issue Mar 9, 2019 · 1 comment
Open

Comments

@mikelzc1990
Copy link

I was wondering if the benchmarked algorithms in the paper (Figure 7) handle isomorphic architectures. To be more specific, when an algorithm creates a new architecture, do you check against all evaluated architectures for isomorphism or not?

The random search and the regularized evolution methods implemented in your NASBench.ipynb seem to only check if an architecture is valid. I did a quick experiment that collects the MD5 hash for each evaluated architecture. The results show that, on average over 11 repeats, more than 98% of the architectures created by regularized evolution are isomorphic.

@aaronkl
Copy link

aaronkl commented Mar 13, 2019

Hi,
no we do not check for isomorphism, since all methods that we used are not able to handle that natively. But I think it would be very interesting future work to somehow exploit isomorphism inside the NAS algorithm.

Sign up for free to subscribe to this conversation on GitHub. Already have an account? Sign in.
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants