Code for search & train WITHOUT weight sharing

[killawhale2]

The authors report the results for the random search baseline without any weight sharing but do not provide the code to reproduce the baseline's result. This would be helpful in fully reproducing their tables in the results section and not just their weight-shared versions.

[liamcli]

This code is provided in a repository linked from the README: https://github.com/liamcli/darts_asha

Note that you won't be able to get exactly reproducible results for the search phase due to asynchronous promotions in the parallel setting.

[killawhale2]

Thanks for the quick reply. I've checked the repo again, and I still only see code for running random search with ASHA. I'm mentioning this because the paper claims the random search baseline takes 4 gpu days while the random search with ASHA takes 9 gpu days for the cnn experiments. I'd like to reproduce the results for the base random search as that is more computationally viable for my resources.

[liamcli]

The random search baseline that takes 4 GPU days is from the original DARTS paper. You can just use the ASHA codebase with no early-stopping to run random search for as long as you want.