Hi, thanks for your great work. After reading your paper, I have some doubts.
1 or n architectures are sampled when using n GPUs for searching?
The ok sampled from Gumbel(0,1) has a range of known values, but the A^k{i,h} may vary a lot when adapting different wd. So what's the best range for A?
Thanks for your interest.
1, n architectures for n GPUs
2, good point. It could be a problem. We initialize A with Gaussin(0, 0.001) and do not constrain it during searching.
Hi, thanks for your great work. After reading your paper, I have some doubts.
Thanks.