(1) Have you tried to do stage1 search by setting num_cells = 8 on CIFAR-10? I ended up with a architecture not having any skip_connects either in the normal or reduction cells and they performed poorly in my test. It seems one-level optimization is more immune to skip_connects.
(2) Would you please share the codes to estimate correlation among a given set of primitive operations?
Dear author
Thanks for sharing this great work!
(1) Have you tried to do stage1 search by setting num_cells = 8 on CIFAR-10? I ended up with a architecture not having any skip_connects either in the normal or reduction cells and they performed poorly in my test. It seems one-level optimization is more immune to skip_connects.
(2) Would you please share the codes to estimate correlation among a given set of primitive operations?
Appreciate any response and thanks again!!
Best, Bryan