Closed thucz closed 2 months ago
I did not.
I did not use it because I think the positional embedding may prompt the classification performance, but may also affect the ability to accept images with different size. Simply interpolating the positional embedding is clearly a suboptimal solution, so I decided to avoid this problem by omitting it.
Thanks for your reply. I get it.
Hi! I'm very curious about the parameter
posembed
. Previous methods such as ViTs use positional embedding. But in your method, you setposembed=False
. Have you done ablation study about this?