adaptivetokensampling/ATS

Has anyone reproduced the results of the paper report?

Opened this issue · 2 comments

Has anyone reproduced the results of the paper report?

I can only get 79.13% acc@1 on ImageNet under off-the-shelf, and I find the average FLOPs is 3.4G instead of 2.9G as reported in the paper (I sum up the calculations on each batch and draw a tie with profile_macs().)

And my model is Deit-S-ATS