adaptivetokensampling/ATS

How to Calculate the FLOPs of DeiT-ATS

ChuanyangZheng opened this issue · 2 comments

Your adaptive token sampling attracts me a lot.
I am curious how you calculate the FLOPs of DeiT-ATS in Table 1, since the number of selected tokens varies for all images as shown in Figure 6 in paper. Do you calculate the mean of the number of selected tokens for all images?
Thanks~

Thanks for your interest!
Yes, as mentioned in the paper, the reported GFLOPs are averaged over the samples of the validation set of each dataset.

Can you share the method that you calculate the FLOPs?I find the FLOPs of DeIT-s-ATS on ImageNet validation is 3.4 instead 2.9 reported in the paper?