Discrepancy on reported flops in paper

[yunfeixie233]

Hello,

First of all, thank you for your work on RegionProxy. It has been quite helpful.

I have a question regarding the FLOPs reported for the ViT-Small model in your paper. In the paper, you mentioned that the FLOPs for ViT-Small is approximately 14.9G (table 1, ADE20k, 512x512 input). However, when using the get_flops script from the mmsegmentation library and the fvcore library, the calculated FLOPs for the same architecture come out to be approximately 31.8G.

Could you kindly clarify the methodology used to calculate FLOPs in your paper?