The issues of Computational Complexity Analysis experiment in paper

[JinluZhang1126]

Hi, Dr. Zheng,

Ventured to disturb.

I have some questions about the paper, could you please give me the answer?

1. In the Computational Complexity Analysis experiment of the paper, I can not get the FLOPs per frame as yours in the paper(101M), I get about 76.5M, the computing details:
   spatial attention: 4*(1732172)=73,984≈0.074M;
   temporal attention: 4(81544812)=28,553,472M≈28.6M
   the others like MLP and linear layer total: about 9.6M
   Th total: 2(0.074+28.6+9.6)≈76.5M
   Could you tell me your computing details?
2. Still in the Computational Complexity Analysis experiment of the paper, the proposed poseformer model achieve the 269 FPS inference speed, but I can not get that number by using your GitHub code. I tried the torch.cuda.Event(enable_timing=True) function to compute the time, and get the number_of_frames by using num_frame() function in the generator class. However, I can still not get the FPS in the paper by computing “time/number_of_frames”(approximately 800 FPS on the single 2080Ti GPU).
   Could you tell me the details of doing it?
   Sorry for taking your time, hope to get a reply!

Best wishes,
Jinlu