Inference speed

[vitae-transformer]

Inference speed

[Annbless]

Hi, thanks for your interest. We will update the detailed comparisons in the paper.
Roughly speaking, in our experiment environments, the ViTPose-b model runs a little faster than HRNet-w32, i.e. over 900fps on a single A100GPU without the flip test.

[levan92]

Alright, will look out for your updated paper 😄 Thank you!

[levan92]

Sorry @Annbless just to clarify, you were saying the ViTPose-b model is over 900FPS on A100? Or is it a typo, 90FPS instead?

[Annbless]

900 fps, with 64 batch size, 256x192 input resolutions, and without flip test, We just iteratively feed the inputs into the model for calculation after several warmup iterations. The average speed is calculated then. We have updated the results in the paper, please refer to arxiv.

The code for the speed test is rather simple as follows,

[levan92]

I see! That makes sense 😄 Thanks for sharing the detailed answer!