EverybodyDanceNow_reproduce_pytorch icon indicating copy to clipboard operation
EverybodyDanceNow_reproduce_pytorch copied to clipboard

Published 20 hours ago • verified publisher icon yanx27

GPU under-utilisation

Open DhavalUnifynd opened this issue 5 years ago • 1 comments

I have tried executing the code on 1060-6GB variant and Tesla K80 12GB GPU. The GPU is heavily under-utilized on both the GPU's. Can you suggest some workaround?

DhavalUnifynd avatar Mar 28 '19 12:03 DhavalUnifynd

Im using the same hardware as you. It will take around 2 days to estimate poses and detect the head in for a 14 min video when running make_target.py on a Tesla K80 GPU. The code looks very sequential. It should be possible to add multiprocessing to do as much ops in CPU while loading GPU with batches. But is a lot of work to refactor the code like this.

slothkong avatar Mar 24 '20 08:03 slothkong