DLC 3.0: Minimal GPU usage #2639

Is there an existing issue for this?

• I have searched the existing issues

Bug description

I'm starting out with some tracking again, now with the new PyTorch backend. I'm analysing some massive videos, 2 hours at 60fps and dimensions w=1280, h=720. Currently I'm getting just over 7fps when I'm analysing videos. The issue is not that PyTorch is not registering the GPU - I had that issue (was getting 1.3fps before when torch.cuda.is_available() returned False, now it returns True), but even with the GPU being accessible, it's not really being used much, sitting at about 10%, and the Python thread running DLC is using less than 1% of that.

Operating System

Windows 10

DeepLabCut version

DLC 3.0.0rc1

DeepLabCut mode

single animal

Device type

Nvidia Quadro M5000, 8GB

Steps To Reproduce

1. Install DLC 3.0.0rc1 as per the installation instructions for Windows.
2. Start tracking videos
3. Watch the GPU usage in the Task Manager

Relevant log output

No response

Anything else?

Here see the total GPU usage:

Here, see that the Python process uses less than 1% GPU:

Again, this time from nvidia-smi, the memory usage is less than 10% (812Mb/8192Mb):

Code of Conduct

• I agree to follow this project's Code of Conduct