frame-semantic-transformer icon indicating copy to clipboard operation
frame-semantic-transformer copied to clipboard

Published 20 hours ago • verified publisher icon chanind

GPU

Open sdspieg opened this issue 2 years ago • 2 comments

Any suggestions for which versions of Tensorflow, CUDA and cuDNN to run?

sdspieg avatar Feb 27 '23 13:02 sdspieg

This library uses PyTorch under the hood rather than Tensorflow, but other than that whatever version of CUDA will work with the version of PyTorch you want to install should be fine. It looks like the most recent version of PyTorch (1.13.1) works with CUDA 11.6 or 11.7 (https://pytorch.org/get-started/locally/#start-locally).

chanind avatar Feb 27 '23 17:02 chanind

Just sharing progress (from a Windows machine with a GPU)

  • conda create --name pytorch python=3.10
  • conda activate pytorch
  • conda install pytorch torchvision torchaudio pytorch-cuda=11.7 -c pytorch -c nvidia

That SEEMS to work - I see that my gpu is using 6GB of VRAM. Not sure about time improvement, though...

sdspieg avatar Mar 20 '23 20:03 sdspieg