GPU support

[Tenuro]

Hello,
did you manage to use Tensorflow with GPU support?
I am currently trying to run VHAP using the GPU but in the "Align and track faces" step, tensorflow always gives me an error:

tensorflow/tsl/cuda/cudart_stub.cc:28] Could not find cuda drivers on your machine, GPU will not be used.

However, torch.cuda.is_available() returns true, so I would assume that it is not an issue with the driver.
I am following your installation instructions leaving me with
CUDA 12.1
tensorflow 2.12

Best regards

[ShenhanQian]

Hi, this repo mainly relies on PyTorch. Tensorflow with gpu support is not required. Can you please make sure you strictly follow the installation instructions from scratch?

[Tenuro]

Just for clarification. It still works. It is not an error but an information.
The preprocessing still works but I assumed that the GPU is not used, due to the message and the time.
Currently it takes approximately 6 hours to preprocess a 1-2 minutes video.

I will try to follow it again from scratch.

[ShenhanQian]

I see. I think you won't get much speed up by reinstallation. PyTorch is the major part, while Tensorflow is only used for Tensorboard. The current repo is not optimized in terms of efficiency. It usually takes 10-20 seconds to optimize on a frame. You may try optimize fewer steps on each frame for acceleration.