TensRT issue

[Alpgirl]

Good day!
I successfully set up the environment. When I launch src/theia/scripts/preprocessing/feature_extraction.py , there is a warning: "tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Could not find TensorRT". I installed TensorRT from https://developer.nvidia.com/tensorrt and via pip, but I still catch the same warning.
Could you please explain how critical is to set up TensorRT and how to install it properly?
Thank you.

[jshang-bdai]

Hi @Alpgirl ,

TLDR: you can safely ignore it.

This warning comes along with Tensorflow. We do not use Tensorflow except for reading Open-X-Embodiment dataset. We explored this dataset at our early study but we didn't use it eventually.

[Alpgirl]

Actually, I am interested in Open-X-Embodiement dataset...
Do you know if TensorRT accelerates computation of features?

[jshang-bdai]

I don't think so because our model is based on PyTorch.

Also a kind note, if you don't plan to use SAM as a teacher model, computing teacher features on-the-fly won't cost much more time. SAM features may take huge storage space for a large dataset like OXE.

[Alpgirl]

@jshang-bdai , thank you!

[jshang-bdai]

Sorry @Alpgirl let me correct my comment on TensorRT. It's not dependent on Tensorflow or Pytorch.

After a brief search, I think you could use it, for example https://huggingface.co/docs/optimum/en/onnxruntime/usage_guides/gpu#accelerated-inference-on-nvidia-gpus. Sorry for the confusion!

[Alpgirl]

@jshang-bdai , thank you for the search. But I have been unsuccessful in installing TensorRT yet. Will try more.