Installing Pytorch3d without Cuda

[YanivHollander]

Hi,

I am trying to install Pytorch3D on a Ubuntu 20.04 LTS which happens to not have GPU. Is it possible to install Pytorch3d for CPU only runs (for debugging purposes, e.g.)? If so, what is the workflow I should take to install this package in a Python 3.8.10 based virtual environment?

[bottler]

Yes, this works. You need to install from source. Start with an environment which has torch and torchvision. I recommend a conda environment. (You can do this with conda install -c pytorch pytorch=1.9.1 torchvision cpuonly or following https://pytorch.org/get-started .

• Verify that import torch; print(torch.cuda.is_enabled()) works in python and gives False.

• Then install fvcore and iopath with pip install fvcore iopath or conda install -c fvcore -c iopath -c conda-forge fvcore iopath.

• Make sure you have compilers (e.g. is there a g++ command available). If not, sudo apt install build-essential might be enough.

• Then install pytorch3d from source with pip install "git+https://github.com/facebookresearch/pytorch3d.git@stable" or by cloning this repository, git checkout stable and pip install -e .

[YanivHollander]

Thank you, this workflow worked! I am now able to import pytorch3d.renderer without a CUDA support. Just a typo correction for reference: torch.cuda.is_available(), instead of what you wrote.