Looking at using pytorch3d in software package I develop. Currently I use conda to install all the dependencies so it runs perfectly in Windows, Mac and Linux. Currently I depend on pytorch and make sure to only update the version when all 3 platforms have new releases.
Pytorch conda support is great, Pytorch :: Anaconda.org , all platforms you could want binaries for are available with conda.
Pytorch3d which builds on-top of pytorch however is very different.
Pytorch3D conda only supports linux, pip only supports mac. Neither supports both and neither supports windows.
Why doesn’t pytorch3D use the same building infrastructure and build for all environments like pytorch itself does and release the binaries through conda for them all?
If you have issues about the PyTorch3D project it will get more attention if you ask them on the pytorch3d github itself.
Basically, the installation story for PyTorch3D is different between operating systems. (1) The team have not invested in the Windows build and built packages are not availble. Help from outside the team on doing windows builds might be welcome, but things are a bit complicated. (2) For Linux, the build is most developed (and most users are on Linux) and a multitude of packages for different PyTorch, Python and CUDA versions is provided via conda. (3) For mac, CUDA is unavailable and so we don’t think PyTorch3D is used much. We provide a very limited range of packages on PyPI. Not having so many versions means that conda is less essential.