Fix poetry lock fails

[nbgit10]

Poetry lock fails with pytorch3d as dependency because the resolver runs in a dedicated, clean venv on every run. Therefore, import torch fails in pytorch3d's setup.py and we need to install it while resolving dependencies.

I believe this is a bug that anyone using poetry would have when specifying pytorch3d as a dependency in their pyproject.toml file - at least when installing directly from source.

[bottler]

As discussed before, e.g. on #1479, we aren't likely to add anything to pytorch3d which automatically installs torch anytime soon.

How this fits with poetry is something we are interested in getting advice about.

[nbgit10]

I am installing pytorch3d from source and have added it to my pyproject.toml simply with pytorch3d = {git = "[email protected]:facebookresearch/pytorch3d.git", tag = "stable"}. When updating/creating the poetry.lock file by running poetry lock Poetry runs through the setup.py of pytorch3d which fails due to torch not being installed. This happens if a new venv needs to be created. If poetry can find a venv of the project and it already has torch installed, it works. However, even if, for example, the system Python environment has torch available, a newly created venv has not.

[nbgit10]

I think in the end it is the same issue as #1419

Switching to a pyproject.toml file and adding torch>=1.10 as a build requirement would solve this and is much less hacky than my change proposed here.

Interestingly, it does not help if I add torch as a build requirement in my project, where pytorch3d is a regular dependency. I am not deep enough in the various PEP standards to know if this behaviour is as designed but it is unfortunately not a suitable workaround. So actually this is independent of poetry but actually affects any install from-source in a clean environment.