Jeremy Reizenstein

I generally call the constructor directly.

The reason we avoid to_object is because of considerations about what should happen when a notebook user redefines a class.

You're right. The ValueError is probably straight from dataclasses because it doesn't like the mutable default. Fixing that with default_factory doesn't satisfy old OmegaConf, so there's no regression there. The...

I imagine lots has changed in theano, keras and tensorflow since that demo was made, making it unlikely to work, and you can learn the idea by reading the code...

All internal calculations use float64 (i.e. double precision). It would make sense for float64 to be the type of all the outputs. It just happens that `sig` was implemented first...

Are you aiming for the centroid of the surface defined by the mesh (like the first reference, and "surface-based" in the second) or the volume (which depends on tetrahedrons and...

In that case, I think this does make sense, but the docstring should be made very clear. The function only makes sense (I think) for watertight meshes whose faces are...

I think this is a bit too specialised to live in the main Meshes object. It would make sense to add a new file, perhaps `pytorch3d/ops/watertight.py` for simple operations which...

The comment in the release note is not relevant here. There are builds for PyTorch 1.12.0 was just the news. We are still supporting PyTorch going back to PyTorch 1.8.0...

Pytorch3D has not been tested with autocast or designed with autocast in mind. The custom pytorch3d ops (written in cuda and c++) could potentially have problems with autocast. ~~But I...