Remove own system in favor of metatensor.torch.atomistic.System

[PicoCentauri]

Our System class was always a temporary solution and is now removed and replaced by the much better metatensor.torch.atomistic.System.

To have an independent torch API the synopsis of the the pure torch MeshPotential class now is

    def compute(
        self,
        types: Union[List[torch.Tensor], torch.Tensor],
        positions: Union[List[torch.Tensor], torch.Tensor],
        cell: Union[List[torch.Tensor], torch.Tensor],
    ) -> Union[torch.Tensor, List[torch.Tensor]]:

which I think this synopsis also clear for people who only live in the torch world.

I also tried to work on the comment from @kvhuguenin and replaced species by the new lab conventions types.

---

📚 Documentation preview 📚: https://meshlode--12.org.readthedocs.build/en/12/

[sirmarcel]

My impression was that we were keeping metatensor as an optional dependency, as far as I can tell this change makes it an integral dependency?

[Luthaf]

To have an independent torch API the synopsis of the the pure torch MeshPotential class now is

Just as a suggestions, you could also do

SystemDict = {
   "types": torch.Tensor,
   "positions": torch.Tensor,
   "cell": torch.Tensor,
}

def compute(
        self,
        systems: Union[List[SystemDict], SystemDict],
    ) -> Union[torch.Tensor, List[torch.Tensor]]:

########################################################################
# alternatively
@torch.jit.script
class System:
    def __init__(self, types, positions, cell):
        self.types = types
        self.positions = positions
        self.cell = cell
        
        # add some shape checks & friends

def compute(
        self,
        systems: Union[List[System], System],
    ) -> Union[torch.Tensor, List[torch.Tensor]]:

[PicoCentauri]

My impression was that we were keeping metatensor as an optional dependency, as far as I can tell this change makes it an integral dependency?

It still is. metatensor is only required for the metatensor branch of the code. The torch API is independent and even more torch like with removing the System and replacing it by the three parameters positions, types and cell as independent parameters.

[PicoCentauri]

Just as a suggestions, you could also do

I am not a huge fan of dictionary. They are usually very hard to document and people abuse them a lot

A custom class was what we had before but I don't want to maintain this 😆

[sirmarcel]

Thanks @PicoCentauri , that's good then! I'll have a look at the changes.

I agree with @PicoCentauri -- not a fan of dict either, as it's mutable and awkward to access. One could use a namedtuple instead of passing the arrays around, but I think this just makes batching awkward. I'm happy with raw arrays here.

[sirmarcel]

Shouldn't this start with an upper-case C?

[PicoCentauri]

Yes!

[sirmarcel]

This is where using type really becomes silly, in my opinion: you couldn't even write this as python code since type is reserved!

[PicoCentauri]

Yes, it is hard but basically the user will never see type since we also work with a list which we name types. However internally this will be interesting...

[PicoCentauri]

I also changed the error message from type -> types

[sirmarcel]

Looks fine by me. I'm pretty opposed to the naming of types as it conflicts heavily with established computer science terminology and type is a restricted keyword in Python -- but I guess that ship has sailed. Feel free to merge.