PyG installation with poetry

[ekosman]

I recently moved to using poetry for managing my environment, and I was wondering what is the correct way to translate the installation.
Until now, I used conda for installing pyg via the command conda install pyg -c pyg. However, it seems that pyg has many dependencies that are not directly translated to poetry.

I tried to simply add torch-geometric to the pyproject.toml file but it doesn't seem to work well. What is your suggestion?

[rusty1s]

I would assume that poetry can also access packages from PyPi, right? In this case, torch-geometric should work. What is the issue you are seeing?

[ekosman]

I would assume that poetry can also access packages from PyPi, right? In this case, torch-geometric should work. What is the issue you are seeing?

Indeed. However, it doesn't automatically install the related dependencies, e.g. torch-scatter, torch-sparse, pyg_lib etc.

[SauravMaheshkar]

I would assume that poetry can also access packages from PyPi, right? In this case, torch-geometric should work. What is the issue you are seeing?

Indeed. However, it doesn't automatically install the related dependencies, e.g. torch-scatter, torch-sparse, pyg_lib etc.

torch-scatter, torch-sparse, pyg_lib, .. aren't related dependencies and are usually installed separately from torch_geometric. For more reference, refer to the Installation section of the README. You are expected to install pyg_lib, torch_scatter and torch_sparse before installing torch_geometric.

A possible solution could be adding the aforementioned repositories to your pyproject.toml or manually using poetry add <>

[rusty1s]

Yeah, this would only effect PyG installation.

[ekosman]

It is also important to note that with the upcoming release, neither of these libraries is required anymore to be installed. This hopefully unblocks this issue :)

Is there any expectation for the release date? :)

[rusty1s]

We are targeting for March 21. For now, you can also install via

pip install git+https://github.com/pyg-team/pytorch_geometric.git

[q-aaronzolnailucas]

adding torch-scatter and torch-sparse to my pyproject.toml doesn't work, I have to define a separate task to install these, since the order seems to matter: torch -> torch_scatter -> torch_sparse -> torch_geometric. Unfortunately my source code directly uses SparseTensor, so I couldn't avoid it. I'd welcome any working examples people have!

[rusty1s]

Yes, this is a current limitation. The only workaround right now is to install from wheels in setup.py, but this is not allowed by PyPi :(