Make package ready and upload on PyPI

[FrancescoCasalegno]

This could be complicated by the fact that the installation of PyTorch-Geometric is not as trivial as pip install torch-geometric.

This is particularly true for the torch-scatter, torch-sparse, ... dependencies of torch-geometric. For these packages, it seems like wheels are not on PyPI but must be found instead on https://data.pyg.org/whl/torch-${TORCH}+${CUDA}.html.

Notice that we do need such dependencies for our project: https://github.com/BlueBrain/morphoclass/blob/72f59fad5dca9eb08a5752a21de7577c47f1011b/install.sh#L54-L57

[FrancescoCasalegno]

Update

PyPI seems to have PyTorch-Geometric and its various torch-scatter ... dependencies:

• torch-geometric: https://pypi.org/project/torch-geometric
• torch-scatter: https://pypi.org/project/torch-scatter
• torch-sparse: https://pypi.org/project/torch-sparse

So:

• [ ] Do they work?
• [ ] If Yes, then why does PyTorch-Geometric tells you to install its dependcies like torch-scatter with -f https://data.pyg.org/whl/torch-${TORCH}+${CUDA}.html instead?
• [ ] Maybe everything works fine on CPU doing pip install ... from PyPI, but it breaks for GPUs?
• [ ] If Yes, then maybe we can have our package on PyPI for CPU only, and then rely on our install.sh for GPU install and usage.

[FrancescoCasalegno]

Update

It turns out that torch-scatter, etc. are indeed on PyPI. But only the source tarball is there, so pip install torch-scatter will take some time to compile the C extensions!

On the other hand, torch-geometric is quick to install with pip install torch-geometric even if only the source tarball is on PyPI because torch-geometric is a pure-Python package.