Install issues on POWER, aka, make mostly-unsued dependencies optional

🐛 Bug

Many of pytorch_geometric’s dependencies are entirely optional for the core library. Searching the codebase seems to indicate that the only “core” dependencies (those that are used anywhere except in a specific dataset or single transform) are

install_requires = [
    'numpy',
    'tqdm',
    'scipy',
    'networkx',
    'jinja2',
]

It would be nice if the other dependencies are specified as one or more extras_require sections, such as a datasets and transforms extras, for example. (Unfortunately, there does not appear to be a way to specify a default set of extras — see here — so the installation instructions would have to be updated correspondingly.) Alternatively, dependencies that are only used for specific datasets can just be wrapped in

try:
   
except ImportError:
    raise "useful message"

guards so that only users who need those dependencies will install them.

This is related to a specific bug: while trying to install pytorch_geometric on a ppc64 architecture cluster, we ran into issues related to the llvmlite dependency of numba, which is a pytorch_geometric dependency. numba is only used in the GDC transform. By commenting out numba in setup.py and its import in gdc.py, we were able to install and import pytorch_geometric.

[simonbatzner]

Vielen Dank Matthias! 😃

[rusty1s]

Should be fixed as well.