Last year, when I was choosing a framework for graph neural networks,
I gave a try PyG, as they call it now (then they said
pytorch-geometric). I don't say it was hard to use, but I had big
problems installing it. After feeling the same pain again, I
restarted my search and stumbled upon DGL, another library hard to
get with ``pip`` (you damned ``conda`` lovers). The bright side of
DGL is its independence from the DL framework, which, I think, is a
must nowadays. There are also some signs that probably it could work
with Ray RLlib, although from what I see now, it has to be tricky, at
least.
