Using import_profiler, I've traced slow imports to networkx, especially:

• networkx.generators (300ms)
• networkx.linalg <- .algebraicconnectivity <- numpy (200ms)
• networkx.readwrite <- .graphml <- xml.etree.cElementTree (100ms)

I'm not using linalg at all, and I rarely need readwrite. I don't think I need generators either. These three trees represent 60% of the import-time for networkx.

When my filesystem is slow, the import-times can be much worse.

I have a couple of ideas for you:

• If possible, avoid importing these unless they are needed.
• Or, depend on and use lazy-import for at least these 3 sub-trees.