Dan Schult

I think it should match what `steiner_tree` does. If that makes sense.

The `pos` node attribute is handled specially when converting using `to_agraph`. This is because for networkx the positions are typically stored as a 2-tuple of numbers. So there is code...

The CI doc build test failure is likely to be unrelated. We've had a couple other PRs recently with this same error. Unfortunately that means we can't look at the...

> This seems to be the only thing driving the optimization. Can we minimize the code changes to the only impacting speed? I think this is the main comment not...

Did you press "save" after changing the title? It looks like the title is still the same one. Otherwise this PR looks good. :)

I think I was looking at the diff of one commit rather than the diff of the whole PR. That's probably causing my confusion. Sorry about that. I still think...

Just what @amcandio said -- and I'll add two other minor slightly subtle points: - `G.edges` without any parentheses returns the `(u, v, ekey)` triples for multigraphs. So you could...

It looks like `from_numpy_array` and `from_pandas_adjacency` should both be treated the same way as far as `nonedge`. We shouldn't need to change `to/from_scipy_sparse_matrix` because 1) they don't have it now,...

Currently `nx.from_numpy_array` currently does not support a `nonedge` parameter. A workaround that ensures creation of a complete graph (without selfloops) involves adding the zero weighted elements manually after the conversion....

Just add a section to the Graph Types reference page (which is in `networkx/classes/__init__.py`)