This seems like a good idea to me! Note that "dense" and "array" are already valid values for format= even though they're not documented: A.asformat("dense") returns a numpy array.

@michaelweinold I'm curious--what's the performance difference between your code for creating a numpy array vs using the original code with format="dense"? If the latter is not noticeably worse, then the code could remain the same and the only thing that would need updated is the docstring.

to_scipy_sparse_array is another function that returns a scipy sparse array and has the format= argument, so it also currently supports "dense" format (but this isn't documented).

The other functions I found that return scipy sparse arrays but don't have a way to specify the format are adjacency_matrix, attr_sparse_matrix, bethe_hessian_matrix, incidence_matrix, laplacian_matrix, normalized_laplacian_matrix, and tournament_matrix. I don't recommend adding format= to these functions in this PR, but if anybody thinks doing so is a good idea then please make a new issue or PR to continue exploring this option.

Hmm... that's odd - I could have sworn that I tried to pass "dense" to the function, even though it was not listed among the options. In any case, I tested my implementation quickly, using two different size graphs:

I believe the speedup is negligible - I can roll back the changes to the function itself and simply leave the updated docstring with the "dense" option.

I would very much like to adapt the from_biadjacency_matrix function though - there, the input really does need to be a sparse array at the moment.

then the code could remain the same and the only thing that would need updated is the docstring.

Thanks for pointing this out @eriknw - I really like this idea! AIUI it's basically just taking advantage of a (new-ish?) feature in scipy to provide users with dense arrays with zero code change!

Depending on your needs, you could try from_numpy_array. It will create a bipartite graph if the input represents a bipartitie graph. The nodes sets are not indicated automatically, but you can find them with top_nodes, bottom_nodes = nx.bipartitie.sets() and set node attributes using G.add_nodes_from(top_nodes, bipartite=0), etc.

I think to update from_bipartitie_matrix to handle numpy arrays you only need to update convert_matrix._generate_weighted_edges using some form of ((int(u), int(v), A[u,v]) for u,v in zip(*A.nonzero())

...I removed the condition (negligible speedup, as mentioned above). The only change now is the 'dense' option mentioned in the function docstring. Ready for review 😎