Generate graph from scipy sparse matrix

[iosonofabio]

This tries to implement #357, i.e. generating a Graph object from a sparse scipy matrix, which is the de facto standard in Pythonlands for such cases.

The key is to bypass GraphBase.Adjacency, which dispatches to a C function expecting a dense matrix. Instead, I reproduce the logic of the various C-level functions depending on the mode (e.g. igraph_i_adjacency_lower, all in the igraph project, src/constructors.c).

A few notes:

• graceful try to import scipy to check if the matrix is sparse
• added a case for (dense) numpy 2D ndarrays (which include numpy.matrix object) with a similar graceful import
• outsourced the sparse case to a private _ function in a new sparse_matrix module to avoid cluttering the main module. Happy to discuss namespacing or alternatives
• the details of the logic use a more Pythonic syntax than their C counterparts, while trying to be somewhat efficient. No efforts were undertaken to actually optimize these operations, and I wouldn't mind getting rid of some black magic if you don't like it (sums over conditional lazy comprehensions).

Funnily enough, the concrete case that prompted me to make this PR is a weighted adjacency sparse matrix... so I would like to implement the weighted case as well, any suggestions how to call the class method in that case? Perhaps just graph.AdjacencyWeighted?

No tests written yet.

edit: Oh I just saw Weighted_Adjacency.

[iosonofabio]

Huh, CI fails in an interesting way and I wouldn't mind some advice. It has to do with inheritance (alas!).

This PR currently has Adjacency as a classmethod of Graph, consistent with the current declaration. If the input is not a sparse matrix, I would like to just use the old call. But it's tricky.

@classmethod def Adjacency(klass, ...): return GraphBase.Adjacency(...) # or return super().Adjacency(...)

does not work, because GraphBase.Adjacency is ALSO a class method, so it will construct whatever object you give it as first argument, which via the dot notation means whatever object comes before the dot. So if we call it like this it will construct a GraphBase, not a Graph. On the other hand, return Graph.Adjacency(...) becomes an infinite recursion.

I could make another copy of GraphBase.Adjacency as GraphBase._Adjacency and then exploit the fact that when subclassing, Python will automatically create Graph._Adjacency, but is that alright or do you guys know of a better method?

[iosonofabio]

I pushed the _Adjacency hack so you guys can see it in action, but am not super happy about it, of course. Any suggestions would be welcome @ntamas @szhorvat @vtraag

[ntamas]

I think that this is a bug in GraphBase.Adjacency(), but let me check.

---

Edit: it seems like I have already started adding test cases for inheritance in test_basic.py (see InheritanceTests), but it does not cover class methods yet. I'll update the test cases to see whether there is indeed a bug there or not.

[ntamas]

@iosonofabio It looks like the correct syntax for calling a class method of the superclass is:

class InheritedGraph(Graph): @classmethod def Adjacency(cls, *args, **kwds): result = super(InheritedGraph, cls).Adjacency(*args, **kwds) return result 

So I would try something like super(Graph, cls).Adjacency() from the Graph implementation to call Adjacency in GraphBase. Let me know if it doesn't work.

[ntamas]

A few minor comments here and there, otherwise I think it's good to go!

[iosonofabio]

I fixed those, but let me add the weighted case too since I'm here

[iosonofabio]

rebased on top of the GitHub actions PR, let's see if this works

[ntamas]

@iosonofabio Can you rebase this once again on top of master to see if the CI checks pass? I can merge it after that.

[jvail]

Sorry for spoiling your PR discussion: It is just a pity this could not make it into the latest release. Would be a very useful feature!

[ntamas]

There will be a 0.9.1 soon(ish) as the API documentation build is currently broken, so I'm hoping to squeeze this into that one.

[ntamas]

Okay, I'm gonna merge this into master as it seems to be okay. If the tests don't pass in master, I'll just fix those there.