Feature/cut variables #548
Conversation
| from dimod.core.composite import ComposedSampler | ||
| from dimod.sampleset import SampleSet | ||
| import dimod | ||
| import networkx as nx |
There was a problem hiding this comment.
See discussion in #539 (comment). However, in this case the functions are potentially more involved, so rewriting might be more trouble than it's worth.
There was a problem hiding this comment.
The nx.articulation_points and nx.biconnected_components aren't too difficult actually, and pure python. It's basically just DFS. So if you'd rather have dimod.articulation_points and dimod.biconnected_components we could do that (although it does seem like a lot of code replication).
There was a problem hiding this comment.
Yeah exactly, e5dd5f3 added bfs so dfs as well is not such a stretch, but at some point we end up just re-implementing NetworkX and there's no point in that. On the other hand, casting BQMs to NetworkX graphs has a non-trivial copy cost so working natively on BQMs can be nice 🤷♂️.
There was a problem hiding this comment.
IMHO, we should try to use nx's function until it's clear there is a drastic slowdown because of copy costs or pure Python.
Glad to learn about nx.articulation_points. I somehow missed the articulation_points function last summer when I was working on this problem. That would have simplified things. 👍
There was a problem hiding this comment.
I tend to agree. It's just frustrating because the BQM and networkx graph are so close in API, for a lot of nx functions the only thing blocking them working on the BQM directly is a decorator checking whether it's a directed graph or not. We could potentially wrap the BQM with some attributes to get around that but I am sure there would be edge cases.
There was a problem hiding this comment.
BTW, I've had great results using Metis in the past to decompose graphs. IIRC, it has a Python interface, but would add another requirement (including C backend 👎 )
There was a problem hiding this comment.
pymetis, I think. Yeah metis is great (although overkill if you just want cut vertices rather a partition of an arbitrary graph). In principle we could extend the functionality in CutVertexComposite to cut sets larger than a single vertex, but the run time grows exponentially with the size of the cut set.
| return tree_decomp.sample(self.child, **parameters) | ||
|
|
||
|
|
||
| def sub_bqm(bqm, variables): |
There was a problem hiding this comment.
Not a comment on this PR per se, but linking to #538 for future reference
| __all__ = ['CutVertexComposite'] | ||
|
|
||
|
|
||
| class CutVertexComposite(ComposedSampler): |
There was a problem hiding this comment.
Alex made a good point that the original name CutVariableComposite is more consistent with dimod.
There was a problem hiding this comment.
To make things easy to search for, would it be too weird to have an alias like CutVertexComposite = CutVariableComposite? Just because "cut vertex" is a more likely search term.
|
|
||
| return T, root | ||
|
|
||
| def sample(self, sampler, **parameters): |
There was a problem hiding this comment.
can we use self.child as a sampler? This requires that BiconnectedTreeDecomposition inherent from ComposedSampler
There was a problem hiding this comment.
Actually, you're already passing self.child of CutVertexComposite to this. Nevermind
|
Closing in favour of dwavesystems/dwave-preprocessing#36 |
Adds CutVertexComposite, which decomposes a connected bqm into biconnected components, solves those components, and merges the solution.
Currently the composite returns a single solution, like the ConnectedComponentsComposite. Eventually it should have the option of sampling.
Closes #547