Port built-in graphs to Python

Created by: gcarleo

In the process it might be good to add wrappers for existing graph libraries (say, networkx)

Created by: VolodyaCO

Although networkx has good support, it is extremely slow, especially for large graphs. I recommend graph-tool (I have used it for learning dynamics of contagion models, and it exponentially outperforms networkx)

Created by: gcarleo

Good suggestion @VolodyaCO thanks, I wasn't aware of this library. However it seems dependent on Boost and it's compiled C++. Does it come in a pip package?

Created by: VolodyaCO

When I installed it I had to do it with a non-standard conda channel. Apparently it is now included in conda-forge. I'm not sure it's in a pip package...

Created by: gcarleo

Alright then it's a no-go for the moment. We can stick to igraph and networkx that have better support. For our tasks they should be relatively fast, since in most cases the lists of edges etc are computed only once.

Created by: VolodyaCO

Well, if it's for small graphs, then surely this is not a critical performance issue and the best decision is to stay with networkx, as it is the most used one (with the best support).

added help wanted label

Created by: VolodyaCO

@gcarleo I could try to help, but I am unfamiliar with the C++ code. If you give me specific tasks I can try to do them. However, maybe a member of the developer team is able to perform this work quicker and better.

Created by: gcarleo

@VolodyaCO a specific first task to start with would be to write an AbstractGraph class in python (say, similar to this style ) that defines what the common interface of the graphs in python should be. In practice, it should have the same methods of the graph classes we already have. Essentially the attributes as defined in here

The pull request should be done to the branch v3.0.

Created by: VolodyaCO

@gcarleo can I write the CustomGraph, and the rest of the graphs using networkx? Also, should I create a new pull request when I write most of the code?

Created by: gcarleo

Yes that sounds good, basically HyperCube and CustomGraph are just special cases of already existing networkx graphs . The only trickier case is Lattice, but maybe we could use some other existing library, for example https://pypi.org/project/pbcpy/ or others

Created by: VolodyaCO

@gcarleo do you want me to check that the nodes are always labeled by integers? Or should I give freedom to label nodes as strings?, e.g. nk.graph.CustomGraph([("cavity 1", "cavity 2", "color 1")])

Created by: gcarleo

I think that at the moment we construct CustomGraph just using a list of (integer) edges... maybe I would leave it this way since it seems easier to handle? I also realized that we actually don't need networkx to implement CustomGraph, it is easier to just work with scipy.sparse.csgraph since it has already all the methods we need!

Created by: VolodyaCO

@gcarleo I already wrote the CustomGraph using networkx, but if you want me to I can wrap the scipy version.

Created by: gcarleo

It was just to avoid introducing the explicit dependence on networkx for the simple cases we are interested in... if it's not too much work I would write CustomGraph with the scipy version, and leave the wrapper you already wrote for networkx as another class/converter to netket.AbstractGraph (maybe call it netket.graph.NetworkX that can be constructed from a networkx graph?)... up to you

Created by: gcarleo

But if you have already done it in networkx it's OK no problem... no need to rewrite it now

Created by: gcarleo

Yes sorry! I realize that there is a few methods missing in the scipy interface anyways... (is_bipartite is not there and finding automorphisms is definitely not there) so indeed it makes more sense to proceed with networkx...

Created by: gcarleo

It's amazing that now newtorkx has also a way to find automorphisms, before we had to use python-igraph...

graph=nx.Graph([[0,1],[1,2],[2,3],[3,0]])

ismags=nx.isomorphism.ISMAGS(graph,graph)
automorphisms=[[iso[i] for i in range(graph.size())] for iso in ismags.isomorphisms_iter(symmetry=False)]

Created by: VolodyaCO

I was checking that before as well... but I was a bit unsure of including something of the sort because in the C++ implementation the automorphisms are an input to the CustomGraph. However, it should be something that gets computed always, right @gcarleo ? i.e. it shouldn't be an argument of the init method.

Created by: gcarleo

Indeed, the old C++ implementation was quite limited in this. Now we can make it automatically computed when one calls .automorphisms() (I wouldn't compute it by default though, but only if one calls that method and wants the automorphisms)...

Also, it would be a good idea to call the main class you are writing just netket.Graph, it doesn't really make sense to call this CustomGraph anymore! (that was a kind of weird name that was left from version 1.0..)

Created by: VolodyaCO

Good, I'll have a couple of free hours tomorrow. I will try to finish the CustomGraph, and to add the Hypercube. Is the Edgeless graph necessary?