Combine two weighted graphs in NetworkX

Question

I'm using python multiprocessing to create multiple different NetworkX graphs and then using the below function to combine the graphs. However, while this function works fine for small graphs, for larger graphs it uses a huge amount of memory and hangs on both my system and a memory-heavy AWS system (using only about a third of the total memory in the system). Is there a more efficient way to perform the following function?

def combine_graphs(graph1, graph2, graph2_weight = 1):
    '''
    Given two graphs of different edge (but same node) structure (and the same type),
    combine the two graphs, summing all edge attributes and multiplying the second one's
    attributes by the desired weights. 

    E.g. if graph1.edge[a][b] = {'a': 1, 'b':2} and 
    graph2.edge[a][b] = {'a': 3, 'c': 4}, 
    with a weight of 1 the final graph edge should be 
    final_graph.edge[a][b] = {'a': 4, 'b': 2, 'c': 4} and with a weight 
    of .5 the final graph edge should be {'a': 2.5, 'b': 2, 'c': 2}.

    Inputs: Two graphs to be combined and a weight to give to the second graph
    '''

    if type(graph1) != type(graph2) or len(set(graph2.nodes()) - set(graph1.nodes())) > 0:
        raise Exception('Graphs must have the same type and graph 2 cannot have nodes that graph 1 does not have.')

    # make a copy of the new graph to ensure that it doesn't change
    new_graph = graph1.copy()

    # iterate over graph2's edges, adding them to graph1
    for node1, node2 in graph2.edges():
        # if that edge already exists, now iterate over the attributes
        if new_graph.has_edge(node1, node2):
            for attr in graph2.edge[node1][node2]:
                # if that attribute exists, sum the values, otherwise, simply copy attrs
                if new_graph.edge[node1][node2].get(attr) is not None:
                    # try adding weighted value: if it fails, it's probably not numeric so add the full value (the only other option is a list)
                    try:
                        new_graph.edge[node1][node2][attr] += graph2.edge[node1][node2][attr] * graph2_weight
                    except:
                        new_graph.edge[node1][node2][attr] += graph2.edge[node1][node2][attr]
                else:
                    try:
                        new_graph.edge[node1][node2][attr] = graph2.edge[node1][node2][attr] * graph2_weight
                    except:
                        new_graph.edge[node1][node2][attr] = graph2.edge[node1][node2][attr]

        # otherwise, add the new edge with all its atributes -- first, iterate through those attributes to weight them
        else:
            attr_dict = graph2.edge[node1][node2]
            for item in attr_dict:
                try:
                    attr_dict[item] = attr_dict[item] * graph2_weight
                except:
                    continue
            new_graph.add_edge(node1, node2, attr_dict = attr_dict)

    return new_graph

Answer 1

There are two places where memory will expand in your code:

1) making a copy of graph1 (perhaps you need to keep a copy though)

2) using graph2.edges() makes a list of all edges in memory, graph2.edges_iter() iterates over the edges without creating a new list

You can probably make it faster by handling the edge data differently too. You can get the data object while iterating over the edges and not have to perform as may dictionary lookups:

def combined_graphs_edges(G, H, weight = 1.0):
    for u,v,hdata in H.edges_iter(data=True):
        # multply attributes of H by weight
        attr = dict( (key, value*weight) for key,value in hdata.items())
        # get data from G or use empty dict if no edge in G
        gdata = G[u].get(v,{})
        # add data from g
        # sum shared items
        shared = set(gdata) & set(hdata)
        attr.update(dict((key, attr[key] + gdata[key]) for key in shared))
        # non shared items
        non_shared = set(gdata) - set(hdata)
        attr.update(dict((key, gdata[key]) for key in non_shared))
        yield u,v,attr
    return


if __name__ == '__main__':
    import networkx as nx
    G = nx.Graph([('a','b', {'a': 1, 'b':2})])
    H = nx.Graph([('a','b', {'a': 3, 'c':4})])
    print list(combined_graphs_edges(G,H,weight=0.5))
    # or to make a new graph 
    graph = G.copy()
    graph.add_edges_from(combined_graphs_edges(G,H,weight=0.5))