I tried graph-tool after some experience with networkx, in search for
better efficiency (attracted by ) ... but unfortunately, it seems
like it all depends on how data is stored!
This is true. If your code essentially depends on python loops,
graph-tool will not provide any advantage.
The (several) networks I play with are dense weighted networks of 1435
nodes. They are defined in terms of their adjacency matrices, which are
pandas dataframes stored in a HDF5 store.
In order to get the relative network with networkx, I would do
dg = nx.from_numpy_matrix(hstore['my_df'].as_matrix(),
which took around 18 seconds. In graph-tool, the best alternative I
found, based on  and , is:
# Since the bottleneck is given by the calls to g.add_edge(),
# numba actually slows down things a bit.
#from numba import autojit
g = Graph()
n = m.shape
edges = list(g.add_vertex(n=n))
c = Counter(every=10, until=n)
for i in xrange(n):
for j in xrange(n):
g = graph_from_adj_matrix(hstore['my_df'].as_matrix())
... which takes around 200 seconds (and is still not even weighted!).
Is there anything trivial (or simply new) I'm missing (apart from the
possibility, mentioned in , of writing my code in C++)?
The ability of inserting many edges without calling add_edge() several
times is indeed lacking...
Because of this, I have now just added a Graph.add_edge_list() to the
git version, which takes a list of edges to be added, which can be a
numpy array. If you have a full adjacency matrix instead of an edge
list, you can do simply:
This should be much faster than the Python loop above.