I've been using networkx to simply create a graph and check the connected
components. The bottleneck of the operation is the creation of the edges.
I've heard that graph-tool is very efficient so I've replaces the code with
a graph-tool graph.
To my surprise, the creation of a graph-tool graph is MUCH slower than that
of a networkx graph.
Am I doing something wrong?
I've created a small sample program (at the bottom of this message) to test
this speed...
it creates a graph in a very natural way: pairs of objects that should be
connected by an edge are iterated through. We happen to give every possible
pair of objects in this example so the complete graph is created.
Here is runsnake image of where the running-time is going:
#!/usr/bin/python
"""
Create graphs in networkx and graph-tool.
"""
import networkx as nx
from graph_tool.all import *
from itertools import combinations
def graph_tool_create():
""" Create a graph_tool graph given a list of pairs. """
G = Graph(directed=False)
objectset = set()
for o1,o2 in get_pairs_of_objects():
if(o1 not in objectset):
u = G.add_vertex()
objectset.add(o1)
if(o2 not in objectset):
v = G.add_vertex()
objectset.add(o2)
G.add_edge(u,v)
def nx_create():
""" Create a graph_tool graph given a list of pairs. """
G = nx.Graph()
for o1,o2 in get_pairs_of_objects():
G.add_edge(o1,o2)
def get_pairs_of_objects():
""" Generate pairs of objects. """
n = 5000
for a,b in combinations(range(n),2):
yield a,b
I posted code with missing lines...
here is the good code:
#!/usr/bin/python
"""
Create graphs in networkx and graph-tool.
"""
import networkx as nx
from graph_tool.all import *
import igraph
from itertools import combinations
def graph_tool_create():
""" Create a graph_tool graph given a list of pairs. """
G = Graph(directed=False)
objectTOv = {}
for o1,o2 in get_pairs_of_objects():
if(o1 in objectTOv):
u = objectTOv[o1]
else:
u = G.add_vertex()
objectTOv[o1] = u
if(o2 in objectTOv):
v = objectTOv[o2]
else:
v = G.add_vertex()
objectTOv[o2]
G.add_edge(u,v)
def nx_create():
""" Create a graph_tool graph given a list of pairs. """
G = nx.Graph()
for o1,o2 in get_pairs_of_objects():
G.add_edge(o1,o2)
def get_pairs_of_objects():
""" Generate pairs of objects. """
n = 3000
for a,b in combinations(range(n),2):
yield a,b
How does the performance change if you create the necessary edges
beforehand?
In graph-tool things are faster than in networkx when they are delegated
to C++, otherwise this should be comparable in speed. In the case of
adding many edges, this is done by using the Graph.add_edge_list()
function, which runs in C++ internally. In your example, this should
provide a massive speed-up.
In this code all the edges and vertices are created by graph-tool and the
result is something much faster...
is this the best I can do?
It's somewhat annoying to have to keep track of the vertices that will be
created like this:
def graph_tool_create_all_at_once():
""" Create a graph_tool graph given a list of pairs. """
G = Graph(directed=False)
objectTOi = {}
vertexpairs = []
counter = 0
for o1,o2 in get_pairs_of_ints():
if(o1 in objectTOi):
u = objectTOi[o1]
else:
u = counter
counter += 1
objectTOi[o1] = u
if(o2 in objectTOi):
v = objectTOi[o2]
else:
v = counter
counter += 1
objectTOi[o2] = v
def add_nodupe_vertex(self,label,*args,**kwargs):
'''Return a node with label. Create node if label is new'''
try:
n = self._nodes[label]
except KeyError:
n = self.add_vertex()
self._nodes[label]=n
return n
def add_nodupe_edge(self, n1_label, n2_label,directed=False):
"""
Get or create edges using get_or_create_node
""" #there may be two if graph is directed but edge isn't
edges = []
n1 = self.add_nodupe_vertex(n1_label)
n2 = self.add_nodupe_vertex(n2_label)
edges.append(self.add_edge(n1,n2))
if self.is_directed() and not directed:
edges.append(self.add_edge(n2,n1))
return edges
Hi offonoffon...
won't the creation of the vertices and edges one at a time end up
dramatically affecting the performance of graph creation (see my original
question at the top of the thread)?
I've found that creating the edges one at a time is MUCH slower than
creating them all at once and that creating the vertices one at a time is a
little slower.
Hi offonoffon...
won't the creation of the vertices and edges one at a time end up
dramatically affecting the performance of graph creation (see my original
question at the top of the thread)?
Yes, and that has been solved already. I was responding to "It's somewhat
annoying to have to keep track of the vertices that will be created like
this", and to the observation that your code is maybe not as
flexible/reuseable as it could be. And as I said, my code could be easily
modified to add the edges as a batch instead of one at a time.
Yes, it is possible to improve this. If your o1 and o2 objects are
always ints (as I gather from get_pairs_of_ints()), then you don't need
these dictionaries at all. The G.add_edge_list() will create missing
nodes as necessary. So you only need to do:
G.add_edge_list(list(get_pairs_of_ints()))
This will be much faster. However it requires your ints to be within
some reasonable range, since missing ints without edges will also be
created. But you are better off guaranteeing this is the case when you
generate the ints in the first place.
Ok. Still, the best option would be not to iterate over all pairs to
construct the mapping to ints, but to iterate only through the
individual objects.
