networkx remove node attribute

Question

i have attached numpy-array to nodes in a networkx graph. How to store the graph in gexf-format on disk? (without the numpy vector, as its just something intermediate...)

def create():
    G = nx.Graph()
    for i in range(256):
        G.add_node(i, vector=np.arange(20))
    for i in range(1,20):
        for j in range(1,256, 10):
            G.add_edge(i,j)

    temp = tempfile.mktemp(suffix=".gexf")
    print("dumping G = (V: %s, E: %s) to disk %s"
        % (len(G.nodes()), len(G.edges()), temp))
    nx.write_gexf(G, temp)

However, this breaks. I'm new to python, but to me it seems like the ndarray is not serializable?! So, how to tell networkx to ignore that node attribute?

File "...lib\site-packages\networkx\readwrite\gexf.py", line 430, in add_attributes
    attr_id = self.get_attr_id(make_str(k), self.xml_type[val_type],
KeyError: <type 'numpy.ndarray'>

Answer 1

Use a library like the native pickle , or h5py for HDF5, to serialize your graph object. For example you can do:

import pickle

with open("pickle_file", "wb") as f:
    pickle.dump(create(), f)

The pickled graph can be loaded back into Python by:

with open("pickle_file", "rb") as f:
    G = pickle.load(f)

Answer 2

I solved this by removing the property "vector" from the data items:

for (n,d) in G.nodes(data=True):
    del d["vector"]

full MWE:

def create():
    G = nx.Graph()
    for i in range(256):
        G.add_node(i, vector=np.arange(20))
    for i in range(1,20):
        for j in range(1,256, 10):
            G.add_edge(i,j)

    temp = tempfile.mktemp(suffix=".gexf")
    print("dumping G = (V: %s, E: %s) to disk %s"
        % (len(G.nodes()), len(G.edges()), temp))
    for (n,d) in G.nodes(data=True):
        del d["vector"]
    nx.write_gexf(G, temp)