def test_directed_laplacian(self):
"Directed Laplacian"
# Graph used as an example in Sec. 4.1 of Langville and Meyer,
# "Google's PageRank and Beyond". The graph contains dangling nodes, so
# the pagerank random walk is selected by directed_laplacian
G = nx.DiGraph()
G.add_edges_from(((1,2), (1,3), (3,1), (3,2), (3,5), (4,5), (4,6),
(5,4), (5,6), (6,4)))
GL = numpy.array([[ 0.9833, -0.2941, -0.3882, -0.0291, -0.0231, -0.0261],
[-0.2941, 0.8333, -0.2339, -0.0536, -0.0589, -0.0554],
[-0.3882, -0.2339, 0.9833, -0.0278, -0.0896, -0.0251],
[-0.0291, -0.0536, -0.0278, 0.9833, -0.4878, -0.6675],
[-0.0231, -0.0589, -0.0896, -0.4878, 0.9833, -0.2078],
[-0.0261, -0.0554, -0.0251, -0.6675, -0.2078, 0.9833]])
assert_almost_equal(nx.directed_laplacian_matrix(G, alpha=0.9), GL, decimal=3)
# Make the graph strongly connected, so we can use a random and lazy walk
G.add_edges_from((((2,5), (6,1))))
GL = numpy.array([[ 1. , -0.3062, -0.4714, 0. , 0. , -0.3227],
[-0.3062, 1. , -0.1443, 0. , -0.3162, 0. ],
[-0.4714, -0.1443, 1. , 0. , -0.0913, 0. ],
[ 0. , 0. , 0. , 1. , -0.5 , -0.5 ],
[ 0. , -0.3162, -0.0913, -0.5 , 1. , -0.25 ],
[-0.3227, 0. , 0. , -0.5 , -0.25 , 1. ]])
assert_almost_equal(nx.directed_laplacian_matrix(G, walk_type='random'), GL, decimal=3)
GL = numpy.array([[ 0.5 , -0.1531, -0.2357, 0. , 0. , -0.1614],
[-0.1531, 0.5 , -0.0722, 0. , -0.1581, 0. ],
[-0.2357, -0.0722, 0.5 , 0. , -0.0456, 0. ],
[ 0. , 0. , 0. , 0.5 , -0.25 , -0.25 ],
[ 0. , -0.1581, -0.0456, -0.25 , 0.5 , -0.125 ],
[-0.1614, 0. , 0. , -0.25 , -0.125 , 0.5 ]])
assert_almost_equal(nx.directed_laplacian_matrix(G, walk_type='lazy'), GL, decimal=3)
def get_graph_laplacian(self,
edge_types: list,
node_list=None,
databases=None):
"""
Returns an adjacency matrix from edges with type specified in :param edge_types: and nodes specified in
:param edge_types: A list of edge types letter codes in ["d", "u", "u_n"]
:param node_list: A list of node names
:return: A csr_matrix sparse adjacency matrix
"""
if "d" in edge_types:
directed = True
elif "u" in edge_types:
directed = False
else:
raise Exception("edge_types must be either 'd' or 'u'")
if hasattr(self, "laplacian_adj"):
laplacian_adj = self.laplacian_adj
else:
if databases is not None:
if directed:
edge_list = [
(u, v)
for u, v, d in self.G.edges(nbunch=self.node_list,
data=True)
if 'database' in d and d['database'] in databases
]
laplacian_adj = nx.directed_laplacian_matrix(
nx.DiGraph(incoming_graph_data=edge_list),
nodelist=self.node_list)
else:
edge_list = [(u, v) for u, v, d in self.G_u.edges(
nbunch=self.node_list, data=True)
if d['type'] in edge_types]
laplacian_adj = nx.normalized_laplacian_matrix(
nx.Graph(incoming_graph_data=edge_list),
nodelist=self.node_list)
elif directed:
laplacian_adj = nx.directed_laplacian_matrix(
self.G.subgraph(nodes=self.node_list),
nodelist=self.node_list)
elif not directed:
laplacian_adj = nx.normalized_laplacian_matrix(
self.G_u.subgraph(nodes=self.node_list),
nodelist=self.node_list)
else:
raise Exception()
self.laplacian_adj = laplacian_adj
if node_list is None or node_list == self.node_list:
return laplacian_adj
elif set(node_list) <= set(self.node_list):
return self.slice_adj(laplacian_adj, node_list, None)
elif not (set(node_list) < set(self.node_list)):
raise Exception("A node in node_l is not in self.node_list.")
return laplacian_adj
def compute(tribes, adj_matrix, conv, precision):
import networkx as nx
spectra = []
pbar = progressbar.ProgressBar()
for tribe in pbar(tribes):
tribe_ids = conv.indices(tribe)
adj_submat = adj_matrix[np.ix_(tribe_ids, tribe_ids)]
G = nx.from_scipy_sparse_matrix(adj_submat, create_using=nx.DiGraph)
# Find the largest connected component of the graph
largest = max(nx.strongly_connected_components(G), key=len)
if len(largest) <= 2: # Needs at least a certain size...
spectra.append([])
else:
# Compute the Chung's laplacian matrix of tribe's largest connected component
L = nx.directed_laplacian_matrix(G.subgraph(largest))
# Find the eigenvalues
eig = scipy.linalg.eig(L)[0]
# Order the non-zero eigenvalues and round to desired precision
spectrum = np.unique(np.round(eig[np.nonzero(eig)], precision))
spectra.append(spectrum)
return spectra
def directed_heat_kernel(G, t):
# Input: DiGraph G and time parameter t
# Output: heat kernel matrix
# Automatically computes directed laplacian matrix and then exponentiates
L = np.asarray(nx.directed_laplacian_matrix(G))
lam, phi = np.linalg.eigh(L)
return heat_kernel(lam, phi, t)
def do(self):
logger.info("SPECTRAL Eigenvalues %i" % (self.spectral_eigenvalues, ))
logger.info("SPECTRAL Dimensions %i" % (self.spectral_dimensions,))
logger.info("SPECTRAL Laplacian %i" % (self.spectral_laplacian,))
logger.info("SPECTRAL Eigenvector %i" % (self.spectral_eigenvector,))
logger.info("SPECTRAL Compute Neuron Graph")
neuron_graph = self.database.db2graph()
logger.info("SPECTRAL Compute Laplacian Matrix")
if self.spectral_laplacian == Config.SPECTRAL_LAPLACIAN_DIRECTED:
neuron_graph_laplacian = networkx.directed_laplacian_matrix(neuron_graph, weight=None,
walk_type='pagerank', alpha=0.95)
elif self.spectral_laplacian == Config.SPECTRAL_LAPLACIAN_UNDIRECTED:
neuron_graph = neuron_graph.to_undirected()
neuron_graph_laplacian = networkx.normalized_laplacian_matrix(neuron_graph, weight=None)
else:
logger.critical("SPECTRAL Laplacian Method not supported")
raise RuntimeError()
logger.info("SPECTRAL Compute Lapalcian Eigenvectors")
if self.spectral_eigenvector == Config.EIGENVECTOR_LEFT:
eigenvalues, eigenvectors = scipy.sparse.linalg.eigen.arpack.eigs(neuron_graph_laplacian,
self.spectral_eigenvalues,
sigma=0, which='LM')
elif self.spectral_eigenvector == Config.EIGENVECTOR_RIGHT:
eigenvalues, eigenvectors = scipy.sparse.linalg.eigen.arpack.eigs(neuron_graph_laplacian,
self.spectral_eigenvalues)
else:
logger.critical("SPECTRAL Eigenvector direction not supported")
raise RuntimeError()
points = eigenvectors.real[:, :self.spectral_dimensions]
database_session = self.database.new_session()
db_populations = {p.id: p for p in
database_session.query(orm.population.Population).filter_by(shadow=False).all()}
self.data = dict()
for neuron, neuron_data in neuron_graph.nodes_iter(data=True):
pop_id = neuron_data['p']
if pop_id not in self.data .keys():
self.data[pop_id] = dict()
self.data[pop_id]['label'] = db_populations[pop_id].label
self.data[pop_id]['neurons'] = db_populations[pop_id].neurons
self.data[pop_id]['neurons_core'] = db_populations[pop_id].neurons_core
self.data[pop_id]['space'] = dict()
self.data[pop_id]['clusters'] = dict()
self.data[pop_id]['nodes'] = dict()
pop_neuron_id = neuron_data['n']
self.data[pop_id]['space'][pop_neuron_id] = points[neuron_graph.nodes().index(neuron)]
return self.data
def algebraicConnectivity(g):
if nx.is_directed(g):
laplacian=nx.directed_laplacian_matrix(g, weight=None)
else:
tmplaplacian=nx.laplacian_matrix(g)
laplacian=tmplaplacian.todense()
#print laplacian
laplacian_eig=np.linalg.eig(laplacian)
eigs=laplacian_eig[0]
return np.min(eigs[eigs> 1e-7 ])
def directed_spectral_worker(G):
#eigs = eigvalsh(nx.normalized_laplacian_matrix(G).todense())
eigs = np.linalg.eigvalsh(
nx.directed_laplacian_matrix(G).tolist().todense())
spectral_pmf, _ = np.histogram(eigs,
bins=200,
range=(-1e-5, 2),
density=False)
spectral_pmf = spectral_pmf / spectral_pmf.sum()
# from scipy import stats
# kernel = stats.gaussian_kde(eigs)
# positions = np.arange(0.0, 2.0, 0.1)
# spectral_density = kernel(positions)
# import pdb; pdb.set_trace()
return spectral_pmf
def networkX_laplacian(adj_graph, sym, norm, weight='weight'):
# Compute the laplacian matrix using NetworkX built in functions.
#
# Can take symmetric or directed graphs (sym='sym' or sym='').
# Can take weighted or unweighted graphs (weight='weight' or weight=None).
# Can create normalized or non-normalized matrices (norm='norm' or norm='').
#
if (sym=='sym') and (norm=='norm'):
B = nx.normalized_laplacian_matrix(adj_graph,weight=weight).A # nx returns is numpy martrix, .A returns array.
elif sym=='' and norm=='norm':
B = nx.directed_laplacian_matrix(adj_graph,weight=weight).A # nx returns is numpy martrix, .A returns array.
else:
B = nx.laplacian_matrix(adj_graph,weight=weight).toarray() # nx returns scipy sparse, toarray() returns a np array.
# Ensures cohesion with nx normalized and nx directed modules,
# and supports np.array indexing conventions.
return B
def run(self, graph, n=2000, c=0.1):
"""
:param graph: not a networkx graph, but an instance of the wrapper class in graphs.py
:param n: number of wave equation time steps
:param c: wave speed
"""
# graph state should be uninitialized; if not, then clear it
graph.state = None
# initialize t = -1, 0 states
init = np.random.uniform(size=(1, len(graph.nodes)))
graph.update_state(np.append(init, init, axis=0))
if graph.is_directed():
laplacian = np.asarray(nx.directed_laplacian_matrix(graph))
else:
laplacian = nx.laplacian_matrix(graph).toarray()
relaxation = tqdm(range(n))
relaxation.set_description('Relaxation: ')
for t in relaxation:
# sum over all neighbors
neighbors = np.dot(
(2 * np.identity(len(laplacian)) - (c**2) * laplacian),
graph.state[t + 1])
# t-1, t-2
relaxation = graph.state[t]
graph.update_state([neighbors - relaxation])
# if a WaveEquationSimulation acts on a single graph at a time, then functions below like
# highlight_clusters, etc only make sense inside WaveEquationSimulation if it maintains a local copy of
# the current graph after running the simulation
self.graph = graph
self.peaks = self.__fft_peaks()
dummy_peaks = {}
mapping = sorted(self.graph.nodes)
for it, key in enumerate(sorted(self.peaks.keys())):
dummy_peaks[mapping[it]] = self.peaks[key]
self.peaks = dummy_peaks
def SVD_perSlice(G_times, directed=True, num_eigen=6, top=True, max_size=500):
Temporal_eigenvalues = []
activity_vecs = [] #eigenvector of the largest eigenvalue
counter = 0
for G in G_times:
if (len(G) < max_size):
for i in range(len(G), max_size):
G.add_node(
-1 *
i) #add empty node with no connectivity (zero padding)
if (directed):
L = nx.directed_laplacian_matrix(G)
else:
L = nx.laplacian_matrix(G)
L = L.asfptype()
#top = False
#compute svd, find diagonal matrix and append the diagonal entries
#only consider 6 eigenvalues for now as the number of graph is small
#num_eigenvalues=6
#k=min(L.shape)-1
if (top):
which = "LM"
else:
which = "SM"
u, s, vh = svds(L, k=num_eigen, which=which)
# u, s, vh = randomized_svd(L, num_eigen)
vals = s
vecs = u
#vals, vecs= LA.eig(L)
max_index = list(vals).index(max(list(vals)))
activity_vecs.append(np.asarray(vecs[max_index]))
Temporal_eigenvalues.append(np.asarray(vals))
print("processing " + str(counter), end="\r")
counter = counter + 1
return (Temporal_eigenvalues, activity_vecs)
def eigenDAG(DAG, DAG_Size, Top_k_Eigenvalue_Number, norm=False):
# Matrix to Sparse
DAG = sp.csr_matrix(DAG)
# Create Graph
G = nx.DiGraph(DAG) # Create The Directed Graph
# Calculate Directed laclacian
Laplacian = nx.directed_laplacian_matrix(G,
nodelist=None,
weight='weight',
walk_type=None,
alpha=0.95)
# Normalize the matrix
if norm:
Laplacian = NormalizeMatrix(Laplacian)
# Eigen value of Laplacian
eigenvalues, eigenvectors = np.linalg.eig(Laplacian)
# Sorting the eigenvalues
np.matrix.sort(eigenvalues)
# Top K EigenValues
Top_k_Eigenvalue = eigenvalues[(DAG_Size -
Top_k_Eigenvalue_Number):DAG_Size]
## If the test is for 2nd Eigen Value then this line will choose the 2nd one otherwise 1st one
Top_k_Eigenvalue = Top_k_Eigenvalue[0]
# Getting the index for Max value
Top_k_Eigenvalue_Index = sorted(range(len(eigenvalues)),
key=lambda i: eigenvalues[i])[-2:]
# List of Top Eigen Vactors
Top_k_Eigenvector = np.zeros
Top_k_Eigenvector = eigenvectors[:, Top_k_Eigenvalue_Index[0]]
for i in range(Top_k_Eigenvalue_Number - 1):
Top_k_Eigenvector = np.column_stack(
(Top_k_Eigenvector, eigenvectors[:,
Top_k_Eigenvalue_Index[i + 1]]))
return Top_k_Eigenvalue, Top_k_Eigenvector, Top_k_Eigenvalue_Index, Laplacian
fig = plt.Figure()
fig.set_canvas(plt.gcf().canvas)
nx.draw_networkx_nodes(B, pos, node_size=200,node_color=col ,cmap=plt.get_cmap('jet'),labels=B.nodes())#,,
nx.draw_networkx_labels(B,pos,font_size=10)
nx.draw_networkx_edges(B, pos, edge_color='b', arrows=True)
os.chdir(fig_path)
plt.savefig('bipartite_Network_1.pdf')
#plt.show()
### Algebraic connectivity
### second-smallest eigenvalue of the Laplacian matrix
L=nx.directed_laplacian_matrix(G)
eigenValues,eigenVectors=np.linalg.eig(L)
idx = eigenValues.argsort()
eigenValues = eigenValues[idx]
eigenVectors = eigenVectors[:,idx]
sec_min=idx[1]
np.set_printoptions(precision=4)
AC=eigenValues[sec_min]
g_ac = "%0.4f" % AC