def predict_communities(self, deg_corr):
if self.is_weighted:
state = gt.minimize_blockmodel_dl(self.coocurence_graph,
overlap=self.allow_overlap,
deg_corr=deg_corr,
layers=True,
state_args=dict(ec=self.weights,
layers=False))
else:
state = gt.minimize_blockmodel_dl(self.coocurence_graph,
overlap=self.allow_overlap,
deg_corr=deg_corr)
state = state.copy(B=self.coocurence_graph.num_vertices())
self.dls_[deg_corr] = [] # description length history
self.vm_[deg_corr] = None # vertex marginals
self.em_[deg_corr] = None # edge marginals
self.h_[deg_corr] = np.zeros(self.coocurence_graph.num_vertices() + 1)
def collect_marginals(s, deg_corr, obj):
obj.vm_[deg_corr] = s.collect_vertex_marginals(obj.vm_[deg_corr])
obj.em_[deg_corr] = s.collect_edge_marginals(obj.em_[deg_corr])
obj.dls_[deg_corr].append(s.entropy())
B = s.get_nonempty_B()
obj.h_[deg_corr][B] += 1
collect_marginals_for_class = lambda s: collect_marginals(
s, deg_corr, self)
# Now we collect the marginal distributions for exactly 200,000 sweeps
gt.mcmc_equilibrate(state,
force_niter=self.n_iters,
mcmc_args=dict(niter=self.n_init_iters),
callback=collect_marginals_for_class,
**self.equlibrate_options)
S_mf = gt.mf_entropy(self.coocurence_graph, self.vm_[deg_corr])
S_bethe = gt.bethe_entropy(self.coocurence_graph,
self.em_[deg_corr])[0]
L = -np.mean(self.dls_[deg_corr])
self.state_[deg_corr] = copy.copy(state)
self.S_bethe_[deg_corr] = copy.copy(S_bethe)
self.S_mf_[deg_corr] = copy.copy(S_mf)
self.L_[deg_corr] = copy.copy(L)
if self.verbose:
print(("Model evidence for deg_corr = %s:" % deg_corr, L + S_mf,
"(mean field),", L + S_bethe, "(Bethe)"))
def collect_marginals(s):
global pv, pe
b = gt.perfect_prop_hash([s.b])[0]
pv = s.collect_vertex_marginals(pv, b=b)
pe = s.collect_edge_marginals(pe)
dls.append(s.entropy())
# Apply MCMC
gt.mcmc_equilibrate(state,
force_niter=10000,
mcmc_args=dict(niter=10),
callback=collect_marginals)
entropy = state.entropy()
S_mf = gt.mf_entropy(g, pv)
S_bethe = gt.bethe_entropy(g, pe)[0]
L = -np.mean(dls)
nClass = len(np.unique(state.get_blocks().a))
print("%d classes, entropy %f, mean_field %f, bethe %f" %
(nClass, entropy, L + S_mf, L + S_bethe))
# Save final graph
fname = "SBM_mcmc_%d_%f" % (nClass, entropy)
write_classes(os.path.join(outdir, fname + ".tsv"), g, state)
pickle.dump([g, state, dls, pv, pe],
open(os.path.join(outdir, fname + ".pickle"), "wb"), -1)
g.save(os.path.join(outdir, fname + ".gt.gz"))
# Find the maximal nClass over all mcmc sweeps
nClassMax = 0
for v in g.vertices():