def _e_step(self):
logging.debug('E-Step')
if not self._lp_init:
self._init_LP()
nprt = len(self.parts)
# individual potentials
logging.debug('Obj func - indiv potentials')
c_indiv = cvx_mat([0.0] * len(self.author_graph) * nprt)
for a in self.author_graph:
c_indiv[a*nprt:(a+1)*nprt] = [self._log_phi(a, p) for p in self.parts]
objF = -cvx_dot(c_indiv, self.alpha)
# pairwise potentials
logging.debug('Obj func - pair potentials')
objF -= cvx_dot(self.c_pair, self.beta2)
lp = cvx_op(objF, self.constraints)
# if logging.getLogger('root').level != logging.DEBUG:
# lp.options['show_progress'] = False
# solve the LP
logging.debug('Solving the LP')
lp.solve('sparse', 'glpk')
logging.debug('Solving the LP Done')
# hard partitions for nodes (authors)
self.partition = {}
for a in self.author_graph:
self.partition[a] = np.argmax(self.alpha.value[a*nprt:(a+1)*nprt])
logging.debug('E-Step Done')
def _e_step(self):
logging.debug('E-Step')
if not self._lp_init:
self._init_LP()
nprt = len(self.parts)
# individual potentials
logging.debug('Obj func - indiv potentials')
c_indiv = cvx_mat([0.0] * len(self.author_graph) * nprt)
for a in self.author_graph:
c_indiv[a * nprt:(a + 1) *
nprt] = [self._log_phi(a, p) for p in self.parts]
objF = -cvx_dot(c_indiv, self.alpha)
# pairwise potentials
logging.debug('Obj func - pair potentials')
objF -= cvx_dot(self.c_pair, self.beta2)
lp = cvx_op(objF, self.constraints)
# if logging.getLogger('root').level != logging.DEBUG:
# lp.options['show_progress'] = False
# solve the LP
logging.debug('Solving the LP')
lp.solve('sparse', 'glpk')
logging.debug('Solving the LP Done')
# hard partitions for nodes (authors)
self.partition = {}
for a in self.author_graph:
self.partition[a] = np.argmax(self.alpha.value[a * nprt:(a + 1) *
nprt])
logging.debug('E-Step Done')
def _init_LP(self):
logging.debug('Init LP')
# add constraints once here
# constraints
self.constraints = []
nprt = len(self.parts)
self.alpha = cvx_var(size=len(self.author_graph) * nprt, name='alpha')
self.beta2 = cvx_var(size=self.author_graph.size(), name='beta2')
self.beta3 = cvx_var(size=self.author_graph.size() * nprt,
name='beta3')
# node ids are integers (handled in the constructor)
logging.debug('Init LP - indiv constraints')
# alphas are indicator vars
for a in self.author_graph:
self.constraints.append(
cvx_sum(self.alpha[a * nprt + p] for p in self.parts) == 1)
for a in self.alpha:
self.constraints.append(a >= 0)
# beta2 is the sum of beta3s
logging.debug('Init LP - pair constraints')
for ei, (a, b) in enumerate(self.author_graph.edges()):
# if self.author_graph[a][b]['denom'] <= 2:
# continue
self.constraints.append(
0.5 * cvx_sum(self.beta3[ei * nprt + p]
for p in self.parts) == self.beta2[ei])
for p in self.parts:
self.constraints.append(
self.alpha[a * nprt + p] -
self.alpha[b * nprt + p] <= self.beta3[ei * nprt + p])
self.constraints.append(
self.alpha[b * nprt + p] -
self.alpha[a * nprt + p] <= self.beta3[ei * nprt + p])
# pre compute pair-wise costs
logging.debug('Init LP - pair costs')
self.c_pair = cvx_mat([0.0] * self.author_graph.size())
s = log(1 - self.TAU) - log(self.TAU)
for ei, (a, b) in enumerate(self.author_graph.edges()):
# if self.author_graph[a][b]['denom'] <= 2:
# continue
self.c_pair[ei] = self.author_graph[a][b]['weight'] * s
self._lp_init = True
logging.debug('Init LP Done')
def _init_LP(self):
logging.debug('Init LP')
# add constraints once here
# constraints
self.constraints = []
nprt = len(self.parts)
self.alpha = cvx_var(size=len(self.author_graph) * nprt, name='alpha')
self.beta2 = cvx_var(size=self.author_graph.size(), name='beta2')
self.beta3 = cvx_var(size=self.author_graph.size() * nprt, name='beta3')
# node ids are integers (handled in the constructor)
logging.debug('Init LP - indiv constraints')
# alphas are indicator vars
for a in self.author_graph:
self.constraints.append(cvx_sum(self.alpha[a*nprt + p] for p in self.parts) == 1)
for a in self.alpha:
self.constraints.append(a >= 0)
# beta2 is the sum of beta3s
logging.debug('Init LP - pair constraints')
for ei, (a, b) in enumerate(self.author_graph.edges()):
# if self.author_graph[a][b]['denom'] <= 2:
# continue
self.constraints.append(0.5 * cvx_sum(self.beta3[ei*nprt + p] for p in self.parts) == self.beta2[ei])
for p in self.parts:
self.constraints.append(self.alpha[a*nprt + p] - self.alpha[b*nprt + p] <= self.beta3[ei*nprt + p])
self.constraints.append(self.alpha[b*nprt + p] - self.alpha[a*nprt + p] <= self.beta3[ei*nprt + p])
# pre compute pair-wise costs
logging.debug('Init LP - pair costs')
self.c_pair = cvx_mat([0.0] * self.author_graph.size())
s = log(1 - self.TAU) - log(self.TAU)
for ei, (a, b) in enumerate(self.author_graph.edges()):
# if self.author_graph[a][b]['denom'] <= 2:
# continue
self.c_pair[ei] = self.author_graph[a][b]['weight'] * s
self._lp_init = True
logging.debug('Init LP Done')
