def backward_pass(self, var, normalize=False):
w = self.weights[var]
if self.get_child_factors(var):
phi = factor_product(*self.get_child_factors(var))
else:
phi = 1.0
if self.children[var]:
phi = factor_product(
*[phi],
*[self.messages[(child, var)] for child in self.children[var]])
if self.parents[var]:
phi = factor_product(*[phi],
*[self.messages[(self.parents[var], var)]])
phi.abs()
for child in self.children[var]:
wc = self.weights[child]
m = self.messages[(child, var)].copy()
self.messages[(var, child)] = factor_product(
m.power(-1.0, inplace=False), phi.power(wc / w, inplace=False))
self.messages[(var, child)].weightedsum(
list(set(phi.variables) - set(m.variables)), wc)
self.messages[(var,
child)].name = "M_" + str(var) + "->" + str(child)
if normalize:
self.messages[(var, child)].normalize()
def update_reparam(self, ovar, method="projected_gradient"):
if method == "projected_gradient":
step_size = self.reparam_step_size
b = {
var: self.get_q(var).marginalize_except([ovar], inplace=False)
for var in self.nvardict[ovar]
}
b_avg = np.sum(
np.array([b[var].values for var in self.nvardict[ovar]]),
0) / len(self.nvardict[ovar])
for var in self.nvardict[ovar]:
temp = b[var].copy()
temp.values = np.exp(-step_size * (temp.values - b_avg))
factor = self.model.get_factor(self.adj_factor_names[var][0])
factor.product(temp)
elif method == "fixedpoint":
b = {
var: self.get_q(var).marginalize_except([ovar], inplace=False)
for var in self.nvardict[ovar]
}
b_avg = factor_product(*[
b[var].power(self.weights[var], inplace=False)
for var in self.nvardict[ovar]
])
for var in self.nvardict[ovar]:
temp = b[var].copy()
temp.values = b_avg.values / b[var].values
temp.power(self.weights[var])
factor = self.model.get_factor(self.adj_factor_names[var][0])
factor.product(temp)
def forward_pass(self, var, normalize=False):
if self.parents[var]:
w = self.weights[var]
self.messages[(var, self.parents[var])] = factor_product(
*self.get_child_factors(var),
*[self.messages[(child, var)] for child in self.children[var]])
self.messages[(var, self.parents[var])].weightedsum(
[self.ovardict[var]], w)
if normalize:
self.messages[(var, self.parents[var])].normalize()
def get_q(self, var):
w = self.weights[var]
if self.get_child_factors(var):
phi = factor_product(*self.get_child_factors(var))
else:
phi = 1.0
if self.children[var]:
phi = factor_product(
*[phi],
*[self.messages[(child, var)] for child in self.children[var]])
if self.parents[var]:
phi = factor_product(*[phi],
*[self.messages[(self.parents[var], var)]])
phi.abs()
phi.power(1 / w)
phi.normalize()
phi.name = "q_" + str(var)
return phi.copy()
def compute_Z(self, add_time=False):
if add_time:
t = time()
for ovar in self.model.variables:
for var in self.nvardict[ovar]:
self.forward_pass(var, normalize=False)
Z = 1.0
for var in self.variables:
w = self.weights[var]
if not self.parents[var]:
phi = factor_product(
*self.get_child_factors(var), *[
self.messages[(child, var)]
for child in self.children[var]
])
Z *= phi.weightedsum([self.ovardict[var]], w,
inplace=False).values
self.Z = Z * np.exp(self.baselogZ)
self.logZ = np.log(Z) + self.baselogZ
if add_time:
self.elapsed_time += time() - t
def __init__(self, model, ibound):
self.model = model.copy()
self.logZ = 0.0
self.Z = 0.0
self.baselogZ = 0.0
self.messages = {}
self.reparam = {
var: np.zeros(model.get_nstate(var))
for var in model.variables
}
self.gauges = {
var: Factor(
"G" + str(var),
[var + "_", var],
[model.get_nstate(var),
model.get_nstate(var)],
np.eye(model.get_nstate(var)),
)
for var in model.variables
}
self.build_cover_graph(ibound)
for var in self.variables:
if self.parents[var]:
self.messages[(var, self.parents[var])] = factor_product(
*self.get_child_factors(var), *[
self.messages[(child, var)]
for child in self.children[var]
])
self.messages[(var, self.parents[var])].marginalize(
[self.ovardict[var]])
self.messages[(var, self.parents[var])].values = np.full_like(
self.messages[(var, self.parents[var])].values, 1.0)
self.messages[(var, self.parents[var])].name = (
"M_" + str(var) + str(self.parents[var]))
for var in reversed(self.variables):
if self.get_child_factors(var):
phi = factor_product(*self.get_child_factors(var))
else:
phi = 1.0
if self.children[var]:
phi = factor_product(
*[phi], *[
self.messages[(child, var)]
for child in self.children[var]
])
if self.parents[var]:
phi = factor_product(
*[phi], *[self.messages[(self.parents[var], var)]])
for child in self.children[var]:
m = self.messages[(child, var)].copy()
self.messages[(var, child)] = factor_product(phi, m)
self.messages[(var, child)].marginalize(
list(set(phi.variables) - set(m.variables)))
self.messages[(var, child)].values = np.full_like(
self.messages[(var, child)].values, 1.0)
self.messages[(
var, child)].name = "M_" + str(var) + "->" + str(child)
self.reparam_step_size = 0.1
self.weight_step_size = 0.1
self.gauge_step_size = 0.005
self.elapsed_time = 0.0
def get_gauges_der_factor(self, ovariable, factor):
factor_der = {
gauge_idx: np.full_like(factor.values, 0.0)
for gauge_idx, gauge_val in np.ndenumerate(
self.gauges[ovariable].values)
}
temp = {
gauge_idx: np.full_like(factor.values, 0.0)
for gauge_idx, gauge_val in np.ndenumerate(
self.gauges[ovariable].values)
}
gauge_list = {}
if factor == self.model.get_adj_factors(ovariable)[0]:
top_factor = factor.copy()
factor_order = copy(factor.variables)
for ovar in factor.variables:
if ovar != ovariable:
if self.model.get_adj_factors(ovar).index(factor) == 0:
gauge = self.gauges[ovar].copy()
elif self.model.get_adj_factors(ovar).index(factor) == 1:
gauge = self.gauges[ovar].invT(inplace=False)
top_factor = factor_product(top_factor, gauge)
top_factor.marginalize([ovar])
top_factor.variables[top_factor.variables.index(
ovar + "_")] = ovar
top_factor.values = top_factor.values.transpose(
[top_factor.variables.index(ovar) for ovar in factor_order])
top_factor.variables = copy(factor_order)
for gauge_idx, gauge_val in np.ndenumerate(
self.gauges[ovar].values):
for factor_idx, factor_val in np.ndenumerate(factor.values):
if gauge_idx[0] == factor_idx[factor.variables.index(
ovariable)]:
top_idx = list(copy(factor_idx))
top_idx[top_factor.variables.index(
ovariable)] = gauge_idx[1]
factor_der[gauge_idx][factor_idx] = top_factor.values[
tuple(top_idx)]
return factor_der
elif factor == self.model.get_adj_factors(ovariable)[1]:
top_factor = factor.copy()
factor_order = copy(factor.variables)
for ovar in factor.variables:
if ovar != ovariable:
if self.model.get_adj_factors(ovar).index(factor) == 0:
gauge = self.gauges[ovar].copy()
elif self.model.get_adj_factors(ovar).index(factor) == 1:
gauge = self.gauges[ovar].invT(inplace=False)
top_factor = factor_product(top_factor, gauge)
top_factor.marginalize([ovar])
top_factor.variables[top_factor.variables.index(
ovar + "_")] = ovar
top_factor.values = top_factor.values.transpose(
[top_factor.variables.index(ovar) for ovar in factor_order])
top_factor.variables = copy(factor_order)
for gauge_idx, gauge_val in np.ndenumerate(
self.gauges[ovar].values):
for factor_idx, factor_val in np.ndenumerate(factor.values):
if gauge_idx[1] == factor_idx[factor.variables.index(
ovariable)]:
top_idx = list(copy(factor_idx))
top_idx[top_factor.variables.index(
ovariable)] = gauge_idx[0]
factor_der[gauge_idx][factor_idx] -= top_factor.values[
tuple(top_idx)]
return factor_der
else:
print("Something wrong")
return False