def setup_factor_graph(self):
self.vecfac = dai.VecFactor()
elem_map = {}
for elem in self.variable_map:
if elem.variable_name not in elem_map:
elem_map[elem.variable_name] = {
elem.variable_type: elem.variable_id
}
else:
elem_map[elem.variable_name][
elem.variable_type] = elem.variable_id
#dai_var_map[elem.variable] = dai.Var(elem.variable, elem.dim)
for cpt in sorted(self.cpt_map.keys(), key=lambda x: x.factor_id):
cpt_value = self.cpt_map[cpt]
var_list = dai.VarSet()
v_list = list(cpt_value.variables)
v_list.sort()
for v_id in v_list:
v = self.variable_map.get_variable_by_id(v_id)
if v not in self.dai_variables:
self.dai_variables[v] = dai.Var(v.variable_id, v.dim)
var_list.append(self.dai_variables[v])
dai_factor = dai.Factor(var_list)
for i, v in enumerate(cpt_value.cpt_linear_table()):
dai_factor[i] = v
self.vecfac.append(dai_factor)
def build_single_node_factor(variables, fixed_variables, var_idx, f):
'''
copied from https://github.com/jkuck/mrf_nesting_ub/blob/master/Factor_Graphs/libdai_ising_model.py
Inputs:
- variables: (list of dai.Var) available variables.
- fixed_variables: (dictionary)
key: (int) 0 to N-1 variable index
value: (int) -1 or 1, the value the variable is fixed to
- var_idx: (int) variable index, 0 to N-1, for this factor's node
- f: (float) local field at this node
Outputs:
- factor: (dai.Factor)
'''
clause_variables = dai.VarSet(variables[var_idx])
factor = dai.Factor(clause_variables)
if var_idx in fixed_variables:
factor[0] = np.exp(fixed_variables[var_idx] *
f) #the single fixed factor value
else:
factor[0] = np.exp(-f) #corresponding to the node taking value -1
factor[1] = np.exp(f) #corresponding to the node taking value 1
return factor
def build_libdaiFactor_fromClause(clause, variables):
'''
Inputs:
- clause (list of integers): literals in the clause
- variables (list of dai.Var): list of all dai variables in the sat problem.
variable i (where i can be in [1, N]) is stored at variables[i-1]
Outputs:
- factor (dai.Factor)
'''
dai_vars_list = [variables[abs(literal) - 1] for literal in clause]
if len(clause) == 1:
dai_vars = dai.VarSet(dai_vars_list[0])
else:
dai_vars = dai.VarSet(dai_vars_list[-1], dai_vars_list[-2])
#have to reverse the order of the variables because of the way libdai
#stores factor potentials with respect to variable ordering
# see https://github.com/dbtsai/libDAI/blob/master/swig/example_sprinkler.py
for var in reversed(dai_vars_list[:-2]):
dai_vars.append(var)
factor = dai.Factor(dai_vars)
#Create a tensor for the 2^state_dimensions states
state = np.zeros([2 for i in range(len(clause))])
#Iterate over the 2^state_dimensions variable assignments and set those to 1 that satisfy the clause
for indices in np.ndindex(state.shape):
set_to_1 = False
for dimension, index_val in enumerate(indices):
if clause[dimension] > 0 and index_val == 1:
set_to_1 = True
elif clause[dimension] < 0 and index_val == 0:
set_to_1 = True
if set_to_1:
state[indices] = 1
else:
state[indices] = 0
for idx, state_val in enumerate(state.flatten()):
factor[idx] = state_val
return factor
def build_pairwise_factor(variables, fixed_variables, var_idx1, var_idx2, c):
'''
copied from https://github.com/jkuck/mrf_nesting_ub/blob/master/Factor_Graphs/libdai_ising_model.py
Inputs:
- variables: (list of dai.Var) available variables.
- fixed_variables: (dictionary)
key: (int) 0 to N-1 variable index
value: (int) -1 or 1, the value the variable is fixed to
- var_idx1: (int) variable index, 0 to N-1, for the first node in this factor
- var_idx2: (int) variable index, 0 to N-1, for the second node in this factor
- c: (float) coupling strength for this factor
Outputs:
- factor: (dai.Factor)
'''
clause_variables = dai.VarSet(variables[var_idx1], variables[var_idx2])
factor = dai.Factor(clause_variables)
#this 'pairwise' factor is over two fixed variables and has only 1 state
if (var_idx1 in fixed_variables) and (var_idx2 in fixed_variables):
factor[0] = np.exp(fixed_variables[var_idx1] *
fixed_variables[var_idx2] * c)
#this 'pairwise' factor is over one fixed variable and one binary variable and has 2 states
elif (var_idx1 in fixed_variables) and (var_idx2 not in fixed_variables):
factor[0] = np.exp(-fixed_variables[var_idx1] * c) # V2 = -1
factor[1] = np.exp(fixed_variables[var_idx1] * c) # V2 = -1
#this 'pairwise' factor is over one fixed variable and one binary variable and has 2 states
elif (var_idx1 not in fixed_variables) and (var_idx2 in fixed_variables):
factor[0] = np.exp(-fixed_variables[var_idx2] * c) # V1 = -1
factor[1] = np.exp(fixed_variables[var_idx2] * c) # V1 = -1
#this pairwise factor is over two binary variables and has 4 states
elif (var_idx1 not in fixed_variables) and (var_idx2
not in fixed_variables):
factor[0] = np.exp(c) # V1 = -1, V2 = -1
factor[1] = np.exp(-c) # V1 = -1, V2 = 1
factor[2] = np.exp(-c) # V1 = 1, V2 = -1
factor[3] = np.exp(c) # V1 = 1, V2 = 1
else:
assert (False), "This shouldn't happen!!?"
return factor
# Copyright (C) 2009 Joris Mooij [joris dot mooij at libdai dot org] # This example program illustrates how to construct a factorgraph # by means of the sprinkler network example discussed at # http://www.cs.ubc.ca/~murphyk/Bayes/bnintro.html # using the SWIG python wrapper of libDAI import dai C = dai.Var(0, 2) # Define binary variable Cloudy (with label 0) S = dai.Var(1, 2) # Define binary variable Sprinkler (with label 1) R = dai.Var(2, 2) # Define binary variable Rain (with label 2) W = dai.Var(3, 2) # Define binary variable Wetgrass (with label 3) # Define probability distribution for C P_C = dai.Factor(C) P_C[0] = 0.5 # C = 0 P_C[1] = 0.5 # C = 1 # Define conditional probability of S given C P_S_given_C = dai.Factor(dai.VarSet(S, C)) P_S_given_C[0] = 0.5 # C = 0, S = 0 P_S_given_C[1] = 0.9 # C = 1, S = 0 P_S_given_C[2] = 0.5 # C = 0, S = 1 P_S_given_C[3] = 0.1 # C = 1, S = 1 # Define conditional probability of R given C P_R_given_C = dai.Factor(dai.VarSet(R, C)) P_R_given_C[0] = 0.8 # C = 0, R = 0 P_R_given_C[1] = 0.2 # C = 1, R = 0 P_R_given_C[2] = 0.2 # C = 0, R = 1
