def true_log_liklihood_per_example(example, var_in_clique, true_parameters,
cardinalities):
log_liklihood = 0
for each_clique in range(len(var_in_clique)):
var_in_this_clique = list(var_in_clique[each_clique])
cardinalities_of_var_in_this_clique = cardinalities[var_in_this_clique]
values_of_tuple = example[var_in_this_clique]
if len(var_in_this_clique) > 1:
index = helper.get_index_given_truth_values(
var_in_this_clique, values_of_tuple,
cardinalities_of_var_in_this_clique)
values_of_conjugate_tuple = values_of_tuple
if true_parameters[each_clique][index] == 0:
true_parameters[each_clique][index] += pow(10, -5)
if values_of_conjugate_tuple[-1] == 0:
values_of_conjugate_tuple[-1] = 1
elif values_of_conjugate_tuple[-1] == 1:
values_of_conjugate_tuple[-1] = 0
index_for_other_child = helper.get_index_given_truth_values(
var_in_this_clique, values_of_conjugate_tuple,
cardinalities_of_var_in_this_clique)
true_parameters[each_clique][index_for_other_child] -= pow(
10, -5)
elif true_parameters[each_clique][index] == 1:
true_parameters[each_clique][index] -= pow(10, -5)
if values_of_conjugate_tuple[-1] == 0:
values_of_conjugate_tuple[-1] = 1
elif values_of_conjugate_tuple[-1] == 1:
values_of_conjugate_tuple[-1] = 0
index_for_other_child = helper.get_index_given_truth_values(
var_in_this_clique, values_of_conjugate_tuple,
cardinalities_of_var_in_this_clique)
true_parameters[each_clique][index_for_other_child] += pow(
10, -5)
log_liklihood += log10(true_parameters[each_clique][index])
else:
if true_parameters[each_clique][0] == 0:
true_parameters[each_clique][0] += pow(10, -5)
true_parameters[each_clique][1] -= pow(10, -5)
elif true_parameters[each_clique][0] == 1:
true_parameters[each_clique][0] -= pow(10, -5)
true_parameters[each_clique][1] += pow(10, -5)
log_liklihood += log10(
true_parameters[each_clique][values_of_tuple[0]])
return log_liklihood
def get_parameters_given_weighted_example(examples, weights, var_in_clique, cardinalities_of_var, number_of_k): parameters_for_network = [[] for i in range(number_of_k)] parameters_for_k = [] # todo change the use of var_in_clique for each val of k cardinalities_of_var = array(cardinalities_of_var) examples = array(examples) var_in_clique_np = array(var_in_clique) for each_val_of_k in range(number_of_k): for each_clique in range(len(var_in_clique[each_val_of_k])): var_in_this_clique = var_in_clique_np[each_val_of_k][each_clique] vals_for_this_clique = examples[:, var_in_this_clique] val_of_k = examples[:, -1] cardinalities_of_var_in_this_clique = cardinalities_of_var[var_in_this_clique] total_tuples = product(cardinalities_of_var_in_this_clique) if len(var_in_this_clique) > 1: number_of_parameters = int(total_tuples / cardinalities_of_var_in_this_clique[-1]) nums = zeros([number_of_parameters, ]) denom = zeros([number_of_parameters, ]) for each_vals_for_this_clique, weights_for_this_example, value_of_k_for_this_example in zip( vals_for_this_clique, weights, val_of_k): if value_of_k_for_this_example == each_val_of_k: # indexing of parameters_for_network is done on the basis of parent variables only index_val_without_child = get_index_given_truth_values(var_in_this_clique[:-1], each_vals_for_this_clique[:-1], cardinalities_of_var_in_this_clique[ :-1]) denom[index_val_without_child] += weights_for_this_example if each_vals_for_this_clique[-1] == 0: nums[index_val_without_child] += weights_for_this_example parameter_val = divide(nums, denom) parameter_val = nan_to_num(parameter_val) parameter_val[parameter_val == 0] = pow(10, -5) parameter_val[parameter_val == 1] = 1 - pow(10, -5) parameters_for_network[each_val_of_k].append(parameter_val) else: count_of_zero = 0 total_count = 0 for each_vals_for_this_clique, weights_for_this_example, value_of_k_for_this_example in zip( vals_for_this_clique, weights, val_of_k): if value_of_k_for_this_example == each_val_of_k: total_count += weights_for_this_example if each_vals_for_this_clique[0] == 0: count_of_zero += weights_for_this_example parameter = nan_to_num(divide(count_of_zero, total_count)) if parameter == 0: parameter = pow(10, -5) elif parameter == 1: parameter = 1 - pow(10, -5) parameters_for_network[each_val_of_k].append(parameter) # parameters_for_network = array(parameters_for_network) nums_for_k = zeros([number_of_k, ]) denom_for_k = 0 for each_val_of_k, weight_for_k in zip(examples[:, -1], weights): nums_for_k[each_val_of_k] += weight_for_k denom_for_k += weight_for_k parameters_for_k = divide(nums_for_k, denom_for_k) return parameters_for_network, parameters_for_k
def train(examples, var_in_clique, markov, cardinalities_of_var):
"""
:param examples:
:param var_in_clique:
:param markov:
:param cardinalities_of_var:
:return: parameters in double form
"""
parameters = []
cardinalities_of_var = array(cardinalities_of_var)
if not markov:
var_in_clique_np = array(var_in_clique)
for each_clique in range(len(var_in_clique)):
var_in_this_clique = var_in_clique_np[each_clique]
vals_for_this_clique = examples[:, var_in_this_clique]
cardinalities_of_var_in_this_clique = cardinalities_of_var[
var_in_this_clique]
total_tuples = product(cardinalities_of_var_in_this_clique)
unique_vals, counts = unique(vals_for_this_clique,
return_counts=True,
axis=0)
if len(var_in_this_clique) > 1:
# indexing of parameters is done on the basis of parent variables only
number_of_parameters = int(
total_tuples / cardinalities_of_var_in_this_clique[-1])
# Using 1-Laplace Smoothing
nums = zeros([
number_of_parameters,
]) + 1
denom = zeros([
number_of_parameters,
]) + cardinalities_of_var_in_this_clique[-1]
# todo need to use parent only for parameters
for each_tuple in zip(unique_vals, counts):
index_val_without_child = helper.get_index_given_truth_values(
var_in_this_clique[:-1], each_tuple[0][:-1],
cardinalities_of_var_in_this_clique[:-1])
denom[index_val_without_child] += each_tuple[1]
temp = each_tuple[0][-1]
if each_tuple[0][-1] == 0:
nums[index_val_without_child] += each_tuple[1]
parameter_val = divide(nums, denom)
parameter_val = nan_to_num(parameter_val)
parameters.append(list(parameter_val))
else:
count_of_zero = 0
total_count = 0
for each_tuple in zip(unique_vals, counts):
each_tuple = array(each_tuple)
vals_for_this_clique = int(each_tuple[0])
total_count += each_tuple[1]
if vals_for_this_clique == 0:
count_of_zero += each_tuple[1]
parameters.append(
nan_to_num(divide(count_of_zero, total_count)))
return parameters
def complete_data(examples, cardinalities_of_var, parameters_for_network, parameters_for_k, var_in_clique):
final_examples = []
final_weights = []
for each_example in examples:
if -1 not in each_example:
weight = 1
final_examples.append(each_example)
final_weights.append(weight)
elif -1 in each_example:
weights_for_completed_examples = []
completed_examples = []
cardinalities_of_var = array(cardinalities_of_var)
var_not_present = len(cardinalities_of_var) - 1
cardinalities_of_var_not_present = cardinalities_of_var[var_not_present]
completions_for_var_not_present = list(range(cardinalities_of_var_not_present))
for each_tuple_missing_var in completions_for_var_not_present:
temp_example = deepcopy(each_example)
put(temp_example, var_not_present, each_tuple_missing_var)
weight_for_current_example = 1
value_of_k = temp_example[-1]
# for each tuple we have diff network thus weight will be different
for each_clique in range(len(var_in_clique[each_tuple_missing_var])):
if True:
values_for_current_clique = temp_example[var_in_clique[each_tuple_missing_var][each_clique]]
if len(var_in_clique[each_tuple_missing_var][each_clique]) > 1:
index_for_current_example = get_index_given_truth_values(
var_in_clique[each_tuple_missing_var][each_clique][:-1],
values_for_current_clique[:-1],
cardinalities_of_var[
var_in_clique[each_tuple_missing_var][each_clique][
:-1]])
if values_for_current_clique[-1] == 0:
weight_for_current_example *= parameters_for_network[value_of_k][each_clique][
index_for_current_example]
elif values_for_current_clique[-1] == 1:
weight_for_current_example *= (1 - parameters_for_network[value_of_k][each_clique][
index_for_current_example])
else:
print("Error")
elif len(var_in_clique[each_tuple_missing_var][each_clique]) == 1:
if values_for_current_clique[0] == 0:
weight_for_current_example *= parameters_for_network[value_of_k][each_clique]
elif values_for_current_clique[0] == 1:
weight_for_current_example *= (1 - parameters_for_network[value_of_k][each_clique])
else:
print("Error")
weight_for_current_example *= parameters_for_k[value_of_k]
weights_for_completed_examples.append(weight_for_current_example)
completed_examples.append(temp_example)
denom = sum(weights_for_completed_examples)
weights_for_completed_examples = divide(weights_for_completed_examples, denom)
for weight, example in zip(weights_for_completed_examples, completed_examples):
final_examples.append(example)
final_weights.append(float(weight))
return final_examples, final_weights
def find_log_liklihood_for_one_example(example, var_in_clique, parameters,
cardinalities):
log_liklihood = 0
cardinalities = array(cardinalities)
var_in_clique = array(var_in_clique)
for each_clique in range(len(var_in_clique)):
var_in_this_clique = list(var_in_clique[each_clique])
cardinalities_of_var_in_this_clique = cardinalities[var_in_this_clique]
values_of_tuple = example[var_in_this_clique]
if len(var_in_this_clique) > 1:
parents = values_of_tuple[:-1]
parents_var_in_this_clique = var_in_this_clique[:-1]
cardinalities_of_parents_in_this_clique = cardinalities_of_var_in_this_clique[:
-1]
child = values_of_tuple[-1]
index_for_parents = helper.get_index_given_truth_values(
parents_var_in_this_clique, parents,
cardinalities_of_parents_in_this_clique)
if parameters[each_clique][index_for_parents] == 0:
parameters[each_clique][index_for_parents] += pow(10, -5)
elif parameters[each_clique][index_for_parents] == 1:
parameters[each_clique][index_for_parents] -= pow(10, -5)
if child == 0:
log_liklihood += log10(
parameters[each_clique][index_for_parents])
elif child == 1:
log_liklihood += log10(
1 - parameters[each_clique][index_for_parents])
else:
print("Error line 37")
elif len(var_in_this_clique) == 1:
if parameters[each_clique] == 0:
parameters[each_clique] += pow(10, -5)
elif parameters[each_clique] == 1:
parameters[each_clique] -= pow(10, -5)
if values_of_tuple[0] == 0:
log_liklihood += log10(parameters[each_clique])
elif values_of_tuple[0] == 1:
log_liklihood += log10(1 - parameters[each_clique])
else:
print("Error line 48")
else:
print("Error line ")
return log_liklihood
def get_parameters_given_weighted_example(examples, weights, var_in_clique, cardinalities_of_var): parameters = [] cardinalities_of_var = array(cardinalities_of_var) examples = array(examples) var_in_clique_np = array(var_in_clique) for each_clique in range(len(var_in_clique)): var_in_this_clique = var_in_clique_np[each_clique] vals_for_this_clique = examples[:, var_in_this_clique] cardinalities_of_var_in_this_clique = cardinalities_of_var[var_in_this_clique] total_tuples = product(cardinalities_of_var_in_this_clique) if len(var_in_this_clique) > 1: number_of_parameters = int(total_tuples / cardinalities_of_var_in_this_clique[-1]) nums = zeros([number_of_parameters, ]) denom = zeros([number_of_parameters, ]) for each_vals_for_this_clique, weights_for_this_example in zip(vals_for_this_clique, weights): # indexing of parameters is done on the basis of parent variables only index_val_without_child = helper.get_index_given_truth_values(var_in_this_clique[:-1], each_vals_for_this_clique[:-1], cardinalities_of_var_in_this_clique[ :-1]) denom[index_val_without_child] += weights_for_this_example if each_vals_for_this_clique[-1] == 0: nums[index_val_without_child] += weights_for_this_example parameter_val = divide(nums, denom) parameter_val = nan_to_num(parameter_val) parameter_val[parameter_val == 0] = pow(10, -5) parameters.append(parameter_val) else: count_of_zero = 0 total_count = 0 for each_vals_for_this_clique, weights_for_this_example in zip(vals_for_this_clique, weights): total_count += weights_for_this_example if each_vals_for_this_clique[0] == 0: count_of_zero += weights_for_this_example parameter = nan_to_num(divide(count_of_zero, total_count)) if parameter == 0: parameter += pow(10, -5) parameters.append(parameter) return parameters
def find_log_liklihood_for_one_example_mixture_model(example, var_in_clique,
parameters,
parameters_for_k,
cardinalities):
log_liklihood = []
cardinalities = array(cardinalities)
var_in_clique = array(var_in_clique)
for each_val_of_k in range(len(parameters_for_k)):
log_liklihood_for_this_val_of_k = 1
for each_clique in range(len(var_in_clique[each_val_of_k])):
var_in_this_clique = var_in_clique[each_val_of_k][each_clique]
cardinalities_of_var_in_this_clique = cardinalities[
var_in_this_clique]
values_of_tuple = example[var_in_this_clique]
if len(var_in_this_clique) > 1:
parents = values_of_tuple[:-1]
parents_var_in_this_clique = var_in_this_clique[:-1]
cardinalities_of_parents_in_this_clique = cardinalities_of_var_in_this_clique[:
-1]
child = values_of_tuple[-1]
index_for_parents = helper.get_index_given_truth_values(
parents_var_in_this_clique, parents,
cardinalities_of_parents_in_this_clique)
if parameters[each_val_of_k][each_clique][
index_for_parents] == 0:
parameters[each_val_of_k][each_clique][
index_for_parents] += pow(10, -5)
elif parameters[each_val_of_k][each_clique][
index_for_parents] == 1:
parameters[each_val_of_k][each_clique][
index_for_parents] -= pow(10, -5)
if child == 0:
log_liklihood_for_this_val_of_k += log10(
parameters[each_val_of_k][each_clique]
[index_for_parents])
elif child == 1:
log_liklihood_for_this_val_of_k += log10(
1 - parameters[each_val_of_k][each_clique]
[index_for_parents])
else:
print("Error line 37")
elif len(var_in_this_clique) == 1:
if parameters[each_val_of_k][each_clique] == 0:
parameters[each_val_of_k][each_clique] += pow(10, -5)
elif parameters[each_val_of_k][each_clique] == 1:
parameters[each_val_of_k][each_clique] -= pow(10, -5)
if values_of_tuple[0] == 0:
log_liklihood_for_this_val_of_k += log10(
parameters[each_val_of_k][each_clique])
elif values_of_tuple[0] == 1:
log_liklihood_for_this_val_of_k += log10(
1 - parameters[each_val_of_k][each_clique])
else:
print("Error line 48")
else:
print("Error line ")
log_liklihood_for_this_val_of_k += log10(
parameters_for_k[each_val_of_k])
log_liklihood.append(log_liklihood_for_this_val_of_k)
log_liklihood_output = helper.log_sum_exp(log_liklihood)
return log_liklihood_output
def sampling_VE(num_of_var, cardinalities, num_of_cliques,
num_of_var_in_clique, var_in_clique, distribution_array,
w_cutset_bound, num_samples):
"""
This is the main function used to do sampling VE
@param num_of_var: The numbre of variables in the PGM
@param cardinalities: The cardialities of the PGM
@param num_of_cliques: The number of cliques in the PGM
@param num_of_var_in_clique: The number of variables in each clique
@param var_in_clique: The variables in each clique
@param distribution_array: The distribution array for given PGM
@param w_cutset_bound: The cutset bound for the PGM
@param num_samples: The number of samples for the algorithm
@return: The predicted value of Z or probability
"""
z = 0
X = w_cutset(num_of_var, var_in_clique, w_cutset_bound)
cardinalities = np.array(cardinalities)
num_of_var_in_x = len(X)
cardinalities_of_x = cardinalities[X]
num_of_cliques_in_X = 1
var_in_clique_X = X
if num_of_var_in_x != 0:
sum_of_log_of_cardinalities = np.sum(np.log10(cardinalities_of_x))
distribution_array_X = dict()
distribution_array_X["FIRST"] = [1 / sum_of_log_of_cardinalities]
uniform_dist = 1 / sum_of_log_of_cardinalities
else:
distribution_array_X = [1]
uniform_dist = 1
num = {}
denom = 0
weights = []
for each_N in range(num_samples):
distribution_array1 = distribution_array.copy()
var_in_clique1 = var_in_clique.copy()
num_of_var_in_clique1 = num_of_var_in_clique.copy()
num_of_var1 = num_of_var
evidence = []
for each_evidence in range(num_of_var_in_x):
evidence.append((var_in_clique_X[each_evidence],
randint(0,
cardinalities_of_x[each_evidence] - 1)))
# evidence = [(each_var, each_val) for (each_var, each_val) in zip(var_in_clique_X, sample)]
var_in_clique1, distribution_array1 = helper.instantiate(
num_of_var1, evidence, cardinalities, var_in_clique1,
distribution_array1)
var_elem_sol = variable_elimination.variable_elimination(
num_of_var1, cardinalities, num_of_cliques, num_of_var_in_clique1,
var_in_clique1, distribution_array1, evidence)
sample = [one_value for (var, one_value) in evidence]
sample_string = sample.__str__()
evidence_tuple = helper.get_index_given_truth_values(
var_in_clique_X, sample, cardinalities_of_x)
if distribution_array_X is list:
Q = uniform_dist
else:
if sample_string in distribution_array_X:
Q = distribution_array_X[sample_string]
else:
Q = uniform_dist
weight = var_elem_sol / Q
weights.append(weight)
denom += weight
if sample_string not in num:
num[sample_string] = weight
else:
num[sample_string] += weight
if each_N % 100 == 0 and each_N != 0 and distribution_array_X is not list:
distribution_array_X = {}
for each in num:
distribution_array_X[each] = num[each] / denom
weights = np.array(weights)
if (weights[1:] == weights[:-1]).all:
z = np.sum(weights)
else:
z = helper.threshold(helper.logsumexp(weights))
"""if weight != float('inf'):
z += weight
else:
z += sys.float_info.max * np.random.uniform(0,2)"""
return z / num_samples
