def estimate(self, xs, max_iterations=10000):
self.n = len(xs)
self.s2 = invgamma.rvs(self.alpha, scale=self.beta)
self.means = []
response_sums = []
self.counts = []
assignments = []
assignments.append(0)
response_sums.append(xs[0])
self.counts.append(1)
self.means.append(self.sample_posterior_mean(response_sums[0], 1,
self.s2, self.m0,
self.s20))
for i in range(1, len(xs)):
weights = self.__create_weight_vector(xs[i], self.means,
self.counts, self.s2,
self.m0, self.s20,
self.m)
# Sample an assignment for each item and update statistics
assignment = multinomial_sampling.sample(weights)
if assignment < len(self.means):
response_sums[assignment] += xs[i]
self.counts[assignment] += 1
# Create a new component
elif assignment == len(self.means):
response_sums.append(xs[i])
self.counts.append(1)
self.means.append(self.sample_posterior_mean(xs[i], 1,
self.s2, self.m0,
self.s20))
else:
print("WWWWW Should never get here. WWWWWW")
assignments.append(assignment)
for i in xrange(max_iterations):
# First sample an assignment for each data item
for j in xrange(len(xs)):
old_assignment = assignments[j]
response_sums[old_assignment] -= xs[j]
self.counts[old_assignment] -= 1
if self.counts[old_assignment] == 0:
self.__fix_hole(assignments, response_sums, self.means,
self.counts, old_assignment)
weights = self.__create_weight_vector(xs[j], self.means,
self.counts,
self.s2, self.m0,
self.s20, self.m)
new_assignment = multinomial_sampling.sample(weights)
# Create a new component
if new_assignment < len(self.means):
response_sums[new_assignment] += xs[j]
self.counts[new_assignment] += 1
elif new_assignment == len(self.means):
response_sums.append(xs[j])
self.counts.append(1)
self.means.append(self.sample_posterior_mean(xs[j], 1,
self.s2,
self.m0,
self.s20))
else:
print("EEEEEEE This should never happenEEEEEEEEEE")
assignments[j] = new_assignment
# Sample new values for the means
for j in xrange(len(self.means)):
self.means[j] = self.sample_posterior_mean(response_sums[j],
self.counts[j],
self.s2, self.m0,
self.s20)
# Sample new value for the component variance
res_sum_squares = self.calculate_res_sum_squares(xs, assignments,
self.means)
self.s2 = self.sample_posterior_var(self.n, res_sum_squares,
self.alpha, self.beta)
return self
def __rv(self, weights):
s = multinomial_sampling.sample(weights)
if s < len(self.means): # Draw from a component
return norm.rvs(self. means[s], scale=sqrt(self.s2))
else: # Draw from the marginal
return norm.rvs(self.m0, scale=sqrt(self.s20 + self.s2))