def estimate(self, model, transactions):
likelihood_loss_function_coefficients = self.likelihood_loss_function_coefficients(
transactions)
new_likelihood = model.log_likelihood_for(transactions)
max_iterations = len(likelihood_loss_function_coefficients)
cpu_time = time_for_optimization(
partial_time=Settings.instance(
).non_linear_solver_partial_time_limit(),
total_time=Settings.instance().solver_total_time_limit(),
profiler=self.profiler())
start_time = time.time()
for _ in range(max_iterations):
old_likelihood = new_likelihood
possible_mnl_model = self.look_for_new_mnl_model(
model, likelihood_loss_function_coefficients)
model.add_new_class_with(possible_mnl_model)
self.update_weights_for(model,
likelihood_loss_function_coefficients)
new_likelihood = model.log_likelihood_for(transactions)
likelihood_does_not_increase = new_likelihood < old_likelihood
likelihood_does_not_increase_enough = abs(
new_likelihood - old_likelihood) / len(transactions) < 1e-7
time_limit = (time.time() - start_time) > cpu_time
if likelihood_does_not_increase or likelihood_does_not_increase_enough or time_limit:
break
return model
def estimate(self, model, transactions):
self.profiler().reset_convergence_criteria()
self.profiler().update_time()
model = self.custom_initial_solution(model, transactions)
cpu_time = time_for_optimization(partial_time=Settings.instance().non_linear_solver_partial_time_limit(),
total_time=Settings.instance().solver_total_time_limit(),
profiler=self.profiler())
start_time = time.time()
while True:
self.profiler().start_iteration()
model = self.one_step(model, transactions)
likelihood = model.log_likelihood_for(transactions)
self.profiler().stop_iteration(likelihood)
if self.profiler().should_stop() or (time.time() - start_time) > cpu_time:
break
return model
def cpu_time(self, profiler):
return time_for_optimization(
partial_time=Settings.instance().linear_solver_partial_time_limit(
),
total_time=Settings.instance().solver_total_time_limit(),
profiler=profiler)