def __init__(self, training_data, dim_no=3, k_order=1, alpha=0.001):
LinearRegressor.__init__(self, training_data, dim_no, k_order, alpha, normalize=False)
def run_linear_regression_test(part):
# Construct data sets
dir_path = os.path.dirname(os.path.realpath(__file__))
data = ConcreteData.ConcreteData(dir_path +
'/../../Data/concrete/train.csv')
# Test data
test_data = ConcreteData.ConcreteData(dir_path +
'/../../Data/concrete/test.csv')
if part != 3:
manual_input = str(
input("Do you wish to input parameters manually? y/n\n"))
analytic_answer = str(
input(
"Do you wish to print out the analytic answer for comparison? y/n\n"
))
if manual_input == "y":
max_iters = int(
input(
"Please enter max number of iterations for gradient descent\n"
))
tolerance = float(
input("Please enter a tolerance for algorithm termination\n"))
step_size = float(
input(
"Please enter a constant step size for each iteration\n"))
else:
max_iters = 10000
tolerance = float(1e-6)
step_size = 0.02
grad_desc = GradientDescent.GradientDescent(data.features, data.output)
if part == 1:
step_size = 0.0145
result = grad_desc.fit(obj_func=lr.objective_function,
grad_func=lr.obj_gradient_function,
args=None,
max_iters=max_iters,
step_function=lambda x: step_size,
tolerance=tolerance)
else:
result = grad_desc.fit_stochastic(
obj_func=lr.objective_function,
grad_func=lr.stoch_gradient_function,
args=None,
max_iters=max_iters,
step_function=lambda x: step_size,
tolerance=tolerance)
print("step size was " + str(step_size))
if result[1] == max_iters + 1:
print("Iteration did not converge after " + str(max_iters) +
" iterations. Resultant vector is:")
print(result[0])
print("Try with another constant step size.")
else:
print("Success! Converged after " + str(result[1]) +
" iterations. Resultant vector is: ")
print(result[0])
print("\nEvaluating at the vector yields train cost of %.16f" %
(lr.objective_function(data.features, data.output, [result[0]])))
t_vals = np.linspace(0, (len(result[2])) * 10, len(result[2]))
print("Using the above, we then have the following test cost:")
test_cost = lr.objective_function(test_data.features, test_data.output,
[result[0]])
print("%.16f" % test_cost)
if analytic_answer == "y":
print("\nAnalytic minimizer is:")
minimizer = lr.analytic_solution(data.features, data.output)
print(minimizer)
print("Cost using analytic minimizer is: %.16f" %
(lr.objective_function(data.features, data.output,
[minimizer])))
minimizer_cost = lr.objective_function(test_data.features,
test_data.output,
[minimizer])
print("Test cost using analytic minimizer is: %.16f" %
minimizer_cost)
print("\n2-norm error in result vector vs. minimizer is %.16f" %
(la.norm(result[0] - minimizer)))
print(
"Absolute error in test cost vs analytic test cost is %.16f" %
(np.abs(minimizer_cost - test_cost)))
print(
"Relative error in test cost vs analytic test cost is %.16f" %
(np.abs((minimizer_cost - test_cost) / minimizer_cost)))
plt.plot(t_vals, result[2], label='Cost')
plt.legend(loc='best')
plt.show()
else:
print("\nAnalytic minimizer is:")
minimizer = lr.analytic_solution(data.features, data.output)
print(minimizer)
print("Cost using analytic minimizer is: %.16f" %
(lr.objective_function(data.features, data.output, [minimizer])))
print("Test cost using analytic minimizer is: %.16f" %
(lr.objective_function(test_data.features, test_data.output,
[minimizer])))
def h_theta(self, theta, x_features):
"""
This is the logistic function that uses the sigmoid
"""
h_theta = LinearRegressor.h_theta(self, theta, x_features)
return 1.0 / (1 + np.power(np.e, -h_theta))
