def estimate_parameters(ann, sample, target, epsilon=0.0001):
original_parameters = extract_parameters(ann)
ann = ml.ForwardFeedNetwork(ann.arch, ann.activation)
updated_parameters = []
for index in range(len(original_parameters)):
dp = np.zeros(len(original_parameters))
dp[index] += epsilon
minus_p = original_parameters - dp
ann = update_ann(ann, minus_p)
minus_errors = ml.quadratic_error(ann.forward_feed(sample), target)
minus_cost = np.mean(minus_errors)
plus_p = original_parameters + dp
ann = update_ann(ann, plus_p)
plus_errors = ml.quadratic_error(ann.forward_feed(sample), target)
plus_cost = np.mean(plus_errors)
derivative = (plus_cost - minus_cost) / (2 * epsilon)
updated_parameters.append(derivative)
return np.asarray(updated_parameters)
def test_quadratic_error_matrix(self):
output = np.asarray([
[1.0, 0.5, 0.25],
[0.5, 1.0, 0.25],
])
target = np.asarray([
[0.1, 1.0, 0.25],
[1.0, 0.1, 0.25],
])
expected_errors = [[0.405, 0.125, 0.0], [0.125, 0.405, 0.0]]
assert np.all(ml.quadratic_error(output, target) == expected_errors)
def test_quadratic_error_vector(self):
output = np.asarray([0.5, 1.0, 0.25])
target = np.asarray([1.0, 0.1, 0.25])
expected_errors = [0.125, 0.405, 0.0]
assert np.all(ml.quadratic_error(output, target) == expected_errors)
def test_quadratic_error_scalar(self):
assert ml.quadratic_error(1, 0.5) == 0.125
