def estimate_logReg_true_test_error(xs_train,
ys_train,
xs_test,
ys_test,
num_features,
best_c_param_value=1e12,
score_type='brier_score'):
logReg = skl_logistic_regression(
C=best_c_param_value, tol=1e-10
) # Initialise the Logistic Regression Classifier, for the Best Regularization C Parameter found
logReg.fit(
xs_train[:, :num_features], ys_train
) # Fit the Logistic Regression Classifier with the Training Set
ys_logReg_predict_prob = logReg.predict_proba(
xs_test[:, :num_features]
)[:,
1] # Predict the Probabilities of the Features of the Testing Set, belongs to a certain Class
logReg_predict_classes_xs_test = logReg.predict(
xs_test
) # Predict and Classify the Values of the Testing Set, with the Logistic Regression Classifier TODO Confirmar
# Estimate the Testing Error, based on a certain type of Scoring
# 1) Brier Scoring
if (score_type == 'brier_score'):
estimated_true_test_error = skl_brier_score_loss(
ys_test, ys_logReg_predict_prob
) # Estimate the Testing Error, related to its Brier Score
# 2) Logistic Regression Scoring
if (score_type == 'logistic_regression_score'):
estimated_accuracy_test = logReg.score(
xs_test, ys_test
) # Compute the Training Set's Accuracy (Score), for the Logistic Regression
estimated_true_test_error = (
1 - estimated_accuracy_test
) # Compute the Training Error, regarding its Accuracy (Score)
num_samples_test_set = len(
xs_test) # The Number of Samples, from the Testing Set
logReg_num_incorrect_predictions = 0 # The Real Number of Incorrect Predictions, regarding the Logistic Regression Classifier
# For each Sample, from the Testing Set
for current_sample_test in range(num_samples_test_set):
# If the Prediction/Classification of the Class for the current Sample, of the Testing Set is different from the Real Class of the same,
# it's considered an Real Error in Prediction/Classification, regarding the Logistic Regression Classifier
if (logReg_predict_classes_xs_test[current_sample_test] !=
ys_test[current_sample_test]):
logReg_num_incorrect_predictions += 1
# Return the Predictions of the Samples,
# the Real Number of Incorrect Predictions and the Estimated True/Test Error, for the Logistic Regression Classifier
return logReg_predict_classes_xs_test, logReg_num_incorrect_predictions, estimated_true_test_error
def compute_logReg_errors(xs,
ys,
train_idx,
valid_idx,
c_param_value,
num_features,
score_type='brier_score'):
# Initialise the Logistic Regression, from the Linear Model of the SciKit-Learn
logReg = skl_logistic_regression(C=c_param_value, tol=1e-10)
# Fit the Logistic Regression
logReg.fit(xs[train_idx, :num_features], ys[train_idx])
# Compute the prediction probabilities of some Features, belonging to a certain Class, due to the
ys_logReg_predict_prob = logReg.predict_proba(xs[:, :num_features])[:, 1]
# Compute the Training and Validation Errors, based on a certain type of Scoring:
# 1) Based on Brier Score
if (score_type == 'brier_score'):
logReg_train_error = skl_brier_score_loss(
ys[train_idx], ys_logReg_predict_prob[train_idx]
) # Compute the Training Error, related to its Brier Score
logReg_valid_error = skl_brier_score_loss(
ys[valid_idx], ys_logReg_predict_prob[valid_idx]
) # Compute the Validation Error, related to its Brier Score
# 2) Based on Logistic Regression Score
if (score_type == 'logistic_regression_score'):
logReg_accuracy_train = logReg.score(
xs[train_idx], ys[train_idx]
) # Compute the Training Set's Accuracy (Score), for the Logistic Regression
logReg_accuracy_valid = logReg.score(
xs[valid_idx], ys[valid_idx]
) # Compute the Validation Set's Accuracy (Score), for the Logistic Regression
logReg_train_error = (
1 - logReg_accuracy_train
) # Compute the Training Error, regarding its Accuracy (Score)
logReg_valid_error = (
1 - logReg_accuracy_valid
) # Compute the Validation Error, regarding its Accuracy (Score)
# Return the Training and Validation Errors, for the Logistic Regression
return logReg_train_error, logReg_valid_error
def fit(self, train_fs, train_labels, valid_fs, valid_labels, test_fs,
test_labels):
self.model = skl_logistic_regression(
penalty=self.params['penalty'],
dual=self.params['dual'],
tol=self.params['tol'],
C=self.params['C'],
verbose=self.params['verbose'],
max_iter=self.params['max_iter'],
solver=self.params['solver'],
n_jobs=self.params['n_jobs'],
multi_class=self.params['multi_class'])
self.model.fit(X=train_fs, y=train_labels)
train_preds = self.model.predict_proba(train_fs)[:, 1]
valid_preds = self.model.predict_proba(valid_fs)[:, 1]
test_preds = self.model.predict_proba(test_fs)[:, 1]
return train_preds, valid_preds, test_preds
def fit(self,
train_fs, train_labels,
valid_fs, valid_labels,
test_fs, test_labels):
self.model = skl_logistic_regression(penalty=self.params['penalty'],
dual=self.params['dual'],
tol=self.params['tol'],
C=self.params['C'],
verbose=self.params['verbose'],
max_iter=self.params['max_iter'],
solver=self.params['solver'],
n_jobs=self.params['n_jobs'],
multi_class=self.params['multi_class'])
self.__lock()
self.model.fit(X=train_fs, y=train_labels)
self.__unlock()
train_preds = self.model.predict_proba(train_fs)[:, 1]
valid_preds = self.model.predict_proba(valid_fs)[:, 1]
test_preds = self.model.predict_proba(test_fs)[:, 1]
return train_preds, valid_preds, test_preds
# Loop to iterate all the Iris' Classes
for ys_class_iris in range(3):
# Create an array of Classes,
# for the current Logistic Regression Classifier
ys_classes_current_logReg = np.zeros(num_samples)
# Set the new Classes, based on the current iterated Iris' Class
ys_classes_current_logReg[ys_classes == ys_class_iris] = 0
ys_classes_current_logReg[ys_classes != ys_class_iris] = 1
# Initialise the current Logistic Regression Classifier,
# for the current Settings
logReg = skl_logistic_regression(C=c_param_value, tol=1e-10)
# Fit the Logistic Regression Classifier with the Training Set
logReg.fit(xs_features, ys_classes_current_logReg)
# Append the current Logistic Regression Classifier,
# for the current Settings
logRegs_classifiers.append(logReg)
# Plot a customized One vs. Rest Algorithm, for the Logistic Regression Classifier,
# in the classification of the Iris' Dataset problem
plot_logistic_regressions(logRegs_classifiers, "one-vs-rest-logistic-regression-iris-dataset.png")
# Initialise the One vs. Rest Algorithm,
# from the SciKit-Learn's Python Library, for Support Vector Machines,