def train_linear_elastic_net(training_predictor_table, training_target_table,
validation_predictor_table,
validation_target_table):
"""Trains linear regression with elastic-net penalty.
:param training_predictor_table: See doc for `utils.read_feature_file`.
:param training_target_table: Same.
:param validation_predictor_table: Same.
:param validation_target_table: Same.
"""
linear_en_model_object = utils.setup_linear_regression(lambda1=1e-5,
lambda2=5.)
_ = utils.train_linear_regression(
model_object=linear_en_model_object,
training_predictor_table=training_predictor_table,
training_target_table=training_target_table)
training_predictions = linear_en_model_object.predict(
training_predictor_table.as_matrix())
mean_training_target_value = numpy.mean(
training_target_table[utils.TARGET_NAME].values)
_ = utils.evaluate_regression(
target_values=training_target_table[utils.TARGET_NAME].values,
predicted_target_values=training_predictions,
mean_training_target_value=mean_training_target_value,
dataset_name='training')
print(MINOR_SEPARATOR_STRING)
validation_predictions = linear_en_model_object.predict(
validation_predictor_table.as_matrix())
_ = utils.evaluate_regression(
target_values=validation_target_table[utils.TARGET_NAME].values,
predicted_target_values=validation_predictions,
mean_training_target_value=mean_training_target_value,
dataset_name='validation')
def l1l2_experiment_testing(lambda1_values, lambda2_values,
validation_mae_skill_matrix,
training_predictor_table, training_target_table,
testing_predictor_table, testing_target_table):
"""Selects and tests model for experiment with L1/L2 regularization.
:param lambda1_values: See doc for `l1l2_experiment_validation`.
:param lambda2_values: Same.
:param validation_mae_skill_matrix: Same.
:param training_predictor_table: See doc for `utils.read_feature_file`.
:param training_target_table: Same.
:param testing_predictor_table: Same.
:param testing_target_table: Same.
"""
best_linear_index = numpy.argmax(numpy.ravel(validation_mae_skill_matrix))
best_lambda1_index, best_lambda2_index = numpy.unravel_index(
best_linear_index, (len(lambda1_values), len(lambda2_values)))
best_lambda1 = lambda1_values[best_lambda1_index]
best_lambda2 = lambda2_values[best_lambda2_index]
best_validation_maess = numpy.max(validation_mae_skill_matrix)
message_string = (
'Best MAE skill score on validation data = {0:.3f} ... corresponding '
'lasso coeff = 10^{1:.1f}, ridge coeff = 10^{2:.1f}').format(
best_validation_maess, numpy.log10(best_lambda1),
numpy.log10(best_lambda2))
print(message_string)
final_model_object = utils.setup_linear_regression(lambda1=best_lambda1,
lambda2=best_lambda2)
_ = utils.train_linear_regression(
model_object=final_model_object,
training_predictor_table=training_predictor_table,
training_target_table=training_target_table)
testing_predictions = final_model_object.predict(
testing_predictor_table.as_matrix())
mean_training_target_value = numpy.mean(
training_target_table[utils.TARGET_NAME].values)
this_evaluation_dict = utils.evaluate_regression(
target_values=testing_target_table[utils.TARGET_NAME].values,
predicted_target_values=testing_predictions,
mean_training_target_value=mean_training_target_value,
dataset_name='testing')
def l1l2_experiment_training(training_predictor_table, training_target_table,
validation_predictor_table,
validation_target_table):
"""Trains models for hyperparameter experiment with L1/L2 regularization.
:param training_predictor_table: See doc for `utils.read_feature_file`.
:param training_target_table: Same.
:param validation_predictor_table: Same.
:param validation_target_table: Same.
"""
lambda1_values = numpy.logspace(-8, -4, num=9)
lambda2_values = numpy.logspace(-4, 1, num=11)
num_lambda1 = len(lambda1_values)
num_lambda2 = len(lambda2_values)
validation_mae_matrix_s01 = numpy.full((num_lambda1, num_lambda2),
numpy.nan)
validation_mse_matrix_s02 = numpy.full((num_lambda1, num_lambda2),
numpy.nan)
validation_mae_skill_matrix = numpy.full((num_lambda1, num_lambda2),
numpy.nan)
validation_mse_skill_matrix = numpy.full((num_lambda1, num_lambda2),
numpy.nan)
mean_training_target_value = numpy.mean(
training_target_table[utils.TARGET_NAME].values)
for i in range(num_lambda1):
for j in range(num_lambda2):
this_message_string = (
'Training model with lasso coeff = 10^{0:.1f}, ridge coeff = '
'10^{1:.1f}...').format(numpy.log10(lambda1_values[i]),
numpy.log10(lambda2_values[j]))
print(this_message_string)
this_model_object = utils.setup_linear_regression(
lambda1=lambda1_values[i], lambda2=lambda2_values[j])
_ = utils.train_linear_regression(
model_object=this_model_object,
training_predictor_table=training_predictor_table,
training_target_table=training_target_table)
these_validation_predictions = this_model_object.predict(
validation_predictor_table.as_matrix())
this_evaluation_dict = utils.evaluate_regression(
target_values=validation_target_table[
utils.TARGET_NAME].values,
predicted_target_values=these_validation_predictions,
mean_training_target_value=mean_training_target_value,
verbose=False,
create_plots=False)
validation_mae_matrix_s01[i,
j] = this_evaluation_dict[utils.MAE_KEY]
validation_mse_matrix_s02[i,
j] = this_evaluation_dict[utils.MSE_KEY]
validation_mae_skill_matrix[i, j] = this_evaluation_dict[
utils.MAE_SKILL_SCORE_KEY]
validation_mse_skill_matrix[i, j] = this_evaluation_dict[
utils.MSE_SKILL_SCORE_KEY]