def _fit_cv(self, X, y, model_no):
# We only look at cv across folds - no storing of models or results with 'cv'
evals = []
i = 0
for train_index, test_index in self.folds_strategy.split(X, y):
# Loop over the different folds.
X_train, X_test, y_train, y_test = split_folds(
train_index, test_index, X, y)
# Fit on each fold for each model.
# We'll add a try condition here for using x_test as a validation set - useful of XGB early stopping.
try:
self.base_estimators[model_no].fit(X_train, y_train, X_test,
y_test)
except:
self.base_estimators[model_no].fit(X_train, y_train)
if self.estimator_type is 'regression':
predicted_y = self.base_estimators[model_no].predict(X_test)
elif self.estimator_type is 'classification':
predicted_y = self.base_estimators[model_no].predict_proba(
X_test)
if self.num_classes is 2 and 'sklearn' in str(
type(self.base_estimators[model_no])):
predicted_y = predicted_y[:, 1]
if self.feval is not None:
fold_score = self.feval(y_test, predicted_y)
evals.append(fold_score)
print('Fold{}: {}'.format(i + 1, evals[i]))
i += 1
print('CV Mean: ', np.mean(evals), ' Std: ', np.std(evals))
return
def _fit_s(self, X, y, model_no):
# Fit a model that stacks for CV folds, predicts the out-of-fold rows for X, and then runs a predict on the
# test set, the final test set prediction is the average from all fold models.
evals = []
fold_fits = {}
i = 0
for train_index, test_index in self.folds_strategy.split(X, y):
# Loop over the different folds.
X_train, X_test, y_train, y_test = split_folds(
train_index, test_index, X, y)
# Fit on each fold for each model.
# We'll add a try condition here for using x_test as a validation set - useful of XGB early stopping.
try:
self.base_estimators[model_no].fit(X_train, y_train, X_test,
y_test)
except:
self.base_estimators[model_no].fit(X_train, y_train)
# Predict on the out of fold set
if self.estimator_type is 'regression':
predicted_y = self.base_estimators[model_no].predict(X_test)
self.stacking_train.ix[
test_index,
self.base_estimators_names[model_no]] = predicted_y
elif self.estimator_type is 'classification':
predicted_y = self.base_estimators[model_no].predict_proba(
X_test)
if self.num_classes == 2:
if 'sklearn' in str(type(self.base_estimators[model_no])):
predicted_y = predicted_y[:, 1]
self.stacking_train.ix[
test_index,
self.base_estimators_names[model_no]] = predicted_y
elif self.num_classes > 2:
self.stacking_train.ix[test_index, [
self.base_estimators_names[model_no] + '_class_' +
str(j) for j in range(self.num_classes)
]] = predicted_y
# Finally save the base_estimator.fit object for each fold of the data set.
# In predict we need to loop through these to get an average prediction for the test set.
# We create a model specific dictionary and we append each fold to this
fold_fits[self.base_estimators_names[model_no] + 'fold' +
str(i)] = self.base_estimators[model_no]
# Evaluate the Folds
if self.feval is not None:
fold_score = self.feval(y_test, predicted_y)
evals.append(fold_score)
print('Fold{}: {}'.format(i + 1, evals[i]))
i += 1
print('CV Mean: ', np.mean(evals), ' Std: ', np.std(evals))
# Last part add to the fold estimators
self.fold_estimators[self.base_estimators_names[model_no]] = fold_fits
return
def _fit_st(self, X, y, model_no):
# Fit a model that stacks for CV folds, predicts the out-of-fold rows for the X and then runs a full fit on
# the data to use for preditions.
evals = []
i = 0
for train_index, test_index in self.folds_strategy.split(X, y):
# Loop over the different folds.
X_train, X_test, y_train, y_test = split_folds(
train_index, test_index, X, y)
# Fit on each fold for each model.
# We'll add a try condition here for using x_test as a validation set - useful of XGB early stopping.
try:
self.base_estimators[model_no].fit(X_train, y_train, X_test,
y_test)
except:
self.base_estimators[model_no].fit(X_train, y_train)
# Predict on the out of fold set
if self.estimator_type is 'regression':
predicted_y = self.base_estimators[model_no].predict(X_test)
self.stacking_train.ix[
test_index,
self.base_estimators_names[model_no]] = predicted_y
elif self.estimator_type is 'classification':
predicted_y = self.base_estimators[model_no].predict_proba(
X_test)
if self.num_classes == 2:
if 'sklearn' in str(type(self.base_estimators[model_no])):
predicted_y = predicted_y[:, 1]
self.stacking_train.ix[
test_index,
self.base_estimators_names[model_no]] = predicted_y
elif self.num_classes > 2:
self.stacking_train.ix[test_index, [
self.base_estimators_names[model_no] + '_class_' +
str(i) for i in range(self.num_classes)
]] = predicted_y
# Evaluate the Folds
if self.feval is not None:
assert (len(y_test) == len(predicted_y))
fold_score = self.feval(y_test, predicted_y)
evals.append(fold_score)
print('Fold{}: {}'.format(i + 1, evals[i]))
i += 1
print('CV Mean: ', np.mean(evals), ' Std: ', np.std(evals))
# Finally fit against all the data
self._fit_t(X, y, model_no)
return