def _generate_oof_predictions(self, X_train, y_train, feature_selection_func=None, **kwargs):
train_n = X_train.shape[0]
cv = KFold(n=train_n, n_folds=10, random_state=123)
cv_scores = []
oof_predictions = np.zeros(shape=(train_n,))
for fold, (tr_idx, te_idx) in enumerate(cv):
X_train_ = X_train.iloc[tr_idx]
y_train_ = self.target_transform.transform(y_train.iloc[tr_idx])
X_test_ = X_train.iloc[te_idx]
y_test_ = y_train.iloc[te_idx]
# select features
self.feature_selector.fit(X_train_, y_train_)
X_train_ = self.feature_selector.transform(X_train_)
X_test_ = self.feature_selector.transform(X_test_)
self.model.fit(X_train_, y_train_)
preds_k = self.target_transform.transform_back(self.model.predict(X_test_))
oof_predictions[te_idx] = preds_k
gini_k = normalized_gini(y_test_, preds_k)
cv_scores.append(gini_k)
print 'Fold %d: %.4f' % (fold + 1, gini_k)
oof_gini = normalized_gini(y_train, oof_predictions)
print 'Final: %.4f' % (oof_gini)
return oof_predictions, oof_gini, cv_scores
def ensemble_predictions(self):
# find optimal weights
self.opt_weights = find_ensemble_weights(
opt_func=self.opt_func,
predictions=self.oof_predictions,
y_true=Y_TRAIN,
w_init=np.array(self.oof_ginis)**0.5,
verbose=self.verbose,
**self.kwargs)['x']
# ensemble oof predictions
self.ensembled_oof_predictions_ = _ensemble_predictions(
predictions=self.oof_predictions,
ensemble_weights=self.opt_weights)
self.ensembled_oof_gini_ = normalized_gini(
Y_TRAIN, self.ensembled_oof_predictions_)
if self.verbose:
print '\nEnsembled CV: {}\n'.format(
np.round(self.ensembled_oof_gini_, 6))
for w, pred_col in zip(self.opt_weights,
self.oof_predictions.columns):
print '{} : {}'.format(np.round(w, 4), pred_col)
# ensemble lb predictions
self.ensembled_lb_predictions_ = _ensemble_predictions(
predictions=self.lb_predictions, ensemble_weights=self.opt_weights)
def ensemble_predictions(self):
# find optimal weights
self.opt_weights = find_ensemble_weights(
opt_func=self.opt_func,
predictions=self.oof_predictions,
y_true=Y_TRAIN,
w_init=np.array(self.oof_ginis)**0.5,
verbose=self.verbose,
**self.kwargs
)['x']
# ensemble oof predictions
self.ensembled_oof_predictions_ = _ensemble_predictions(
predictions=self.oof_predictions,
ensemble_weights=self.opt_weights
)
self.ensembled_oof_gini_ = normalized_gini(Y_TRAIN, self.ensembled_oof_predictions_)
if self.verbose:
print '\nEnsembled CV: {}\n'.format(np.round(self.ensembled_oof_gini_, 6))
for w, pred_col in zip(self.opt_weights, self.oof_predictions.columns):
print '{} : {}'.format(np.round(w, 4), pred_col)
# ensemble lb predictions
self.ensembled_lb_predictions_ = _ensemble_predictions(
predictions=self.lb_predictions,
ensemble_weights=self.opt_weights
)
def ensemble_predictions(self):
# ensemble oof predictions
self.ensembled_oof_predictions_ = self.oof_predictions.apply(self.ensemble_func, axis=1)
self.ensembled_oof_gini_ = normalized_gini(Y_TRAIN, self.ensembled_oof_predictions_)
if self.verbose:
print '\nEnsembled CV: {}\n'.format(np.round(self.ensembled_oof_gini_, 6))
# ensemble lb predictions
self.ensembled_lb_predictions_ = self.lb_predictions.apply(self.ensemble_func, axis=1)
def ensemble_predictions(self):
# ensemble oof predictions
self.ensembled_oof_predictions_ = self.oof_predictions.apply(
self.ensemble_func, axis=1)
self.ensembled_oof_gini_ = normalized_gini(
Y_TRAIN, self.ensembled_oof_predictions_)
if self.verbose:
print '\nEnsembled CV: {}\n'.format(
np.round(self.ensembled_oof_gini_, 6))
# ensemble lb predictions
self.ensembled_lb_predictions_ = self.lb_predictions.apply(
self.ensemble_func, axis=1)
def fit(self, X, y):
for col in X.columns:
gini = normalized_gini(y, X[col])
if abs(gini) > self.min_gini:
self.current_features.append(col)
self.selected_features += self.current_features
def fit(self, X, y):
for col in X.columns:
gini = normalized_gini(y, X[col])
if abs(gini) > self.min_gini:
self.current_features.append(col)
self.selected_features += self.current_features
