def objective(args):
max_depth, min_child_weight, row_subsample, min_loss_reduction, column_subsample = args
clf = XGBoost(max_depth=max_depth, min_child_weight=min_child_weight,
row_subsample=row_subsample, min_loss_reduction=min_loss_reduction,
column_subsample=column_subsample, verbose=False)
score = utils.hold_out_evaluation(clf, train, labels, calibrate=False)
print 'max_depth, min_child_weight, row_subsample, min_loss_reduction, column_subsample, logloss'
print args, score
return score
def objective(args):
max_depth, min_child_weight, row_subsample, min_loss_reduction, column_subsample = args
clf = XGBoost(max_depth=max_depth, min_child_weight=min_child_weight,
row_subsample=row_subsample, min_loss_reduction=min_loss_reduction,
column_subsample=column_subsample, verbose=False)
score = utils.hold_out_evaluation(clf, train, labels, calibrate=False)
print ('max_depth, min_child_weight, row_subsample, min_loss_reduction, column_subsample, logloss')
print (args, score)
return score
def objective(args):
c, gamma = args
clf = OneVsRestClassifier(svm.SVC(C=c, kernel='rbf', tol=.001, gamma=gamma,
probability=True, random_state=23))
score1 = 0
score2 = utils.hold_out_evaluation(clf, train, labels, calibrate=False)
score = log_loss(valid_labels, clf.predict_proba(valid))
print 'C=%f, gamma=%f, score1=%f, score2=%f, score=%f' % (c, gamma, score1, score2, score)
return score
def objective(args):
c, gamma = args
clf = OneVsRestClassifier(
svm.SVC(C=c,
kernel='rbf',
tol=.001,
gamma=gamma,
probability=True,
random_state=23))
score1 = 0
score2 = utils.hold_out_evaluation(clf, train, labels, calibrate=False)
score = log_loss(valid_labels, clf.predict_proba(valid))
print 'C=%f, gamma=%f, score1=%f, score2=%f, score=%f' % (
c, gamma, score1, score2, score)
return score
row_subsample=.9134478530382129, min_loss_reduction=.5132278416508804,
column_subsample=.730128689911957, step_size=.009)
if MODE == 'cv':
scores, predictions = utils.make_blender_cv(clf, train, labels, calibrate=False)
print ('CV:', scores, 'Mean log loss:', np.mean(scores))
utils.write_blender_data(consts.BLEND_PATH, MODEL_NAME + '.csv', predictions)
elif MODE == 'submission':
clf.fit(train, labels)
predictions = clf.predict_proba(test)
utils.save_submission(consts.DATA_SAMPLE_SUBMISSION_PATH,
os.path.join(consts.ENSEMBLE_PATH, MODEL_NAME + '.csv'),
predictions)
elif MODE == 'holdout':
train, labels, _, _ = utils.stratified_split(train, labels, test_size=.7)
score = utils.hold_out_evaluation(clf, train, labels, calibrate=False)
print ('Log loss:', score)
elif MODE == 'tune':
# Objective function
def objective(args):
max_depth, min_child_weight, row_subsample, min_loss_reduction, column_subsample = args
clf = XGBoost(max_depth=max_depth, min_child_weight=min_child_weight,
row_subsample=row_subsample, min_loss_reduction=min_loss_reduction,
column_subsample=column_subsample, verbose=False)
score = utils.hold_out_evaluation(clf, train, labels, calibrate=False)
print ('max_depth, min_child_weight, row_subsample, min_loss_reduction, column_subsample, logloss')
print (args, score)
return score
# Searching space
space = (
hp.quniform('max_depth', 2, 14, 1),
train, labels, test, _, _ = utils.load_data()
# transform counts to TFIDF features
tfidf = feature_extraction.text.TfidfTransformer(smooth_idf=False)
train = tfidf.fit_transform(train).toarray()
test = tfidf.transform(test).toarray()
# encode labels
lbl_enc = preprocessing.LabelEncoder()
labels = lbl_enc.fit_transform(labels)
# train classifier
clf = ensemble.ExtraTreesClassifier(n_jobs=4, n_estimators=2000, max_features=20, min_samples_split=3,
bootstrap=False, verbose=3, random_state=23)
if MODE == 'cv':
scores, predictions = utils.make_blender_cv(clf, train, labels, calibrate=False)
print('CV:', scores, 'Mean log loss:', np.mean(scores))
utils.write_blender_data(consts.BLEND_PATH, MODEL_NAME + '.csv', predictions)
elif MODE == 'submission':
clf.fit(train, labels)
predictions = clf.predict_proba(test)
utils.save_submission(consts.DATA_SAMPLE_SUBMISSION_PATH,
os.path.join(consts.ENSEMBLE_PATH, MODEL_NAME + '.csv'),
predictions)
elif MODE == 'holdout':
score = utils.hold_out_evaluation(clf, train, labels, calibrate=False)
print('Log loss:', score)
else:
print('Unknown mode')
print (train.shape)
# encode labels
lbl_enc = preprocessing.LabelEncoder()
labels = lbl_enc.fit_transform(labels)
# train classifier
linear_clf = linear_model.LogisticRegression(C=1, penalty='l1',
fit_intercept=True, random_state=23)
clf = ensemble.BaggingClassifier(base_estimator=linear_clf, n_estimators=40,
max_samples=1., max_features=1., bootstrap=True,
n_jobs=5, verbose=True, random_state=23)
if MODE == 'cv':
scores, predictions = utils.make_blender_cv(clf, train, labels)
print ('CV:', scores, 'Mean log loss:', np.mean(scores))
utils.write_blender_data(consts.BLEND_PATH, MODEL_NAME + '.csv', predictions)
elif MODE == 'submission':
clf.fit(train, labels)
predictions = clf.predict_proba(test)
utils.save_submission(consts.DATA_SAMPLE_SUBMISSION_PATH,
os.path.join(consts.ENSEMBLE_PATH, MODEL_NAME + '.csv'),
predictions)
elif MODE == 'holdout':
score = utils.hold_out_evaluation(clf, train, labels)
print ('Log loss:', score)
else:
print ('Unknown mode')
print 'CV:', scores, 'Mean log loss:', np.mean(scores)
utils.write_blender_data(consts.BLEND_PATH, MODEL_NAME + '.csv',
predictions)
elif MODE == 'submission':
calibrated_classifier = CalibratedClassifierCV(clf,
method='isotonic',
cv=utils.get_cv(labels))
fitted_classifier = calibrated_classifier.fit(train, labels)
predictions = fitted_classifier.predict_proba(test)
utils.save_submission(
consts.DATA_SAMPLE_SUBMISSION_PATH,
os.path.join(consts.ENSEMBLE_PATH, MODEL_NAME + '.csv'), predictions)
elif MODE == 'holdout':
score = utils.hold_out_evaluation(clf,
train,
labels,
calibrate=False,
test_size=0.9)
print 'Log loss:', score
elif MODE == 'tune':
train, labels, valid, valid_labels = utils.stratified_split(train,
labels,
test_size=.8)
from sklearn.metrics import log_loss
# Objective function
def objective(args):
c, gamma = args
clf = OneVsRestClassifier(
svm.SVC(C=c,
kernel='rbf',
clf = OneVsRestClassifier(svm.SVC(C=4.919646+2., kernel='rbf', tol=.001,
verbose=True, probability=True, gamma=0.646508+.3, random_state=23))
if MODE == 'cv':
scores, predictions = utils.make_blender_cv(clf, train, labels, calibrate=True)
print 'CV:', scores, 'Mean log loss:', np.mean(scores)
utils.write_blender_data(consts.BLEND_PATH, MODEL_NAME + '.csv', predictions)
elif MODE == 'submission':
calibrated_classifier = CalibratedClassifierCV(clf, method='isotonic', cv=utils.get_cv(labels))
fitted_classifier = calibrated_classifier.fit(train, labels)
predictions = fitted_classifier.predict_proba(test)
utils.save_submission(consts.DATA_SAMPLE_SUBMISSION_PATH,
os.path.join(consts.ENSEMBLE_PATH, MODEL_NAME + '.csv'),
predictions)
elif MODE == 'holdout':
score = utils.hold_out_evaluation(clf, train, labels, calibrate=False, test_size=0.9)
print 'Log loss:', score
elif MODE == 'tune':
train, labels, valid, valid_labels = utils.stratified_split(train, labels, test_size=.8)
from sklearn.metrics import log_loss
# Objective function
def objective(args):
c, gamma = args
clf = OneVsRestClassifier(svm.SVC(C=c, kernel='rbf', tol=.001, gamma=gamma,
probability=True, random_state=23))
score1 = 0
score2 = utils.hold_out_evaluation(clf, train, labels, calibrate=False)
score = log_loss(valid_labels, clf.predict_proba(valid))
print 'C=%f, gamma=%f, score1=%f, score2=%f, score=%f' % (c, gamma, score1, score2, score)
return score
# Searching space