def create_best_gradboost(data_path, output_path):
X, y, X_test, _ = prepare_data(data_path)
X_train, X_valid, y_train, y_valid = train_test_split(X,
y,
test_size=0.10,
random_state=42)
pipe = Pipeline([('feature_selection', SelectKBest(f_classif)),
('clf', GradientBoostingClassifier(random_state=2))])
params = {
'feature_selection__k': [12, 8, 4],
'clf__n_estimators': [50, 100, 500]
}
grid_search = GridSearchCV(pipe, param_grid=params, scoring='roc_auc')
#logging.INFO(str(grid_search.fit(X_train, y_train).best_params_))
grid_search.fit(X_train, y_train)
tn, fp, fn, tp = confusion_matrix(
y_valid, grid_search.best_estimator_.predict(X_valid)).ravel()
#logging.INFO("= Gradient Boosting Classifier = ")
#logging.INFO("Accuracy: {}".format( str(((tp + tn) / (tn + fp + fn + tp)))))
#logging.INFO("Recall: {}".format(((tp) / (fn + tp))))
#logging.INFO("Specificity: {}".format(((tn) / (tn + fp))))
#logging.INFO("F1 Score: {}".format(f1_score(y_valid, grid_search.best_estimator_.predict(X_valid))))
#logging.INFO("AUC: {}".format(roc_auc_score(y_valid, grid_search.best_estimator_.predict(X_valid))))
model_path = "{0}{1}".format(output_path, 'gradboost.pkl')
with open(model_path, 'wb') as file:
pickle.dump(grid_search.best_estimator_, file=file)
return model_path, roc_auc_score(
y_valid, grid_search.best_estimator_.predict(X_valid))
def build_data_loaders(args, tokenizer):
snli, mnli = prepare_data(tokenizer)
injected_collate = partial(collate,
span_drop=args.span_drop,
max_spans=args.max_spans,
padding_token_id=tokenizer.pad_token_id)
train_loader_factory = TrainLoaderFactory(args.batch_size, snli, mnli,
injected_collate)
train_all = train_loader_factory.get_loader(MODE.TRAIN_ALL)
train_snli = train_loader_factory.get_loader(MODE.TRAIN_SNLI)
train_mnli = train_loader_factory.get_loader(MODE.TRAIN_MNLI)
test_snli = train_loader_factory.get_loader(MODE.TEST_SNLI)
test_mnli_matched = train_loader_factory.get_loader(MODE.TEST_MNLI_MATCHED)
test_mnli_mismatched = train_loader_factory.get_loader(
MODE.TEST_MNLI_MISMATCHED)
if args.dataset == 'all':
return train_all, {
'snli': test_snli,
"mnli_matched": test_mnli_matched,
"mnli_mismatched": test_mnli_mismatched
}
elif args.dataset == 'snli':
return train_snli, {'snli': test_snli}
elif args.dataset == 'mnli':
return train_mnli, {
"mnli_matched": test_mnli_matched,
"mnli_mismatched": test_mnli_mismatched
}
else:
raise ValueError("--dataset has to contain one of: all, mnli or snli")
def get_predictions( data_path="", model_path=""):
champion_path = "{0}{1}".format(model_path, 'champion.pkl')
print("0 "+champion_path)
champion = pickle.load(open(champion_path, 'rb'))
# Future extension: Data path or data retrieval process for consistently updated data
_, _, X_test, test_id = prepare_data(data_path)
# Future extension: Write to another file, database, etc.
print(champion.predict(X_test))
submission_path = "{0}{1}".format(model_path, "submission.csv")
print(submission_path)
submission = pd.DataFrame({'PassengerId': test_id, 'Survived': champion.predict(X_test)})
submission.to_csv(submission_path)