def create_dataset(db, run_config, aws_config=None):
"""
Create a dataset and add it to the ModelHub database.
db: initialized Database object
run_config: RunConfig object describing the dataset to create
aws_config: optional. AWS credentials for downloading data from S3.
"""
# download data to the local filesystem to extract metadata
train_local, test_local = download_data(run_config.train_path,
run_config.test_path, aws_config)
# create the name of the dataset from the path to the data
name = os.path.basename(train_local)
name = name.replace("_train.csv", "").replace(".csv", "")
# process the data into the form ATM needs and save it to disk
meta = MetaData(run_config.class_column, train_local, test_local)
# enter dataset into database
dataset = db.create_dataset(name=name,
description=run_config.data_description,
train_path=run_config.train_path,
test_path=run_config.test_path,
class_column=run_config.class_column,
n_examples=meta.n_examples,
k_classes=meta.k_classes,
d_features=meta.d_features,
majority=meta.majority,
size_kb=old_div(meta.size, 1000))
return dataset
def create_dataset(db, label_column, train_path, test_path=None,
data_description=None):
"""
Create a dataset and add it to the ModelHub database.
db: initialized Database object
label_column: name of csv column representing the label
train_path: path to raw training data
test_path: path to raw test data
data_description: description of the dataset (max 1000 chars)
"""
# create the name of the dataset from the path to the data
name = os.path.basename(train_path)
name = name.replace("_train.csv", "").replace(".csv", "")
# process the data into the form ATM needs and save it to disk
meta = MetaData(label_column, train_path, test_path)
# enter dataset into database
dataset = db.create_dataset(name=name,
description=data_description,
train_path=train_path,
test_path=test_path,
label_column=label_column,
n_examples=meta.n_examples,
k_classes=meta.k_classes,
d_features=meta.d_features,
majority=meta.majority,
size_kb=meta.size / 1000)
return dataset
def create_dataset(self):
"""
Create a dataset and add it to the ModelHub database.
"""
# download data to the local filesystem to extract metadata
train_local, test_local = download_data(self.run_conf.train_path,
self.run_conf.test_path,
self.aws_conf)
# create the name of the dataset from the path to the data
name = os.path.basename(train_local)
name = name.replace("_train.csv", "").replace(".csv", "")
# process the data into the form ATM needs and save it to disk
meta = MetaData(self.run_conf.class_column, train_local, test_local)
# enter dataset into database
dataset = self.db.create_dataset(
name=name,
description=self.run_conf.data_description,
train_path=self.run_conf.train_path,
test_path=self.run_conf.test_path,
class_column=self.run_conf.class_column,
n_examples=meta.n_examples,
k_classes=meta.k_classes,
d_features=meta.d_features,
majority=meta.majority,
size_kb=old_div(meta.size, 1000))
return dataset
def train_test(self, train_path, test_path=None):
# load train and (maybe) test data
metadata = MetaData(class_column=self.class_column,
train_path=train_path,
test_path=test_path)
self.num_classes = metadata.k_classes
self.num_features = metadata.d_features
# if necessary, cast judgment metric into its binary/multiary equivalent
if self.num_classes == 2:
if self.judgment_metric in [Metrics.F1_MICRO, Metrics.F1_MACRO]:
self.judgment_metric = Metrics.F1
elif self.judgment_metric in [
Metrics.ROC_AUC_MICRO, Metrics.ROC_AUC_MACRO
]:
self.judgment_metric = Metrics.ROC_AUC
else:
if self.judgment_metric == Metrics.F1:
self.judgment_metric = Metrics.F1_MACRO
elif self.judgment_metric == Metrics.ROC_AUC:
self.judgment_metric = Metrics.ROC_AUC_MACRO
# load training data
train_data = self.load_data(train_path)
# if necessary, generate permanent train/test split
if test_path is not None:
test_data = self.load_data(test_path)
all_data = pd.concat([train_data, test_data])
else:
all_data = train_data
train_data, test_data = train_test_split(
train_data,
test_size=self.testing_ratio,
random_state=self.random_state)
# extract feature matrix and labels from raw data
self.encoder = DataEncoder(class_column=self.class_column)
self.encoder.fit(all_data)
X_train, y_train = self.encoder.transform(train_data)
X_test, y_test = self.encoder.transform(test_data)
# create and cross-validate pipeline
self.make_pipeline()
cv_scores = self.cross_validate(X_train, y_train)
# train and test the final model
self.pipeline.fit(X_train, y_train)
test_scores = self.test_final_model(X_test, y_test)
return {'cv': cv_scores, 'test': test_scores}