def test_do_almost_nothing():
X, target = get_random_data()
afreg = AutoFeatRegressor(verbose=1, feateng_steps=0, featsel_runs=0)
df = afreg.fit_transform(pd.DataFrame(X, columns=["x1", "x2", "x3"]),
target)
assert list(df.columns) == ["x1", "x2", "x3"], "Only original columns"
df = afreg.transform(pd.DataFrame(X, columns=["x1", "x2", "x3"]))
assert list(df.columns) == ["x1", "x2", "x3"], "Only original columns"
def test_regular_df_X_y():
# autofeat with df without column names
X, target = get_random_data()
afreg = AutoFeatRegressor(verbose=1, feateng_steps=3)
df = afreg.fit_transform(pd.DataFrame(X), pd.DataFrame(target))
# score once with original, once with transformed data
assert afreg.score(pd.DataFrame(X), target) >= 0.999, "R^2 should be 1."
assert afreg.score(df, target) >= 0.999, "R^2 should be 1."
assert list(df.columns)[:3] == ["0", "1", "2"], "Wrong column names"
def test_regular_X_y():
# autofeat with numpy arrays
X, target = get_random_data()
afreg = AutoFeatRegressor(verbose=1, feateng_steps=3)
df = afreg.fit_transform(X, target)
assert afreg.score(X, target) >= 0.999, "R^2 should be 1."
assert afreg.score(df, target) >= 0.999, "R^2 should be 1."
assert list(df.columns)[:3] == ["x000", "x001",
"x002"], "Wrong column names"
def test_units():
np.random.seed(15)
x1 = np.random.rand(1000)
x2 = np.random.randn(1000)
x3 = np.random.rand(1000)
target = 2 + 15 * x1 + 3 / (x2 - 1 / x3) + 5 * (x2 * np.log(x1))**3
X = np.vstack([x1, x2, x3]).T
units = {"x2": "m/sec", "x3": "min/mm"}
afreg = AutoFeatRegressor(verbose=1, units=units, feateng_steps=3)
_ = afreg.fit_transform(pd.DataFrame(X, columns=["x1", "x2", "x3"]),
target)
assert afreg.score(pd.DataFrame(X, columns=["x1", "x2", "x3"]),
target) >= 0.999, "R^2 should be 1."
def test_categorical_cols():
np.random.seed(15)
x1 = np.random.rand(1000)
x2 = np.random.randn(1000)
x3 = np.random.rand(1000)
x4 = np.array(200 * [4] + 300 * [5] + 500 * [2], dtype=int)
target = 2 + 15 * x1 + 3 / (x2 - 1 / x3) + 5 * (x2 + np.log(x1))**3 + x4
X = np.vstack([x1, x2, x3, x4]).T
afreg = AutoFeatRegressor(verbose=1,
categorical_cols=["x4", "x5"],
feateng_steps=3)
try:
df = afreg.fit_transform(
pd.DataFrame(X, columns=["x1", "x2", "x3", "x4"]), target)
except ValueError:
pass
else:
raise AssertionError(
"categorical_cols not in df should throw an error")
afreg = AutoFeatRegressor(verbose=1,
categorical_cols=["x4"],
feateng_steps=3)
df = afreg.fit_transform(pd.DataFrame(X, columns=["x1", "x2", "x3", "x4"]),
target)
assert list(df.columns)[3:6] == [
"cat_x4_2.0", "cat_x4_4.0", "cat_x4_5.0"
], "categorical_cols were not transformed correctly"
assert "x4" not in df.columns, "categorical_cols weren't deleted from df"
df = afreg.transform(pd.DataFrame(X, columns=["x1", "x2", "x3", "x4"]))
assert list(df.columns)[3:6] == [
"cat_x4_2.0", "cat_x4_4.0", "cat_x4_5.0"
], "categorical_cols were not transformed correctly"
assert "x4" not in df.columns, "categorical_cols weren't deleted from df"
assert afreg.score(pd.DataFrame(X, columns=["x1", "x2", "x3", "x4"]),
target) >= 0.999, "R^2 should be 1."
def test_nans():
# nans are ok in transform but not fit or predict (due to sklearn model)
X, target = get_random_data()
X[998, 0] = np.nan
X[999, 1] = np.nan
afreg = AutoFeatRegressor(verbose=1, feateng_steps=3)
try:
_ = afreg.fit_transform(pd.DataFrame(X, columns=["x 1.1", 2, "x/3"]),
pd.DataFrame(target))
except ValueError:
pass
else:
raise AssertionError("fit with NaNs should throw an error")
_ = afreg.fit_transform(pd.DataFrame(X[:900], columns=["x 1.1", 2, "x/3"]),
pd.DataFrame(target[:900]))
try:
_ = afreg.predict(pd.DataFrame(X[900:], columns=["x 1.1", 2, "x/3"]))
except ValueError:
pass
else:
raise AssertionError("predict with NaNs should throw an error")
df = afreg.transform(pd.DataFrame(X, columns=["x 1.1", 2, "x/3"]))
assert all([pd.isna(df.iloc[998, 0]),
pd.isna(df.iloc[999, 1])]), "Original features should be NaNs"
assert np.sum(
np.array(pd.isna(df.iloc[998]),
dtype=int)) >= 2, "There should be at least 2 NaNs in row 998"
assert np.sum(
np.array(pd.isna(df.iloc[999]),
dtype=int)) >= 2, "There should be at least 3 NaNs in row 999"
def test_weird_colnames():
# autofeat with df with weird column names
X, target = get_random_data()
afreg = AutoFeatRegressor(verbose=1, feateng_steps=3)
df = afreg.fit_transform(pd.DataFrame(X, columns=["x 1.1", 2, "x/3"]),
pd.DataFrame(target))
assert afreg.score(pd.DataFrame(X, columns=["x 1.1", 2, "x/3"]),
target) >= 0.999, "R^2 should be 1."
assert list(df.columns)[:3] == ["x 1.1", "2", "x/3"], "Wrong column names"
# error if the column names aren't the same as before
try:
afreg.score(pd.DataFrame(X, columns=["x 11", 2, "x/3"]), target)
except ValueError:
pass
else:
raise AssertionError("Should throw error on mismatch column names")
def test_feateng_cols():
X, target = get_random_data()
afreg = AutoFeatRegressor(verbose=1,
feateng_cols=["x1", "x3", "x4"],
feateng_steps=3)
try:
df = afreg.fit_transform(pd.DataFrame(X, columns=["x1", "x2", "x3"]),
target)
except ValueError:
pass
else:
raise AssertionError("feateng_cols not in df should throw an error")
afreg = AutoFeatRegressor(verbose=1,
feateng_cols=["x1", "x3"],
feateng_steps=3)
df = afreg.fit_transform(pd.DataFrame(X, columns=["x1", "x2", "x3"]),
target)
for c in df.columns[3:]:
assert "x2" not in c, "only feateng_cols should occur in engineered features"
def regression(self):
X_train, X_test, y_train, y_test, categoricalCols = \
self.dataFunction(preprocessed=True, specifics="AUTOFEAT", trainSize=self.trainSize, steps=self.steps, nDataPoints=self.nDataPoints)
featureEngColumns = None
# If feature engineering not wanted for categorical values, uncomment
# if categoricalCols is not None:
# featureEngColumns = X_train.columns.values.tolist()
# featureEngColumns = [i for i in featureEngColumns + categoricalCols if i not in featureEngColumns or i not in categoricalCols]
# Measure runtime
start_time = time.time()
print(f"Start time: {start_time}")
# Automated feature engineering with autofeat
model = AutoFeatRegressor(verbose=1,
feateng_steps=self.feateng_steps,
featsel_runs=self.featuresel_steps,
categorical_cols=categoricalCols,
feateng_cols=featureEngColumns)
# Fit model and get transformed dataframe with additional features
x_train_extended = model.fit_transform(X_train, y_train)
total_time = int(divmod(time.time() - start_time, 60)[0])
print(f"Time: {total_time}")
# Export model
dump(
model,
f"{self.savePath}/feng{model.feateng_steps}_fsel{model.featsel_runs}_time{total_time}_model.joblib"
)
x_test_extended = model.transform(X_test)
# Predictions
predictions = {}
predictionModel = DecisionTreeRegressor()
predictionModel.fit(x_train_extended, y_train)
predictions["DecisionTree"] = mean_squared_error(
y_test, predictionModel.predict(x_test_extended))
print(f"Final MSE prediction score: {predictions['DecisionTree']}")
predictionModel = RandomForestRegressor(n_estimators=10)
predictionModel.fit(x_train_extended, y_train)
predictions["RandomForest"] = mean_squared_error(
y_test, predictionModel.predict(x_test_extended))
print(f"Final MSE prediction score: {predictions['RandomForest']}")
predictionModel = LinearRegression()
predictionModel.fit(x_train_extended, y_train)
predictions["LinearRegression"] = mean_squared_error(
y_test, predictionModel.predict(x_test_extended))
print(f"Final MSE prediction score: {predictions['LinearRegression']}")
predictionModel = LassoLarsCV(cv=5)
predictionModel.fit(x_train_extended, y_train)
predictions["LassoLarsCV"] = mean_squared_error(
y_test, predictionModel.predict(x_test_extended))
print(f"Final MSE prediction score: {predictions['LassoLarsCV']}")
# Additionally save transformations steps since not saved in joblib file
predictions["new_features"] = model.new_feat_cols_
# Export predictions
with open(
f"{self.savePath}/feng{model.feateng_steps}_fsel{model.featsel_runs}_performance.pkl",
'wb') as file:
pickle.dump(predictions, file)
return model
# autofeat with numpy arrays but as classification
X, target = get_random_data()
target = np.array(target > target.mean(), dtype=int)
afreg = AutoFeatClassifier(verbose=1, feateng_steps=3)
df = afreg.fit_transform(X, target)
assert afreg.score(X, target) >= 0.9999, "Accuracy should be 1."
assert afreg.score(df, target) >= 0.9999, "Accuracy should be 1."
assert list(df.columns)[:3] == ["x000", "x001",
"x002"], "Wrong column names"
if __name__ == '__main__':
print("## Running sklearn Regressor tests")
# we allow for nan in transform
successful_tests = set(["check_estimators_nan_inf"])
for estimator, check in check_estimator(AutoFeatRegressor(
feateng_steps=1, featsel_runs=1, always_return_numpy=True),
generate_only=True):
if check.func.__name__ not in successful_tests:
print(check.func.__name__)
successful_tests.add(check.func.__name__)
check(estimator)
# additionally check the class, but don't run all the other tests
for estimator, check in check_estimator(AutoFeatRegressor,
generate_only=True):
if check.func.__name__ not in successful_tests:
print(check.func.__name__)
successful_tests.add(check.func.__name__)
check(estimator)
print("## Running sklearn Classifier tests")
# we allow for nan in transform
import pandas as pd
import matplotlib.pyplot as plt
from autofeat import FeatureSelector, AutoFeatRegressor
from sklearn.pipeline import make_pipeline
import pickle
def main():
# Get the dataset from the users GitHub repository
dataset_path = "https://raw.githubusercontent.com/" + os.environ["GITHUB_REPOSITORY"] +"/master/dataset.csv"
df = pd.read_csv(dataset_path)
print()
print(df.describe())
for steps in range(5):
np.random.seed(55)
print("### AutoFeat with %i feateng_steps" % steps)
afreg = AutoFeatRegressor(verbose=1, feateng_steps=steps)
df = afreg.fit_transform(df_org, target)
r2 = afreg.score(df_org, target)
print("## Final R^2: %.4f" % r2)
plt.figure()
plt.scatter(afreg.predict(df_org), target, s=2);
plt.title("%i FE steps (R^2: %.4f; %i new features)" % (steps, r2, len(afreg.new_feat_cols_)))
afreg = AutoFeatRegressor(verbose=1, feateng_steps=3)
# train on noisy data
df = afreg.fit_transform(df_org, target_noisy)
# test on real targets
print("Final R^2: %.4f" % afreg.score(df, target))
plt.figure()
plt.scatter(afreg.predict(df), target, s=2);
afreg = AutoFeatRegressor(verbose=1, feateng_steps=3)
# train on noisy data