def test_out_of_time_and_space_splitter():
result, logs = out_of_time_and_space_splitter(
sample_data,
2,
'2015-05-05',
time_column='time',
space_column='space',
holdout_gap=timedelta(days=31))
assert len(result) == 2
train_1 = sample_data.iloc[result[0][0]]
test_1 = sample_data.iloc[result[0][1][0]]
train_2 = sample_data.iloc[result[1][0]]
test_2 = sample_data.iloc[result[1][1][0]]
# there must be no overlap in space between folds
assert len(train_1[train_1.space.isin(train_2.space)]) == 0
assert len(train_2[train_2.space.isin(train_1.space)]) == 0
# the training sets must have no dates after 2015-05-05
assert len(train_1[train_1.time > '2015-05-05']) == 0
assert len(train_2[train_2.time > '2015-05-05']) == 0
# the test sets must have no dates before 2015-05-05 + 31 days
assert len(test_1[test_1.time <= '2015-06-05']) == 0
assert len(test_2[test_2.time <= '2015-06-05']) == 0
# all rows with time before '2015-05-05' must be in a training set
assert len(train_1) + len(train_2) == len(
sample_data[sample_data.time <= '2015-05-05'])
def test_grid_search_tuner(tmpdir):
train_set = pd.DataFrame({
'id': ["id1", "id2", "id3", "id3"],
'date':
pd.to_datetime(
["2016-01-01", "2016-02-01", "2016-03-01", "2016-04-01"]),
'x': [.2, .9, .3, .3],
'target': [0, 1, 0, 1]
})
eval_fn = auc_evaluator(target_column="target")
space = {
'learning_rate': lambda: [1e-3, 1e-2, 1e-1],
'num_estimators': lambda: [1, 2],
'silent': lambda: [True]
}
@curry
def param_train_fn(space, train_set):
return xgb_classification_learner(
features=["x"],
target="target",
learning_rate=space["learning_rate"],
num_estimators=space["num_estimators"])(train_set)
split_fn = out_of_time_and_space_splitter(n_splits=2,
in_time_limit="2016-05-01",
space_column="id",
time_column="date")
tuning_log = grid_search_cv(space=space,
train_set=train_set,
param_train_fn=param_train_fn,
split_fn=split_fn,
eval_fn=eval_fn)
assert len(tuning_log) == 3 * 2
space = {
'learning_rate': lambda: [1e-3, 1e-2, 1e-1, 1],
'num_estimators': lambda: [1, 2],
'silent': lambda: [True]
}
tuning_log = grid_search_cv(space=space,
train_set=train_set,
param_train_fn=param_train_fn,
split_fn=split_fn,
eval_fn=eval_fn)
assert len(tuning_log) == 4 * 2
def test_extract():
boston = load_boston()
df = pd.DataFrame(boston['data'], columns=boston['feature_names'])
df['target'] = boston['target']
df['time'] = pd.date_range(start='2015-01-01', periods=len(df))
np.random.seed(42)
df['space'] = np.random.randint(0, 100, size=len(df))
# Define train function
train_fn = linear_regression_learner(
features=boston['feature_names'].tolist(), target="target")
# Define evaluator function
base_evaluator = combined_evaluators(evaluators=[
r2_evaluator(target_column='target', prediction_column='prediction'),
spearman_evaluator(target_column='target',
prediction_column='prediction')
])
splitter = split_evaluator(eval_fn=base_evaluator,
split_col='RAD',
split_values=[4.0, 5.0, 24.0])
temporal_week_splitter = temporal_split_evaluator(eval_fn=base_evaluator,
time_col='time',
time_format='%Y-%W')
temporal_year_splitter = temporal_split_evaluator(eval_fn=base_evaluator,
time_col='time',
time_format='%Y')
eval_fn = combined_evaluators(evaluators=[base_evaluator, splitter])
temporal_week_eval_fn = combined_evaluators(
evaluators=[base_evaluator, temporal_week_splitter])
temporal_year_eval_fn = combined_evaluators(
evaluators=[base_evaluator, temporal_year_splitter])
# Define splitters
cv_split_fn = out_of_time_and_space_splitter(n_splits=5,
in_time_limit='2016-01-01',
time_column='time',
space_column='space')
tlc_split_fn = time_learning_curve_splitter(
training_time_limit='2016-01-01', time_column='time', min_samples=0)
sc_split_fn = stability_curve_time_splitter(
training_time_limit='2016-01-01', time_column='time', min_samples=0)
fw_sc_split_fn = forward_stability_curve_time_splitter(
training_time_start="2015-01-01",
training_time_end="2016-01-01",
holdout_gap=timedelta(days=30),
holdout_size=timedelta(days=30),
step=timedelta(days=30),
time_column='time')
# Validate results
cv_results = validator(df, cv_split_fn, train_fn, eval_fn)['validator_log']
tlc_results = validator(df, tlc_split_fn, train_fn,
eval_fn)['validator_log']
sc_results = validator(df, sc_split_fn, train_fn, eval_fn)['validator_log']
fw_sc_results = validator(df, fw_sc_split_fn, train_fn,
eval_fn)['validator_log']
# temporal evaluation results
predict_fn, _, _ = train_fn(df)
temporal_week_results = temporal_week_eval_fn(predict_fn(df))
temporal_year_results = temporal_year_eval_fn(predict_fn(df))
# Define extractors
base_extractors = combined_evaluator_extractor(base_extractors=[
evaluator_extractor(evaluator_name="r2_evaluator__target"),
evaluator_extractor(evaluator_name="spearman_evaluator__target")
])
splitter_extractor = split_evaluator_extractor(
split_col='RAD',
split_values=[4.0, 5.0, 24.0],
base_extractor=base_extractors)
temporal_week_splitter_extractor = temporal_split_evaluator_extractor(
time_col='time', time_format='%Y-%W', base_extractor=base_extractors)
temporal_year_splitter_extractor = temporal_split_evaluator_extractor(
time_col='time', time_format='%Y', base_extractor=base_extractors)
assert extract(cv_results, base_extractors).shape == (5, 9)
assert extract(cv_results, splitter_extractor).shape == (15, 10)
assert extract(tlc_results, base_extractors).shape == (12, 9)
assert extract(tlc_results, splitter_extractor).shape == (36, 10)
assert extract(sc_results, base_extractors).shape == (5, 9)
assert extract(sc_results, splitter_extractor).shape == (15, 10)
assert extract(fw_sc_results, base_extractors).shape == (3, 9)
assert extract(fw_sc_results, splitter_extractor).shape == (9, 10)
n_time_week_folds = len(df['time'].dt.strftime('%Y-%W').unique())
n_time_year_folds = len(df['time'].dt.strftime('%Y').unique())
assert temporal_week_splitter_extractor(temporal_week_results).shape == (
n_time_week_folds, 3)
assert temporal_year_splitter_extractor(temporal_year_results).shape == (
n_time_year_folds, 3)