def test_drlearner_clipping(self):
X = np.linspace(0, 1, 200).reshape(-1, 1)
T = np.random.binomial(1, X)
Y = np.random.normal(size=T.shape)
X[0] = -1000 # one split will have only X values between 0 and 1,
# so the predicted propensity for this point will be extremely low
learner = DRLearner()
learner.fit(Y, T, X)
effect = learner.const_marginal_effect(np.array([[0.5]]))
assert not(np.any(np.isnan(effect)))
def test_drlearner_all_attributes(self):
from sklearn.ensemble import GradientBoostingClassifier, GradientBoostingRegressor, RandomForestRegressor
from sklearn.linear_model import LinearRegression, LogisticRegression
from econml.utilities import StatsModelsLinearRegression
import scipy.special
np.random.seed(123)
controls = np.random.uniform(-1, 1, size=(5000, 3))
T = np.random.binomial(2, scipy.special.expit(controls[:, 0]))
sigma = 0.01
y = (1 + .5 * controls[:, 0]) * T + controls[:,
0] + np.random.normal(0, sigma, size=(5000,))
for X in [controls]:
for W in [None, controls]:
for sample_weight in [None, 1 + np.random.randint(10, size=X.shape[0])]:
for sample_var in [None, 1 + np.random.randint(10, size=X.shape[0])]:
for featurizer in [None, PolynomialFeatures(degree=2, include_bias=False)]:
for models in [(GradientBoostingClassifier(), GradientBoostingRegressor(),
RandomForestRegressor(n_estimators=100,
max_depth=5, min_samples_leaf=50)),
(GradientBoostingClassifier(), GradientBoostingRegressor(),
RandomForestRegressor(n_estimators=100,
max_depth=5, min_samples_leaf=50)),
(LogisticRegression(solver='lbfgs', multi_class='auto'),
LinearRegression(), StatsModelsLinearRegression())]:
for multitask_model_final in [False, True]:
if (not isinstance(models, StatsModelsLinearRegression))\
and (sample_var is not None):
continue
with self.subTest(X=X, W=W, sample_weight=sample_weight, sample_var=sample_var,
featurizer=featurizer, models=models,
multitask_model_final=multitask_model_final):
est = DRLearner(model_propensity=models[0],
model_regression=models[1],
model_final=models[2],
featurizer=featurizer,
multitask_model_final=multitask_model_final)
if (X is None) and (W is None):
with pytest.raises(AttributeError) as e_info:
est.fit(y, T, X=X, W=W,
sample_weight=sample_weight, sample_var=sample_var)
continue
est.fit(
y, T, X=X, W=W, sample_weight=sample_weight, sample_var=sample_var)
np.testing.assert_allclose(est.effect(X[:3], T0=0, T1=1), 1 + .5 * X[:3, 0],
rtol=0, atol=.15)
np.testing.assert_allclose(est.const_marginal_effect(X[:3]),
np.hstack(
[1 + .5 * X[:3, [0]],
2 * (1 + .5 * X[:3, [0]])]),
rtol=0, atol=.15)
for t in [1, 2]:
np.testing.assert_allclose(est.marginal_effect(t, X[:3]),
np.hstack([1 + .5 * X[:3, [0]],
2 * (1 + .5 * X[:3, [0]])]),
rtol=0, atol=.15)
assert isinstance(est.score_, float)
assert isinstance(
est.score(y, T, X=X, W=W), float)
feat_names = ['A', 'B', 'C']
out_feat_names = feat_names
if featurizer is not None:
out_feat_names = featurizer.fit(
X).get_feature_names(feat_names)
np.testing.assert_array_equal(
est.featurizer.n_input_features_, 3)
np.testing.assert_array_equal(est.cate_feature_names(feat_names),
out_feat_names)
if isinstance(models[0], GradientBoostingClassifier):
np.testing.assert_array_equal(np.array([mdl.feature_importances_
for mdl
in est.models_regression]).shape,
[2, 2 + X.shape[1] +
(W.shape[1] if W is not None else 0)])
np.testing.assert_array_equal(np.array([mdl.feature_importances_
for mdl
in est.models_propensity]).shape,
[2, X.shape[1] +
(W.shape[1] if W is not None else 0)])
else:
np.testing.assert_array_equal(np.array([mdl.coef_
for mdl
in est.models_regression]).shape,
[2, 2 + X.shape[1] +
(W.shape[1] if W is not None else 0)])
np.testing.assert_array_equal(np.array([mdl.coef_
for mdl
in est.models_propensity]).shape,
[2, 3, X.shape[1] +
(W.shape[1] if W is not None else 0)])
if multitask_model_final:
if isinstance(models[2], RandomForestRegressor):
np.testing.assert_equal(np.argsort(
est.multitask_model_cate.feature_importances_)[-1], 0)
else:
true_coef = np.zeros(
(2, len(out_feat_names)))
true_coef[:, 0] = [.5, 1]
np.testing.assert_allclose(
est.multitask_model_cate.coef_, true_coef, rtol=0, atol=.15)
np.testing.assert_allclose(
est.multitask_model_cate.intercept_, [1, 2], rtol=0, atol=.15)
else:
for t in [1, 2]:
if isinstance(models[2], RandomForestRegressor):
np.testing.assert_equal(np.argsort(
est.model_cate(T=t).feature_importances_)[-1], 0)
else:
true_coef = np.zeros(
len(out_feat_names))
true_coef[0] = .5 * t
np.testing.assert_allclose(
est.model_cate(T=t).coef_, true_coef, rtol=0, atol=.15)
np.testing.assert_allclose(
est.model_cate(T=t).intercept_, t, rtol=0, atol=.15)