def _train_local_classifier(
estimator,
samples,
similarity,
predict_proba,
expand_factor=10,
test_size=0.3,
random_state=None,
):
# type: (Any, Any, np.ndarray, Callable[[Any], np.ndarray], int, float, Any) -> Dict[str, float]
rng = check_random_state(random_state)
y_proba = predict_proba(samples)
(X_train, X_test, similarity_train, similarity_test, y_proba_train,
y_proba_test) = train_test_split(samples,
similarity,
y_proba,
test_size=test_size,
random_state=rng)
# XXX: in the original lime code instead of a probabilitsic classifier
# they build several regression models which try to output probabilities.
#
# XXX: Probability information is helpful because it could be hard
# to get enough examples of all classes automatically, so we're fitting
# classifier to produce the same probabilities, not only the same
# best answer.
# TODO: feature selection
# Ideally, it should be supported as a Pipeline (i.e. user should
# be able to configure it).
fit_proba(estimator,
X_train,
y_proba_train,
expand_factor=expand_factor,
sample_weight=similarity_train,
random_state=rng)
y_proba_test_pred = estimator.predict_proba(X_test)
return {
'mean_KL_divergence':
mean_kl_divergence(y_proba_test_pred,
y_proba_test,
sample_weight=similarity_test),
'score':
score_with_sample_weight(estimator,
X_test,
y_proba_test.argmax(axis=1),
sample_weight=similarity_test)
}
def test_fit_proba():
X = np.array([
[0.0, 0.8],
[0.0, 0.5],
[1.0, 0.1],
[0.9, 0.2],
[0.7, 0.3],
])
y_proba = np.array([
[0.0, 1.0],
[0.1, 0.9],
[1.0, 0.0],
[0.55, 0.45],
[0.4, 0.6],
])
y_bin = y_proba.argmax(axis=1)
# fit on binary labels
clf = LogisticRegression(C=10, random_state=42)
clf.fit(X, y_bin)
y_pred = clf.predict_proba(X)[:, 1]
mae = mean_absolute_error(y_proba[:, 1], y_pred)
print(y_pred, mae)
# fit on probabilities
clf2 = LogisticRegression(C=10, random_state=42)
fit_proba(clf2, X, y_proba, expand_factor=200, random_state=42)
y_pred2 = clf2.predict_proba(X)[:, 1]
mae2 = mean_absolute_error(y_proba[:, 1], y_pred2)
print(y_pred2, mae2)
assert mae2 * 1.2 < mae
# let's get 3th example really right
sample_weight = np.array([0.1, 0.1, 0.1, 10.0, 0.1])
clf3 = LogisticRegression(C=10, random_state=42)
fit_proba(clf3, X, y_proba, expand_factor=200, sample_weight=sample_weight,
random_state=42)
y_pred3 = clf3.predict_proba(X)[:, 1]
print(y_pred3)
val = y_proba[3][1]
assert abs(y_pred3[3] - val) * 1.5 < abs(y_pred2[3] - val)
assert abs(y_pred3[3] - val) * 1.5 < abs(y_pred[3] - val)
# without expand_factor it is just clf.fit
clf4 = LogisticRegression(C=10, random_state=42)
fit_proba(clf4, X, y_proba, expand_factor=None,
random_state=42)
y_pred4 = clf4.predict_proba(X)[:, 1]
assert np.allclose(y_pred, y_pred4)
# it should work the same with sparse data
X_sparse = sp.csr_matrix(X)
clf4 = LogisticRegression(C=10, random_state=42)
fit_proba(clf4, X_sparse, y_proba, expand_factor=200, random_state=42)
y_pred4 = clf4.predict_proba(X)[:, 1]
assert np.allclose(y_pred2, y_pred4)
def _train_local_classifier(estimator,
samples,
similarity, # type: np.ndarray
y_proba, # type: np.ndarray
expand_factor=10, # type: int
test_size=0.3, # type: float
random_state=None,
):
# type: (...) -> Dict[str, float]
rng = check_random_state(random_state)
(X_train, X_test,
similarity_train, similarity_test,
y_proba_train, y_proba_test) = train_test_split(samples,
similarity,
y_proba,
test_size=test_size,
random_state=rng)
# XXX: in the original lime code instead of a probabilitsic classifier
# they build several regression models which try to output probabilities.
#
# XXX: Probability information is helpful because it could be hard
# to get enough examples of all classes automatically, so we're fitting
# classifier to produce the same probabilities, not only the same
# best answer.
# TODO: feature selection
# Ideally, it should be supported as a Pipeline (i.e. user should
# be able to configure it).
fit_proba(estimator, X_train, y_proba_train,
expand_factor=expand_factor,
sample_weight=similarity_train,
random_state=rng)
y_proba_test_pred = estimator.predict_proba(X_test)
if y_proba_test_pred.shape != y_proba_test.shape:
# Sometimes generated training labels may contain only a subset of
# target classes; it means it could happen that dimensions
# of predicted probability matrices don't match.
#
# XXX: the fix is not complete; to explain predictions
# of the fitted estimator one still have to take care of target_names.
if not hasattr(estimator, 'classes_'):
raise ValueError("Result dimensions don't match and estimator"
"doesn't provide 'classes_' attribute; can't"
"figure out how are columns related.")
seen_classes = estimator.classes_
complete_classes = np.arange(y_proba.shape[1])
y_proba_test_pred = fix_multiclass_predict_proba(
y_proba=y_proba_test_pred,
seen_classes=seen_classes,
complete_classes=complete_classes
)
return {
'mean_KL_divergence': mean_kl_divergence(
y_proba_test_pred,
y_proba_test,
sample_weight=similarity_test
),
'score': score_with_sample_weight(estimator,
X_test,
y_proba_test.argmax(axis=1),
sample_weight=similarity_test)
}