def test_save_AnchorImage(ai_explainer, mnist_predictor):
X = np.random.rand(28, 28, 1)
exp0 = ai_explainer.explain(X)
with tempfile.TemporaryDirectory() as temp_dir:
ai_explainer.save(temp_dir)
ai_explainer1 = load_explainer(temp_dir, predictor=mnist_predictor)
assert isinstance(ai_explainer1, AnchorImage)
assert ai_explainer.meta == ai_explainer1.meta
exp1 = ai_explainer1.explain(X)
assert exp0.meta == exp1.meta
def test_anchor_tabular():
skmodel = SKLearnServer(IRIS_MODEL_URI)
skmodel.load()
with tempfile.TemporaryDirectory() as alibi_model_dir:
make_anchor_tabular(alibi_model_dir)
alibi_model = load_explainer(predictor=skmodel.predict,
path=alibi_model_dir)
anchor_tabular = AnchorTabular(alibi_model)
test_data = np.array([[5.964, 4.006, 2.081, 1.031]])
explanation = anchor_tabular.explain(test_data)
explanation_json = json.loads(explanation.to_json())
assert explanation_json["meta"]["name"] == "AnchorTabular"
def load(cls, path: Union[str, os.PathLike],
predictor: Any) -> "Explainer":
"""
Load an explainer from disk.
Parameters
----------
path
Path to a directory containing the saved explainer.
predictor
Model or prediction function used to originally initialize the explainer.
Returns
-------
An explainer instance.
"""
return load_explainer(path, predictor)
def test_save_TreeShap(tree_explainer, rf_classifier, iris_data):
X = iris_data['X_test']
exp0 = tree_explainer.explain(X)
with tempfile.TemporaryDirectory() as temp_dir:
tree_explainer.save(temp_dir)
tree_explainer1 = load_explainer(temp_dir, predictor=rf_classifier)
assert isinstance(tree_explainer1, TreeShap)
assert tree_explainer.meta == tree_explainer1.meta
exp1 = tree_explainer1.explain(X)
assert exp0.meta == exp1.meta
# TreeShap is deterministic
assert_allclose(exp0.shap_values[0], exp1.shap_values[0])
def test_save_KernelShap(kshap_explainer, lr_classifier, adult_data):
predictor = predict_fcn(predict_type='proba',
clf=lr_classifier,
preproc=adult_data['preprocessor'])
X = adult_data['X_test'][:2]
exp0 = kshap_explainer.explain(X)
with tempfile.TemporaryDirectory() as temp_dir:
kshap_explainer.save(temp_dir)
kshap_explainer1 = load_explainer(temp_dir, predictor=predictor)
assert isinstance(kshap_explainer1, KernelShap)
assert kshap_explainer.meta == kshap_explainer1.meta
exp1 = kshap_explainer.explain(X)
assert exp0.meta == exp1.meta
def test_save_AnchorTabular(atab_explainer, lr_classifier, adult_data):
predictor = predict_fcn(predict_type='class',
clf=lr_classifier,
preproc=adult_data['preprocessor'])
X = adult_data['X_test'][0]
exp0 = atab_explainer.explain(X)
with tempfile.TemporaryDirectory() as temp_dir:
atab_explainer.save(temp_dir)
atab_explainer1 = load_explainer(temp_dir, predictor=predictor)
assert isinstance(atab_explainer1, AnchorTabular)
assert atab_explainer.meta == atab_explainer1.meta
exp1 = atab_explainer1.explain(X)
assert exp0.meta == exp1.meta
def test_save_IG(ig_explainer, ffn_classifier, iris_data):
X = iris_data['X_test']
target = iris_data['y_test']
exp0 = ig_explainer.explain(X, target=target)
with tempfile.TemporaryDirectory() as temp_dir:
ig_explainer.save(temp_dir)
ig_explainer1 = load_explainer(temp_dir, predictor=ffn_classifier)
assert isinstance(ig_explainer1, IntegratedGradients)
assert ig_explainer.meta == ig_explainer1.meta
exp1 = ig_explainer.explain(X, target=target)
assert exp0.meta == exp1.meta
# IG is deterministic
assert np.all(exp0.attributions[0] == exp1.attributions[0])
def test_save_ALE(ale_explainer, lr_classifier, iris_data):
X = iris_data['X_test']
exp0 = ale_explainer.explain(X)
with tempfile.TemporaryDirectory() as temp_dir:
ale_explainer.save(temp_dir)
ale_explainer1 = load_explainer(temp_dir,
predictor=lr_classifier.predict_proba)
assert isinstance(ale_explainer1, ALE)
# TODO: cannot pass as meta updated after explain
# assert ale_explainer.meta == ale_explainer1.meta
exp1 = ale_explainer1.explain(X)
# ALE explanations are deterministic
assert exp0.meta == exp1.meta
# assert exp0.data == exp1.data # cannot compare as many types instide TODO: define equality for explanations?
# or compare pydantic schemas?
assert np.all(exp0.ale_values[0] == exp1.ale_values[0])
def get_persisted_explainer(dirname, predict_fn: Callable) -> Explainer:
logging.info(f"Loading Alibi model from {dirname}")
return load_explainer(predictor=predict_fn, path=dirname)