class Anchors2(FeatureImportance):
"""
Feature importance method by [RIB]_.
References
----------
.. [RIB] Ribeiro, et al, "Anchors: High-precision model-agnostic explanations",
Proceedings of the AAAI Conference on Artificial Intelligence, Volume 32, 2018.
"""
def __init__(self, model: Any, seed: int = SEED):
super().__init__(seed=seed)
self._model = assign_model(model=model)
self._explainer = None
def fit(self, X: Any) -> None:
self._explainer = AnchorTabularExplainer(
class_names=['0', '1'],
feature_names=list(range(X.shape[1])),
train_data=X,
# discretizer='quartile')
discretizer='decile')
def _compute_anchors_per_sample(self, X: np.ndarray, idx: int) -> List:
result = self._explainer.explain_instance(
data_row=X[idx, :], classifier_fn=self._model.predict)
return result.exp_map['feature']
@staticmethod
def _calculate_importance(anchors: List, output_shape: Tuple) -> np.ndarray:
importance = np.zeros(shape=output_shape)
for k, anchor in enumerate(anchors):
if isinstance(anchor, list):
importance[k, anchor] = 1
else:
importance[anchor] = 1
return importance
def _compute_anchors(self, X: np.ndarray, num_jobs: int) -> List:
return Parallel(n_jobs=num_jobs)(
delayed(self._compute_anchors_per_sample)(X, sample_idx)
for sample_idx in range(X.shape[0]))
def explain(self, X: np.ndarray, sample_idx: int) -> np.ndarray:
anchors = self._compute_anchors_per_sample(X=X, idx=sample_idx)
return self._calculate_importance(anchors=anchors, output_shape=(X.shape[1],))
def explain_batch(self, X: np.ndarray, num_jobs: int = 2) -> np.ndarray:
np.random.seed(self._seed)
anchors = self._compute_anchors(X=X, num_jobs=num_jobs)
return self._calculate_importance(anchors=anchors, output_shape=X.shape)
class XDeepAnchorTabularExplainer(Explainer):
def __init__(self,
class_names,
feature_names,
data,
categorical_names=None):
"""Init function.
# Arguments
class_names: List. A list of class names, ordered according to whatever the classifier is using.
feature_names: List. list of names (strings) corresponding to the columns in the training data.
data: Array. Full data including train data, validation_data and test data.
categorical_names: Dict. A dict which maps from int to list of names, where categorical_names[x][y] represents the name of the yth value of column x.
"""
Explainer.__init__(self, None, class_names)
self.feature_names = feature_names
self.data = data
self.categorical_names = categorical_names
# Initialize explainer
self.set_parameters()
def set_parameters(self, **kwargs):
"""Parameter setter for anchor_tabular.
# Arguments
**kwargs: Parameters setter. For more detail, please check https://lime-ml.readthedocs.io/en/latest/index.html.
"""
class_names = kwargs.pop("class_names", self.class_names)
feature_names = kwargs.pop("feature_names", self.feature_names)
data = kwargs.pop("data", self.data)
categorical_names = kwargs.pop("categorical_names",
self.categorical_names)
self.explainer = AnchorTabularExplainer(
class_names,
feature_names,
data=data,
categorical_names=categorical_names)
def get_anchor_encoder(self,
train_data,
train_labels,
validation_data,
validation_labels,
discretizer='quartile'):
"""Get encoder for tabular data.
If you want to use 'anchor' to explain tabular classifier. You need to get this encoder to encode your data, and train another model.
# Arguments
train_data: Array. Train data.
train_labels: Array. Train labels.
validation_data: Array. Validation set.
validation_labels: Array. Validation labels.
discretizer: Str. Discretizer for data. Please choose between 'quartile' and 'decile'.
# Return
A encoder object which has function 'transform'.
"""
self.explainer.fit(train_data,
train_labels,
validation_data,
validation_labels,
discretizer=discretizer)
return self.explainer.encoder
def set_anchor_predict_proba(self, predict_proba):
"""Predict function setter.
# Arguments
predict_proba: Function. A classifier prediction probability function.
"""
self.predict_proba = predict_proba
def explain(self, instance, top_labels=None, labels=(1, ), **kwargs):
"""Generate explanation for a prediction using certain method.
Anchor does not use top_labels and labels.
# Arguments
instance: Array. One row of tabular data to be explained.
**kwarg: Parameters setter. For more detail, please check https://lime-ml.readthedocs.io/en/latest/index.html.
"""
if self.predict_proba is None:
raise XDeepError(
"Please call set_anchor_predict_proba to pass in your new predict function."
)
def predict_label(x):
return np.argmax(self.predict_proba(x), axis=1)
self.instance = instance
self.labels = self.predict_proba(
self.explainer.encoder.transform([instance])).argsort()[0][-1:]
self.explanation = self.explainer.explain_instance(
instance, predict_label, **kwargs)
def show_explanation(self, show_in_note_book=True, verbose=True):
"""Visualization of explanation of anchor_tabular.
# Arguments
show_in_note_book: Boolean. Whether show in jupyter notebook.
verbose: Boolean. Whether print out examples and counter examples.
"""
Explainer.show_explanation(self)
exp = self.explanation
print()
print("Anchor Explanation")
print()
if verbose:
print()
print(
'Examples where anchor applies and model predicts same to instance:'
)
print()
print(exp.examples(only_same_prediction=True))
print()
print(
'Examples where anchor applies and model predicts different with instance:'
)
print()
print(exp.examples(only_different_prediction=True))
label = self.labels[0]
print('Prediction: {}'.format(label))
print('Anchor: %s' % (' AND '.join(exp.names())))
print('Precision: %.2f' % exp.precision())
print('Coverage: %.2f' % exp.coverage())
if show_in_note_book:
exp.show_in_notebook()
shap_plot = shap.force_plot(explainer.expected_value[i],
shap_values[i],
user_encoded,
feature_names=feature_names,
out_names=class_names[i])
shap_plot_reprs.append(shap_plot._repr_html_())
shap_html_repr = "".join(shap_plot_reprs)
bundle_path = Path(shap.__file__).parent / 'plots' / 'resources' / "bundle.js"
with bundle_path.open('r') as f:
bundle_data = f.read()
shape_js = f"<script charset='utf-8'>{bundle_data}</script>"
componenets.html(f"{shape_js}{shap_html_repr}", height=420)
st.header("Anchors")
X_encoded = clf[:-1].transform(X)
anchor_explainer = AnchorTabularExplainer(class_names,
feature_names,
X_encoded,
categorical_names={
0: metadata['island_categories'],
1: metadata['gender_categories']
})
exp = anchor_explainer.explain_instance(user_encoded[0, :], clf[-1].predict)
exp_html = exp.as_html()
componenets.html(exp_html, height=700)