def explain(self, X: np.ndarray, Y: np.ndarray = None, verbose: bool = False) -> Explanation:
"""
Explain instance and return PP or PN with metadata.
Parameters
----------
X
Instances to attack
Y
Labels for X
verbose
Print intermediate results of optimization if True
Returns
-------
explanation
`Explanation` object containing the PP or PN with additional metadata as attributes.
"""
if X.shape[0] != 1:
logger.warning('Currently only single instance explanations supported (first dim = 1), '
'but first dim = %s', X.shape[0])
if Y is None:
if self.model:
Y = self.sess.run(self.predict(tf.convert_to_tensor(X, dtype=tf.float32)))
else:
Y = self.predict(X)
Y_ohe = np.zeros(Y.shape)
Y_ohe[np.arange(Y.shape[0]), np.argmax(Y, axis=1)] = 1
Y = Y_ohe.copy()
# find best PP or PN
self.best_attack = False
best_attack, grads = self.attack(X, Y=Y, verbose=verbose)
# output explanation dictionary
data = copy.deepcopy(DEFAULT_DATA_CEM)
data['X'] = X
data['X_pred'] = np.argmax(Y, axis=1)[0]
if not self.best_attack:
logger.warning('No {} found!'.format(self.mode))
# create explanation object
explanation = Explanation(meta=copy.deepcopy(self.meta), data=data)
return explanation
data[self.mode] = best_attack
if self.model:
Y_pert = self.sess.run(self.predict(tf.convert_to_tensor(best_attack, dtype=tf.float32)))
else:
Y_pert = self.predict(best_attack)
data[self.mode + '_pred'] = np.argmax(Y_pert, axis=1)[0]
data['grads_graph'], data['grads_num'] = grads[0], grads[1]
# create explanation object
explanation = Explanation(meta=copy.deepcopy(self.meta), data=data)
return explanation
def build_explanation(self, text: str, result: dict, predicted_label: int,
params: dict) -> Explanation:
""" Uses the metadata returned by the anchor search algorithm together with
the instance to be explained to build an explanation object.
Parameters
----------
text
Instance to be explained.
result
Dictionary containing the search result and metadata.
predicted_label
Label of the instance to be explained. Inferred if not received.
params
Parameters passed to `explain`
"""
result['instance'] = text
result['prediction'] = predicted_label
exp = AnchorExplanation('text', result)
# output explanation dictionary
data = copy.deepcopy(DEFAULT_DATA_ANCHOR)
data.update(anchor=exp.names(),
precision=exp.precision(),
coverage=exp.coverage(),
raw=exp.exp_map)
# create explanation object
explanation = Explanation(meta=copy.deepcopy(self.meta), data=data)
# params passed to explain
explanation.meta['params'].update(params)
return explanation
def build_explanation(self,
X: List[np.ndarray],
baselines: List[np.ndarray],
target: Optional[List[int]],
attributions: List[np.ndarray]) -> Explanation:
data = copy.deepcopy(DEFAULT_DATA_INTGRAD)
data.update(X=X,
baselines=baselines,
target=target,
attributions=attributions)
# calculate predictions
predictions = self.model(X).numpy()
data.update(predictions=predictions)
# calculate convergence deltas
deltas = _compute_convergence_delta(self.model,
self.input_dtypes,
attributions,
baselines,
X,
target,
self._has_inputs)
data.update(deltas=deltas)
return Explanation(meta=copy.deepcopy(self.meta), data=data)
def build_explanation(self,
ale_values: List[np.ndarray],
ale0: List[np.ndarray],
constant_value: float,
feature_values: List[np.ndarray],
feature_deciles: List[np.ndarray],
feature_names: np.ndarray) -> Explanation:
"""
Helper method to build the Explanation object.
"""
# TODO decide on the format for these lists of arrays
# Currently each list element relates to a feature and each column relates to an output dimension,
# this is different from e.g. SHAP but arguably more convenient for ALE.
data = copy.deepcopy(DEFAULT_DATA_ALE)
data.update(
ale_values=ale_values,
ale0=ale0,
constant_value=constant_value,
feature_values=feature_values,
feature_names=feature_names,
target_names=self.target_names,
feature_deciles=feature_deciles
)
return Explanation(meta=copy.deepcopy(self.meta), data=data)
def explain(self, X: np.ndarray) -> Explanation:
"""
Explain an instance and return the counterfactual with metadata.
Parameters
----------
X
Instance to be explained
Returns
-------
`Explanation` object containing the counterfactual with additional metadata as attributes.
"""
# TODO change init parameters on the fly
if X.shape[0] != 1:
logger.warning(
'Currently only single instance explanations supported (first dim = 1), '
'but first dim = %s', X.shape[0])
# make a prediction
Y = self.predict_fn(X)
pred_class = Y.argmax(axis=1).item()
pred_prob = Y.max(axis=1).item()
self.return_dict['orig_class'] = pred_class
self.return_dict['orig_proba'] = pred_prob
logger.debug('Initial prediction: %s with p=%s', pred_class, pred_prob)
# define the class-specific prediction function
self.predict_class_fn, t_class = _define_func(self.predict_fn,
pred_class,
self.target_class)
# initialize with an instance
X_init = self._initialize(X)
# minimize loss iteratively
self._minimize_loss(X, X_init, Y)
return_dict = self.return_dict.copy()
self.instance_dict = dict.fromkeys(
['X', 'distance', 'lambda', 'index', 'class', 'proba', 'loss'])
self.return_dict = {
'cf': None,
'all': {i: []
for i in range(self.max_lam_steps)},
'orig_class': None,
'orig_proba': None
}
# create explanation object
explanation = Explanation(meta=copy.deepcopy(self.meta),
data=return_dict)
return explanation
def test_explanation():
exp = Explanation(meta=valid_meta, data=valid_data)
assert exp.meta == valid_meta
assert exp.data == valid_data
assert isinstance(exp, Explanation)
# test that a warning is raised if accessing attributes as dict keys
with pytest.warns(None) as record:
_ = exp['anchor']
assert len(record) == 1
def build_explanation(self, X: Union[np.ndarray, pd.DataFrame,
sparse.spmatrix],
shap_values: List[np.ndarray],
expected_value: List) -> Explanation:
"""
Create an explanation object.
Parameters
----------
X
Array of instances to be explained.
shap_values
Each entry is a n_instances x n_features array, and the length of the list equals the dimensionality
of the predictor output. The rows of each array correspond to the shap values for the instances with
the corresponding row index in X
expected_value
A list containing the expected value of the prediction for each class.
Returns
-------
An explanation containing a meta field with basic classifier metadata
# TODO: Plotting default should be same space as the explanation? How do we figure out what space they
# explain in?
"""
# TODO: DEFINE COMPLETE SCHEMA FOR THE METADATA (ONGOING)
raw_predictions = self._explainer.linkfv(self.predictor(X))
argmax_pred = np.argmax(np.atleast_2d(raw_predictions), axis=1)
importances = self.rank_by_importance(shap_values)
if isinstance(X, sparse.spmatrix):
X = X.toarray()
else:
X = np.array(X)
# output explanation dictionary
data = copy.deepcopy(DEFAULT_DATA_SHAP)
data.update(shap_values=shap_values,
expected_value=expected_value,
link=self.link,
categorical_names=self.categorical_names,
feature_names=self.feature_names)
data['raw'].update(raw_prediction=raw_predictions,
prediction=argmax_pred,
instances=X,
importances=importances)
return Explanation(meta=copy.deepcopy(self.meta), data=data)
def build_explanation(self, X: np.ndarray, baselines: np.ndarray,
target: list,
attributions: np.ndarray) -> Explanation:
data = copy.deepcopy(DEFAULT_DATA_INTGRAD)
data.update(X=X,
baselines=baselines,
target=target,
attributions=attributions)
# calculate predictions
predictions = self.model(X).numpy()
data.update(predictions=predictions)
# calculate convergence deltas
deltas = _compute_convergence_delta(self.model, attributions,
baselines, X, target)
data.update(deltas=deltas)
return Explanation(meta=copy.deepcopy(self.meta), data=data)
def build_explanation(self, image: np.ndarray, result: dict, predicted_label: int, params: dict) -> Explanation:
"""
Uses the metadata returned by the anchor search algorithm together with
the instance to be explained to build an explanation object.
Parameters
----------
image
Instance to be explained.
result
Dictionary containing the search anchor and metadata.
predicted_label
Label of the instance to be explained.
params
Parameters passed to `explain`
"""
result['instance'] = image
result['instances'] = np.expand_dims(image, 0)
result['prediction'] = np.array([predicted_label])
# overlay image with anchor mask
anchor = self.overlay_mask(image, self.segments, result['feature'])
exp = AnchorExplanation('image', result)
# output explanation dictionary
data = copy.deepcopy(DEFAULT_DATA_ANCHOR_IMG)
data.update(
anchor=anchor,
segments=self.segments,
precision=exp.precision(),
coverage=exp.coverage(),
raw=exp.exp_map
)
# create explanation object
explanation = Explanation(meta=copy.deepcopy(self.meta), data=data)
# params passed to explain
explanation.meta['params'].update(params)
return explanation
def build_explanation(self, X: np.ndarray, result: dict,
predicted_label: int, params: dict) -> Explanation:
"""
Preprocess search output and return an explanation object containing metdata
Parameters
----------
X:
Instance to be explained.
result:
Dictionary with explanation search output and metadata.
predicted_label:
Label of the instance to be explained (inferred if not given).
params
Parameters passed to `explain`
Return
------
`Explanation` object containing human readable explanation, metadata, and precision/coverage
info as attributes.
"""
self.add_names_to_exp(result)
result['prediction'] = np.array([predicted_label])
result['instance'] = X
result['instances'] = np.atleast_2d(X)
exp = AnchorExplanation('tabular', result)
# output explanation dictionary
data = copy.deepcopy(DEFAULT_DATA_ANCHOR)
data.update(anchor=exp.names(),
precision=exp.precision(),
coverage=exp.coverage(),
raw=exp.exp_map)
# create explanation object
explanation = Explanation(meta=copy.deepcopy(self.meta), data=data)
# params passed to explain
explanation.meta['params'].update(params)
return explanation
def test_serialize_deserialize_explanation():
exp = Explanation(meta=valid_meta, data=valid_data)
jrep = exp.to_json()
exp2 = Explanation.from_json(jrep)
assert exp == exp2
