def describe_cf_instance(X,
explanation,
class_names,
cat_vars_ohe=None,
category_map=None,
feature_names=None,
eps=1e-2):
print("Instance Outcomes and Probabilities")
print("-" * 48)
max_len = len(max(feature_names, key=len))
print('{}: {}\r\n{} {}'.format('original'.rjust(max_len),
class_names[explanation.orig_class],
" " * max_len,
explanation.orig_proba[0]))
if explanation.cf != None:
print('{}: {}\r\n{} {}'.format('counterfactual'.rjust(max_len),
class_names[explanation.cf['class']],
" " * max_len,
explanation.cf['proba'][0]))
print("\r\nCategorical Feature Counterfactual Perturbations")
print("-" * 48)
X_orig_ord = ohe_to_ord(X, cat_vars_ohe)[0]
try:
X_cf_ord = ohe_to_ord(explanation.cf['X'], cat_vars_ohe)[0]
except:
X_cf_ord = ohe_to_ord(explanation.cf['X'].transpose(),
cat_vars_ohe)[0].transpose()
delta_cat = {}
for _, (i, v) in enumerate(category_map.items()):
cat_orig = v[int(X_orig_ord[0, i])]
cat_cf = v[int(X_cf_ord[0, i])]
if cat_orig != cat_cf:
delta_cat[feature_names[i]] = [cat_orig, cat_cf]
if delta_cat:
for k, v in delta_cat.items():
print("\t{}: {} --> {}".format(k.rjust(max_len), v[0],
v[1]))
print("\r\nNumerical Feature Counterfactual Perturbations")
print("-" * 48)
num_idxs = [i for i in list(range(0,len(feature_names)))\
if i not in category_map.keys()]
delta_num = X_cf_ord[0, num_idxs] - X_orig_ord[0, num_idxs]
for i in range(delta_num.shape[0]):
if np.abs(delta_num[i]) > eps:
print("\t{}: {:.2f} --> {:.2f}".format(
feature_names[i].rjust(max_len), X_orig_ord[0, i],
X_cf_ord[0, i]))
else:
print("\tNO COUNTERFACTUALS")
def test_tf_keras_adult_explainer(disable_tf2, adult_data,
tf_keras_adult_explainer, use_kdtree, k,
d_type):
model, cf = tf_keras_adult_explainer
X_train = adult_data['preprocessor'].transform(
adult_data['X_train']).toarray()
# instance to be explained
x = X_train[0].reshape(1, -1)
pred_class = np.argmax(model.predict(x))
not_pred_class = np.argmin(model.predict(x))
# test fit
cf.fit(X_train, d_type=d_type)
# checked ragged tensor shape
n_cat = len(list(cf.cat_vars_ord.keys()))
max_key = max(cf.cat_vars_ord, key=cf.cat_vars_ord.get)
max_cat = cf.cat_vars_ord[max_key]
assert cf.d_abs_ragged.shape == (n_cat, max_cat)
if use_kdtree: # k-d trees
assert len(cf.kdtrees) == cf.classes # each class has a k-d tree
n_by_class = 0
for c in range(cf.classes):
n_by_class += cf.X_by_class[c].shape[0]
assert n_by_class == X_train.shape[
0] # all training instances are stored in the trees
# test explanation
explanation = cf.explain(x, k=k)
if use_kdtree:
assert cf.id_proto != pred_class
assert np.argmax(model.predict(
explanation.cf['X'])) == explanation.cf['class']
num_shape = (1, 12)
assert explanation.cf['grads_num'].shape == explanation.cf[
'grads_graph'].shape == num_shape
assert explanation.meta.keys() == DEFAULT_META_CFP.keys()
assert explanation.data.keys() == DEFAULT_DATA_CFP.keys()
# test gradient shapes
y = np.zeros((1, cf.classes))
np.put(y, pred_class, 1)
cf.predict = cf.predict.predict # make model black box
# convert instance to numerical space
x_ord = ohe_to_ord(x, cf.cat_vars)[0]
x_num = ord_to_num(x_ord, cf.d_abs)
# check gradients
grads = cf.get_gradients(x_num, y, num_shape[1:], cf.cat_vars_ord)
assert grads.shape == num_shape
def test_mapping_fn():
# ohe_to_ord_shape
assert mp.ohe_to_ord_shape(shape_ohe, cat_vars_ohe,
is_ohe=True) == shape_ord
# ohe_to_ord
X_ohe_to_ord, cat_vars_ohe_to_ord = mp.ohe_to_ord(X_ohe, cat_vars_ohe)
assert (X_ohe_to_ord
== X_ord).all() and cat_vars_ohe_to_ord == cat_vars_ord
# ord_to_ohe
X_ord_to_ohe, cat_vars_ord_to_ohe = mp.ord_to_ohe(X_ord, cat_vars_ord)
assert (X_ord_to_ohe
== X_ohe).all() and cat_vars_ohe == cat_vars_ord_to_ohe
# ord_to_num
X_ord_to_num = mp.ord_to_num(X_ord, dist)
assert (X_num == X_ord_to_num).all()
# num_to_ord
X_num_to_ord = mp.num_to_ord(X_num, dist)
assert (X_ord == X_num_to_ord).all()
