def test_tree_shap():
np.random.seed(0)
alibi_model = make_tree_shap()
tree_shap = TreeShap(alibi_model)
adult = fetch_adult()
X_test = adult.data[30001:, :]
explanation = tree_shap.explain(X_test[0:1].tolist())
exp_json = json.loads(explanation.to_json())
assert exp_json["meta"]["name"] == "TreeShap"
def test_tree_shap():
os.environ.clear()
alibi_model = os.path.join(kfserving.Storage.download(ADULT_EXPLAINER_URI),
EXPLAINER_FILENAME)
with open(alibi_model, "rb") as f:
alibi_model = dill.load(f)
tree_shap = TreeShap(alibi_model)
adult = fetch_adult()
X_test = adult.data[30001:, :]
np.random.seed(0)
explanation = tree_shap.explain(X_test[0:1].tolist())
exp_json = json.loads(explanation.to_json())
print(exp_json)
def __init__(self, ):
adult = fetch_adult()
data = adult.data
target = adult.target
self.feature_names = adult.feature_names
self.category_map = adult.category_map
np.random.seed(0)
data_perm = np.random.permutation(np.c_[data, target])
data = data_perm[:, :-1]
target = data_perm[:, -1]
idx = 30000
self.X_train, self.Y_train = data[:idx, :], target[:idx]
self.X_test, self.Y_test = data[idx + 1:, :], target[idx + 1:]
def test_adult(return_X_y):
data = fetch_adult(return_X_y=return_X_y)
if return_X_y:
assert len(data) == 2
X, y = data
else:
assert len(data) == 5
X = data.data
y = data.target
assert X.ndim == ADULT_DIM
assert X.shape[1] == ADULT_FEATURES
assert len(X) == len(y)
assert len(set(y)) == ADULT_CLASSES
def adult_dataset():
"""
Loads and preprocesses Adult dataset.
"""
# load raw data
adult = fetch_adult()
data = adult.data
target = adult.target
feature_names = adult.feature_names
category_map = adult.category_map
# split it
idx = 30000
X_train, Y_train = data[:idx, :], target[:idx]
X_test, Y_test = data[idx + 1:, :], target[idx + 1:]
# Create feature transformation pipeline
ordinal_features = [
x for x in range(len(feature_names))
if x not in list(category_map.keys())
]
ordinal_transformer = Pipeline(
steps=[('imputer',
SimpleImputer(strategy='median')), ('scaler',
StandardScaler())])
categorical_features = list(category_map.keys())
categorical_transformer = Pipeline(
steps=[('imputer', SimpleImputer(strategy='median')
), ('onehot', OneHotEncoder(handle_unknown='ignore'))])
preprocessor = ColumnTransformer(
transformers=[('num', ordinal_transformer, ordinal_features
), ('cat', categorical_transformer,
categorical_features)])
preprocessor.fit(X_train)
return {
'X_train': X_train,
'y_train': Y_train,
'X_test': X_test,
'y_test': Y_test,
'preprocessor': preprocessor,
'metadata': {
'feature_names': feature_names,
'category_map': category_map,
'name': 'adult'
}
}
def test_kernel_shap():
os.environ.clear()
alibi_model = os.path.join(kfserving.Storage.download(ADULT_EXPLAINER_URI),
EXPLAINER_FILENAME)
with open(alibi_model, "rb") as f:
skmodel = SKLearnServer(ADULT_MODEL_URI)
skmodel.load()
alibi_model = dill.load(f)
kernel_shap = KernelShap(skmodel.predict, alibi_model)
adult = fetch_adult()
X_test = adult.data[30001:, :]
np.random.seed(0)
explanation = kernel_shap.explain(X_test[0:1].tolist())
exp_json = json.loads(explanation.to_json())
print(exp_json)
def tf_keras_adult(tf_keras_adult_model):
# fetch data
adult = fetch_adult()
X = adult.data
X_ord = np.c_[X[:, 1:8], X[:, 11], X[:, 0], X[:, 8:11]]
y = adult.target
# scale numerical features
X_num = X_ord[:, -4:].astype(np.float32, copy=False)
xmin, xmax = X_num.min(axis=0), X_num.max(axis=0)
rng = (-1., 1.)
X_num_scaled = (X_num - xmin) / (xmax - xmin) * (rng[1] - rng[0]) + rng[0]
# OHE categorical features
X_cat = X_ord[:, :-4].copy()
ohe = OneHotEncoder()
ohe.fit(X_cat)
X_cat_ohe = ohe.transform(X_cat)
# combine categorical and numerical data
X_comb = np.c_[X_cat_ohe.todense(), X_num_scaled].astype(np.float32,
copy=False)
# split in train and test set
idx = 30000
X_train, y_train = X_comb[:idx, :], y[:idx]
assert X_train.shape[1] == 57
# set random seed
np.random.seed(1)
tf.set_random_seed(1)
model = tf_keras_adult_model
model.fit(X_train,
to_categorical(y_train),
batch_size=128,
epochs=5,
verbose=0)
# create categorical variable dict
cat_vars_ord = {}
n_categories = 8
for i in range(n_categories):
cat_vars_ord[i] = len(np.unique(X_ord[:, i]))
cat_vars_ohe = ord_to_ohe(X_ord, cat_vars_ord)[1]
return X_train, model, cat_vars_ohe
def test_anchor_tabular():
alibi_model = os.path.join(
kfserving.Storage.download(ADULT_EXPLAINER_URI), EXPLAINER_FILENAME
)
with open(alibi_model, "rb") as f:
skmodel = SKLearnServer(ADULT_MODEL_URI)
skmodel.load()
alibi_model = dill.load(f)
anchor_tabular = AnchorTabular(skmodel.predict, alibi_model)
adult = fetch_adult()
X_test = adult.data[30001:, :]
np.random.seed(0)
explanation = anchor_tabular.explain(X_test[0:1].tolist())
exp_json = json.loads(explanation.to_json())
assert exp_json["data"]["anchor"][0] == "Marital Status = Never-Married"
def test_adult(return_X_y):
try:
data = fetch_adult(return_X_y=return_X_y)
except RequestException:
pytest.skip('Adult dataset URL down')
if return_X_y:
assert len(data) == 2
X, y = data
else:
assert len(data) == 5
X = data.data
y = data.target
assert X.ndim == ADULT_DIM
assert X.shape[1] == ADULT_FEATURES
assert len(X) == len(y)
assert len(set(y)) == ADULT_CLASSES
def test_anchor_tabular():
os.environ.clear()
alibi_model = os.path.join(kserve.Storage.download(ADULT_EXPLAINER_URI),
EXPLAINER_FILENAME)
with open(alibi_model, "rb") as f:
skmodel = SKLearnModel("adult", ADULT_MODEL_URI)
skmodel.load()
predictor = Predictor(skmodel)
alibi_model = dill.load(f)
anchor_tabular = AnchorTabular(predictor.predict_fn, alibi_model)
adult = fetch_adult()
X_test = adult.data[30001:, :]
np.random.seed(0)
explanation = anchor_tabular.explain(X_test[0:1].tolist())
exp_json = json.loads(explanation.to_json())
assert exp_json["data"]["anchor"][0] == "Relationship = Own-child" or \
exp_json["data"]["anchor"][0] == "Age <= 28.00"
def make_tree_shap(dirname: Optional[Path] = None) -> TreeShap:
np.random.seed(0)
# get X_train for explainer fit
adult = fetch_adult()
data = adult.data
target = adult.target
data_perm = np.random.permutation(np.c_[data, target])
data = data_perm[:, :-1]
target = data_perm[:, -1]
idx = 30000
X_train, y_train = data[:idx, :], target[:idx]
X_test, y_test = data[idx + 1:, :], target[idx + 1:]
d_train = xgboost.DMatrix(X_train, label=y_train)
d_test = xgboost.DMatrix(X_test, label=y_test)
params = {
"eta": 0.01,
"objective": "binary:logistic",
"subsample": 0.5,
"base_score": np.mean(y_train),
"eval_metric": "logloss",
}
model = xgboost.train(
params,
d_train,
5000,
evals=[(d_test, "test")],
verbose_eval=100,
early_stopping_rounds=20,
)
tree_explainer = TreeShap(model, model_output="raw", task="classification")
tree_explainer.fit(X_train)
if dirname is not None:
tree_explainer.save(dirname)
return tree_explainer
def make_anchor_tabular_income(
dirname: Optional[Path] = None) -> AnchorTabular:
# adapted from:
# https://docs.seldon.io/projects/alibi/en/latest/examples/anchor_tabular_adult.html
np.random.seed(0)
# prepare data
adult = fetch_adult()
data = adult.data
target = adult.target
feature_names = adult.feature_names
category_map = adult.category_map
data_perm = np.random.permutation(np.c_[data, target])
data = data_perm[:, :-1]
target = data_perm[:, -1]
# build model
idx = 30000
X_train, Y_train = data[:idx, :], target[:idx]
X_test, Y_test = data[idx + 1:, :], target[idx + 1:]
ordinal_features = [
x for x in range(len(feature_names))
if x not in list(category_map.keys())
]
ordinal_transformer = Pipeline(steps=[
("imputer", SimpleImputer(strategy="median")),
("scaler", StandardScaler()),
])
categorical_features = list(category_map.keys())
categorical_transformer = Pipeline(steps=[
("imputer", SimpleImputer(strategy="median")),
("onehot", OneHotEncoder(handle_unknown="ignore")),
])
preprocessor = ColumnTransformer(transformers=[
("num", ordinal_transformer, ordinal_features),
("cat", categorical_transformer, categorical_features),
])
clf = RandomForestClassifier(n_estimators=50)
model_pipeline = Pipeline(steps=[
("preprocess", preprocessor),
("classifier", clf),
])
model_pipeline.fit(X_train, Y_train)
explainer = AnchorTabular(model_pipeline.predict,
feature_names,
categorical_names=category_map,
seed=1)
explainer.fit(X_train, disc_perc=[25, 50, 75])
if dirname is not None:
explainer.save(dirname)
return explainer
# See the License for the specific language governing permissions and # limitations under the License. import numpy as np from sklearn.ensemble import RandomForestClassifier from sklearn.compose import ColumnTransformer from sklearn.impute import SimpleImputer from sklearn.preprocessing import StandardScaler, OneHotEncoder from alibi.datasets import fetch_adult import joblib import dill from sklearn.pipeline import Pipeline import alibi # load data adult = fetch_adult() data = adult.data targets = adult.target feature_names = adult.feature_names category_map = adult.category_map # define train and test set np.random.seed(0) data_perm = np.random.permutation(np.c_[data, targets]) data = data_perm[:, :-1] labels = data_perm[:, -1] idx = 30000 X_train, Y_train = data[:idx, :], targets[:idx] X_test, Y_test = data[idx + 1:, :], targets[idx + 1:]
