def train_explainer(regressor: LogisticRegression, feature_names: List[str], X_train: np.ndarray, X_test: np.ndarray, y_test: np.ndarray):
predict_fn = lambda x: regressor.predict(x)
explainer = AnchorTabular(predict_fn, feature_names)
explainer.fit(X_train)
file_path=""
with open("explainer.dill", "wb") as file:
dill.dump(explainer, file)
file_path = file.name
mlflow.log_artifact("explainer.dill", "model")
print(np.where(y_test == 1)[0])
probe = np. array([40.316667556762695, 0.5605325219195545, 0.350, 0, 3, 1, 5], dtype=float)
#probe = np. array(X_test[700], dtype=float)
explanation = explainer.explain(probe)
print('Anchor: %s' % (' AND '.join(explanation['names'])))
print('Precision: %.2f' % explanation['precision'])
print('Coverage: %.2f' % explanation['coverage'])
print(explanation)
return explainer
# kedro install
# kedro run
# kedro viz
def atab_explainer(lr_classifier, adult_data):
predictor = predict_fcn(predict_type='class',
clf=lr_classifier,
preproc=adult_data['preprocessor'])
atab = AnchorTabular(
predictor=predictor,
feature_names=adult_data['metadata']['feature_names'],
categorical_names=adult_data['metadata']['category_map'])
atab.fit(adult_data['X_train'], disc_perc=(25, 50, 75))
return atab
def train_explainer(artifacts_folder: str, data: AdultData, model: RandomForestClassifier) -> AnchorTabular:
def predict_fn(x):
return model.predict(x)
explainer = AnchorTabular(predict_fn, data.feature_names, categorical_names=data.category_map, seed=1)
explainer.fit(data.X_train, disc_perc=(25, 50, 75))
with open(f"{artifacts_folder}/{EXPLAINER_FOLDER}" + "/explainer.dill", "wb") as f:
explainer.predictor = None
explainer.samplers[0].predictor = None
dill.dump(explainer, f)
return explainer
class Anchors(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 = AnchorTabular(
predictor=self._model.predict_proba,
feature_names=list(range(X.shape[1])), seed=self._seed)
self._explainer.fit(train_data=X)
# disc_perc=(0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9))
# disc_perc=(0.1, 0.3, 0.5, 0.7, 0.9))
# disc_perc=(0.2, 0.4, 0.6, 0.8))
def _compute_anchors_per_sample(self, X: np.ndarray, idx: int) -> List:
result = self._explainer.explain(X=X[idx, :])
return result.data['raw']['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:
anchors = self._compute_anchors(X=X, num_jobs=num_jobs)
return self._calculate_importance(anchors=anchors, output_shape=X.shape)
def at_iris_explainer(get_iris_dataset, rf_classifier, request):
"""
Instantiates and fits an AnchorTabular explainer for the Iris dataset.
"""
predict_type = request.param
data = get_iris_dataset
clf, _ = rf_classifier # preprocessor not necessary
# instantiate and fit explainer
pred_fn = predict_fcn(predict_type, clf)
explainer = AnchorTabular(pred_fn, data['metadata']['feature_names'])
explainer.fit(data['X_train'], disc_perc=(25, 50, 75))
return data['X_test'], explainer, pred_fn, predict_type
def at_adult_explainer(get_adult_dataset, rf_classifier, request):
"""
Instantiates and fits an AnchorTabular explainer for the Adult dataset.
"""
# fit random forest classifier
predict_type = request.param
data = get_adult_dataset
clf, preprocessor = rf_classifier
# instantiate and fit explainer
pred_fn = predict_fcn(predict_type, clf, preprocessor)
explainer = AnchorTabular(
pred_fn,
data['metadata']['feature_names'],
categorical_names=data['metadata']['category_map'])
explainer.fit(data['X_train'], disc_perc=(25, 50, 75))
return data['X_test'], explainer, pred_fn, predict_type
def fit(self, x, y):
self.dim = x.shape[1]
# clf = sklearn.svm.SVC(kernel=self.kernel, probability=True)
clf = RandomForestClassifier()
clf.fit(x, y)
y_pred = clf.predict(x)
print("Clf model accuracy: [{:.4f}]".format(
sklearn.metrics.accuracy_score(y, y_pred)))
self.ano_idx = np.where(y == 1)[0]
print(self.ano_idx.shape)
n_f = x.shape[1]
feature_names = ["A" + str(i) for i in range(n_f)]
# use anchor
predict_fn = lambda xx: clf.predict_proba(xx)
explainer = AnchorTabular(predict_fn, feature_names)
explainer.fit(x, disc_perc=(25, 50, 75))
exp_sub_lst = []
for i in tqdm(range(len(self.ano_idx))):
ano = x[self.ano_idx[i]]
explanation = explainer.explain(ano, threshold=0.95)
anchor = explanation['anchor']
f_sub = []
for a in anchor:
for item in a.split(" "):
if item.startswith("A"):
item = int(item[1:])
f_sub.append(item)
# print(anchor, f_sub)
if len(f_sub) == 0:
f_sub = np.arange(n_f)
exp_sub_lst.append(f_sub)
return exp_sub_lst
def make_anchor_tabular(dirname: Optional[Path] = None) -> AnchorTabular:
# train model
iris_data = load_iris()
clf = LogisticRegression(solver="liblinear", multi_class="ovr")
clf.fit(iris_data.data, iris_data.target)
# create explainer
explainer = AnchorTabular(clf.predict,
feature_names=iris_data.feature_names)
explainer.fit(iris_data.data, disc_perc=(25, 50, 75))
if dirname is not None:
explainer.save(dirname)
return explainer
def retrain_classifier_final(self, args, nn_model_ref):
nn_model_ref.epochs = args.num_epch_2
nn_model_ref.batch_size_2 = args.batch_size_2
nn_model_ref.net.freeze()
X_train_proba_feat, X_eval_proba_feat = nn_model_ref.all_intermediaire, nn_model_ref.all_intermediaire_val
Y_train_proba = nn_model_ref.Y_train_nn_binaire
Y_eval_proba = nn_model_ref.Y_val_nn_binaire
print("START RETRAIN LINEAR NN GOHR ")
print()
"""net_retrain, h = train_speck_distinguisher(args, X_train_proba_feat.shape[1], X_train_proba_feat,
Y_train_proba, X_eval_proba_feat, Y_eval_proba,
bs=args.batch_size_2,
epoch=args.num_epch_2, name_ici="retrain_nn_gohr",
wdir=self.path_save_model)"""
from alibi.explainers import AnchorTabular
#from alibi.explainers import AnchorImage
from sklearn.ensemble import RandomForestClassifier
clf = RandomForestClassifier(n_estimators=50)
clf.fit(X_train_proba_feat, Y_train_proba)
predict_fn = lambda x: clf.predict_proba(x)
feature_names = [i for i in range(X_train_proba_feat.shape[1])]
explainer = AnchorTabular(predict_fn, feature_names)
idx = 0
explainer.fit(X_train_proba_feat, disc_perc=(25))
print('Prediction: ',
explainer.predictor(X_eval_proba_feat[idx].reshape(1, -1))[0])
#print('Prediction: ', explainer.predict_fn(X_eval_proba_feat[idx].reshape(1, -1))[0])
explanation = explainer.explain(X_eval_proba_feat[idx], threshold=0.8)
print('Anchor: %s' % (' AND '.join(explanation['names'])))
print('Precision: %.2f' % explanation['precision'])
print('Coverage: %.2f' % explanation['coverage'])
print(ok)
return net_retrain
import numpy as np
from sklearn.datasets import load_iris
from alibi.explainers import AnchorTabular
import requests
dataset = load_iris()
feature_names = dataset.feature_names
iris_data = dataset.data
model_url = "http://localhost:8003/seldon/seldon/iris/api/v1.0/predictions"
def predict_fn(X):
data = {"data": {"ndarray": X.tolist()}}
r = requests.post(model_url, json={"data": {"ndarray": [[1, 2, 3, 4]]}})
return np.array(r.json()["data"]["ndarray"])
explainer = AnchorTabular(predict_fn, feature_names)
explainer.fit(iris_data, disc_perc=(25, 50, 75))
explainer.save("./explainer/")
clf = DecisionTreeClassifier(random_state=42)
else:
clf = RandomForestClassifier(random_state=42)
# st.sidebar.write(selected_model)
clf.fit(X_train, y_train)
y_pred = clf.predict(X_test)
st.write("""### Metrics:""")
st.write('Train accuracy:', accuracy_score(y_train,clf.predict(X_train)))
st.write('Test accuracy:', accuracy_score(y_test, clf.predict(X_test)))
confusion_matrix(y_test, y_pred)
st.write('Confusion matrix:')
plot_confusion_matrix(clf, X_test, y_test)
st.pyplot()
# st.write(classification_report(y_test, y_pred))
predict_fn = lambda x: clf.predict_proba(x)
explainer = AnchorTabular(predict_fn, feature_names)
explainer.fit(X_train)
idx = st.sidebar.slider(label='Select an instance:',min_value=1,max_value=len(y_test))
st.write("""### Selected instance:""")
st.write(X_test_df.iloc[[idx-1]], height=150)
print(y_train_df.iloc[[idx-1]])
st.write('Prediction: ', class_names[explainer.predictor(X_test[idx-1].reshape(1, -1))[0]])
st.write("""### Prediction Explained:""")
with st.spinner('Calculating'):
explanation = explainer.explain(X_test[idx-1], threshold=0.70)
st.write('Anchor (instance explanation): %s' % (' AND '.join(explanation.anchor)))
st.write('Precision: %.2f' % explanation.precision)
st.write('Coverage: %.2f' % explanation.coverage)
# st.write("""### Trust score:""")
ts = TrustScore(k_filter=10,
alpha=.05,
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
def fit(self, X: Any) -> None:
self._explainer = AnchorTabular(
predictor=self._model.predict_proba,
feature_names=list(range(X.shape[1])), seed=self._seed)
self._explainer.fit(train_data=X)
def anchors_connector(self, *arg):
query_instance = dict(s.split(':') for s in arg)
#anchor instance to model instance. Input: Numpy. Output: Pandas df. Turns numbers into categories.
def adapter(n):
d = pd.DataFrame(data=n, columns=self.featureNames)
categories = self.getCategoricalFeatures()
for c in categories:
d[c] = d[c].map(self.dictionary[c]["values"])
#d['Sex'] = d['Sex'].map({0:'Male', 1: 'Female'})
#d['Embarked'] = d['Embarked'].map({0: 'Southampton', 1: 'Cherbourg', 2: 'Queenstown'})
#d['Pclass'] = d['Pclass'].map({0: 'First', 1: 'Second', 2: 'Third'})
return d
#model instance to anchor instance. Input: Pandas df. Output: Numpy. Turns categories into numbers.
def reverse_adapter(p):
d = p.copy()
categories = self.getCategoricalFeatures()
for c in categories:
d[c] = d[c].map(
{v: k
for k, v in self.dictionary[c]["values"].items()})
#d['Sex'] = d['Sex'].map({'Male': 0, 'Female': 1})
#d['Embarked'] = d['Embarked'].map({'Southampton': 0, 'Cherbourg': 1, 'Queenstown': 2})
#d['Pclass'] = d['Pclass'].map({'First': 0, 'Second': 1, 'Third': 2})
n = d.to_numpy().astype(np.float)
return (n)
predict_fn = lambda x: self.model.predict(adapter(x))
#create the category map
categories = self.getCategoricalFeatures()
category_map = {}
for i in range(len(self.featureNames)):
if self.featureNames[i] in categories:
category_map[i] = [
str(k) for k in list(self.dictionary[self.featureNames[i]]
["values"].values())
]
#category_map = {0: ['First', 'Second', 'Third'], 1: ['Male','Female'], 4: ['Southampton', 'Cherbourg', 'Queenstown']}
print("-------")
print(query_instance)
print(reverse_adapter(pd.DataFrame([query_instance])))
#sort query_instance
sorted_query_instance = {}
for f in self.featureNames:
sorted_query_instance[f] = query_instance[f]
print(sorted_query_instance)
print(reverse_adapter(pd.DataFrame([sorted_query_instance])))
explainer = AnchorTabular(predict_fn,
feature_names=self.featureNames,
categorical_names=category_map)
anchor_training = reverse_adapter(self.X_train)
explainer.fit(anchor_training, disc_perc=[25, 50, 75])
explanation = explainer.explain(reverse_adapter(
pd.DataFrame([sorted_query_instance])),
threshold=0.90,
max_anchor_size=3,
batch_size=2000)
print('Anchor: %s' % (' AND '.join(explanation['data']['anchor'])))
print('Precision: %.2f' % explanation['precision'])
print('Coverage: %.2f' % explanation['coverage'])
#build rule
rule = ""
names = explanation['data']['anchor']
precision = np.asarray(explanation['raw']['precision'])
precision[1:] -= precision[:-1].copy()
precision = [round(elem, 2) for elem in precision.tolist()]
for i in range(0, len(names)):
rule = rule + names[i]
importance = round(precision[i] / sum(precision) * 100, 2)
rule = rule + " (" + str(importance) + "%)"
if (i < len(names) - 1):
rule = rule + " AND "
self.explanation = 'I generated the following rule for you. It describes the boundaries under which the current prediction remains stable: <br> <br> <big>' + rule + '</big>. <br> <br> Each rule condition has an importance score which shows how critical the condition is for the prediction outcome to stay stable.'
self.certainty = 'I tested the rule on many sample data instances. The condition applies on %.2f' % explanation[
'coverage'] + ' of the instances. In these cases, the rule was accurate in %.2f' % explanation[
'precision'] + ' of the cases.'
return (True)
def main(unused_args):
##Read hypertuning values from file
if args.component == 'training':
timestamp = str(args.timestamp)
filename = "/mnt/Model_Blerssi/hpv-" + timestamp + ".txt"
f = open(filename, "r")
args.tf_batch_size = int(f.readline())
args.learning_rate = float(f.readline())
print("****************")
print("Optimized Hyper paramater value")
print("Batch-size = " + str(args.tf_batch_size))
print("Learning rate = " + str(args.learning_rate))
print("****************")
# Feature columns
COLUMNS = list(BLE_RSSI.columns)
FEATURES = COLUMNS[2:]
LABEL = [COLUMNS[0]]
b3001 = tf.feature_column.numeric_column(key='b3001', dtype=tf.float64)
b3002 = tf.feature_column.numeric_column(key='b3002', dtype=tf.float64)
b3003 = tf.feature_column.numeric_column(key='b3003', dtype=tf.float64)
b3004 = tf.feature_column.numeric_column(key='b3004', dtype=tf.float64)
b3005 = tf.feature_column.numeric_column(key='b3005', dtype=tf.float64)
b3006 = tf.feature_column.numeric_column(key='b3006', dtype=tf.float64)
b3007 = tf.feature_column.numeric_column(key='b3007', dtype=tf.float64)
b3008 = tf.feature_column.numeric_column(key='b3008', dtype=tf.float64)
b3009 = tf.feature_column.numeric_column(key='b3009', dtype=tf.float64)
b3010 = tf.feature_column.numeric_column(key='b3010', dtype=tf.float64)
b3011 = tf.feature_column.numeric_column(key='b3011', dtype=tf.float64)
b3012 = tf.feature_column.numeric_column(key='b3012', dtype=tf.float64)
b3013 = tf.feature_column.numeric_column(key='b3013', dtype=tf.float64)
feature_columns = [
b3001, b3002, b3003, b3004, b3005, b3006, b3007, b3008, b3009, b3010,
b3011, b3012, b3013
]
df_full = pd.read_csv("/opt/iBeacon_RSSI_Labeled.csv") #Labeled dataset
# Input Data Preprocessing
df_full = df_full.drop(['date'], axis=1)
df_full[FEATURES] = (df_full[FEATURES]) / (-200)
#Output Data Preprocessing
dict = {
'O02': 0,
'P01': 1,
'P02': 2,
'R01': 3,
'R02': 4,
'S01': 5,
'S02': 6,
'T01': 7,
'U02': 8,
'U01': 9,
'J03': 10,
'K03': 11,
'L03': 12,
'M03': 13,
'N03': 14,
'O03': 15,
'P03': 16,
'Q03': 17,
'R03': 18,
'S03': 19,
'T03': 20,
'U03': 21,
'U04': 22,
'T04': 23,
'S04': 24,
'R04': 25,
'Q04': 26,
'P04': 27,
'O04': 28,
'N04': 29,
'M04': 30,
'L04': 31,
'K04': 32,
'J04': 33,
'I04': 34,
'I05': 35,
'J05': 36,
'K05': 37,
'L05': 38,
'M05': 39,
'N05': 40,
'O05': 41,
'P05': 42,
'Q05': 43,
'R05': 44,
'S05': 45,
'T05': 46,
'U05': 47,
'S06': 48,
'R06': 49,
'Q06': 50,
'P06': 51,
'O06': 52,
'N06': 53,
'M06': 54,
'L06': 55,
'K06': 56,
'J06': 57,
'I06': 58,
'F08': 59,
'J02': 60,
'J07': 61,
'I07': 62,
'I10': 63,
'J10': 64,
'D15': 65,
'E15': 66,
'G15': 67,
'J15': 68,
'L15': 69,
'R15': 70,
'T15': 71,
'W15': 72,
'I08': 73,
'I03': 74,
'J08': 75,
'I01': 76,
'I02': 77,
'J01': 78,
'K01': 79,
'K02': 80,
'L01': 81,
'L02': 82,
'M01': 83,
'M02': 84,
'N01': 85,
'N02': 86,
'O01': 87,
'I09': 88,
'D14': 89,
'D13': 90,
'K07': 91,
'K08': 92,
'N15': 93,
'P15': 94,
'I15': 95,
'S15': 96,
'U15': 97,
'V15': 98,
'S07': 99,
'S08': 100,
'L09': 101,
'L08': 102,
'Q02': 103,
'Q01': 104
}
df_full['location'] = df_full['location'].map(dict)
df_train = df_full.sample(frac=0.8, random_state=200)
df_valid = df_full.drop(df_train.index)
location_counts = BLE_RSSI.location.value_counts()
x1 = np.asarray(df_train[FEATURES])
y1 = np.asarray(df_train['location'])
x2 = np.asarray(df_valid[FEATURES])
y2 = np.asarray(df_valid['location'])
def formatFeatures(features):
formattedFeatures = {}
numColumns = features.shape[1]
for i in range(0, numColumns):
formattedFeatures["b" + str(3001 + i)] = features[:, i]
return formattedFeatures
trainingFeatures = formatFeatures(x1)
trainingCategories = y1
testFeatures = formatFeatures(x2)
testCategories = y2
# Train Input Function
def train_input_fn():
dataset = tf.data.Dataset.from_tensor_slices((trainingFeatures, y1))
dataset = dataset.repeat(args.epochs).batch(args.tf_batch_size)
return dataset
# Test Input Function
def eval_input_fn():
dataset = tf.data.Dataset.from_tensor_slices((testFeatures, y2))
return dataset.repeat(args.epochs).batch(args.tf_batch_size)
# Provide list of GPUs should be used to train the model
distribution = tf.distribute.experimental.ParameterServerStrategy()
print('Number of devices: {}'.format(distribution.num_replicas_in_sync))
# Configuration of training model
config = tf.estimator.RunConfig(train_distribute=distribution,
model_dir=args.tf_model_dir,
save_summary_steps=100,
save_checkpoints_steps=100)
# Build 3 layer DNN classifier
model = tf.estimator.DNNClassifier(hidden_units=[13, 65, 110],
feature_columns=feature_columns,
optimizer=tf.train.AdamOptimizer(
learning_rate=args.learning_rate,
beta1=args.beta1,
beta2=args.beta2),
model_dir=args.tf_model_dir,
n_classes=105,
config=config)
export_final = tf.estimator.FinalExporter(
args.tf_export_dir,
serving_input_receiver_fn=serving_input_receiver_fn)
train_spec = tf.estimator.TrainSpec(input_fn=train_input_fn,
max_steps=args.tf_train_steps)
eval_spec = tf.estimator.EvalSpec(input_fn=eval_input_fn,
steps=100,
exporters=export_final,
throttle_secs=1,
start_delay_secs=1)
# Train and Evaluate the model
tf.estimator.train_and_evaluate(model, train_spec, eval_spec)
MODEL_EXPORT_PATH = args.tf_model_dir
def predict(request):
"""
Define custom predict function to be used by local prediction
and explainer. Set anchor_tabular predict function so it always returns predicted class
"""
# Get model exporter path
for dir in os.listdir(args.tf_model_dir):
if re.match('[0-9]', dir):
exported_path = os.path.join(args.tf_model_dir, dir)
break
else:
raise Exception("Model path not found")
# Prepare model input data
feature_cols = [
"b3001", "b3002", "b3003", "b3004", "b3005", "b3006", "b3007",
"b3008", "b3009", "b3010", "b3011", "b3012", "b3013"
]
input = {
'b3001': [],
'b3002': [],
'b3003': [],
'b3004': [],
'b3005': [],
'b3006': [],
'b3007': [],
'b3008': [],
'b3009': [],
'b3010': [],
'b3011': [],
'b3012': [],
'b3013': []
}
X = request
if np.ndim(X) != 2:
for i in range(len(X)):
input[feature_cols[i]].append(X[i])
else:
for i in range(len(X)):
for j in range(len(X[i])):
input[feature_cols[j]].append(X[i][j])
# Open a Session to predict
with tf.Session() as sess:
tf.saved_model.loader.load(sess,
[tf.saved_model.tag_constants.SERVING],
exported_path)
predictor = tf.contrib.predictor.from_saved_model(
exported_path, signature_def_key='predict')
output_dict = predictor(input)
sess.close()
output = {}
output["predictions"] = {
"probabilities": output_dict["probabilities"].tolist()
}
return np.asarray(output['predictions']["probabilities"])
#Initialize and fit
feature_cols = [
"b3001", "b3002", "b3003", "b3004", "b3005", "b3006", "b3007", "b3008",
"b3009", "b3010", "b3011", "b3012", "b3013"
]
explainer = AnchorTabular(predict, feature_cols)
explainer.fit(x1, disc_perc=(25, 50, 75))
#Save Explainer file
#Save explainer file with .dill extension. It will be used when creating the InferenceService
if not os.path.exists(args.explainer_dir):
os.mkdir(args.explainer_dir)
with open("%s/explainer.dill" % args.explainer_dir, 'wb') as f:
dill.dump(explainer, f)
