def __init__(self,
name: str,
predictor_host: str,
method: ExplainerMethod,
config: Mapping,
explainer: object = None,
protocol: Protocol = Protocol.seldon_grpc,
tf_data_type: str = None,
keras_model: keras.Model = None ):
super().__init__(name)
self.predictor_host = predictor_host
logging.info("Predict URL set to %s", self.predictor_host)
self.method = method
self.protocol = protocol
self.tf_data_type = tf_data_type
logging.info("Protocol is %s",str(self.protocol))
# Add type for first value to help pass mypy type checks
if self.method is ExplainerMethod.anchor_tabular:
self.wrapper:ExplainerWrapper = AnchorTabular(self._predict_fn, explainer, **config)
elif self.method is ExplainerMethod.anchor_images:
self.wrapper = AnchorImages(self._predict_fn, explainer, **config)
elif self.method is ExplainerMethod.anchor_text:
self.wrapper = AnchorText(self._predict_fn, explainer, **config)
elif self.method is ExplainerMethod.kernel_shap:
self.wrapper = KernelShap(self._predict_fn, explainer, **config)
elif self.method is ExplainerMethod.integrated_gradients:
self.wrapper = IntegratedGradients(keras_model, **config)
else:
raise NotImplementedError
def __init__(
self,
name: str,
predict_fn: Callable,
method: ExplainerMethod,
config: Mapping,
explainer: object = None,
protocol: Protocol = Protocol.seldon_grpc,
keras_model: keras.Model = None,
) -> None:
super().__init__(name)
self.method = method
self.protocol = protocol
logging.info("Protocol is %s", str(self.protocol))
# Add type for first value to help pass mypy type checks
if self.method is ExplainerMethod.anchor_tabular:
self.wrapper: ExplainerWrapper = AnchorTabular(explainer, **config)
elif self.method is ExplainerMethod.anchor_images:
self.wrapper = AnchorImages(explainer, **config)
elif self.method is ExplainerMethod.anchor_text:
self.wrapper = AnchorText(predict_fn, explainer, **config)
elif self.method is ExplainerMethod.kernel_shap:
self.wrapper = KernelShap(explainer, **config)
elif self.method is ExplainerMethod.integrated_gradients:
self.wrapper = IntegratedGradients(keras_model, **config)
elif self.method is ExplainerMethod.tree_shap:
self.wrapper = TreeShap(explainer, **config)
elif self.method is ExplainerMethod.ale:
self.wrapper = ALE(explainer, **config)
else:
raise NotImplementedError
def test_integrated_gradients():
with tempfile.TemporaryDirectory() as model_dir:
download_from_gs(IMDB_KERAS_MODEL_URI, model_dir)
keras_model_path = os.path.join(model_dir, KERAS_MODEL_FILENAME)
keras_model = keras.models.load_model(keras_model_path)
integrated_gradients = IntegratedGradients(keras_model, layer=1)
max_features = 10000
maxlen = 100
(x_train, y_train), (x_test,
y_test) = imdb.load_data(num_words=max_features)
x_train = sequence.pad_sequences(x_train, maxlen=maxlen)
x_test = sequence.pad_sequences(x_test, maxlen=maxlen)
explanation = integrated_gradients.explain(x_test[0:1].tolist())
attrs = explanation["attributions"]
assert len(attrs) > 0
def test_integrated_gradients():
keras_model_path = os.path.join(
kfserving.Storage.download(IMDB_KERAS_MODEL_URI), KERAS_MODEL_FILENAME)
keras_model = keras.models.load_model(keras_model_path)
integrated_gradients = IntegratedGradients(keras_model, layer=1)
max_features = 10000
maxlen = 100
(x_train, y_train), (x_test,
y_test) = imdb.load_data(num_words=max_features)
print(len(x_train), 'train sequences')
print(len(x_test), 'test sequences')
print('Pad sequences (samples x time)')
x_train = sequence.pad_sequences(x_train, maxlen=maxlen)
x_test = sequence.pad_sequences(x_test, maxlen=maxlen)
print('x_train shape:', x_train.shape)
print('x_test shape:', x_test.shape)
explanation = integrated_gradients.explain(x_test[0:1].tolist())
attrs = explanation["attributions"]
assert len(attrs) > 0
print(explanation)
class AlibiExplainer(ExplainerModel):
def __init__(self,
name: str,
predictor_host: str,
method: ExplainerMethod,
config: Mapping,
explainer: object = None,
protocol: Protocol = Protocol.seldon_grpc,
tf_data_type: str = None,
keras_model: keras.Model = None ):
super().__init__(name)
self.predictor_host = predictor_host
logging.info("Predict URL set to %s", self.predictor_host)
self.method = method
self.protocol = protocol
self.tf_data_type = tf_data_type
logging.info("Protocol is %s",str(self.protocol))
# Add type for first value to help pass mypy type checks
if self.method is ExplainerMethod.anchor_tabular:
self.wrapper:ExplainerWrapper = AnchorTabular(self._predict_fn, explainer, **config)
elif self.method is ExplainerMethod.anchor_images:
self.wrapper = AnchorImages(self._predict_fn, explainer, **config)
elif self.method is ExplainerMethod.anchor_text:
self.wrapper = AnchorText(self._predict_fn, explainer, **config)
elif self.method is ExplainerMethod.kernel_shap:
self.wrapper = KernelShap(self._predict_fn, explainer, **config)
elif self.method is ExplainerMethod.integrated_gradients:
self.wrapper = IntegratedGradients(keras_model, **config)
else:
raise NotImplementedError
def load(self):
pass
def _predict_fn(self, arr: Union[np.ndarray, List]) -> np.ndarray:
print(arr)
if type(arr) == list:
arr = np.array(arr)
if self.protocol == Protocol.seldon_grpc:
return self._grpc(arr)
elif self.protocol == Protocol.seldon_http:
payload = seldon.create_request(arr, seldon.SeldonPayload.NDARRAY)
response_raw = requests.post(SELDON_PREDICTOR_URL_FORMAT.format(self.predictor_host), json=payload)
if response_raw.status_code == 200:
rh = seldon.SeldonRequestHandler(response_raw.json())
response_list = rh.extract_request()
return np.array(response_list)
else:
raise Exception(
"Failed to get response from model return_code:%d" % response_raw.status_code)
elif self.protocol == Protocol.tensorflow_http:
instances = []
for req_data in arr:
if isinstance(req_data, np.ndarray):
instances.append(req_data.tolist())
else:
instances.append(req_data)
request = {"instances": instances}
response = requests.post(
KFSERVING_PREDICTOR_URL_FORMAT.format(self.predictor_host, self.name),
json.dumps(request)
)
if response.status_code != 200:
raise Exception(
"Failed to get response from model return_code:%d" % response.status_code)
return np.array(response.json()["predictions"])
def explain(self, request: Dict) -> Any:
if self.method is ExplainerMethod.anchor_tabular or self.method is ExplainerMethod.anchor_images or \
self.method is ExplainerMethod.anchor_text or self.method is ExplainerMethod.kernel_shap or \
self.method is ExplainerMethod.integrated_gradients:
if self.protocol == Protocol.tensorflow_http:
explanation: Explanation = self.wrapper.explain(request["instances"])
else:
rh = seldon.SeldonRequestHandler(request)
response_list = rh.extract_request()
explanation = self.wrapper.explain(response_list)
explanationAsJsonStr = explanation.to_json()
logging.info("Explanation: %s", explanationAsJsonStr)
return json.loads(explanationAsJsonStr)
else:
raise NotImplementedError
def _grpc(self, arr: np.array) -> np.array:
options = [
('grpc.max_send_message_length', GRPC_MAX_MSG_LEN),
('grpc.max_receive_message_length', GRPC_MAX_MSG_LEN)]
channel = grpc.insecure_channel(self.predictor_host, options)
stub = prediction_pb2_grpc.SeldonStub(channel)
if self.tf_data_type is not None:
datadef = prediction_pb2.DefaultData(
tftensor=tf.make_tensor_proto(arr, self.tf_data_type))
else:
datadef = prediction_pb2.DefaultData(
tftensor=tf.make_tensor_proto(arr))
request = prediction_pb2.SeldonMessage(data=datadef)
response = stub.Predict(request=request)
arr_resp = tf.make_ndarray(response.data.tftensor)
return arr_resp