def extractRow(i: int, requestMsg: prediction_pb2.SeldonMessage,
req_datatype: str, req_features: np.ndarray,
req_datadef: "prediction_pb2.SeldonMessage.data"):
datatyReq = "ndarray"
dataType = "tabular"
if len(req_features.shape) == 2:
dataReq = array_to_grpc_datadef(
datatyReq, np.expand_dims(req_features[i], axis=0),
req_datadef.names)
else:
if len(req_features.shape) > 2:
dataType = "image"
else:
dataType = "text"
req_features = np.char.decode(req_features.astype('S'), "utf-8")
dataReq = array_to_grpc_datadef(
datatyReq, np.expand_dims(req_features[i], axis=0),
req_datadef.names)
requestMsg2 = prediction_pb2.SeldonMessage(data=dataReq,
meta=requestMsg.meta)
reqJson = {}
reqJson["payload"] = seldon_message_to_json(requestMsg2)
# setting dataType here temporarily so calling method will be able to access it
# don't want to set it at the payload level
reqJson["dataType"] = dataType
return reqJson
def test_proto_array_to_tftensor():
arr = np.array([[1, 2, 3], [4, 5, 6]])
datadef = scu.array_to_grpc_datadef("tftensor", arr, [])
print(datadef)
assert datadef.tftensor.tensor_shape.dim[0].size == 2
assert datadef.tftensor.tensor_shape.dim[1].size == 3
assert datadef.tftensor.dtype == 9
def predict_raw(self, request):
print("Predict raw")
default_data_type = request.data.WhichOneof("data_oneof")
print(default_data_type)
if default_data_type == "tftensor" and self.grpc:
tfrequest = predict_pb2.PredictRequest()
tfrequest.model_spec.name = self.model_name
tfrequest.model_spec.signature_name = self.signature_name
tfrequest.inputs[self.model_input].CopyFrom(request.data.tftensor)
result = self.stub.Predict(tfrequest)
print(result)
datadef = prediction_pb2.DefaultData(
tftensor=result.outputs[self.model_output])
return prediction_pb2.SeldonMessage(data=datadef)
elif default_data_type == "jsonData":
predictions = self.predict(request.jsonData, features_names=[])
return prediction_pb2.SeldonMessage(jsonData=predictions)
else:
features = get_data_from_proto(request)
datadef = request.data
data_type = request.WhichOneof("data_oneof")
predictions = self.predict(features, datadef.names)
predictions = np.array(predictions)
if data_type == "data":
default_data_type = request.data.WhichOneof("data_oneof")
else:
default_data_type = "tensor"
class_names = []
data = array_to_grpc_datadef(predictions, class_names,
default_data_type)
return prediction_pb2.SeldonMessage(data=data)
def extractRow(i:int,requestMsg: prediction_pb2.SeldonMessage,req_datatype: str,req_features: np.ndarray,req_datadef: "prediction_pb2.SeldonMessage.data"):
datatyReq = "ndarray"
dataType = "tabular"
if len(req_features.shape) == 2:
dataReq = array_to_grpc_datadef(datatyReq, np.expand_dims(req_features[i], axis=0), req_datadef.names)
else:
if len(req_features.shape) > 2:
dataType="image"
else:
dataType="text"
req_features= np.char.decode(req_features.astype('S'),"utf-8")
dataReq = array_to_grpc_datadef(datatyReq, req_features, req_datadef.names)
requestMsg2 = prediction_pb2.SeldonMessage(data=dataReq, meta=requestMsg.meta)
reqJson = seldon_message_to_json(requestMsg2)
reqJson["dataType"] = dataType
return reqJson
def test_create_grpc_reponse_ndarray():
user_model = UserObject()
request_data = np.array([[5, 6, 7]])
datadef = scu.array_to_grpc_datadef("ndarray", request_data)
request = prediction_pb2.SeldonMessage(data=datadef)
raw_response = np.array([[1, 2, 3]])
sm = scu.construct_response(user_model, True, request, raw_response)
assert sm.data.WhichOneof("data_oneof") == "ndarray"
def test_create_grpc_response_strdata():
user_model = UserObject()
request_data = np.array([[5, 6, 7]])
datadef = scu.array_to_grpc_datadef("ndarray", request_data)
request = prediction_pb2.SeldonMessage(data=datadef)
raw_response = "hello world"
sm = scu.construct_response(user_model, True, request, raw_response)
assert sm.data.WhichOneof("data_oneof") == None
assert len(sm.strData) > 0
def extractRow(i: int, requestMsg: prediction_pb2.SeldonMessage,
req_datatype: str, req_features: np.ndarray,
req_datadef: prediction_pb2.SeldonMessage.data):
if req_datatype == "data":
datatyReq = requestMsg.data.WhichOneof("data_oneof")
else:
datatyReq = "ndarray"
if len(req_features.shape) == 2:
dataReq = array_to_grpc_datadef(
datatyReq, np.expand_dims(req_features[i], axis=0),
req_datadef.names)
else:
dataReq = array_to_grpc_datadef(datatyReq, req_features,
req_datadef.names)
requestMsg2 = prediction_pb2.SeldonMessage(data=dataReq,
meta=requestMsg.meta)
reqJson = seldon_message_to_json(requestMsg2)
return reqJson
def test_create_grpc_reponse_binary():
user_model = UserObject()
request_data = np.array([[5, 6, 7]])
datadef = scu.array_to_grpc_datadef("tensor", request_data)
request = prediction_pb2.SeldonMessage(data=datadef)
raw_response = b"binary"
sm = scu.construct_response(user_model, True, request, raw_response)
assert sm.data.WhichOneof("data_oneof") == None
assert len(sm.strData) == 0
assert len(sm.binData) > 0
def test_create_grpc_response_jsondata():
user_model = UserObject()
request_data = np.array([[5, 6, 7]])
datadef = scu.array_to_grpc_datadef("ndarray", request_data)
request = prediction_pb2.SeldonMessage(data=datadef)
raw_response = {"output": "data"}
sm = scu.construct_response(user_model, True, request, raw_response)
assert sm.data.WhichOneof("data_oneof") == None
emptyValue = Value()
assert sm.jsonData != emptyValue
def test_create_grpc_response_customdata():
user_model = UserObject()
request_data = np.array([[5, 6, 7]])
datadef = scu.array_to_grpc_datadef("ndarray", request_data)
request = prediction_pb2.SeldonMessage(data=datadef)
raw_response = any_pb2.Any(value=b"testdata")
sm = scu.construct_response(user_model, True, request, raw_response)
assert sm.data.WhichOneof("data_oneof") is None
emptyValue = Value()
assert sm.customData != emptyValue
def test_create_grpc_reponse_text_ndarray():
user_model = UserObject()
request_data = np.array([["hello", "world"], ["hello", "another", "world"]])
datadef = scu.array_to_grpc_datadef("ndarray", request_data)
request = prediction_pb2.SeldonMessage(data=datadef)
(features, meta, datadef, data_type) = scu.extract_request_parts(request)
raw_response = np.array([["hello", "world"], ["here", "another"]])
sm = scu.construct_response(user_model, True, request, raw_response)
assert sm.data.WhichOneof("data_oneof") == "ndarray"
assert type(features[0]) == list
assert np.array_equal(sm.data.ndarray, raw_response)
assert datadef == request.data
assert np.array_equal(features, request_data)
assert data_type == "data"
def predict_grpc(self, request):
"""
predict_grpc will be called only when there is a GRPC call to the server
which in this case, the request will be sent to the TFServer directly.
"""
log.debug("Preprocessing contents for predict function")
request_data_type = request.WhichOneof("data_oneof")
default_data_type = request.data.WhichOneof("data_oneof")
log.debug(
f"Request data type: {request_data_type}, Default data type: {default_data_type}"
)
if request_data_type != "data":
raise Exception("strData, binData and jsonData not supported.")
tfrequest = predict_pb2.PredictRequest()
tfrequest.model_spec.name = self.model_name
tfrequest.model_spec.signature_name = self.signature_name
# For GRPC case, if we have a TFTensor message we can pass it directly
if default_data_type == "tftensor":
tfrequest.inputs[self.model_input].CopyFrom(request.data.tftensor)
result = self.stub.Predict(tfrequest)
datadef = prediction_pb2.DefaultData(
tftensor=result.outputs[self.model_output])
return prediction_pb2.SeldonMessage(data=datadef)
else:
data_arr = grpc_datadef_to_array(request.data)
tfrequest.inputs[self.model_input].CopyFrom(
tf.contrib.util.make_tensor_proto(data_arr.tolist(),
shape=data_arr.shape))
result = self.stub.Predict(tfrequest)
result_arr = numpy.array(
result.outputs[self.model_output].float_val)
if len(result_arr.shape) == 1:
result_arr = numpy.expand_dims(result_arr, axis=0)
class_names = []
data = array_to_grpc_datadef(default_data_type, result_arr,
class_names)
return prediction_pb2.SeldonMessage(data=data)
def predict_raw(self, request):
log.debug("Predict raw")
request_data_type = request.WhichOneof("data_oneof")
default_data_type = request.data.WhichOneof("data_oneof")
log.debug(str(request_data_type), str(default_data_type))
if default_data_type == "tftensor" and self.grpc:
tfrequest = predict_pb2.PredictRequest()
tfrequest.model_spec.name = self.model_name
tfrequest.model_spec.signature_name = self.signature_name
tfrequest.inputs[self.model_input].CopyFrom(request.data.tftensor)
result = self.stub.Predict(tfrequest)
log.debug(result)
datadef = prediction_pb2.DefaultData(
tftensor=result.outputs[self.model_output])
return prediction_pb2.SeldonMessage(data=datadef)
elif request_data_type == "jsonData":
features = get_data_from_proto(request)
predictions = self.predict(features, features_names=[])
try:
sm = json_to_seldon_message({"jsonData": predictions})
except ParseError as e:
sm = prediction_pb2.SeldonMessage(strData=predictions)
return sm
else:
features = get_data_from_proto(request)
datadef = request.data
predictions = self.predict(features, datadef.names)
predictions = np.array(predictions)
if request_data_type is not "data":
default_data_type = "tensor"
class_names = []
data = array_to_grpc_datadef(predictions, class_names,
default_data_type)
return prediction_pb2.SeldonMessage(data=data)
def mocked_requests_post_success(url, *args, **kwargs):
data = np.random.rand(1, 1)
datadef = array_to_grpc_datadef("tensor", data)
request = prediction_pb2.SeldonMessage(data=datadef)
json = seldon_message_to_json(request)
return MockResponse(json, 200, text="{}")
