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.make_tensor_proto(data_arr.tolist(), shape=data_arr.shape))
result = self.stub.Predict(tfrequest)
datadef = prediction_pb2.DefaultData(
tftensor=result.outputs[self.model_output])
return prediction_pb2.SeldonMessage(data=datadef)
def test_proto_tftensor_to_array():
names = ["a", "b"]
array = np.array([[1, 2], [3, 4]])
datadef = prediction_pb2.DefaultData(names=names,
tftensor=tf.make_tensor_proto(array))
array2 = scu.grpc_datadef_to_array(datadef)
assert array.shape == array2.shape
assert np.array_equal(array, array2)
def _predict_grpc_data(self, request, default_data_type):
tfrequest = predict_pb2.PredictRequest()
# handle input
if default_data_type == "tftensor":
# For GRPC case, if we have a TFTensor message we can pass it directly
tfrequest.inputs[self.model_input].CopyFrom(request.data.tftensor)
else:
data_arr = grpc_datadef_to_array(request.data)
tfrequest.inputs[self.model_input].CopyFrom(
tf.make_tensor_proto(data_arr.tolist(), shape=data_arr.shape)
)
# handle prediction
tfresponse = self._handle_grpc_prediction(tfrequest)
# handle result
datadef = prediction_pb2.DefaultData(
tftensor=tfresponse.outputs[self.model_output]
)
return prediction_pb2.SeldonMessage(data=datadef)
