support multiple named tensors in seldon protocol and seldon-core client

[chengcheng-pei]

I have data like:
[{
e_input: ...,
e_time_input: ...,
r_input: ...
}]

I can use tensorflow protocol and send the requests like:

channel = grpc.insecure_channel('{host}:{port}'.format(host="tf-grpc.stage.mycompany.net", port=80))
stub = prediction_service_pb2_grpc.PredictionServiceStub(channel)
tfrequest = predict_pb2.PredictRequest()
tfrequest.model_spec.name = 'classifier'
ind = 0
tfrequest.inputs["e_input"].CopyFrom(make_tensor_proto(d2[ind]["e_input"], dtype=types_pb2.DT_FLOAT, shape=[1, 10]))
tfrequest.inputs["e_time_input"].CopyFrom(make_tensor_proto(d2[ind]["e_time_input"], dtype=types_pb2.DT_FLOAT, shape=[1, 100]))
tfrequest.inputs["r_input"].CopyFrom(make_tensor_proto(d2[ind]["r_input"], dtype=types_pb2.DT_FLOAT, shape=[1, 58]))
result = stub.Predict(tfrequest, metadata=(('seldon', 'tf-grpc'), ('namespace', 'team-xxxx')))

But how to send the above request in seldon-core client and seldon protocol?

[ukclivecox]

The Seldon protocol needs update to make multi-headed models easier to use. You can send arbitrary tensors if you are willing to split them yourself into individual inputs. You can also use binData and jsonData to send what you want. The current protocol will not work with the Tensorflow proxy.

Your best best at present is to use the Tensorflow protocol for these models. In future we will be introducing a new V2 protocol shared amongst many projects which will also explicitly allow multiple inputs and outputs.