Can I use ContribOperators in my onnx helper for new nodes?

[maccam912]

I'm in a hackathon-like event and ONNX is new to me, so bear with me.

I am trying to use a contrib op, in this example com.microsoft.MurmurHash3, in a tiny custom graph I'm building with the onnx helper in python. Code snippet below:

X = helper.make_tensor_value_info('X', TensorProto.STRING, [1])
Y = helper.make_tensor_value_info('Y', TensorProto.UINT32, [1])

node_def = helper.make_node(
    'com.microsoft.MurmurHash3', # node name
    ['X'], # inputs
    ['Y'], # outputs
)

graph_def = helper.make_graph(
    [node_def],
    "test-model",
    [X],
    [Y],
)

model_def = helper.make_model(graph_def,
                              producer_name='onnx-example')

#onnx.checker.check_model(model_def)

onnx.save(model_def, "test_model.onnx")

1. My first question is that the second-to-last line, commented out check_model, fails because the node name 'com.microsoft.MurmurHash3' is not registered. Is there a way to register it so the check_model works?

2. When trying to load and run the model later on with onnxruntime, I get [ONNXRuntimeError] : 10 : INVALID_GRAPH : Load model from test_model.onnx failed:This is an invalid model. Error in Node: : No Op registered for com.microsoft.MurmurHash3 with domain_version of 13 so it seems I need to tell onnxruntime somehow that I am using contrib ops? The code to load and run the model is:

sess = rt.InferenceSession("test_model.onnx")
input_names = [x.name for x in sess.get_inputs()]
pred_onx = sess.run(None, {"X": ["Hello World"]})
print(pred_onx)

Thanks for any help anyone can provide here.

[maccam912]

After some more trial and error I think I need to set the register_custom_ops_library(...) for question 2. My new code looks like

so = rt.SessionOptions()
so.register_custom_ops_library("")
sess = rt.InferenceSession("test_model.onnx", sess_options=so)
...

and I'm working on trying to find the contrib ops library, if I have it already somewhere.

[maccam912]

Closing my own question with an answer: The node name is just "MurmurHash3" or whatever op you want, but add in the domain argument and set it to com.microsoft in this case. No need to register a custom ops library any more, it worked first try!

node_def = helper.make_node(
    'MurmurHash3', # node name
    ['X'], # inputs
    ['Y'], # outputs
    domain="com.microsoft",
)