E2E tutorial/example to show how tf2onnx convert model with custom op

[EmmaNingMS]

Do we have E2E tutorial/example to show how tf2onnx convert model with custom op? Similar to this example from pytorch exporter.

[hwangdeyu]

Now we can support the custom ops by parameter --custom-ops. If a model contains ops not recognized by onnx runtime, it can tag these ops with a custom op domain so that the runtime can still open the model.

Consider to add a new tutorial doc file.

[guschmue]

--custom-ops is supported since years - it lets you resolve the unknown op in python, ie. you can compose it with opther ops. There is a section for custom ops in the readme and it says there is an example here: https://github.com/onnx/tensorflow-onnx/blob/master/examples/custom_op_via_python.py

By default, tf2onnx will leave unknown ops in the graph with the same signature, aka it just prints a warning but still produces the onnx file so one just need to add that op to onnxruntime with the same signature that the tensorflow op has.

[EmmaNingMS]

Thanks for the info. The existing example helps a little. But would love to see an E2E tutorial covering all the parts from tf2onnx to ORT, similar to this Pytorch exporter example