Add used opset into opset_imports argument to create an onnx model.

[fatcat-z]

In latest ONNX version, there is an enhancement that if the used domain of a node doesn't exist in the model.opset_import, it will result in an exception and stop processing.

To align with this enhancement, we add current used domain and opset version into model.opset_import property by appending them into "opset_imports" argument in make_model() method.

Signed-off-by: Jay Zhang jiz@microsoft.com

[lgtm-com]

This pull request introduces 1 alert when merging 20850a6 into 872a96b - view on LGTM.com

new alerts:

• 1 for Unused import

[guschmue]

why add onnx.ml as domain if no op in the graph uses it ? It would be super rare for tf2onnx to use a onnx.ml op

[fatcat-z]

why add onnx.ml as domain if no op in the graph uses it ? It would be super rare for tf2onnx to use a onnx.ml op

The op LookupTableFindV2 needs this domain.

[guschmue]

yes, only add the domain when we use onnx.ml. The idea behind the domains is that you tell the runtime what is needed to run the model (ie. one can load the support for domains on demand).

[fatcat-z]

yes, only add the domain when we use onnx.ml. The idea behind the domains is that you tell the runtime what is needed to run the model (ie. one can load the support for domains on demand).

Adding this built-in domain in model.opset_import field is to align with a latest check enhancement of onnx. It won't impact the way how runtime runs the model because the runtime will only consider the domain of each op node.
For the pair of domain and version in model.opset_import field, runtime will only check if they are supported.