Jakub Bachurski

@gramalingam Would you mind viewing the proposal if this sort of interface would work? In the future it could be extended to support the other arguments, they could be optional...

> I think it might be useful to promote an API like this as a wrapper around the raw C++ method (that is, make if part of the public API...

@gramalingam I pushed some changes: ```cpp std::unordered_map CallNodeInferenceFunction( OpSchema* schema, const py::bytes& nodeBytes, std::unordered_map valueTypesByNameBytes, std::unordered_map inputDataByNameBytes, std::unordered_map inputSparseDataByNameBytes) { ``` - Now returns a map, since the inputs are...

Thanks! Should be ready for merge, then?

Are you sure this fully fixes the issue? It still types `AttributeProto` as `Any` for me: ```py import onnx reveal_type(onnx.helper.AttributeProto) ``` ``` $ mypy test.py test.py:2: note: Revealed type is...

Hi, I think the issue might be indeed related to different architectures/Python versions. But I think I found a fix that also properly lists the attributes of the proto object...

All good!

Please see the draft PR for a proposal - it's not finished yet, as I saw some complications. - Currently I'm passing in arguments to construct an `InferenceContextImpl` from Python....

I think a lot of arguments have already been said here and I don't want to reiterate. Even though shape inference is 'best-effort', I think ONNX should strive to make...

My personal hunch says it should be more general, because as @negiyas says the implementation is the same. Avoiding an explicit reshape might be more efficient. Based on this [resource](https://github.com/onnx/onnx/blob/main/docs/AddNewOp.md#step-1-proposing-a-new-operatorfunction-)...