[Disscussion] Would mechanism like onnx "register_schema" be helpful in optimizing models with user-defined operators?

[SStarver]

Recently, I have discovered that the ONNX project has introduced the "onnx.defs.register_schema(op_schema)" interface, enabling support for shape inference in models that utilize user-defined operators. This addition is not only beneficial but crucial in preventing errors that arise during the model verification process, especially when dealing with customized operators.

Given this, I am intrigued to explore whether this functionality could aid in the optimization of models that contain customized operations. Currently, attempts to optimize models with such operators often result in failures, often due to unrecognized operators.

• case without register_schema
  With "onnx.defs.register_schema(op_schema)" unnoted, following errors would be gone. Futhermore, shape_inference would also gone futher.