How to handle deprecated operators? about sigs

[winnietsang]

Handle the op with the previous definition and print out a warning message state that this op is deprecated at this opset version and advise user to use the "new" op to replace this in the future.

[chinhuang007]

The recommendation is also that the frontend converters do not add deprecated ops, with respect to the target opset version, into onnx graph. Use the new/replacement op, if existing, instead.

[ebarsoum]

Regarding this "Use the new/replacement op, if existing, instead.", replacement must exist or the deprecation was a bug. We can't deprecate operator without an equivalent.

[chinhuang007]

Good to know it is a mandate to have replacement when deprecating an op. Will document accordingly.

[winnietsang]

After I ran some test, I found out backend converter don't need to handle any deprecated op at all.

If an user want to convert a deprecated op to a framework, the user will get an error message from the onnx.checker.check_node(node).
onnx.onnx_cpp2py_export.checker.ValidationError: Op registered for Upsample is deprecated in domain_version of 11

Therefore, this node will never get pass to any backend converter.
On the other hand the frontend converter need to make sure it doesn't use any deprecated op in the converted onnx model

[chinhuang007]

I would recommend all backend converters run checker before conversion, as it is in the onnx base class. If not the case, the converter should handle deprecated ops as described earlier in the thread.