BatchNormalization layer with training_mode=1 results in incorrect ONNX model

[bas-aarts]

converting the attached ONNX model with the following script:

from onnxscript import ir
from onnxscript.optimizer import optimize

ir_model = ir.load("input.onnx")
optimized_model = optimize(ir_model)
ir.save(optimized_model, "output.onnx")

generates an ONNX model that does not abide by the ONNX spec.

BatchNormalization layer "/bn1/BatchNormalization" has training_mode=1 set. It has 3 outputs, but the last two are not used. Because they are unused, https://github.com/onnx/ir-py/blob/v0.2.1/src/onnx_ir/passes/common/unused_removal.py#L46-L50 clears the output names and training_mode attribute. While I don't like the notion of unnamed outputs to mean unused, it is an accepted way in ONNX. Later on during optimization, at https://github.com/microsoft/onnxscript/blob/v0.7.0/onnxscript/optimizer/_optimizer.py#L68, new names get reassigned to these outputs, resulting in the following graph snippet:

This graph now has a BatchNormaliztion node, with training_mode=0, with 3 outputs. This is incorrect per spec: "When training_mode=False, extra outputs are invalid." in https://onnx.ai/onnx/operators/onnx__BatchNormalization.html#batchnormalization-15.

I filed this bug with onnxscript, even though the name-clearing happens in onnx-ir.

relevant versions:

onnx                          1.21.0
onnx-ir                       0.2.1
onnxscript                    0.7.1

input.onnx.zip

[bas-aarts]

FWIW, I'm experimenting locally by replacing
https://github.com/onnx/ir py/blob/v0.2.1/src/onnx_ir/passes/common/unused_removal.py#L46-L49 with
node.resize_outputs(1)
It's behaving well for now, But I'm not sure it is actually correct.