Pass "eliminate_consecutive_idempotent_ops" succeeds but model gives an incompatible shape error upon run.

While optimizing a number of Mask and Faster R-CNN models on opset=12 from the ONNX Model Hub, the optimizer generated the models, but I encountered the following error when applying the eliminate_consecutive_idempotent_ops error when I run them:

[ONNXRuntimeError] : 1 : FAIL : Node (<NODE>) Op (Reshape) [ShapeInferenceError] Dimension could not be inferred: incompatible shapes

Where is:
2953 for the Mask models: Mask R-CNN R-50-FPN-int8, Mask R-CNN R-50-FPN-qdq, Mask R-CNN R-50-FPN-fp32.
2794 for the Faster models: Faster R-CNN R-50-FPN-int8, Faster R-CNN R-50-FPN-qdq, and Faster R-CNN R-50-FPN-fp32.

The original models run without problems. By visual inspection using Netron, I believe something is mishandled with the concatenation operations preceding these nodes in the models.

Reproduce the failure by applying eliminate_consecutive_idempotent_ops to the listed Mask R-CNN and Faster R-CNN R-50-FPN models at opset 12, then run them with ONNX Runtime. Inspect nodes 2953 and 2794 and the preceding concatenation operations, using Netron as suggested. Done means the optimized models run without the incompatible-shapes Reshape error.