[Bug] Checker misses data type mismatch for Max

[ganler]

Bug Report

Is the issue related to model conversion?

Partially. PyTorch generates a wrong ONNX model but still onnx.checker did not check it carefully (I use onnx.checker.check_model(mode, full_check=True) to skip wrong models).

Describe the bug

The minimized model looks like this:

"""
graph par1 (
  %i0[DOUBLE, 2x2x1x1]
  %/mlist.0/Mul_1_output_0[FLOAT, scalar]
) {
  %/Max_output_0 = Max(%i0, %/mlist.0/Mul_1_output_0)
  return %/Max_output_0
}
"""

Basically, it is Max(<double>[2,2,1,1], <float>[scalar]). The checker returns successfully when checking the model. Maybe there is some missed cases/conditions in the checker.

System information

• ONNX version: 1.12
• Python version: 3.9
• OS: Ubuntu 22.04

Reproduction instructions

import onnx

X = onnx.helper.make_tensor_value_info("X", onnx.TensorProto.DOUBLE, (2, 2, 1, 1))
Y = onnx.helper.make_tensor_value_info("Y", onnx.TensorProto.FLOAT, ())
Z = onnx.helper.make_tensor_value_info("Z", onnx.TensorProto.DOUBLE, (2, 2, 1, 1))
model = onnx.helper.make_model(
    onnx.helper.make_graph(
        nodes=[
            onnx.helper.make_node("Max", ["X", "Y"], ["Z"]),
        ],
        name="bar",
        inputs=[X, Y],
        outputs=[Z],
        initializer=[],
        value_info=[],
    ),
    producer_name="foo",
)

model = onnx.shape_inference.infer_shapes(model, strict_mode=True)
onnx.checker.check_model(model, full_check=True)

Expected behavior

Should fail but did not. But later this error occurs when using ONNXRuntime.

cc: @jcwchen Thanks!

[jcwchen]

Hi @ganler,
Thank you for catching this. I can repro this by your example. Yes, I think ideally checker should catch this type mismatch. I will submit a PR to fix this.