[Shape Inference] Robustness of ConstantOfShape operator

[shaoyuyoung]

Bug Report

Description

The following model is a ConstantOfShape op with a constant input [-10].
Of course, ONNX does not allow ConstantOfShape input to be negative (https://onnx.ai/onnx/operators/onnx__ConstantOfShape.html#inputs).
However, in this case, this model can pass the check, it outputs an invalid dim: -10.
I think it should output at least be unknown :(

Script

import onnx
from onnx import helper, TensorProto, numpy_helper
import numpy as np
import onnxruntime as ort


opset_imports = [helper.make_opsetid("", 20)]
graph = helper.make_graph(
    nodes=[
        helper.make_node(
            "ConstantOfShape",
            inputs=['std.constant'],
            outputs=['output'],
            name='invalid_node'
        )
    ],
    name='ConstantOfShapeGraph',
    inputs=[],
    outputs=[helper.make_tensor_value_info('output', TensorProto.FLOAT, [None])],
    initializer=[numpy_helper.from_array(np.array([-10], dtype=np.int64), name='std.constant')]
)

model = helper.make_model(graph, opset_imports=opset_imports, producer_name='pytorch', ir_version=9)
onnx.checker.check_model(model, full_check=True)
model_shapes = onnx.shape_inference.infer_shapes(model)
onnx.checker.check_model(model_shapes, full_check=True)
print(model_shapes.graph.output)
onnx.save(model_shapes,"model_shapes.onnx")

output

[name: "output"
type {
  tensor_type {
    elem_type: 1
    shape {
      dim {
        dim_value: -10
      }
    }
  }
}
]

Acatually, in some cases, this output will cause all shape inferences of downstream operators to fail

[justinchuby]

Thank you! We are happy to take a pull request.

[shaoyuyoung]

cc @jcwchen .

Chun-Wei, can you take a look at this?

It's supposed to be something you're good at