[frontend][onnx]fix pad constant value is none

[chengven027]

hi, cc @AndrewZhaoLuo
When I import my onnx model. I found my model get a error with pad. because the pad constant value is none. shown as follow:

It`s just a very little bug. So I fix it and no unnit test.

[AndrewZhaoLuo]

Reading the docs https://github.com/onnx/onnx/blob/main/docs/Operators.md#Pad we should also be checking if input[2] is also an empty string or False.

Pythonically we can just do and input[2] (though I would add a comment since this handles multiple types)

[leandron]

Hi @chengven027-intellif, thanks for the PR. Any chance we can have a test case for this change?

[chengven027]

Dose this PR need to be reviewed again? @AndrewZhaoLuo @leandron

[leandron]

Is setting False here (as opposed to None) reproducing the same issue this patch is fixing?

[chengven027]

No, But I can't add a test case with None.
First, In the doc test_forward.py test_pad, The value is a default number 0.0,can't set None.
def verify_pad_v11(indata, pads, mode="constant", value=0.0)
Then, the input node constant_value is a TensorProto.FLOAT. not None too.
Can you give me some suggestions to add a test case with 'None'? Thanks very much.

[AndrewZhaoLuo]

Dose this PR need to be reviewed again? @AndrewZhaoLuo @leandron

Yeah I'll leave it to Leandron B)

[chengven027]

I am sorry, I really can't create an unit test with empty string of constant_value.
@leandron @AndrewZhaoLuo

[AndrewZhaoLuo]

I am sorry, I really can't create an unit test with empty string of constant_value. @leandron @AndrewZhaoLuo

In verify_pad_v11 I can seem to make the following:

          graph = helper.make_graph(
                [node],
                "pad_test",
                inputs=[
                    helper.make_tensor_value_info("input", TensorProto.FLOAT, list(indata.shape)),
                    helper.make_tensor_value_info("pads", TensorProto.INT64, (len(pads),)),
                    helper.make_tensor_value_info("constant_value", TensorProto.STRING, (1,)),
                ],
                initializer=[
                    helper.make_tensor("pads", TensorProto.INT64, (len(pads),), pads),
                    helper.make_tensor("constant_value", TensorProto.STRING, (1,), [b""]),
                ],
                outputs=[
                    helper.make_tensor_value_info("output", TensorProto.FLOAT, list(outdata.shape))
                ],
            )

However, my version of ONNXRT can't seem to accept the string type. What have you tried?

[chengven027]

@AndrewZhaoLuo Yes, I tried many times but I got the same error as yours.
onnxruntime.capi.onnxruntime_pybind11_state.InvalidGraph: [ONNXRuntimeError] : 10 : INVALID_GRAPH : This is an invalid model. Type Error: Type 'tensor(string)' of input parameter (constant_value) of operator (Pad) in node () is invalid.

[AndrewZhaoLuo]

Can you share your model by any chance? Can you also run your model via ONNX-RT?

[chengven027]

Can you share your model by any chance? Can you also run your model via ONNX-RT?

coat_model.zip

1. model have been uploaded.
2. The model can pass by ONNX-RT. And the floating point comparison between tvm and onnx-RT result can also pass.

[AndrewZhaoLuo]

Sorry, I'll take a look tomorrow.

Edit: On Friday

[chengven027]

Hey, @AndrewZhaoLuo Have you seen this PR, and any update with me? thanks.

[AndrewZhaoLuo]

Oops, I am so sorry, I have been very busy lately. I will try to get to this this weekend