Get tensor value: xxx must be Const

[Exlsunshine]

Describe the bug

I have noticed that xxx must be const is a known issue, but I also found a PR which says ONNX has supported dynamic padding, but I still got the following error:

OP=Pad
Name=some/path/to/op/Pad
Inputs:
        some/path/to/op/concat:0=Concat, [-1, -1], 1
        some/path/to/op/Pad/paddings_Concat__113:0=Concat, [2, 2], 6
Outpus:
        some/path/to/op/Pad:0=[-1, -1], 1
Traceback (most recent call last):
  File "D:\Python\Python36\lib\site-packages\tf2onnx\tfonnx.py", line 352, in tensorflow_onnx_mapping
    func(g, node, **kwargs)
  File "D:\Python\Python36\lib\site-packages\tf2onnx\onnx_opset\nn.py", line 401, in version_1
    paddings = np.array(node.inputs[1].get_tensor_value()).transpose().flatten()
  File "D:\Python\Python36\lib\site-packages\tf2onnx\graph.py", line 256, in get_tensor_value
    raise ValueError("get tensor value: {} must be Const".format(self.name))
ValueError: get tensor value: some/path/to/op/Pad/paddings_Concat__113 must be Const
2019-11-21 15:52:39,788 - VERBOSE - tf2onnx.tfonnx: Mapping TF node to ONNX node(s)
2019-11-21 15:52:40,098 - ERROR - tf2onnx.tfonnx: Failed to convert node some/path/to/op/Fill

Does tf2onnx plan to support dynamic padding?

System information

• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Win10
• Tensorflow Version: 1.13.1
• Python version: 3.6.6
• ONNX: 1.6.0
• ONNXRUNTIME: 1.1.10
• tf2onnx: 1.5.3

[guschmue]

try --opset 11 in the command line.
We have not pushed this to pypi yet so you'd need to user master.

[Exlsunshine]

try --opset 11 in the command line.
We have not pushed this to pypi yet so you'd need to user master.

Hi @guschmue , I have tried opset 11 and install tf2onnx from master, but still got the same error. And I forget to mention that there's another op keeps telling me the same error, I always get the same result regardless of using opset 11/10/9/8/7 or installing tf2onnx from pypi/master:

2019-11-22 10:02:17,020 - VERBOSE - tf2onnx.tfonnx: Mapping TF node to ONNX node(s)
2019-11-22 10:02:17,195 - ERROR - tf2onnx.tfonnx: Failed to convert node parallel_0/while/Fill
OP=ConstantOfShape
Name=parallel_0/while/Fill
Inputs:
        parallel_0/while/Fill__674:0=Cast, [1], 7
        parallel_0/while/strided_slice_21__666:0=Squeeze, [], 6
Outpus:
        parallel_0/while/Fill:0=[-1], 6
Traceback (most recent call last):
  File "tensorflow-onnx\tf2onnx\tfonnx.py", line 354, in tensorflow_onnx_mapping
    func(g, node, **kwargs)
  File "tensorflow-onnx\tf2onnx\onnx_opset\generator.py", line 100, in version_9
    value = np.array([node.inputs[1].get_tensor_value()]).astype(utils.map_onnx_to_numpy_type(dtype))
  File "tensorflow-onnx\tf2onnx\graph.py", line 260, in get_tensor_value
    raise ValueError("get tensor value: {} must be Const".format(self.name))
ValueError: get tensor value: parallel_0/while/strided_slice_21__666 must be Const

[jignparm]

The error after --opset 11 is slightly different.

Could you attach the model you are converting (to ensure the solution works for your scenario)? The Fill operator should be refactor-able to work with dynamic inputs.

[jignparm]

@Exlsunshine, thanks for sharing the model offline.

The Fill operator is updated by #748, and the Expand_Dims operator is addressed by #753

The other fix required is not in TF2ONNX, but in OnnxRuntime, for the Reshape operator. This requires a change in the ONNX spec to make Reshape behave similarly to NumPy and TensorFlow. The current spec has an idiosyncrasy which causes the wrong shape to be produced (e.g. if a tensor of shape [0,1] is reshaped to [1,0], it will end up as [1,1] instead, which is not intuitive/correct).

The ONNX issue is raised here onnx/onnx#2507. Once that is approved by the community, the required code changes will be made in ONNXRuntime via PR here microsoft/onnxruntime#2656

[guschmue]

I assume this is fixed with opset-13.