Error converting ONNX to TF file: Cannot take the length of shape with unknown rank.

Question

Error converting ONNX to TF file: Cannot take the length of shape with unknown rank.

ferqui opened this issue 2 years ago · 5 comments

Describe the bug

Hi, I want to convert an onnx model to Tensorflow, but when I try to convert it, using either the CLI or by script, it through the following error.

File "/usr/local/lib/python3.7/dist-packages/onnx_tf/backend_tf_module.py", line 99, in __call__  *
        output_ops = self.backend._onnx_node_to_tensorflow_op(onnx_node,
    File "/usr/local/lib/python3.7/dist-packages/onnx_tf/backend.py", line 347, in _onnx_node_to_tensorflow_op  *
        return handler.handle(node, tensor_dict=tensor_dict, strict=strict)
    File "/usr/local/lib/python3.7/dist-packages/onnx_tf/handlers/handler.py", line 59, in handle  *
        return ver_handle(node, **kwargs)
    File "/usr/local/lib/python3.7/dist-packages/onnx_tf/handlers/backend/conv.py", line 15, in version_11  *
        return cls.conv(node, kwargs["tensor_dict"])
    File "/usr/local/lib/python3.7/dist-packages/onnx_tf/handlers/backend/conv_mixin.py", line 30, in conv  *
        x_rank = len(x.get_shape())

    ValueError: Cannot take the length of shape with unknown rank.

To Reproduce

To reproduce this error I created a Google Colab notebook

ONNX model file

The ONNX file can be downloaded directly on the google colab script

Python, ONNX, ONNX-TF, Tensorflow version

Python version:
3.7.13 (default, Apr 24 2022, 01:04:09)
[GCC 7.5.0]
ONNX version:
1.12.0
ONNX-TF version:
1.10.0
Tensorflow version:
2.8.2

Answer 1 · 2022-09-29T06:28:47.000Z

Describe the bug

Hi, I want to convert an onnx model to Tensorflow, but when I try to convert it, using either the CLI or by script, it through the following error.
File "/usr/local/lib/python3.7/dist-packages/onnx_tf/backend_tf_module.py", line 99, in __call__  *
        output_ops = self.backend._onnx_node_to_tensorflow_op(onnx_node,
    File "/usr/local/lib/python3.7/dist-packages/onnx_tf/backend.py", line 347, in _onnx_node_to_tensorflow_op  *
        return handler.handle(node, tensor_dict=tensor_dict, strict=strict)
    File "/usr/local/lib/python3.7/dist-packages/onnx_tf/handlers/handler.py", line 59, in handle  *
        return ver_handle(node, **kwargs)
    File "/usr/local/lib/python3.7/dist-packages/onnx_tf/handlers/backend/conv.py", line 15, in version_11  *
        return cls.conv(node, kwargs["tensor_dict"])
    File "/usr/local/lib/python3.7/dist-packages/onnx_tf/handlers/backend/conv_mixin.py", line 30, in conv  *
        x_rank = len(x.get_shape())

    ValueError: Cannot take the length of shape with unknown rank.
To Reproduce

To reproduce this error I created a Google Colab notebook

ONNX model file

The ONNX file can be downloaded directly on the google colab script

Python, ONNX, ONNX-TF, Tensorflow version

Python version:
3.7.13 (default, Apr 24 2022, 01:04:09)
[GCC 7.5.0]

ONNX version:
1.12.0

ONNX-TF version:
1.10.0

Tensorflow version:
2.8.2

Hello, I am currently facing the same problem as you, in the process of converting a model from pytorch to onnx then to tflite. I have tried the solution for a similar situation that I have seen in the issue so far, but the problem still exists, do you have any good suggestions?

Answer 2 · 2022-10-29T20:14:06.000Z

Having the same issue with similar setup. Looking a bit more into this I see that the output from prepare does not seem right: tf_rep.tensor_dict and tf_rep.graph are just empty. This is weird because tf_rep.inputs and tf_rep.outputs both list correct names, tf_rep.onnx_op_list is giving the right stats and so on...

Totally stuck here so any help will be appreciated!

Python version:
3.8.5
ONNX version:
1.12.0
ONNX-TF version:
1.10.0
Tensorflow version:
2.10

Answer 3 · 2022-10-29T21:44:17.000Z

Seems like I was able to solve this problem by explicitly setting opset_version and dynamic_axes when saving from torch to onnx:
torch.onnx.export(
model,
image,
onnx_file,
input_names=["input0"],
output_names=["output0", "output1"],
export_params=True,
opset_version=12,
do_constant_folding=True,
dynamic_axes={
"input0": {0: "batch_size"},
"output0": {0: "batch_size"},
"output1": {0: "batch_size"},
},
)

My torch version is 1.12

Answer 4 · 2022-11-10T14:24:25.000Z

I have confirmed that I can convert successfully with this tool I am creating. onnx2tf
I am only interested in the transformation of the model and not in checking if the inference results of the model are degraded or normal.
https://github.com/PINTO0309/onnx2tf
https://github.com/PINTO0309/onnx2tf/releases/tag/1.1.18

replace.json

{
  "format_version": 1,
  "operations": [
    {
      "op_name": "Reshape_448",
      "param_target": "outputs",
      "param_name": "onnx::Add_817",
      "post_process_transpose_perm": [0,2,3,1]
    }
  ]
}

Run

onnx2tf -i vae_encoder.onnx -prf replace.json

Converted tflite
https://github.com/PINTO0309/onnx2tf/releases/download/1.1.15/model_float32.tflite

Answer 5 · 2022-12-09T09:01:40.000Z

Having the same issue with similar setup and any help will be appreciated!
https://github.com/snakers4/silero-vad/blob/v3.1/files/silero_vad.onnx

Python version:
3.8.10
ONNX version:
1.12.0
ONNX-TF version:
1.10.0
Tensorflow version:
2.10