Attempting to add a duplicate function when exporting to TF Lite

[monatis]

I want to deploy Conformer Subwords model on RPi for a PoC work. When exporting the latest pretrained model with tflite_subword_conformer.py and tf-nightly v2.5.0.dev20201106, I'm getting the following exception. Anyone getting the same?

  File "tflite_subword_conformer.py", line 70, in <module>
    tflite_model = converter.convert()
  File "D:\coding\asr\examples\conformer\env\lib\site-packages\tensorflow\lite\python\lite.py", line 1118, in convert
    return super(TFLiteConverterV2, self).convert()
  File "D:\coding\asr\examples\conformer\env\lib\site-packages\tensorflow\lite\python\lite.py", line 922, in convert
    self._funcs[0], lower_control_flow=False))
  File "D:\coding\asr\examples\conformer\env\lib\site-packages\tensorflow\python\framework\convert_to_constants.py", line 1111, in convert_variables_to_constants_v2_as_graph
    converted_input_indices)
  File "D:\coding\asr\examples\conformer\env\lib\site-packages\tensorflow\python\framework\convert_to_constants.py", line 1003, in _construct_concrete_function
    new_output_names)
  File "D:\coding\asr\examples\conformer\env\lib\site-packages\tensorflow\python\eager\wrap_function.py", line 650, in function_from_graph_def
    wrapped_import = wrap_function(_imports_graph_def, [])
  File "D:\coding\asr\examples\conformer\env\lib\site-packages\tensorflow\python\eager\wrap_function.py", line 630, in wrap_function
    signature=signature)
  File "D:\coding\asr\examples\conformer\env\lib\site-packages\tensorflow\python\eager\wrap_function.py", line 229, in __init__
    context.context().add_function_def(f)
  File "D:\coding\asr\examples\conformer\env\lib\site-packages\tensorflow\python\eager\context.py", line 1138, in add_function_def
    len(fdef_string))
tensorflow.python.framework.errors_impl.InvalidArgumentError: Attempting to add a duplicate function with name: __inference_standard_lstm_20496 where the previous and current definitions differ. Previous definiton: signature {
...
}

Please note that I omited the signature definition for brevity, but I can post it separately if it matters.

[nglehuy]

@monatis the issue is reproduced, I'm working on it.

[monatis]

Hi @usimarit, thanks. I was able to export to TF Lite with tf-nightly v2.4.0.dev20200901, but I'm getting the following error during inference. I'm investigating this one now:

Traceback (most recent call last):
  File "conformer.py", line 54, in <module>
    tflitemodel.invoke()
  File "D:\coding\asr\env\lib\site-packages\tensorflow\lite\python\interpreter.py", line 540, in invoke
    self._interpreter.Invoke()
RuntimeError: tensorflow/lite/kernels/reshape.cc:58 stretch_dim != -1 (0 != -1)Node number 199 (RESHAPE) failed to prepare.

[nglehuy]

@monatis There's some bugs in conformer encoder in tf-nightly (it works with tf 2.3). I'm still working on this.

[nglehuy]

@monatis I fixed the issue: the tf.squeeze is not supported in tf-nightly so I have to use pure tf.reshape. Please pull the newest commits on branch main and try again. If there still has bugs, just raise them and I'll continue to fix 😄

[monatis]

Hi @usimarit, I was able to export with tf-nightly v2.4.0.dev20200901 but needed to upgrade to v2.5.0.dev20201108 for inference due to a runtime error with unsupported flex ops in the former. Now that the issue is resolved, I'm closing it, but attached is the converted TF Lite models --you may want to share them on Drive alongside the saved models. Thanks.

[nglehuy]

@monatis I uploaded to the drive, thank you for sharing 😄