ValueError: Importing a SavedModel with tf.saved_model.load requires a tags= argument if there is more than one MetaGraph. Got tags=None, but there are 0 MetaGraphs in the SavedModel with tag sets: []. Pass a tags= argument to load this SavedModel.

[S-hyun]

Dear NNEF programers,

This error occurred while converting from ONNX to TF.

Enviroment

In Google Colab, we installed these packages.

!pip install onnx
!git clone https://github.com/KhronosGroup/NNEF-Tools.git
!cd /content/NNEF-Tools && python setup.py install
!cd /content/NNEF-Tools/parser/python && python setup.py install
!pip install future typing six numpy protobuf torch
!pip install future typing six numpy protobuf onnx onnx-simplifier onnxruntime
!pip install onnx==1.8.1
!pip install torch
!pip install torchvision
!pip install numpy
!pip install onnx
!pip install onnx2keras

Convert

With onnx file, we tried to convert from ONNX to TF.

!python -m nnef_tools.convert --input-format onnx --output-format nnef --input-model /content/pytorch2onnx.onnx --output-model onnx2nnef.nnef --fold-constants

!python -m nnef_tools.convert --input-format nnef --output-format tf --input-model onnx2nnef.nnef --output-model nnef2tf.pb

Load models

After conversion, I got the following error message while loading the TF model. Import statements that are not important are skipped.

import tensorflow as tf

load_model = tf.keras.models.load_model('/content/')
print(list(loaded.signatures.keys()))

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-10-0f33b0d08d70> in <module>
      3 #loaded_model.summary()
      4 
----> 5 loaded = tf.saved_model.load("/content/")
      6 print(list(loaded.signatures.keys()))

4 frames
/usr/local/lib/python3.7/dist-packages/tensorflow/python/saved_model/load.py in load(export_dir, tags, options)
    934     ValueError: If `tags` don't match a MetaGraph in the SavedModel.
    935   """
--> 936   result = load_internal(export_dir, tags, options)["root"]
    937   return result
    938 

/usr/local/lib/python3.7/dist-packages/tensorflow/python/saved_model/load.py in load_internal(export_dir, tags, options, loader_cls, filters)
    992                        " version) cannot be loaded with node filters.")
    993     with ops.init_scope():
--> 994       root = load_v1_in_v2.load(export_dir, tags)
    995       root.graph_debug_info = debug_info
    996 

/usr/local/lib/python3.7/dist-packages/tensorflow/python/saved_model/load_v1_in_v2.py in load(export_dir, tags)
    280   metrics.IncrementReadApi(_LOAD_V1_V2_LABEL)
    281   loader = _EagerSavedModelLoader(export_dir)
--> 282   result = loader.load(tags=tags)
    283   metrics.IncrementRead(write_version="1")
    284   return result

/usr/local/lib/python3.7/dist-packages/tensorflow/python/saved_model/load_v1_in_v2.py in load(self, tags)
    205   def load(self, tags):
    206     """Creates an object from the MetaGraph identified by `tags`."""
--> 207     meta_graph_def = self.get_meta_graph_def_from_tags(tags)
    208     load_shared_name_suffix = "_load_{}".format(ops.uid())
    209     functions = function_deserialization.load_function_def_library(

/usr/local/lib/python3.7/dist-packages/tensorflow/python/saved_model/load_v1_in_v2.py in get_meta_graph_def_from_tags(self, tags)
     78                     for mg in self._saved_model.meta_graphs]
     79         raise ValueError(
---> 80             "Importing a SavedModel with `tf.saved_model.load` requires a "
     81             "`tags=` argument if there is more than one MetaGraph. Got "
     82             f"`tags=None`, but there are {len(self._saved_model.meta_graphs)} "

ValueError: Importing a SavedModel with `tf.saved_model.load` requires a `tags=` argument if there is more than one MetaGraph. Got `tags=None`, but there are 0 MetaGraphs in the SavedModel with tag sets: []. Pass a `tags=` argument to load this SavedModel.

I am asking how I can solve this problem.

Thank you very much for your time.

Sincerely,

[gyenesvi]

Hi, It seems that the code you are using to read the model expects a 'saved-model' format as input, which is a newer format of TensorFlow, while the NNEF tools work with the older frozen graph format. The converter can only output this older format, and hence your code is not able to read it. Note that the NNEF converter tools are not really meant to serve as a conversion tool from ONNX to TF, although it happens to be a possible usage path. For that purpose, ONNX has its own tools, and NNEF is not necessary as an intermediate step. May I ask why you want to use NNEF this way instead of converting from ONNX to TF?

[gyenesvi]

@S-hyun any comment on this? Can I close the issue?