apply_saved_models for Tensorflow 2.7 models

[agrimabahl]

Hi 👋

My application uses the apply_saved_models api within tf-transform to run inference on tensorflow models.
For a TF2 model from tf-hub, when running apply_saved_models I get the following error -

---------------------------------------------------------------------------
NotImplementedError                       Traceback (most recent call last)
<ipython-input-30-594dff228f07> in <module>()
     19       constant_graph_def,
     20       input_map=inputs,
---> 21       return_elements=output_tensor_names + loaded_initializer_op_names)
     22

/usr/local/lib/python3.7/site-packages/tensorflow/python/util/deprecation.py in new_func(*args, **kwargs)
    550                 'in a future version' if date is None else ('after %s' % date),
    551                 instructions)
--> 552       return func(*args, **kwargs)
    553
    554     doc = _add_deprecated_arg_notice_to_docstring(

/usr/local/lib/python3.7/site-packages/tensorflow/python/framework/importer.py in import_graph_def(***failed resolving arguments***)
    410       return_elements=return_elements,
    411       name=name,
--> 412       producer_op_list=producer_op_list)
    413
    414

/usr/local/lib/python3.7/site-packages/tensorflow/python/framework/importer.py in _import_graph_def_internal(graph_def, input_map, return_elements, validate_colocation_constraints, name, producer_op_list)
    492   options = scoped_options.options
    493   _PopulateTFImportGraphDefOptions(options, prefix, input_map, return_elements,
--> 494                                    validate_colocation_constraints)
    495
    496   # _ProcessNewOps mutates the new operations. _mutation_lock ensures a

/usr/local/lib/python3.7/site-packages/tensorflow/python/framework/importer.py in _PopulateTFImportGraphDefOptions(options, prefix, input_map, return_elements, validate_colocation_constraints)
    227       src_name, src_idx = _ParseTensorName(input_src)
    228       src_name = compat.as_str(src_name)
--> 229       dst_output = input_dst._as_tf_output()  # pylint: disable=protected-access
    230       c_api.TF_ImportGraphDefOptionsAddInputMapping(options, src_name, src_idx,
    231                                                     dst_output)

/usr/local/lib/python3.7/site-packages/tensorflow/python/framework/ops.py in _as_tf_output(self)
   1330   def _as_tf_output(self):
   1331     raise NotImplementedError(
-> 1332         "_as_tf_output not supported when eager execution is enabled.")
   1333
   1334   def eval(self, feed_dict=None, session=None):

NotImplementedError: _as_tf_output not supported when eager execution is enabled.

is there an alternative API more suited to TF2 models since apply_saved_models can only be used with TF2 behavior disabled and TF2 does not support GraphDefs.

Thank you!

[zoyahav]

Could you share with us the code snippet you're running?
And which API you're using, is it pretrained_models.apply_saved_model, and how is it used?

Since you're running into an eager execution error, it's not clear to me how tensorflow-transform is used here.