-
Notifications
You must be signed in to change notification settings - Fork 220
Open
Description
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!