'EmbeddingVariable' object has no attribute '_is_primary' When using import_meta_graph

[Lihengwannafly]

Describe the current behavior

  File "/root/workspace/rec-rank-train/vmax/estimator/estimator_v2.py", line 122, in export_big_model
    self.estimator_core.export_big_model(server, checkpoint_path=checkpoint_path)
  File "/root/workspace/rec-rank-train/vmax/core/estimator_core_v2.py", line 415, in export_big_model
    tf.train.import_meta_graph(meta_graph_or_file='/tmp/saved_model/tmp.meta')
  File "/usr/local/lib/python3.8/dist-packages/tensorflow_core/python/training/saver.py", line 1695, in import_meta_graph
    return _import_meta_graph_with_return_elements(meta_graph_or_file,
  File "/usr/local/lib/python3.8/dist-packages/tensorflow_core/python/training/saver.py", line 1723, in _import_meta_graph_with_return_elements
    saver = _create_saver_from_imported_meta_graph(meta_graph_def, import_scope,
  File "/usr/local/lib/python3.8/dist-packages/tensorflow_core/python/training/saver.py", line 1744, in _create_saver_from_imported_meta_graph
    return Saver()
  File "/usr/local/lib/python3.8/dist-packages/tensorflow_core/python/training/saver.py", line 1033, in __init__
    self.build()
  File "/usr/local/lib/python3.8/dist-packages/tensorflow_core/python/training/saver.py", line 1045, in build
    self._build(self._filename, build_save=True, build_restore=True)
  File "/usr/local/lib/python3.8/dist-packages/tensorflow_core/python/training/saver.py", line 1112, in _build
    self.saver_def = self._builder._build_internal(  # pylint: disable=protected-access
  File "/usr/local/lib/python3.8/dist-packages/tensorflow_core/python/training/saver.py", line 656, in _build_internal
    restore_op = self._AddRestoreOps(filename_tensor, saveables,
  File "/usr/local/lib/python3.8/dist-packages/tensorflow_core/python/training/saver.py", line 491, in _AddRestoreOps
    assign_ops.append(saveable.restore(saveable_tensors, shapes))
  File "/usr/local/lib/python3.8/dist-packages/tensorflow_core/python/training/saving/saveable_object_util.py", line 185, in restore
    with ops.control_dependencies(None if self.var._is_primary else [self.var._primary.initializer]):
AttributeError: 'EmbeddingVariable' object has no attribute '_is_primary'

Code to reproduce the issue

  meta_graph_def = tf.train.export_meta_graph()
  meta_graph_def.meta_info_def.meta_graph_version = str(int(time.time()))
  self.logger.info('meta_graph_version = %s' %
                   meta_graph_def.meta_info_def.meta_graph_version)
  tf.reset_default_graph()
  tf.train.import_meta_graph(meta_graph_def)

and

  meta_graph_def = tf.train.export_meta_graph(filename='/tmp/saved_model/tmp.meta')
  meta_graph_def.meta_info_def.meta_graph_version = str(int(time.time()))
  self.logger.info('meta_graph_version = %s' %
                   meta_graph_def.meta_info_def.meta_graph_version)
  tf.reset_default_graph()
  tf.train.import_meta_graph(meta_graph_or_file='/tmp/saved_model/tmp.meta')

[Lihengwannafly]

But if I import_meta_graph from checkpoint.meta, it works.

tf.reset_default_graph()
saver = tf.train.import_meta_graph('{}.meta'.format(latest_checkpoint_path))

[liutongxuan]

Fixed by #128, PR merged. @Lihengwannafly please help to verify.