Function lift_to_graph

tensorflow/python/eager/lift_to_graph.py:205–359

Copies the tensor and all its inputs recursively to the outer graph. Args: tensors: The Tensors to lift. graph: The graph to lift to. sources: Optional sequence of nodes to start from. If omitted the whole subgraph which feeds into `init_tensor` is lifted. disallowed_placeho

(tensors,
                  graph,
                  sources=None,
                  disallowed_placeholders=None,
                  add_sources=False,
                  handle_captures=False,
                  base_graph=None,
                  op_map=None)

def lift_to_graph(tensors,
                  graph,
                  sources=None,
                  disallowed_placeholders=None,
                  add_sources=False,
                  handle_captures=False,
                  base_graph=None,
                  op_map=None):
  """Copies the tensor and all its inputs recursively to the outer graph.

  Args:
    tensors: The Tensors to lift.
    graph: The graph to lift to.
    sources: Optional sequence of nodes to start from. If omitted the whole
      subgraph which feeds into `init_tensor` is lifted.
    disallowed_placeholders: An optional set of ops which may not appear in the
      lifted graph. Defaults to all placeholders.
    add_sources: A boolean indicating whether placeholders which are not in
      sources should be allowed.
    handle_captures: A boolean indicating whether to re-capture s in the new
      graph or simply create a vanilla placeholder.
    base_graph: The graph from which to lift ops. This will be inferred if not
      specified.
    op_map: A map contains all the existing nodes that have been lifted to the
      destination graph, so they won't be lifted and copied again.

  Returns:
    A mapping from ops in the current default graph to ops in `graph`.

  Raises:
    UnliftableError: If a placeholder blocks lifting.
  """
  variable_init_tensors = []
  init_tensors = []
  for tensor in tensors:
    if isinstance(tensor, resource_variable_ops.ResourceVariable):
      variable_init_tensors.append(tensor)
    else:
      init_tensors.append(tensor)
  base_graph = base_graph or init_tensors[0].graph
  op_map = op_map or object_identity.ObjectIdentityDictionary()

  # Check that the initializer does not depend on any placeholders.
  sources = object_identity.ObjectIdentitySet(sources or [])
  visited_ops = set([x.op for x in sources])
  op_outputs = collections.defaultdict(set)

  # First we extract the subgraph between init_tensors and sources.
  for init_tensor in init_tensors:
    sources.update(op_selector.map_subgraph(
        init_tensor=init_tensor,
        sources=sources,
        disallowed_placeholders=disallowed_placeholders,
        visited_ops=visited_ops,
        op_outputs=op_outputs,
        add_sources=add_sources))

  # Try to topologically sort the nodes we've extracted. Now we know how many of