Function _reroute_ts

tensorflow/contrib/graph_editor/reroute.py:135–205

Reroute the end of the tensors in each pair (t0,t1) in ts0 x ts1. This function is the back-bone of the Graph-Editor. It is essentially a thin wrapper on top of the tf.Operation._update_input. Given a pair of tensor t0, t1 in ts0 x ts1, this function re-route the end of t0 and t1 in three

(ts0, ts1, mode, can_modify=None, cannot_modify=None)

def _reroute_ts(ts0, ts1, mode, can_modify=None, cannot_modify=None):
  """Reroute the end of the tensors in each pair (t0,t1) in ts0 x ts1.

  This function is the back-bone of the Graph-Editor. It is essentially a thin
  wrapper on top of the tf.Operation._update_input.

  Given a pair of tensor t0, t1 in ts0 x ts1, this function re-route the end
  of t0 and t1 in three possible ways:
  1) The reroute mode is "a<->b" or "b<->a": the tensors&#x27; end are swapped. After
  this operation, the previous consumers of t0 are now consumers of t1 and
  vice-versa.
  2) The reroute mode is "a->b": the tensors&#x27; end of t0 are re-routed to the
  tensors&#x27;s end of t1 (which are left dangling). After this operation, the
  previous consumers of t0 are still consuming t0 but the previous consumers of
  t1 are not also consuming t0. The tensor t1 has no consumer.
  3) The reroute mode is "b->a": this mode is the symmetric of the "a->b" mode.

  Note that this function is re-routing the end of two tensors, not the start.
  Re-routing the start of two tensors is not supported by this library. The
  reason for that is the following: TensorFlow, by design, creates a strong bond
  between an op and its output tensor. This Graph editor follows this design and
  treats an operation A and its generating tensors {t_i} as an entity which
  cannot be broken. In other words, an op cannot be detached from any of its
  output tensors, ever. But it is possible to detach an op from its input
  tensors, which is what this function concerns itself with.

  Warning: this function is directly manipulating the internals of the tf.Graph.

  Args:
    ts0: an object convertible to a list of `tf.Tensor`.
    ts1: an object convertible to a list of `tf.Tensor`.
    mode: what to do with those tensors: "a->b" or "b<->a" for swaping and
      "a->b" or "b->a" for one direction re-routing.
    can_modify: iterable of operations which can be modified. Any operation
      outside within_ops will be left untouched by this function.
    cannot_modify: iterable of operations which cannot be modified.
      Any operation within cannot_modify will be left untouched by this
      function.
  Returns:
    The number of individual modifications made by the function.
  Raises:
    TypeError: if `ts0` or `ts1` cannot be converted to a list of `tf.Tensor`.
    TypeError: if `can_modify` or `cannot_modify` is not `None` and cannot be
      converted to a list of `tf.Operation`.
  """
  a2b, b2a = _RerouteMode.check(mode)
  ts0 = _util.make_list_of_t(ts0)
  ts1 = _util.make_list_of_t(ts1)
  _check_ts_compatibility(ts0, ts1)
  if cannot_modify is not None:
    cannot_modify = frozenset(_util.make_list_of_op(cannot_modify))
  if can_modify is not None:
    can_modify = frozenset(_util.make_list_of_op(can_modify))
  nb_update_inputs = 0
  precomputed_consumers = []
  # precompute consumers to avoid issue with repeated tensors:
  for t0, t1 in zip(ts0, ts1):
    consumers0 = set(t0.consumers())