Missed TensorFlow World? Check out the recap. Learn more

tf.contrib.checkpoint.UniqueNameTracker

View source on GitHub

Class UniqueNameTracker

Adds dependencies on trackable objects with name hints.

Useful for creating dependencies with locally unique names.

Example usage:

class SlotManager(tf.contrib.checkpoint.Checkpointable):

  def __init__(self):
    # Create a dependency named "slotdeps" on the container.
    self.slotdeps = tf.contrib.checkpoint.UniqueNameTracker()
    slotdeps = self.slotdeps
    slots = []
    slots.append(slotdeps.track(tf.Variable(3.), "x"))  # Named "x"
    slots.append(slotdeps.track(tf.Variable(4.), "y"))
    slots.append(slotdeps.track(tf.Variable(5.), "x"))  # Named "x_1"

__init__

View source

__init__()

Properties

layers

losses

Aggregate losses from any Layer instances.

non_trainable_variables

non_trainable_weights

trainable

trainable_variables

trainable_weights

updates

Aggregate updates from any Layer instances.

variables

weights

Methods

__eq__

View source

__eq__(other)

track

View source

track(
    trackable,
    base_name
)

Add a dependency on trackable.

Args:

  • trackable: An object to add a checkpoint dependency on.
  • base_name: A name hint, which is uniquified to determine the dependency name.

Returns:

trackable, for chaining.

Raises:

  • ValueError: If trackable is not a trackable object.