Apply to speak at TensorFlow World. Deadline April 23rd. Propose talk

Module: tf.contrib.checkpoint

Defined in tensorflow/contrib/checkpoint/

Tools for working with object-based checkpoints.

Visualization and inspection:

Managing dependencies:

Checkpointable data structures:

Checkpoint management:

Saving and restoring Python state:


class CheckpointManager: Deletes old checkpoints.

class Checkpointable: Manages dependencies on other objects.

class CheckpointableBase: Base class for Checkpointable objects without automatic dependencies.

class CheckpointableObjectGraph: A ProtocolMessage

class List: An append-only sequence type which is checkpointable.

class Mapping: An append-only checkpointable mapping data structure with string keys.

class NoDependency: Allows attribute assignment to Checkpointable objects with no dependency.

class NumpyState: A checkpointable object whose NumPy array attributes are saved/restored.

class PythonStateWrapper: Wraps a Python object for storage in an object-based checkpoint.

class UniqueNameTracker: Adds dependencies on checkpointable objects with name hints.


capture_dependencies(...): Capture variables created within this scope as Template dependencies.

dot_graph_from_checkpoint(...): Visualizes an object-based checkpoint (from tf.train.Checkpoint).

list_objects(...): Traverse the object graph and list all accessible objects.

object_metadata(...): Retrieves information about the objects in a checkpoint.

split_dependency(...): Creates multiple dependencies with a synchronized save/restore.