Training checkpoints

View on Run in Google Colab View source on GitHub Download notebook

The phrase "Saving a TensorFlow model" typically means one of two things:

  1. Checkpoints, OR
  2. SavedModel.

Checkpoints capture the exact value of all parameters (tf.Variable objects) used by a model. Checkpoints do not contain any description of the computation defined by the model and thus are typically only useful when source code that will use the saved parameter values is available.

The SavedModel format on the other hand includes a serialized description of the computation defined by the model in addition to the parameter values (checkpoint). Models in this format are independent of the source code that created the model. They are thus suitable for deployment via TensorFlow Serving, TensorFlow Lite, TensorFlow.js, or programs in other programming languages (the C, C++, Java, Go, Rust, C# etc. TensorFlow APIs).

This guide covers APIs for writing and reading checkpoints.


import tensorflow as tf
class Net(tf.keras.Model):
  """A simple linear model."""

  def __init__(self):
    super(Net, self).__init__()
    self.l1 = tf.keras.layers.Dense(5)

  def call(self, x):
    return self.l1(x)
net = Net()

Saving from tf.keras training APIs

See the tf.keras guide on saving and restoring.

tf.keras.Model.save_weights saves a TensorFlow checkpoint.


Writing checkpoints

The persistent state of a TensorFlow model is stored in tf.Variable objects. These can be constructed directly, but are often created through high-level APIs like tf.keras.layers or tf.keras.Model.

The easiest way to manage variables is by attaching them to Python objects, then referencing those objects.

Subclasses of tf.train.Checkpoint, tf.keras.layers.Layer, and tf.keras.Model automatically track variables assigned to their attributes. The following example constructs a simple linear model, then writes checkpoints which contain values for all of the model's variables.

You can easily save a model-checkpoint with Model.save_weights.

Manual checkpointing


To help demonstrate all the features of tf.train.Checkpoint, define a toy dataset and optimization step:

def toy_dataset():
  inputs = tf.range(10.)[:, None]
  labels = inputs * 5. + tf.range(5.)[None, :]
    dict(x=inputs, y=labels)).repeat().batch(2)
def train_step(net, example, optimizer):
  """Trains `net` on `example` using `optimizer`."""
  with tf.GradientTape() as tape:
    output = net(example['x'])
    loss = tf.reduce_mean(tf.abs(output - example['y']))
  variables = net.trainable_variables
  gradients = tape.gradient(loss, variables)
  optimizer.apply_gradients(zip(gradients, variables))
  return loss

Create the checkpoint objects

Use a tf.train.Checkpoint object to manually create a checkpoint, where the objects you want to checkpoint are set as attributes on the object.

A tf.train.CheckpointManager can also be helpful for managing multiple checkpoints.

opt = tf.keras.optimizers.Adam(0.1)
dataset = toy_dataset()
iterator = iter(dataset)
ckpt = tf.train.Checkpoint(step=tf.Variable(1), optimizer=opt, net=net, iterator=iterator)
manager = tf.train.CheckpointManager(ckpt, './tf_ckpts', max_to_keep=3)

Train and checkpoint the model

The following training loop creates an instance of the model and of an optimizer, then gathers them into a tf.train.Checkpoint object. It calls the training step in a loop on each batch of data, and periodically writes checkpoints to disk.

def train_and_checkpoint(net, manager):
  if manager.latest_checkpoint:
    print("Restored from {}".format(manager.latest_checkpoint))
    print("Initializing from scratch.")

  for _ in range(50):
    example = next(iterator)
    loss = train_step(net, example, opt)
    if int(ckpt.step) % 10 == 0:
      save_path =
      print("Saved checkpoint for step {}: {}".format(int(ckpt.step), save_path))
      print("loss {:1.2f}".format(loss.numpy()))
train_and_checkpoint(net, manager)

Restore and continue training

After the first training cycle you can pass a new model and manager, but pick up training exactly where you left off:

opt = tf.keras.optimizers.Adam(0.1)
net = Net()
dataset = toy_dataset()
iterator = iter(dataset)
ckpt = tf.train.Checkpoint(step=tf.Variable(1), optimizer=opt, net=net, iterator=iterator)
manager = tf.train.CheckpointManager(ckpt, './tf_ckpts', max_to_keep=3)

train_and_checkpoint(net, manager)

The tf.train.CheckpointManager object deletes old checkpoints. Above it's configured to keep only the three most recent checkpoints.

print(manager.checkpoints)  # List the three remaining checkpoints

These paths, e.g. './tf_ckpts/ckpt-10', are not files on disk. Instead they are prefixes for an index file and one or more data files which contain the variable values. These prefixes are grouped together in a single checkpoint file ('./tf_ckpts/checkpoint') where the CheckpointManager saves its state.

ls ./tf_ckpts

Loading mechanics

TensorFlow matches variables to checkpointed values by traversing a directed graph with named edges, starting from the object being loaded. Edge names typically come from attribute names in objects, for example the "l1" in self.l1 = tf.keras.layers.Dense(5). tf.train.Checkpoint uses its keyword argument names, as in the "step" in tf.train.Checkpoint(step=...).

The dependency graph from the example above looks like this:

Visualization of the dependency graph for the example training loop

The optimizer is in red, regular variables are in blue, and the optimizer slot variables are in orange. The other nodes—for example, representing the tf.train.Checkpoint—are in black.

Slot variables are part of the optimizer's state, but are created for a specific variable. For example, the 'm' edges above correspond to momentum, which the Adam optimizer tracks for each variable. Slot variables are only saved in a checkpoint if the variable and the optimizer would both be saved, thus the dashed edges.

Calling restore on a tf.train.Checkpoint object queues the requested restorations, restoring variable values as soon as there's a matching path from the Checkpoint object. For example, you can load just the bias from the model you defined above by reconstructing one path to it through the network and the layer.

to_restore = tf.Variable(tf.zeros([5]))
print(to_restore.numpy())  # All zeros
fake_layer = tf.train.Checkpoint(bias=to_restore)
fake_net = tf.train.Checkpoint(l1=fake_layer)
new_root = tf.train.Checkpoint(net=fake_net)
status = new_root.restore(tf.train.latest_checkpoint('./tf_ckpts/'))
print(to_restore.numpy())  # This gets the restored value.

The dependency graph for these new objects is a much smaller subgraph of the larger checkpoint you wrote above. It includes only the bias and a save counter that tf.train.Checkpoint uses to number checkpoints.

Visualization of a subgraph for the bias variable

restore returns a status object, which has optional assertions. All of the objects created in the new Checkpoint have been restored, so status.assert_existing_objects_matched passes.


There are many objects in the checkpoint which haven't matched, including the layer's kernel and the optimizer's variables. status.assert_consumed only passes if the checkpoint and the program match exactly, and would throw an exception here.

Deferred restorations

Layer objects in TensorFlow may defer the creation of variables to their first call, when input shapes are available. For example, the shape of a Dense layer's kernel depends on both the layer's input and output shapes, and so the output shape required as a constructor argument is not enough information to create the variable on its own. Since calling a Layer also reads the variable's value, a restore must happen between the variable's creation and its first use.

To support this idiom, tf.train.Checkpoint defers restores which don't yet have a matching variable.

deferred_restore = tf.Variable(tf.zeros([1, 5]))
print(deferred_restore.numpy())  # Not restored; still zeros
fake_layer.kernel = deferred_restore
print(deferred_restore.numpy())  # Restored

Manually inspecting checkpoints

tf.train.load_checkpoint returns a CheckpointReader that gives lower level access to the checkpoint contents. It contains mappings from each variable's key, to the shape and dtype for each variable in the checkpoint. A variable's key is its object path, like in the graphs displayed above.

reader = tf.train.load_checkpoint('./tf_ckpts/')
shape_from_key = reader.get_variable_to_shape_map()
dtype_from_key = reader.get_variable_to_dtype_map()


So if you're interested in the value of net.l1.kernel you can get the value with the following code:

key = 'net/l1/kernel/.ATTRIBUTES/VARIABLE_VALUE'

print("Shape:", shape_from_key[key])
print("Dtype:", dtype_from_key[key].name)

It also provides a get_tensor method allowing you to inspect the value of a variable:


Object tracking

Checkpoints save and restore the values of tf.Variable objects by "tracking" any variable or trackable object set in one of its attributes. When executing a save, variables are gathered recursively from all of the reachable tracked objects.

As with direct attribute assignments like self.l1 = tf.keras.layers.Dense(5), assigning lists and dictionaries to attributes will track their contents.

save = tf.train.Checkpoint()
save.listed = [tf.Variable(1.)]
save.mapped = {'one': save.listed[0]}
save.mapped['two'] = save.listed[1]
save_path ='./tf_list_example')

restore = tf.train.Checkpoint()
v2 = tf.Variable(0.)
assert 0. == v2.numpy()  # Not restored yet
restore.mapped = {'two': v2}
assert 2. == v2.numpy()

You may notice wrapper objects for lists and dictionaries. These wrappers are checkpointable versions of the underlying data-structures. Just like the attribute based loading, these wrappers restore a variable's value as soon as it's added to the container.

restore.listed = []
print(restore.listed)  # ListWrapper([])
v1 = tf.Variable(0.)
restore.listed.append(v1)  # Restores v1, from restore() in the previous cell
assert 1. == v1.numpy()

Trackable objects include tf.train.Checkpoint, tf.Module and its subclasses (e.g. keras.layers.Layer and keras.Model), and recognized Python containers:

  • dict (and collections.OrderedDict)
  • list
  • tuple (and collections.namedtuple, typing.NamedTuple)

Other container types are not supported, including:

  • collections.defaultdict
  • set

All other Python objects are ignored, including:

  • int
  • string
  • float


TensorFlow objects provide an easy automatic mechanism for saving and restoring the values of variables they use.