tf.contrib.eager.Network

View source on GitHub

Represents the composition of a set of Layers.

Inherits From: Layer

tf.contrib.eager.Network(
    name=None
)

Deprecated. Please inherit from tf.keras.Model, and see its documentation for details. tf.keras.Model should be a drop-in replacement for tfe.Network in most cases, but note that track_layer is no longer necessary or supported. Instead, Layer instances are tracked on attribute assignment (see the section of tf.keras.Model's documentation on subclassing). Since the output of track_layer is often assigned to an attribute anyway, most code can be ported by simply removing the track_layer calls.

tf.keras.Model works with all TensorFlow Layer instances, including those from tf.layers, but switching to the tf.keras.layers versions along with the migration to tf.keras.Model is recommended, since it will preserve variable names. Feel free to import it with an alias to avoid excess typing :).

Network implements the Layer interface and adds convenience methods for managing sub-Layers, such as listing variables.

Layers (including other Networks) should be added via track_layer. They can then be used when overriding the Network.call method:

class TwoLayerNetwork(tfe.Network):

  def __init__(self, name):
    super(TwoLayerNetwork, self).__init__(name=name)
    self.layer_one = self.track_layer(tf.compat.v1.layers.Dense(16,
    input_shape=(8,)))
    self.layer_two = self.track_layer(tf.compat.v1.layers.Dense(1,
    input_shape=(16,)))

  def call(self, inputs):
    return self.layer_two(self.layer_one(inputs))

After constructing an object and calling the Network, a list of variables created by tracked Layers is available via Network.variables:

net = TwoLayerNetwork(name="net")
output = net(tf.ones([1, 8]))
print([v.name for v in net.variables])

This example prints variable names, one kernel and one bias per tf.compat.v1.layers.Dense layer:

['net/dense/kernel:0',
 'net/dense/bias:0',
 'net/dense_1/kernel:0',
 'net/dense_1/bias:0']

These variables can be passed to a Saver (tf.compat.v1.train.Saver, or tf.contrib.eager.Saver when executing eagerly) to save or restore the Network, typically alongside a global step and tf.compat.v1.train.Optimizer variables when checkpointing during training.

Note that the semantics of calling a Network with graph execution (i.e. not executing eagerly) may change slightly in the future. Currently stateful ops are pruned from the graph unless they or something that depends on them is executed in a session, but this behavior is not consistent with eager execution (where stateful ops are executed eagerly). Layers from tf.layers do not depend on this pruning and so will not be affected, but Networks which rely on stateful ops being added to the graph but not executed (e.g. via custom Layers which manage stateful ops) may break with this change.

Methods

get_layer

View source

get_layer(
    name=None, index=None
)

Get a contained tf.compat.v1.layers.Layer either by name or index.

track_layer

View source

track_layer(
    layer
)

Track a Layer in this Network.

Network requires that all Layers used in call() be tracked so that the Network can export a complete list of variables.