Layer weight initializers

Usage of initializers

Initializers define the way to set the initial random weights of Keras layers.

The keyword arguments used for passing initializers to layers depends on the layer. Usually, it is simply kernel_initializer and bias_initializer:

from tensorflow.keras import layers
from tensorflow.keras import initializers

layer = layers.Dense(
    units=64,
    kernel_initializer=initializers.RandomNormal(stddev=0.01),
    bias_initializer=initializers.Zeros()
)

All built-in initializers can also be passed via their string identifier:

layer = layers.Dense(
    units=64,
    kernel_initializer='random_normal',
    bias_initializer='zeros'
)

Available initializers

The following built-in initializers are available as part of the tf.keras.initializers module:

[source]

RandomNormal class

tf.keras.initializers.RandomNormal(mean=0.0, stddev=0.05, seed=None)

Initializer that generates tensors with a normal distribution.

Also available via the shortcut function tf.keras.initializers.random_normal.

Examples

>>> # Standalone usage:
>>> initializer = tf.keras.initializers.RandomNormal(mean=0., stddev=1.)
>>> values = initializer(shape=(2, 2))

>>> # Usage in a Keras layer:
>>> initializer = tf.keras.initializers.RandomNormal(mean=0., stddev=1.)
>>> layer = tf.keras.layers.Dense(3, kernel_initializer=initializer)

Arguments

• mean: a python scalar or a scalar tensor. Mean of the random values to generate.
• stddev: a python scalar or a scalar tensor. Standard deviation of the random values to generate.
• seed: A Python integer. Used to make the behavior of the initializer deterministic. Note that a seeded initializer will produce the same random values across multiple calls.

[source]

RandomUniform class

tf.keras.initializers.RandomUniform(minval=-0.05, maxval=0.05, seed=None)

Initializer that generates tensors with a uniform distribution.

Also available via the shortcut function tf.keras.initializers.random_uniform.

Examples

>>> # Standalone usage:
>>> initializer = tf.keras.initializers.RandomUniform(minval=0., maxval=1.)
>>> values = initializer(shape=(2, 2))

>>> # Usage in a Keras layer:
>>> initializer = tf.keras.initializers.RandomUniform(minval=0., maxval=1.)
>>> layer = tf.keras.layers.Dense(3, kernel_initializer=initializer)

Arguments

• minval: A python scalar or a scalar tensor. Lower bound of the range of random values to generate (inclusive).
• maxval: A python scalar or a scalar tensor. Upper bound of the range of random values to generate (exclusive).
• seed: A Python integer. Used to make the behavior of the initializer deterministic. Note that a seeded initializer will produce the same random values across multiple calls.

[source]

TruncatedNormal class

tf.keras.initializers.TruncatedNormal(mean=0.0, stddev=0.05, seed=None)

Initializer that generates a truncated normal distribution.

Also available via the shortcut function tf.keras.initializers.truncated_normal.

The values generated are similar to values from a tf.keras.initializers.RandomNormal initializer except that values more than two standard deviations from the mean are discarded and re-drawn.

Examples

>>> # Standalone usage:
>>> initializer = tf.keras.initializers.TruncatedNormal(mean=0., stddev=1.)
>>> values = initializer(shape=(2, 2))

>>> # Usage in a Keras layer:
>>> initializer = tf.keras.initializers.TruncatedNormal(mean=0., stddev=1.)
>>> layer = tf.keras.layers.Dense(3, kernel_initializer=initializer)

Arguments

• mean: a python scalar or a scalar tensor. Mean of the random values to generate.
• stddev: a python scalar or a scalar tensor. Standard deviation of the random values to generate before truncation.
• seed: A Python integer. Used to make the behavior of the initializer deterministic. Note that a seeded initializer will produce the same random values across multiple calls.

[source]

Zeros class

tf.keras.initializers.Zeros()

Initializer that generates tensors initialized to 0.

Also available via the shortcut function tf.keras.initializers.zeros.

Examples

>>> # Standalone usage:
>>> initializer = tf.keras.initializers.Zeros()
>>> values = initializer(shape=(2, 2))

>>> # Usage in a Keras layer:
>>> initializer = tf.keras.initializers.Zeros()
>>> layer = tf.keras.layers.Dense(3, kernel_initializer=initializer)

[source]

Ones class

tf.keras.initializers.Ones()

Initializer that generates tensors initialized to 1.

Also available via the shortcut function tf.keras.initializers.ones.

Examples

>>> # Standalone usage:
>>> initializer = tf.keras.initializers.Ones()
>>> values = initializer(shape=(2, 2))

>>> # Usage in a Keras layer:
>>> initializer = tf.keras.initializers.Ones()
>>> layer = tf.keras.layers.Dense(3, kernel_initializer=initializer)

[source]

GlorotNormal class

tf.keras.initializers.GlorotNormal(seed=None)

The Glorot normal initializer, also called Xavier normal initializer.

Also available via the shortcut function tf.keras.initializers.glorot_normal.

Draws samples from a truncated normal distribution centered on 0 with stddev = sqrt(2 / (fan_in + fan_out)) where fan_in is the number of input units in the weight tensor and fan_out is the number of output units in the weight tensor.

Examples

>>> # Standalone usage:
>>> initializer = tf.keras.initializers.GlorotNormal()
>>> values = initializer(shape=(2, 2))

>>> # Usage in a Keras layer:
>>> initializer = tf.keras.initializers.GlorotNormal()
>>> layer = tf.keras.layers.Dense(3, kernel_initializer=initializer)

Arguments

• seed: A Python integer. Used to make the behavior of the initializer deterministic. Note that a seeded initializer will not produce the same random values across multiple calls, but multiple initializers will produce the same sequence when constructed with the same seed value.

References

• Glorot et al., 2010

[source]

GlorotUniform class

tf.keras.initializers.GlorotUniform(seed=None)

The Glorot uniform initializer, also called Xavier uniform initializer.

Also available via the shortcut function tf.keras.initializers.glorot_uniform.

Draws samples from a uniform distribution within [-limit, limit], where limit = sqrt(6 / (fan_in + fan_out)) (fan_in is the number of input units in the weight tensor and fan_out is the number of output units).

Examples

>>> # Standalone usage:
>>> initializer = tf.keras.initializers.GlorotUniform()
>>> values = initializer(shape=(2, 2))

>>> # Usage in a Keras layer:
>>> initializer = tf.keras.initializers.GlorotUniform()
>>> layer = tf.keras.layers.Dense(3, kernel_initializer=initializer)

Arguments

• seed: A Python integer. Used to make the behavior of the initializer deterministic. Note that a seeded initializer will not produce the same random values across multiple calls, but multiple initializers will produce the same sequence when constructed with the same seed value.

References

• Glorot et al., 2010

[source]

HeNormal class

tf.keras.initializers.HeNormal(seed=None)

He normal initializer.

Also available via the shortcut function tf.keras.initializers.he_normal.

It draws samples from a truncated normal distribution centered on 0 with stddev = sqrt(2 / fan_in) where fan_in is the number of input units in the weight tensor.

Examples

>>> # Standalone usage:
>>> initializer = tf.keras.initializers.HeNormal()
>>> values = initializer(shape=(2, 2))

>>> # Usage in a Keras layer:
>>> initializer = tf.keras.initializers.HeNormal()
>>> layer = tf.keras.layers.Dense(3, kernel_initializer=initializer)

Arguments

• seed: A Python integer. Used to make the behavior of the initializer deterministic. Note that a seeded initializer will not produce the same random values across multiple calls, but multiple initializers will produce the same sequence when constructed with the same seed value.

References

• He et al., 2015

[source]

HeUniform class

tf.keras.initializers.HeUniform(seed=None)

He uniform variance scaling initializer.

Also available via the shortcut function tf.keras.initializers.he_uniform.

Draws samples from a uniform distribution within [-limit, limit], where limit = sqrt(6 / fan_in) (fan_in is the number of input units in the weight tensor).

Examples

>>> # Standalone usage:
>>> initializer = tf.keras.initializers.HeUniform()
>>> values = initializer(shape=(2, 2))

>>> # Usage in a Keras layer:
>>> initializer = tf.keras.initializers.HeUniform()
>>> layer = tf.keras.layers.Dense(3, kernel_initializer=initializer)

Arguments

• seed: A Python integer. Used to make the behavior of the initializer deterministic. Note that a seeded initializer will not produce the same random values across multiple calls, but multiple initializers will produce the same sequence when constructed with the same seed value.

References

• He et al., 2015

[source]

Identity class

tf.keras.initializers.Identity(gain=1.0)

Initializer that generates the identity matrix.

Also available via the shortcut function tf.keras.initializers.identity.

Only usable for generating 2D matrices.

Examples

>>> # Standalone usage:
>>> initializer = tf.keras.initializers.Identity()
>>> values = initializer(shape=(2, 2))

>>> # Usage in a Keras layer:
>>> initializer = tf.keras.initializers.Identity()
>>> layer = tf.keras.layers.Dense(3, kernel_initializer=initializer)

Arguments

• gain: Multiplicative factor to apply to the identity matrix.

[source]

Orthogonal class

tf.keras.initializers.Orthogonal(gain=1.0, seed=None)

Initializer that generates an orthogonal matrix.

Also available via the shortcut function tf.keras.initializers.orthogonal.

If the shape of the tensor to initialize is two-dimensional, it is initialized with an orthogonal matrix obtained from the QR decomposition of a matrix of random numbers drawn from a normal distribution. If the matrix has fewer rows than columns then the output will have orthogonal rows. Otherwise, the output will have orthogonal columns.

If the shape of the tensor to initialize is more than two-dimensional, a matrix of shape (shape[0] * ... * shape[n - 2], shape[n - 1]) is initialized, where n is the length of the shape vector. The matrix is subsequently reshaped to give a tensor of the desired shape.

Examples

>>> # Standalone usage:
>>> initializer = tf.keras.initializers.Orthogonal()
>>> values = initializer(shape=(2, 2))

>>> # Usage in a Keras layer:
>>> initializer = tf.keras.initializers.Orthogonal()
>>> layer = tf.keras.layers.Dense(3, kernel_initializer=initializer)

Arguments

• gain: multiplicative factor to apply to the orthogonal matrix
• seed: A Python integer. Used to make the behavior of the initializer deterministic. Note that a seeded initializer will produce the same random values across multiple calls.

References

• Saxe et al., 2014

[source]

Constant class

tf.keras.initializers.Constant(value=0)

Initializer that generates tensors with constant values.

Also available via the shortcut function tf.keras.initializers.constant.

Only scalar values are allowed. The constant value provided must be convertible to the dtype requested when calling the initializer.

Examples

>>> # Standalone usage:
>>> initializer = tf.keras.initializers.Constant(3.)
>>> values = initializer(shape=(2, 2))

>>> # Usage in a Keras layer:
>>> initializer = tf.keras.initializers.Constant(3.)
>>> layer = tf.keras.layers.Dense(3, kernel_initializer=initializer)

Arguments

• value: A Python scalar.

[source]

VarianceScaling class

tf.keras.initializers.VarianceScaling(
    scale=1.0, mode="fan_in", distribution="truncated_normal", seed=None
)

Initializer capable of adapting its scale to the shape of weights tensors.

Also available via the shortcut function tf.keras.initializers.variance_scaling.

With distribution="truncated_normal" or "untruncated_normal", samples are drawn from a truncated/untruncated normal distribution with a mean of zero and a standard deviation (after truncation, if used) stddev = sqrt(scale / n), where n is:

• number of input units in the weight tensor, if mode="fan_in"
• number of output units, if mode="fan_out"
• average of the numbers of input and output units, if mode="fan_avg"

With distribution="uniform", samples are drawn from a uniform distribution within [-limit, limit], where limit = sqrt(3 * scale / n).

Examples

>>> # Standalone usage:
>>> initializer = tf.keras.initializers.VarianceScaling(
... scale=0.1, mode='fan_in', distribution='uniform')
>>> values = initializer(shape=(2, 2))

>>> # Usage in a Keras layer:
>>> initializer = tf.keras.initializers.VarianceScaling(
... scale=0.1, mode='fan_in', distribution='uniform')
>>> layer = tf.keras.layers.Dense(3, kernel_initializer=initializer)

Arguments

• scale: Scaling factor (positive float).
• mode: One of "fan_in", "fan_out", "fan_avg".
• distribution: Random distribution to use. One of "truncated_normal", "untruncated_normal" and "uniform".
• seed: A Python integer. Used to make the behavior of the initializer deterministic. Note that a seeded initializer will produce the same random values across multiple calls.

Creating custom initializers

Simple callables

You can pass a custom callable as initializer. It must take the arguments shape (shape of the variable to initialize) and dtype (dtype of generated values):

def my_init(shape, dtype=None):
    return tf.random.normal(shape, dtype=dtype)

layer = Dense(64, kernel_initializer=my_init)

Initializer subclasses

If you need to configure your initializer via various arguments (e.g. stddev argument in RandomNormal), you should implement it as a subclass of tf.keras.initializers.Initializer.

Initializers should implement a __call__ method with the following signature:

def __call__(self, shape, dtype=None)`:
    # returns a tensor of shape `shape` and dtype `dtype`
    # containing values drawn from a distribution of your choice.

Optionally, you an also implement the method get_config and the class method from_config in order to support serialization -- just like with any Keras object.

Here's a simple example: a random normal initializer.

import tensorflow as tf

class ExampleRandomNormal(tf.keras.initializers.Initializer):

    def __init__(self, mean, stddev):
      self.mean = mean
      self.stddev = stddev

    def __call__(self, shape, dtype=None)`:
      return tf.random.normal(
          shape, mean=self.mean, stddev=self.stddev, dtype=dtype)

    def get_config(self):  # To support serialization
      return {'mean': self.mean, 'stddev': self.stddev}

Note that we don't have to implement from_config in the example above since the constructor arguments of the class the keys in the config returned by get_config are the same. In this case, the default from_config works fine.