Regularizers allow you to apply penalties on layer parameters or layer activity during optimization. These penalties are summed into the loss function that the network optimizes.
Regularization penalties are applied on a per-layer basis. The exact API will
depend on the layer, but many layers (e.g. Dense
, Conv1D
, Conv2D
and
Conv3D
) have a unified API.
These layers expose 3 keyword arguments:
kernel_regularizer
: Regularizer to apply a penalty on the layer's kernelbias_regularizer
: Regularizer to apply a penalty on the layer's biasactivity_regularizer
: Regularizer to apply a penalty on the layer's outputfrom tensorflow.keras import layers
from tensorflow.keras import regularizers
layer = layers.Dense(
units=64,
kernel_regularizer=regularizers.L1L2(l1=1e-5, l2=1e-4),
bias_regularizer=regularizers.L2(1e-4),
activity_regularizer=regularizers.L2(1e-5)
)
The value returned by the activity_regularizer
object gets divided by the input
batch size so that the relative weighting between the weight regularizers and
the activity regularizers does not change with the batch size.
You can access a layer's regularization penalties by calling layer.losses
after calling the layer on inputs:
layer = layers.Dense(units=5,
kernel_initializer='ones',
kernel_regularizer=regularizers.L1(0.01),
activity_regularizer=regularizers.L2(0.01))
tensor = tf.ones(shape=(5, 5)) * 2.0
out = layer(tensor)
# The kernel regularization term is 0.25
# The activity regularization term (after dividing by the batch size) is 5
print(tf.math.reduce_sum(layer.losses)) # 5.25 (= 5 + 0.25)
The following built-in regularizers are available as part of the tf.keras.regularizers
module:
L1
classtf.keras.regularizers.L1(l1=0.01, **kwargs)
A regularizer that applies a L1 regularization penalty.
The L1 regularization penalty is computed as:
loss = l1 * reduce_sum(abs(x))
L1 may be passed to a layer as a string identifier:
>>> dense = tf.keras.layers.Dense(3, kernel_regularizer='l1')
In this case, the default value used is l1=0.01
.
Arguments
L2
classtf.keras.regularizers.L2(l2=0.01, **kwargs)
A regularizer that applies a L2 regularization penalty.
The L2 regularization penalty is computed as:
loss = l2 * reduce_sum(square(x))
L2 may be passed to a layer as a string identifier:
>>> dense = tf.keras.layers.Dense(3, kernel_regularizer='l2')
In this case, the default value used is l2=0.01
.
Arguments
L1L2
classtf.keras.regularizers.L1L2(l1=0.0, l2=0.0)
A regularizer that applies both L1 and L2 regularization penalties.
The L1 regularization penalty is computed as:
loss = l1 * reduce_sum(abs(x))
The L2 regularization penalty is computed as
loss = l2 * reduce_sum(square(x))
L1L2 may be passed to a layer as a string identifier:
>>> dense = tf.keras.layers.Dense(3, kernel_regularizer='l1_l2')
In this case, the default values used are l1=0.01
and l2=0.01
.
Arguments
OrthogonalRegularizer
classtf.keras.regularizers.OrthogonalRegularizer(factor=0.01, mode="rows")
Regularizer that encourages input vectors to be orthogonal to each other.
It can be applied to either the rows of a matrix (mode="rows"
) or its
columns (mode="columns"
). When applied to a Dense
kernel of shape
(input_dim, units)
, rows mode will seek to make the feature vectors
(i.e. the basis of the output space) orthogonal to each other.
Arguments
factor
times the mean of the dot products between
the L2-normalized rows (if mode="rows"
, or columns if
mode="columns"
) of the inputs, excluding the product of each
row/column with itself. Defaults to 0.01.{"rows", "columns"}
. Defaults to "rows"
. In rows
mode, the regularization effect seeks to make the rows of the input
orthogonal to each other. In columns mode, it seeks to make the columns
of the input orthogonal to each other.Example
>>> regularizer = tf.keras.regularizers.OrthogonalRegularizer(factor=0.01)
>>> layer = tf.keras.layers.Dense(units=4, kernel_regularizer=regularizer)
A weight regularizer can be any callable that takes as input a weight tensor
(e.g. the kernel of a Conv2D
layer), and returns a scalar loss. Like this:
def my_regularizer(x):
return 1e-3 * tf.reduce_sum(tf.square(x))
Regularizer
subclassesIf you need to configure your regularizer via various arguments
(e.g. l1
and l2
arguments in l1_l2
),
you should implement it as a subclass of tf.keras.regularizers.Regularizer
.
Here's a simple example:
class MyRegularizer(regularizers.Regularizer):
def __init__(self, strength):
self.strength = strength
def __call__(self, x):
return self.strength * tf.reduce_sum(tf.square(x))
Optionally, you can also implement the method get_config
and the class
method from_config
in order to support serialization -- just like with
any Keras object. Example:
class MyRegularizer(regularizers.Regularizer):
def __init__(self, strength):
self.strength = strength
def __call__(self, x):
return self.strength * tf.reduce_sum(tf.square(x))
def get_config(self):
return {'strength': self.strength}