"""
Regularization methods designed to work with spiking layers.
"""
import tensorflow as tf
[docs]class L1L2(tf.keras.regularizers.L1L2):
"""
A version of ``tf.keras.regularizers.L1L2`` that allows the user to specify a
nonzero target output.
Parameters
----------
l1 : float
Weight on L1 regularization penalty.
l2 : float
Weight on L2 regularization penalty.
target : float
Target output value (values will be penalized based on their distance from
this point).
"""
def __init__(self, l1=0.0, l2=0.0, target=0, **kwargs):
super().__init__(l1=l1, l2=l2, **kwargs)
self.target = target
def __call__(self, x):
return super().__call__(x - self.target)
[docs] def get_config(self):
"""Return config (for serialization during model saving/loading)."""
cfg = super().get_config()
cfg["target"] = self.target
return cfg
[docs]class L1(L1L2):
"""
A version of ``tf.keras.regularizers.L1`` that allows the user to specify a
nonzero target output.
Parameters
----------
l1 : float
Weight on L1 regularization penalty.
target : float
Target output value (values will be penalized based on their distance from
this point).
"""
def __init__(self, l1=0.01, target=0, **kwargs):
super().__init__(l1=l1, target=target, **kwargs)
[docs]class L2(L1L2):
"""
A version of ``tf.keras.regularizers.L2`` that allows the user to specify a
nonzero target output.
Parameters
----------
l2 : float
Weight on L2 regularization penalty.
target : float
Target output value (values will be penalized based on their distance from
this point).
"""
def __init__(self, l2=0.01, target=0, **kwargs):
super().__init__(l2=l2, target=target, **kwargs)
