Module ainshamsflow.regularizers
Regularizers Module.
In this Module, we include our Regularizers such as L1 and L2.
Expand source code
"""Regularizers Module.
In this Module, we include our Regularizers such as L1 and L2.
"""
import numpy as np
from ainshamsflow.utils.asf_errors import BaseClassError, NameNotFoundError
def get(reg_name):
"""Get any Regularizer in this Module by name."""
regs = [L2, L1, L1_L2]
for reg in regs:
if reg.__name__.lower() == reg_name.lower():
return reg()
else:
raise NameNotFoundError(reg_name, __name__)
class Regularizer:
"""Regularizer Base Class.
Used to define Regularizer Interface.
To create a new Regularizer, create a class that inherits from
this class.
You then have to add any extra parameters in your constructor
(while still calling this class' constructor) and redefining
the __call__() and diff() methods.
"""
def __init__(self, lambd=0.01):
"""Initialize Lambda."""
self.lambd = lambd
def __call__(self, weights_list, m):
raise BaseClassError
def diff(self, weights_list, m):
raise BaseClassError
class L2(Regularizer):
"""L2 Reguarizer."""
__name__ = 'L2'
def __call__(self, weights_list, m):
if isinstance(weights_list, list):
ans = 0
for weights in weights_list:
ans += self.__call__(weights, m)
return ans
else:
return self.lambd * np.sum(np.square(weights_list)) / (2*m)
def diff(self, weights_list, m):
if isinstance(weights_list, list):
ans = []
for weights in weights_list:
ans.append(self.diff(weights, m))
return ans
else:
return self.lambd * np.divide(weights_list, m)
class L1(Regularizer):
"""L1 Regularizer."""
__name__ = 'L1'
def __call__(self, weights_list, m):
if isinstance(weights_list, list):
ans = 0
for weights in weights_list:
ans += self.__call__(weights, m)
return ans
else:
return self.lambd * np.sum(np.abs(weights_list)) / m
def diff(self, weights_list, m):
if isinstance(weights_list, list):
ans = []
for weights in weights_list:
ans.append(self.diff(weights, m))
return ans
else:
return self.lambd * np.divide(np.where(weights_list > 0, 1, -1), m)
class L1_L2(Regularizer):
"""L1_L2 Regularizer."""
__name__ = 'L1_L2'
def __init__(self, l1=0.01, l2=0.01):
super().__init__()
self.l1_reg = L1(l1)
self.l2_reg = L2(l2)
def __call__(self, weights_list, m):
return self.l1_reg(weights_list, m) + self.l2_reg(weights_list, m)
def diff(self, weights_list, m):
if isinstance(weights_list, list):
ans = []
for weights in weights_list:
ans.append(self.diff(weights, m))
return ans
else:
return self.l1_reg.diff(weights_list, m) + self.l2_reg.diff(weights_list, m)Functions
def get(reg_name)-
Get any Regularizer in this Module by name.
Expand source code
def get(reg_name): """Get any Regularizer in this Module by name.""" regs = [L2, L1, L1_L2] for reg in regs: if reg.__name__.lower() == reg_name.lower(): return reg() else: raise NameNotFoundError(reg_name, __name__)
Classes
class L1 (lambd=0.01)-
L1 Regularizer.
Initialize Lambda.
Expand source code
class L1(Regularizer): """L1 Regularizer.""" __name__ = 'L1' def __call__(self, weights_list, m): if isinstance(weights_list, list): ans = 0 for weights in weights_list: ans += self.__call__(weights, m) return ans else: return self.lambd * np.sum(np.abs(weights_list)) / m def diff(self, weights_list, m): if isinstance(weights_list, list): ans = [] for weights in weights_list: ans.append(self.diff(weights, m)) return ans else: return self.lambd * np.divide(np.where(weights_list > 0, 1, -1), m)Ancestors
Methods
def diff(self, weights_list, m)-
Expand source code
def diff(self, weights_list, m): if isinstance(weights_list, list): ans = [] for weights in weights_list: ans.append(self.diff(weights, m)) return ans else: return self.lambd * np.divide(np.where(weights_list > 0, 1, -1), m)
class L1_L2 (l1=0.01, l2=0.01)-
L1_L2 Regularizer.
Initialize Lambda.
Expand source code
class L1_L2(Regularizer): """L1_L2 Regularizer.""" __name__ = 'L1_L2' def __init__(self, l1=0.01, l2=0.01): super().__init__() self.l1_reg = L1(l1) self.l2_reg = L2(l2) def __call__(self, weights_list, m): return self.l1_reg(weights_list, m) + self.l2_reg(weights_list, m) def diff(self, weights_list, m): if isinstance(weights_list, list): ans = [] for weights in weights_list: ans.append(self.diff(weights, m)) return ans else: return self.l1_reg.diff(weights_list, m) + self.l2_reg.diff(weights_list, m)Ancestors
Methods
def diff(self, weights_list, m)-
Expand source code
def diff(self, weights_list, m): if isinstance(weights_list, list): ans = [] for weights in weights_list: ans.append(self.diff(weights, m)) return ans else: return self.l1_reg.diff(weights_list, m) + self.l2_reg.diff(weights_list, m)
class L2 (lambd=0.01)-
L2 Reguarizer.
Initialize Lambda.
Expand source code
class L2(Regularizer): """L2 Reguarizer.""" __name__ = 'L2' def __call__(self, weights_list, m): if isinstance(weights_list, list): ans = 0 for weights in weights_list: ans += self.__call__(weights, m) return ans else: return self.lambd * np.sum(np.square(weights_list)) / (2*m) def diff(self, weights_list, m): if isinstance(weights_list, list): ans = [] for weights in weights_list: ans.append(self.diff(weights, m)) return ans else: return self.lambd * np.divide(weights_list, m)Ancestors
Methods
def diff(self, weights_list, m)-
Expand source code
def diff(self, weights_list, m): if isinstance(weights_list, list): ans = [] for weights in weights_list: ans.append(self.diff(weights, m)) return ans else: return self.lambd * np.divide(weights_list, m)
class Regularizer (lambd=0.01)-
Regularizer Base Class.
Used to define Regularizer Interface.
To create a new Regularizer, create a class that inherits from this class. You then have to add any extra parameters in your constructor (while still calling this class' constructor) and redefining the call() and diff() methods.
Initialize Lambda.
Expand source code
class Regularizer: """Regularizer Base Class. Used to define Regularizer Interface. To create a new Regularizer, create a class that inherits from this class. You then have to add any extra parameters in your constructor (while still calling this class' constructor) and redefining the __call__() and diff() methods. """ def __init__(self, lambd=0.01): """Initialize Lambda.""" self.lambd = lambd def __call__(self, weights_list, m): raise BaseClassError def diff(self, weights_list, m): raise BaseClassErrorSubclasses
Methods
def diff(self, weights_list, m)-
Expand source code
def diff(self, weights_list, m): raise BaseClassError