-
Notifications
You must be signed in to change notification settings - Fork 1
Expand file tree
/
Copy pathfrn.py
More file actions
158 lines (136 loc) · 6.51 KB
/
frn.py
File metadata and controls
158 lines (136 loc) · 6.51 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
# based on the following manuscript
# https://arxiv.org/abs/1911.09737
# https://github.com/amirbar/FilterResponseNormalization/blob/master/frn.py
import tensorflow as tf
from keras.layers import Layer, InputSpec
from keras import initializers, regularizers, constraints
class FRN(Layer):
"""
Filter Response Normalization
"""
def __init__(self,
axis=-1,
epsilon=1e-6,
learnable_epsilon=False,
beta_initializer='zeros',
gamma_initializer='ones',
beta_regularizer=None,
gamma_regularizer=None,
epsilon_regularizer=None,
beta_constraint=None,
gamma_constraint=None,
epsilon_constraint=None,
**kwargs):
'''
:param axis: channels axis
:param epsilon: for numeric stability (should be initialized to 1e-4 if learnable, or set to 1e-6 otherwise, cf. paper)
:param learnable_epsilon: turn epsilon to trainable
'''
super(FRN, self).__init__(**kwargs)
self.supports_masking = True
self.axis = axis
self.epsilon = epsilon
self.learnable_epsilon = learnable_epsilon
self.beta_initializer = initializers.get(beta_initializer)
self.gamma_initializer = initializers.get(gamma_initializer)
self.beta_regularizer = regularizers.get(beta_regularizer)
self.gamma_regularizer = regularizers.get(gamma_regularizer)
self.epsilon_regularizer = regularizers.get(epsilon_regularizer)
self.beta_constraint = constraints.get(beta_constraint)
self.gamma_constraint = constraints.get(gamma_constraint)
self.epsilon_constraint = constraints.get(epsilon_constraint)
def build(self, input_shape):
dim = input_shape[self.axis]
if dim is None:
raise ValueError('Axis ' + str(self.axis) + ' of '
'input tensor should have a defined dimension '
'but the layer received an input with shape ' +
str(input_shape) + '.')
self.input_spec = InputSpec(ndim=len(input_shape),
axes={self.axis: dim})
shape = (dim,)
self.gamma = self.add_weight(shape=shape,
name='gamma',
initializer=self.gamma_initializer,
regularizer=self.gamma_regularizer,
constraint=self.gamma_constraint)
self.beta = self.add_weight(shape=shape,
name='beta',
initializer=self.beta_initializer,
regularizer=self.beta_regularizer,
constraint=self.beta_constraint)
self.epsilon_l = self.add_weight(shape=(1,),
name='epsilon_l',
initializer=initializers.Constant(self.epsilon),
regularizer=self.epsilon_regularizer,
constraint=self.epsilon_constraint,
trainable=self.learnable_epsilon)
self.built = True
def call(self, x, **kwargs):
nu2 = tf.reduce_mean(tf.square(x), axis=list(range(1, x.shape.ndims - 1)), keepdims=True)
# Perform FRN.
x = x * tf.rsqrt(nu2 + tf.abs(self.epsilon_l))
return self.gamma * x + self.beta
def get_config(self):
config = {
'epsilon': self.epsilon,
'learnable_epsilon': self.learnable_epsilon,
'beta_initializer': initializers.serialize(self.beta_initializer),
'gamma_initializer': initializers.serialize(self.gamma_initializer),
'beta_regularizer': regularizers.serialize(self.beta_regularizer),
'gamma_regularizer': regularizers.serialize(self.gamma_regularizer),
'epsilon_regularizer': regularizers.serialize(self.epsilon_regularizer),
'beta_constraint': constraints.serialize(self.beta_constraint),
'gamma_constraint': constraints.serialize(self.gamma_constraint),
'epsilon_constraint': constraints.serialize(self.epsilon_constraint),
}
base_config = super(FRN, self).get_config()
return dict(list(base_config.items()) + list(config.items()))
def compute_output_shape(self, input_shape):
return input_shape
class TLU(Layer):
"""
Thresholded Linear Unit: augmented ReLU with a learned threshold (tau)
"""
def __init__(self,
axis=-1,
tau_initializer='zeros',
tau_regularizer=None,
tau_constraint=None,
**kwargs):
'''
:param axis: channels axis
'''
super(TLU, self).__init__(**kwargs)
self.axis = axis
self.tau_initializer = initializers.get(tau_initializer)
self.tau_regularizer = regularizers.get(tau_regularizer)
self.tau_constraint = constraints.get(tau_constraint)
def build(self, input_shape):
dim = input_shape[self.axis]
if dim is None:
raise ValueError('Axis ' + str(self.axis) + ' of '
'input tensor should have a defined dimension '
'but the layer received an input with shape ' +
str(input_shape) + '.')
self.input_spec = InputSpec(ndim=len(input_shape),
axes={self.axis: dim})
shape = (dim,)
self.tau = self.add_weight(shape=shape,
name='tau',
initializer=self.tau_initializer,
regularizer=self.tau_regularizer,
constraint=self.tau_constraint)
self.built = True
def call(self, x, **kwargs):
return tf.maximum(x, self.tau)
def get_config(self):
config = {
'tau_initializer': initializers.serialize(self.tau_initializer),
'tau_regularizer': regularizers.serialize(self.tau_regularizer),
'tau_constraint': constraints.serialize(self.tau_constraint)
}
base_config = super(TLU, self).get_config()
return dict(list(base_config.items()) + list(config.items()))
def compute_output_shape(self, input_shape):
return input_shape