-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathcustom_layers.py
186 lines (145 loc) · 7.6 KB
/
custom_layers.py
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
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
import tensorflow as tf
import tensorflow.keras.backend as K
class ReflectionPadding2D(tf.keras.layers.Layer):
def __init__(self, padding=(1, 1), **kwargs):
self.padding = tuple(padding)
super(ReflectionPadding2D, self).__init__(**kwargs)
def compute_output_shape(self, input_shape):
return (input_shape[0], input_shape[1] + 2 * self.padding[0], input_shape[2] + 2 * self.padding[1], input_shape[3])
def call(self, input_tensor, mask=None):
padding_width, padding_height = self.padding
return tf.pad(input_tensor, [[0,0], [padding_height, padding_height], [padding_width, padding_width], [0,0] ], 'REFLECT')
class SymmetricPadding2D(tf.keras.layers.Layer):
def __init__(self, padding=(1, 1), **kwargs):
self.padding = tuple(padding)
super(SymmetricPadding2D, self).__init__(**kwargs)
def compute_output_shape(self, input_shape):
return (input_shape[0], input_shape[1] + 2 * self.padding[0], input_shape[2] + 2 * self.padding[1], input_shape[3])
def call(self, input_tensor, mask=None):
padding_width, padding_height = self.padding
return tf.pad(input_tensor, [[0,0], [padding_height, padding_height], [padding_width, padding_width], [0,0] ], 'SYMMETRIC')
class ResizeLayer(tf.keras.layers.Layer):
def __init__(self, **kwargs):
super(ResizeLayer, self).__init__(**kwargs)
def call(self, inputs, shape):
return tf.image.resize(inputs, shape, method='nearest')
class AdaIN(tf.keras.layers.Layer):
def __init__(self, epsilon=1e-5, **kwargs):
self.eps = epsilon
super(AdaIN, self).__init__(**kwargs)
def get_config(self):
config = super().get_config().copy()
config.update({ 'epsilon': self.eps })
return config
def call(self, inputs):
content, style, alpha = inputs
axes = [1, 2]
c_mean, c_var = tf.nn.moments(content, axes=axes, keepdims=True)
s_mean, s_var = tf.nn.moments(style, axes=axes, keepdims=True)
c_std, s_std = tf.math.sqrt(c_var + self.eps), tf.math.sqrt(s_var + self.eps)
adain = s_std * (content - c_mean) / c_std + s_mean
return alpha * adain + (1-alpha) * content
class ResidualBlock(tf.keras.layers.Layer):
def __init__(self, filters=32, **kwargs):
super(ResidualBlock, self).__init__(**kwargs)
self.conv1 = tf.keras.layers.Conv2D(filters, kernel_size=3, strides=1, padding='valid', activation='selu')
self.rpad1 = ReflectionPadding2D((1,1))
self.conv2 = tf.keras.layers.Conv2D(filters, kernel_size=3, strides=1, padding='valid', activation='selu')
self.rpad2 = ReflectionPadding2D((1,1))
self.add = tf.keras.layers.Add()
@tf.function
def call(self, inputs):
x = self.conv1(inputs)
x = self.rpad1(x)
x = self.conv2(x)
x = self.rpad2(x)
x = self.add([x, inputs])
return x
class PUPSBlock(tf.keras.layers.Layer):
def __init__(self, filter_size, **kwargs):
super(PUPSBlock, self).__init__(**kwargs)
self.concat = tf.keras.layers.Concatenate(axis=-1)
self.conv = tf.keras.layers.Conv2D(filter_size, 1, strides=1)
self.resblock = ResidualBlock(filters=filter_size)
# self.upsample = tf.keras.layers.UpSampling2D((2, 2), interpolation='bilinear')
self.upsample = ResizeLayer()
self.outputs = tf.keras.layers.Conv2D(output_shape, 3, padding='valid', activation='selu')
@tf.function
def call(self, inputs, shape):
x = self.concat(inputs)
x = self.conv(x)
x = self.resblock(x)
x = self.upsample(x, shape)
x = self.outputs(x)
return x
class SqueezeExciteBlock(tf.keras.layers.Layer):
def __init__(self, ratio=8, axis=-1, **kwargs):
super(SqueezeExciteBlock, self).__init__(**kwargs)
self.ratio = ratio
self.axis = axis
def build(self, input_shape):
super(SqueezeExciteBlock, self).build(input_shape)
self.GlobalAvgPooling = tf.keras.layers.GlobalAveragePooling2D()
self.GlobalAvgPooling.build(input_shape)
self.filters = input_shape[self.axis]
self.dense1 = tf.keras.layers.Dense(self.filters // self.ratio, activation='relu', use_bias=False)
self.dense2 = tf.keras.layers.Dense(self.filters, activation='sigmoid', use_bias=False)
def call(self, inputs, **kwargs):
avg = self.GlobalAvgPooling(inputs)
se_shape = (inputs.shape[0], 1, 1, self.filters)
se = tf.reshape(avg, se_shape)
d1 = self.dense1(se)
d2 = self.dense2(d1)
x = tf.multiply(inputs, d2)
return x
class BatchAttNorm(tf.keras.layers.BatchNormalization):
def __init__(self, momentum=0.99, epsilon=0.001, axis=-1, **kwargs):
super(BatchAttNorm, self).__init__(momentum=momentum, epsilon=epsilon, axis=axis, center=False, scale=False, **kwargs)
# if self.axis == -1: self.data_format = 'channels_last'
# else: self.data_format = 'channel_first'
def build(self, input_shape):
if len(input_shape) != 4: raise ValueError('expected 4D input (got {}D input)'.format(input_shape))
super(BatchAttNorm, self).build(input_shape)
self.GlobalAvgPooling = tf.keras.layers.GlobalAveragePooling2D("channels_last")
self.GlobalAvgPooling.build(input_shape)
self.weight = self.add_weight(name='weight', shape=input_shape[-1], initializer=tf.keras.initializers.Constant(1), trainable=True)
self.bias = self.add_weight(name='bias', shape=input_shape[-1], initializer=tf.keras.initializers.Constant(0), trainable=True)
self.weight_readjust = self.add_weight(name='weight_readjust', shape=input_shape[-1], initializer=tf.keras.initializers.Constant(0), trainable=True)
self.bias_readjust = self.add_weight(name='bias_readjust', shape=input_shape[-1], initializer=tf.keras.initializers.Constant(-1), trainable=True)
def call(self, inputs):
avg = self.GlobalAvgPooling(inputs)
attention = K.sigmoid(avg * self.weight_readjust + self.bias_readjust)
bn_weights = self.weight * attention
out_bn = super(BatchAttNorm, self).call(inputs)
if K.int_shape(inputs)[0] is None or K.int_shape(inputs)[0] > 1:
bn_weights = bn_weights[:, None, None, :]
# self.bias = self.bias[None, None, None, :]
return out_bn * bn_weights + self.bias
class BatchInstanceNormalization(tf.keras.layers.Layer):
"""Batch Instance Normalization Layer (https://arxiv.org/abs/1805.07925)."""
def __init__(self, epsilon=1e-5):
super(BatchInstanceNormalization, self).__init__()
self.epsilon = epsilon
def build(self, input_shape):
self.rho = self.add_weight(
name='rho',
shape=input_shape[-1:],
initializer='ones',
constraint=lambda x: tf.clip_by_value(x, clip_value_min=0.0, clip_value_max=1.0),
trainable=True)
self.gamma = self.add_weight(
name='gamma',
shape=input_shape[-1:],
initializer='ones',
trainable=True)
self.beta = self.add_weight(
name='beta',
shape=input_shape[-1:],
initializer='zeros',
trainable=True)
def call(self, x):
batch_mean, batch_sigma = tf.nn.moments(x, axes=[0, 1, 2], keepdims=True)
x_batch = (x - batch_mean) * (tf.math.rsqrt(batch_sigma + self.epsilon))
ins_mean, ins_sigma = tf.nn.moments(x, axes=[1, 2], keepdims=True)
x_ins = (x - ins_mean) * (tf.math.rsqrt(ins_sigma + self.epsilon))
return (self.rho * x_batch + (1 - self.rho) * x_ins) * self.gamma + self.beta