fast-style-transfer/model.py at master · DreamAndDead/fast-style-transfer · GitHub

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
import tensorflow as tf
from tensorflow import keras


class InstanceNorm(keras.layers.Layer):
    def __init__(self, **kwargs):
        super(InstanceNorm, self).__init__(**kwargs)

    def build(self, input_shape):
        b, h, w, c = input_shape

        self.shift = self.add_weight(name='shift',
                                     shape=(c,),
                                     initializer='zeros',
                                     trainable=True)
        self.scale = self.add_weight(name='scale',
                                     shape=(c,),
                                     initializer='ones',
                                     trainable=True)
        self.epsilon = 1e-3

        self.built = True

    def call(self, inputs):
        mu, sigma_sq = tf.nn.moments(x=inputs, axes=[1, 2], keepdims=True) # (T, 1, 1, C)
        normalized = (inputs - mu) / ((sigma_sq + self.epsilon) ** (0.5))
        return self.scale * normalized + self.shift


class Conv2D(keras.layers.Layer):
    def __init__(self, filter_num, filter_size, stride, padding='valid', relu=True, **kwargs):
        super(Conv2D, self).__init__(**kwargs)
        self.filter_num = filter_num
        self.filter_size = filter_size
        self.stride = stride
        self.padding = padding
        self.relu_activation = relu

    def build(self, input_shape):
        self.conv = keras.layers.Conv2D(self.filter_num, self.filter_size, self.stride, self.padding)
        self.norm = InstanceNorm()
        #self.norm = keras.layers.BatchNormalization()
        self.relu = keras.layers.ReLU()

        self.built = True

    def call(self, inputs):
        x = self.conv(inputs)
        x = self.norm(x)
        if self.relu_activation:
            x = self.relu(x)

        return x


class Conv2DTranspose(keras.layers.Layer):
    def __init__(self, filter_num, filter_size, stride, padding='valid', relu=True, **kwargs):
        super(Conv2DTranspose, self).__init__(**kwargs)
        self.filter_num = filter_num
        self.filter_size = filter_size
        self.stride = stride
        self.padding = padding
        self.relu_activation = relu

    def build(self, input_shape):
        self.conv_t = keras.layers.Conv2DTranspose(self.filter_num, self.filter_size, self.stride, self.padding)
        self.norm = InstanceNorm()
        #self.norm = keras.layers.BatchNormalization()
        self.relu = keras.layers.ReLU()

        self.built = True

    def call(self, inputs):
        x = self.conv_t(inputs)
        x = self.norm(x)
        if self.relu_activation:
            x = self.relu(x)

        return x


class Res(keras.layers.Layer):
    def __init__(self, filter_num, filter_size, stride, **kwargs):
        super(Res, self).__init__(**kwargs)
        self.filter_num = filter_num
        self.filter_size = filter_size
        self.stride = stride

    def build(self, input_shape):
        self.conv1 = Conv2D(self.filter_num, self.filter_size, self.stride)
        self.conv2 = Conv2D(self.filter_num, self.filter_size, self.stride, relu=False)

        self.built = True

    def call(self, inputs):
        _, h1, w1, _ = inputs.shape

        x = self.conv1(inputs)
        x = self.conv2(x)

        _, h2, w2, _ = x.shape

        h_diff = (h1 - h2) // 2
        w_diff = (w1 - w2) // 2

        crop_inputs = inputs[:, h_diff:-h_diff, w_diff:-w_diff, :]

        return crop_inputs + x


class FastTransferModel(keras.models.Model):
    def __init__(self):
        super(FastTransferModel, self).__init__()

    def build(self, input_shape):
        self.conv1 = Conv2D(32, 9, 1, 'same')
        self.conv2 = Conv2D(64, 3, 2, 'same')
        self.conv3 = Conv2D(128, 3, 2, 'same')

        self.resid1 = Res(128, 3, 1)
        self.resid2 = Res(128, 3, 1)
        self.resid3 = Res(128, 3, 1)
        self.resid4 = Res(128, 3, 1)
        self.resid5 = Res(128, 3, 1)

        self.conv_t1 = Conv2DTranspose(64, 3, 2, 'same')
        self.conv_t2 = Conv2DTranspose(32, 3, 2, 'same')
        self.conv4 = Conv2D(3, 9, 1, 'same', relu=False)

        self.tanh1 = keras.layers.Activation('tanh')
        self.lamb1 = keras.layers.Lambda(lambda x: x * 150.0 + 255.0/2)

    def call(self, inputs):
        paddings = tf.constant([
            [0, 0],
            [40, 40],
            [40, 40],
            [0, 0]
        ])
        inputs = tf.pad(inputs, paddings, "REFLECT")

        x = self.conv1(inputs)
        x = self.conv2(x)
        x = self.conv3(x)

        x = self.resid1(x)
        x = self.resid2(x)
        x = self.resid3(x)
        x = self.resid4(x)
        x = self.resid5(x)

        x = self.conv_t1(x)
        x = self.conv_t2(x)
        x = self.conv4(x)
        x = self.tanh1(x)
        x = self.lamb1(x)

        return x