/
model.py
167 lines (126 loc) · 5.6 KB
/
model.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
import keras
from keras import backend as K
from keras import layers
import criterion
class ConvBnRelu(object):
def __init__(self, filters):
self.conv = layers.Conv2D(filters, (3, 3), padding="same")
self.bn = layers.BatchNormalization()
self.relu = layers.ReLU()
def __call__(self, inputs):
return self.relu(self.bn(self.conv(inputs)))
class EncoderBlock(object):
def __init__(self, filters):
self.conv_bn_relu = ConvBnRelu(filters)
self.pool = layers.MaxPooling2D(pool_size=(2, 2), strides=(2, 2))
def __call__(self, inputs):
h = self.conv_bn_relu(inputs)
return h, self.pool(h)
class DecoderBlock(object):
def __init__(self, filters):
self.concat = layers.Concatenate()
self.deconv = layers.Conv2DTranspose(filters // 2, (3, 3), strides=(2, 2), padding="same")
self.conv_bn_relu = ConvBnRelu(filters)
def __call__(self, input_, feature_map):
inputs = self.concat([input_, feature_map])
h = self.conv_bn_relu(inputs)
output = self.deconv(h)
return output, h
class Encoder(object):
def __init__(self, layers_filters):
self.encoder_blocks = []
for filters in layers_filters:
self.encoder_blocks.append(EncoderBlock(filters))
def __call__(self, inputs):
pooled = inputs
feature_maps = []
for encoder_block in self.encoder_blocks:
feature_map, pooled = encoder_block(pooled)
feature_maps.append(feature_map)
return pooled, feature_maps
class Decoder(object):
def __init__(self, layers_filters):
self.decoder_blocks = []
for filters in layers_filters:
self.decoder_blocks.append(DecoderBlock(filters))
def __call__(self, inputs, feature_maps):
for feature_map, decoder_block in zip(feature_maps, self.decoder_blocks):
inputs, h = decoder_block(inputs, feature_map)
return h
class Q(object):
def __init__(self, encoder_layers_filters, latent_size):
self.concat = layers.Concatenate()
self.flatten = layers.Flatten()
self.dense1 = layers.Dense(latent_size, name="mean")
self.dense2 = layers.Dense(latent_size, name="log_var")
self.encoder = Encoder(encoder_layers_filters)
def __call__(self, input_, context_input):
inputs = self.concat([input_, context_input])
h, _ = self.encoder(inputs)
flat = self.flatten(h)
mean = self.dense1(flat)
log_var = self.dense2(flat)
return mean, log_var
class P(object):
def __init__(self, encoder_layers_filters, decoder_layers_filters, latent_size):
self.concat = layers.Concatenate()
self.flatten = layers.Flatten()
self.dense = layers.Dense(latent_size)
self.reshape = layers.Reshape((1, 1, latent_size))
self.deconv = layers.Conv2DTranspose(1024, (3, 3), strides=(2, 2), padding="same")
self.encoder = Encoder(encoder_layers_filters)
self.decoder = Decoder(decoder_layers_filters)
def __call__(self, z, context_input):
pooled, feature_maps = self.encoder(context_input)
flat = self.flatten(pooled)
merged = self.concat([flat, z])
inputs = self.reshape(self.dense(merged))
output = self.decoder(self.deconv(inputs), reversed(feature_maps))
return output
def sample_z(args):
(noise, mean, log_var) = args
return K.exp(log_var / 2) * noise + mean
class VAE(object):
def __init__(self, lr=0.001):
latent_size = 512
input_shape = (256, 256, 3)
context_shape = (256, 256, 6)
encoder_layers_filters = [64, 128, 256, 512, 512, 512, 512, 512]
decoder_layers_filters = encoder_layers_filters[::-1]
decoder_layers_filters[-1] = 3
q = Q(encoder_layers_filters, latent_size)
p = P(encoder_layers_filters, decoder_layers_filters, latent_size)
input_ = layers.Input(shape=input_shape, name="input_frame")
context_input = layers.Input(shape=context_shape, name="input_ctx")
noise = layers.Input(shape=(latent_size,), name="input_noise")
z_test = layers.Input(shape=(latent_size,), name="z_test")
mean, log_var = q(input_, context_input)
z = layers.Lambda(sample_z, name="z")([noise, mean, log_var])
pred_train = p(z, context_input)
z_param = layers.Concatenate(axis=-2, name="z_params")([
layers.Reshape((1, latent_size))(mean),
layers.Reshape((1, latent_size))(log_var)
])
pred_test = p(z_test, context_input)
model_train = keras.models.Model(inputs=[input_, noise, context_input],
outputs=[pred_train, z_param])
# keras.utils.plot_model(model_train, to_file="model.png")
# model_train.summary()
model_test = keras.models.Model(inputs=[z_test, context_input],
outputs=[pred_test])
# losses = [criterion.mse, criterion.kl]
losses = [criterion.l1, criterion.kl]
# losses = [criterion.neg_ssim, criterion.kl]
metrics = [criterion.scaled_mse, criterion.psnr, criterion.ssim]
optimizer = keras.optimizers.Adam(lr)
model_train.compile(optimizer=optimizer, loss=losses)
model_test.compile(optimizer=optimizer, loss=criterion.mse, metrics=metrics)
self.model_train = model_train
self.model_test = model_test
def save(self, filename):
self.model_train.save_weights(filename)
def load(self, filename):
self.model_train.load_weights(filename)
self.model_test.load_weights(filename, by_name=True)
if __name__ == "__main__":
vae = VAE()