This repository has been archived by the owner on Jan 20, 2021. It is now read-only.
-
Notifications
You must be signed in to change notification settings - Fork 2
/
generator.py
247 lines (186 loc) · 7.44 KB
/
generator.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
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
from keras.models import load_model
import numpy as np
from keras.optimizers import Adam
from keras.models import Model
from keras.layers import Dense, Conv2DTranspose, Conv2D, BatchNormalization, \
Activation, Concatenate, Input, MaxPool2D,\
UpSampling2D, ZeroPadding2D, Lambda, Add
from keras.callbacks import ModelCheckpoint
from keras import backend as K
import keras
import cv2
import os
import librosa
import scipy
from keras.utils import plot_model
import tensorflow as tf
from keras.utils import multi_gpu_model
from discriminator import contrastive_loss
class ModelMGPU(Model):
def __init__(self, ser_model, gpus):
pmodel = multi_gpu_model(ser_model, gpus)
self.__dict__.update(pmodel.__dict__)
self._smodel = ser_model
def __getattribute__(self, attrname):
'''Override load and save methods to be used from the serial-model. The
serial-model holds references to the weights in the multi-gpu model.
'''
# return Model.__getattribute__(self, attrname)
if 'load' in attrname or 'save' in attrname:
return getattr(self._smodel, attrname)
return super(ModelMGPU, self).__getattribute__(attrname)
def conv_block(x, num_filters, kernel_size=3, strides=1, padding='same', act=True):
x = Conv2D(filters=num_filters, kernel_size= kernel_size,
strides=strides, padding=padding)(x)
x = BatchNormalization(momentum=.8)(x)
if act:
x = Activation('relu')(x)
return x
def conv_t_block(x, num_filters, kernel_size=3, strides=2, padding='same'):
x = Conv2DTranspose(filters=num_filters, kernel_size= kernel_size,
strides=strides, padding=padding)(x)
x = BatchNormalization(momentum=.8)(x)
x = Activation('relu')(x)
return x
def create_model(args, mel_step_size):
############# encoder for face/identity
input_face = Input(shape=(args.img_size, args.img_size, 6), name="input_face")
identity_mapping = conv_block(input_face, 32, kernel_size=11) # 96x96
x1_face = conv_block(identity_mapping, 64, kernel_size=7, strides=2) # 48x48
x2_face = conv_block(x1_face, 128, 5, 2) # 24x24
x3_face = conv_block(x2_face, 256, 3, 2) #12x12
x4_face = conv_block(x3_face, 512, 3, 2) #6x6
x5_face = conv_block(x4_face, 512, 3, 2) #3x3
x6_face = conv_block(x5_face, 512, 3, 1, padding='valid')
x7_face = conv_block(x6_face, 256, 1, 1)
############# encoder for audio
input_audio = Input(shape=(80, mel_step_size, 1), name="input_audio")
x = conv_block(input_audio, 32)
x = conv_block(x, 32)
x = conv_block(x, 32)
x = conv_block(x, 64, strides=3) #27X9
x = conv_block(x, 64)
x = conv_block(x, 64)
x = conv_block(x, 128, strides=(3, 1)) #9X9
x = conv_block(x, 128)
x = conv_block(x, 128)
x = conv_block(x, 256, strides=3) #3X3
x = conv_block(x, 256)
x = conv_block(x, 256)
x = conv_block(x, 512, strides=1, padding='valid') #1X1
x = conv_block(x, 512, 1, 1)
embedding = Concatenate(axis=3)([x7_face, x])
############# decoder
x = conv_block(embedding, 512, 1)
x = conv_t_block(embedding, 512, 3, 3)# 3x3
x = Concatenate(axis=3) ([x5_face, x])
x = conv_t_block(x, 512) #6x6
x = Concatenate(axis=3) ([x4_face, x])
x = conv_t_block(x, 256) #12x12
x = Concatenate(axis=3) ([x3_face, x])
x = conv_t_block(x, 128) #24x24
x = Concatenate(axis=3) ([x2_face, x])
x = conv_t_block(x, 64) #48x48
x = Concatenate(axis=3) ([x1_face, x])
x = conv_t_block(x, 32) #96x96
x = Concatenate(axis=3) ([identity_mapping, x])
x = conv_block(x, 16) #96x96
x = conv_block(x, 16) #96x96
x = Conv2D(filters=3, kernel_size=1, strides=1, padding="same") (x)
prediction = Activation("sigmoid", name="prediction")(x)
model = Model(inputs=[input_face, input_audio], outputs=prediction)
model.summary()
if args.n_gpu > 1:
model = ModelMGPU(model , args.n_gpu)
model.compile(loss='mae', optimizer=(Adam(lr=args.lr) if hasattr(args, 'lr') else 'adam'))
return model
def create_model_residual(args, mel_step_size):
def residual_block(inp, num_filters):
x = conv_block(inp, num_filters)
x = conv_block(x, num_filters)
x = Add()([x, inp])
x = Activation('relu') (x)
return x
############# encoder for face/identity
input_face = Input(shape=(args.img_size, args.img_size, 6), name="input_face")
identity_mapping = conv_block(input_face, 32, kernel_size=7) # 96x96
x1_face = conv_block(identity_mapping, 64, kernel_size=5, strides=2) # 48x48
x1_face = residual_block(x1_face, 64)
x1_face = residual_block(x1_face, 64)
x2_face = conv_block(x1_face, 128, 3, 2) # 24x24
x2_face = residual_block(x2_face, 128)
x2_face = residual_block(x2_face, 128)
x2_face = residual_block(x2_face, 128)
x3_face = conv_block(x2_face, 256, 3, 2) #12x12
x3_face = residual_block(x3_face, 256)
x3_face = residual_block(x3_face, 256)
x4_face = conv_block(x3_face, 512, 3, 2) #6x6
x4_face = residual_block(x4_face, 512)
x4_face = residual_block(x4_face, 512)
x5_face = conv_block(x4_face, 512, 3, 2) #3x3
x6_face = conv_block(x5_face, 512, 3, 1, padding='valid')
x7_face = conv_block(x6_face, 512, 1, 1)
############# encoder for audio
input_audio = Input(shape=(80, mel_step_size, 1), name="input_audio")
x = conv_block(input_audio, 32)
x = residual_block(x, 32)
x = residual_block(x, 32)
x = conv_block(x, 64, strides=3) #27X9
x = residual_block(x, 64)
x = residual_block(x, 64)
x = conv_block(x, 128, strides=(3, 1)) #9X9
x = residual_block(x, 128)
x = residual_block(x, 128)
x = conv_block(x, 256, strides=3) #3X3
x = residual_block(x, 256)
x = residual_block(x, 256)
x = conv_block(x, 512, strides=1, padding='valid') #1X1
x = conv_block(x, 512, 1, 1)
embedding = Concatenate(axis=3)([x7_face, x])
############# decoder
x = conv_t_block(embedding, 512, 3, 3)# 3x3
x = Concatenate(axis=3) ([x5_face, x])
x = conv_t_block(x, 512) #6x6
x = residual_block(x, 512)
x = residual_block(x, 512)
x = Concatenate(axis=3) ([x4_face, x])
x = conv_t_block(x, 256) #12x12
x = residual_block(x, 256)
x = residual_block(x, 256)
x = Concatenate(axis=3) ([x3_face, x])
x = conv_t_block(x, 128) #24x24
x = residual_block(x, 128)
x = residual_block(x, 128)
x = Concatenate(axis=3) ([x2_face, x])
x = conv_t_block(x, 64) #48x48
x = residual_block(x, 64)
x = residual_block(x, 64)
x = Concatenate(axis=3) ([x1_face, x])
x = conv_t_block(x, 32) #96x96
x = Concatenate(axis=3) ([identity_mapping, x])
x = conv_block(x, 16) #96x96
x = conv_block(x, 16) #96x96
x = Conv2D(filters=3, kernel_size=1, strides=1, padding="same") (x)
prediction = Activation("sigmoid", name="prediction")(x)
model = Model(inputs=[input_face, input_audio], outputs=prediction)
model.summary()
if args.n_gpu > 1:
model = ModelMGPU(model , args.n_gpu)
model.compile(loss='mae', optimizer=(Adam(lr=args.lr) if hasattr(args, 'lr') else 'adam'))
return model
def create_combined_model(generator, discriminator, args, mel_step_size):
input_face = Input(shape=(args.img_size, args.img_size, 6), name="input_face_comb")
input_audio = Input(shape=(80, mel_step_size, 1), name="input_audio_comb")
fake_face = generator([input_face, input_audio])
discriminator.trainable = False
d = discriminator([fake_face, input_audio])
model = Model([input_face, input_audio], [fake_face, d])
if args.n_gpu > 1:
model = ModelMGPU(model , args.n_gpu)
model.compile(loss=['mae', contrastive_loss],
optimizer=(Adam(lr=args.lr) if hasattr(args, 'lr') else 'adam'),
loss_weights=[1., .01])
return model
if __name__ == '__main__':
model = create_model_residual()
#plot_model(model, to_file='model.png', show_shapes=True)