/
model-noaudio_inferenceGenerator.py
189 lines (149 loc) · 11 KB
/
model-noaudio_inferenceGenerator.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
import tensorlayer as tl
import os
from tensorlayer.layers import *
from data_input import DataInput
from utils import norm_img, denorm_img
import argparse
from PIL import Image
DEFAULT_DATA_FACES_PATH = "/storage/dataset"
DEFAULT_DATA_AUDIOS_PATH = "/storage/dataset_videos/cropped_videos/outputb"
DEFAULT_LOG_DIR = "/storage/logs"
DEFAULT_CHECKPOINT_DIR = "/storage/checkpoints"
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
def restore_model(sess, checkpoint_path):
# Get the state of the checkpoint and then restore using ckpt path
ckpt = tf.train.get_checkpoint_state(checkpoint_path)
if checkpoint_path is not None:
restorer = tf.train.Saver(tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope="generator"))
restorer.restore(sess, ckpt.model_checkpoint_path)
def generator(z, reuse, hidden_number=64, kernel=3):
w_init = tf.random_normal_initializer(stddev=0.02)
with tf.variable_scope("generator", reuse=reuse):
tl.layers.set_name_reuse(reuse)
# DECODER BEGINS
# hidden_number = n = 128
# exponential linear units output convolutions
# Each layer is repeated a number of times (typically 2). We observed that more repetitions led to
# even better visual results
# Down-sampling is implemented as sub-sampling with stride 2 and up- sampling is done by nearest neighbor.
x = InputLayer(z, name="in")
x = DenseLayer(x, n_units=8 * 8 * hidden_number, name='Generator/dense2')
arguments = {'shape': [-1, 8, 8, hidden_number], 'name': 'Generator/reshape1'}
x = LambdaLayer(x, fn=tf.reshape, fn_args=arguments)
x = Conv2dLayer(x, shape=[kernel, kernel, hidden_number, hidden_number], strides=[1, 1, 1, 1], padding='SAME',
W_init=w_init, act=tf.nn.elu, name='Generator/conv1')
x = Conv2dLayer(x, shape=[kernel, kernel, hidden_number, hidden_number], strides=[1, 1, 1, 1], padding='SAME',
W_init=w_init, act=tf.nn.elu, name='Generator/conv2')
x = UpSampling2dLayer(x, size=[2, 2], is_scale=True, method=1, name='Generator/UpSampling1') # method= 1 NN
x = Conv2dLayer(x, shape=[kernel, kernel, hidden_number, hidden_number], strides=[1, 1, 1, 1], padding='SAME',
W_init=w_init, act=tf.nn.elu, name='Generator/conv3')
x = Conv2dLayer(x, shape=[kernel, kernel, hidden_number, hidden_number], strides=[1, 1, 1, 1], padding='SAME',
W_init=w_init, act=tf.nn.elu, name='Generator/conv4')
x = UpSampling2dLayer(x, size=[2, 2], is_scale=True, method=1, name='Encoder/UpSampling2') # method= 1 NN
x = Conv2dLayer(x, shape=[kernel, kernel, hidden_number, hidden_number], strides=[1, 1, 1, 1], padding='SAME',
W_init=w_init, act=tf.nn.elu, name='Generator/conv5')
x = Conv2dLayer(x, shape=[kernel, kernel, hidden_number, hidden_number], strides=[1, 1, 1, 1], padding='SAME',
W_init=w_init, act=tf.nn.elu, name='Generator/conv6')
x = UpSampling2dLayer(x, size=[2, 2], is_scale=True, method=1, name='Generator/UpSampling3') # method= 1 NN
x = Conv2dLayer(x, shape=[kernel, kernel, hidden_number, hidden_number], strides=[1, 1, 1, 1],
padding='SAME',
W_init=w_init, act=tf.nn.elu, name='Generator/conv7')
x = Conv2dLayer(x, shape=[kernel, kernel, hidden_number, hidden_number], strides=[1, 1, 1, 1], padding='SAME',
W_init=w_init, act=tf.nn.elu, name='Generator/conv8')
x = Conv2dLayer(x, shape=[kernel, kernel, hidden_number, 3], strides=[1, 1, 1, 1], padding='SAME',
W_init=w_init, name='Generator/convLAST')
return x
def discriminator(disc_input, reuse, z_num=64, hidden_number=64, kernel=3):
w_init = tf.random_normal_initializer(stddev=0.02)
with tf.variable_scope("discriminator", reuse=reuse):
tl.layers.set_name_reuse(reuse)
# Encoder
# Down-sampling is implemented as sub-sampling with stride 2
x = InputLayer(disc_input, name='in') # [1, height = 64, width = 64, 3 ]
x = Conv2dLayer(x, shape=[kernel, kernel, 3, hidden_number], strides=[1, 1, 1, 1], padding='SAME',
W_init=w_init, act=tf.nn.elu, name='Discriminator/Encoder/conv1')
x = Conv2dLayer(x, shape=[kernel, kernel, hidden_number, hidden_number], strides=[1, 1, 1, 1], padding='SAME',
W_init=w_init, act=tf.nn.elu, name='Discriminator/Encoder/conv2')
x = Conv2dLayer(x, shape=[kernel, kernel, hidden_number, 2 * hidden_number], strides=[1, 1, 1, 1],
padding='SAME',
W_init=w_init, act=tf.nn.elu, name='Discriminator/Encoder/conv3')
x = Conv2dLayer(x, shape=[kernel, kernel, 2 * hidden_number, 2 * hidden_number], strides=[1, 2, 2, 1],
padding='SAME',
W_init=w_init, act=tf.nn.elu, name='Discriminator/Encoder/subsampling1')
# [1, height = 32, width = 32, 2*hidden_number]
x = Conv2dLayer(x, shape=[kernel, kernel, 2 * hidden_number, 2 * hidden_number], strides=[1, 1, 1, 1],
padding='SAME',
W_init=w_init, act=tf.nn.elu, name='Discriminator/Encoder/conv4')
x = Conv2dLayer(x, shape=[kernel, kernel, 2 * hidden_number, 3 * hidden_number], strides=[1, 1, 1, 1],
padding='SAME', W_init=w_init, act=tf.nn.elu, name='Discriminator/Encoder/conv5')
x = Conv2dLayer(x, shape=[kernel, kernel, 3 * hidden_number, 3 * hidden_number], strides=[1, 2, 2, 1],
padding='SAME', W_init=w_init, act=tf.nn.elu, name='Discriminator/Encoder/subsampling2')
# [1, height = 16, width = 16, 3*hidden_number]
x = Conv2dLayer(x, shape=[kernel, kernel, 3 * hidden_number, 3 * hidden_number], strides=[1, 1, 1, 1],
padding='SAME',
W_init=w_init, act=tf.nn.elu, name='Discriminator/Encoder/conv6')
x = Conv2dLayer(x, shape=[kernel, kernel, 3 * hidden_number, 4 * hidden_number], strides=[1, 1, 1, 1],
padding='SAME', W_init=w_init, act=tf.nn.elu, name='Discriminator/Encoder/conv7')
x = Conv2dLayer(x, shape=[kernel, kernel, 4 * hidden_number, 4 * hidden_number], strides=[1, 2, 2, 1],
padding='SAME', W_init=w_init, act=tf.nn.elu, name='Discriminator/Encoder/subsampling3')
# [1, height = 8, width = 8, 4*hidden_number]
x = Conv2dLayer(x, shape=[kernel, kernel, 4 * hidden_number, 4 * hidden_number], strides=[1, 1, 1, 1],
padding='SAME', W_init=w_init, act=tf.nn.elu, name='Discriminator/Encoder/conv8')
x = Conv2dLayer(x, shape=[kernel, kernel, 4 * hidden_number, 4 * hidden_number], strides=[1, 1, 1, 1],
padding='SAME', W_init=w_init, act=tf.nn.elu, name='Discriminator/Encoder/conv9')
x = FlattenLayer(x, name='Discriminator/Encoder/flatten')
z = DenseLayer(x, n_units=z_num, name='Discriminator/Encoder/Dense')
# Decoder
x = DenseLayer(x, n_units=8 * 8 * hidden_number, name='Generator/dense2')
arguments = {'shape': [-1, 8, 8, hidden_number], 'name': 'Generator/reshape1'}
x = LambdaLayer(x, fn=tf.reshape, fn_args=arguments)
x = Conv2dLayer(x, shape=[kernel, kernel, hidden_number, hidden_number], strides=[1, 1, 1, 1], padding='SAME',
W_init=w_init, act=tf.nn.elu, name='Generator/conv1')
x = Conv2dLayer(x, shape=[kernel, kernel, hidden_number, hidden_number], strides=[1, 1, 1, 1], padding='SAME',
W_init=w_init, act=tf.nn.elu, name='Generator/conv2')
x = UpSampling2dLayer(x, size=[2, 2], is_scale=True, method=1, name='Generator/UpSampling1') # method= 1 NN
x = Conv2dLayer(x, shape=[kernel, kernel, hidden_number, hidden_number], strides=[1, 1, 1, 1], padding='SAME',
W_init=w_init, act=tf.nn.elu, name='Generator/conv3')
x = Conv2dLayer(x, shape=[kernel, kernel, hidden_number, hidden_number], strides=[1, 1, 1, 1], padding='SAME',
W_init=w_init, act=tf.nn.elu, name='Generator/conv4')
x = UpSampling2dLayer(x, size=[2, 2], is_scale=True, method=1, name='Encoder/UpSampling2') # method= 1 NN
x = Conv2dLayer(x, shape=[kernel, kernel, hidden_number, hidden_number], strides=[1, 1, 1, 1], padding='SAME',
W_init=w_init, act=tf.nn.elu, name='Generator/conv5')
x = Conv2dLayer(x, shape=[kernel, kernel, hidden_number, hidden_number], strides=[1, 1, 1, 1], padding='SAME',
W_init=w_init, act=tf.nn.elu, name='Generator/conv6')
x = UpSampling2dLayer(x, size=[2, 2], is_scale=True, method=1, name='Generator/UpSampling3') # method= 1 NN
x = Conv2dLayer(x, shape=[kernel, kernel, hidden_number, hidden_number], strides=[1, 1, 1, 1],
padding='SAME',
W_init=w_init, act=tf.nn.elu, name='Generator/conv7')
x = Conv2dLayer(x, shape=[kernel, kernel, hidden_number, hidden_number], strides=[1, 1, 1, 1], padding='SAME',
W_init=w_init, act=tf.nn.elu, name='Generator/conv8')
x = Conv2dLayer(x, shape=[kernel, kernel, hidden_number, 3], strides=[1, 1, 1, 1], padding='SAME',
W_init=w_init, name='Generator/convLAST')
return x, z
def train(batch_size):
# ##========================== DEFINE INPUT DATA ============================###
z = tf.placeholder('float32', [None, 64], name='t_noise_generator')
# ##========================== DEFINE MODEL ============================###
net_gen = generator(z=z, reuse=False)
output_gen = denorm_img(net_gen.outputs) # Denormalization
with tf.Session() as sess:
if args.resume == "True":
print("Restoring model from checkpoint")
restore_model(sess, args.checkpoint_dir)
for iteration in range(0, 10):
input_z = np.random.uniform(0, 0, size=[batch_size, 64])
print "Input vector: {}".format(input_z[0, :50])
output_image = sess.run(output_gen, feed_dict={z: input_z})[0]
ima = Image.fromarray(output_image.astype(np.uint8), 'RGB')
ima.save("test_image_{}.png".format(iteration))
iteration += 1
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='Predict script')
parser.add_argument('-dataset_faces_folder', default=DEFAULT_DATA_FACES_PATH, help='Path to the images file')
parser.add_argument('-dataset_audios_folder', default=DEFAULT_DATA_AUDIOS_PATH, help='Path to the audios file')
parser.add_argument('-checkpoint_dir', default=DEFAULT_CHECKPOINT_DIR, help='Model checkpoint to use')
parser.add_argument('-log_dir', default=DEFAULT_LOG_DIR, help='Model checkpoint to use')
parser.add_argument('-resume', default="True", help='Resume training ("True" or "False")')
args = parser.parse_args()
train(batch_size=1)