-
Notifications
You must be signed in to change notification settings - Fork 6
/
Copy pathvgan.py
200 lines (139 loc) · 7.18 KB
/
vgan.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
import tensorflow as tf
import numpy as np
import h5py
import os
from layers import *
z_size = 100
batch_size = 10 IS_TRAIN = True
lr = 0.0002
beta1 = 0.5
train_iters = 10000
A,B,C = 64,64,32
channel = 3
img_size = 64*64
model_file_name = "results/vgan_applyeyemakeup"
tf.reset_default_graph()
def next_batch(data_array):
length=data_array.shape[0] #assuming the data array to be a np arry
permutations=np.random.permutation(length)
idxs=permutations[0:batch_size]
batch=np.zeros([batch_size, C, img_size*channel], dtype=np.float32)
for i in range(len(idxs)):
batch[i,:]=data_array[idxs[i]].reshape(C,img_size*channel)
return batch / 255.
def static_net(z):
static_input = tf.reshape(z,[batch_size, 1, 1, z_size])
static_dconv1 = deconv2d(static_input,[4,4,512,z_size],[batch_size, 4,4,512],[1,1,1,1],'VALID',name='st_deconv1')
static_dconv1 = batch_norm(static_dconv1, scope='st_bn_dconv1', is_training=IS_TRAIN)
static_dconv1 = tf.nn.relu(static_dconv1)
static_dconv2 = deconv2d(static_dconv1,[4,4,256,512],[batch_size, 8,8,256],[1,2,2,1],'SAME',name='st_deconv2')
static_dconv2 = batch_norm(static_dconv2, scope='st_bn_dconv2', is_training=IS_TRAIN)
static_dconv2 = tf.nn.relu(static_dconv2)
static_dconv3 = deconv2d(static_dconv2,[4,4,128,256],[batch_size, 16,16,128],[1,2,2,1],'SAME',name='st_deconv3')
static_dconv3 = batch_norm(static_dconv3, scope='st_bn_dconv3', is_training=IS_TRAIN)
static_dconv3 = tf.nn.relu(static_dconv3)
static_dconv4 = deconv2d(static_dconv3,[4,4,64,128],[batch_size, 32, 32,64],[1,2,2,1],'SAME',name='st_deconv4')
static_dconv4 = batch_norm(static_dconv4, scope='st_bn_dconv4', is_training=IS_TRAIN)
static_dconv4 = tf.nn.relu(static_dconv4)
static_dconv5 = deconv2d(static_dconv4,[4,4,3,64],[batch_size, 64,64, 3],[1,2,2,1],'SAME',name='st_deconv5')
static_dconv5 = tf.nn.tanh(static_dconv5)
return static_dconv5
def video_net_and_mask(z):
video_input = tf.reshape(z,[batch_size, 1, 1, 1, z_size])
video_dconv1 = deconv3d(video_input,[2,4,4,512,z_size],[batch_size, 2, 4,4,512],[1,1,1,1,1],'VALID',name='video_deconv1')
video_dconv1 = batch_norm(video_dconv1, scope='vd_bn_dconv1', is_training=IS_TRAIN)
video_dconv1 = tf.nn.relu(video_dconv1)
video_dconv2 = deconv3d(video_dconv1,[4,4,4,256,512],[batch_size, 4, 8, 8, 256],[1,2,2,2,1],'SAME',name='video_deconv2')
video_dconv2 = batch_norm(video_dconv2, scope='vd_bn_dconv2', is_training=IS_TRAIN)
video_dconv2 = tf.nn.relu(video_dconv2)
video_dconv3 = deconv3d(video_dconv2,[4,4,4,128,256],[batch_size, 8, 16, 16, 128],[1,2,2,2,1],'SAME',name='video_deconv3')
video_dconv3 = batch_norm(video_dconv3, scope='vd_bn_dconv3', is_training=IS_TRAIN)
video_dconv3 = tf.nn.relu(video_dconv3)
video_dconv4 = deconv3d(video_dconv3,[4,4,4,64,128],[batch_size, 16, 32, 32, 64],[1,2,2,2,1],'SAME',name='video_deconv4')
video_dconv4 = batch_norm(video_dconv4, scope='vd_bn_dconv4', is_training=IS_TRAIN)
video_dconv4 = tf.nn.relu(video_dconv4)
video_dconv5 = deconv3d(video_dconv4,[4,4,4,3,64],[batch_size, 32, 64, 64, 3],[1,2,2,2,1],'SAME',name='video_deconv5')
video_dconv5 = tf.nn.tanh(video_dconv5)
# mast out... (for the mast net)
mask_deconv5, mask_deconv5_weights = deconv3d(video_dconv4,[4,4,4,1,64], [batch_size, 32, 64, 64, 1],[1,2,2,2,1],'SAME',name='mask_deconv5',with_w=True)
mask_deconv5 = tf.nn.sigmoid(mask_deconv5)
l1_regularizer = tf.contrib.layers.l1_regularizer(scale=0.1, scope='mask_l1')
mask_penalty = tf.contrib.layers.apply_regularization(l1_regularizer, [mask_deconv5_weights])
return video_dconv5,mask_deconv5,mask_penalty
def discriminator_net(input,reuse=None):
disc_conv1 = conv3d(input,[4,4,4,3,64],[1,2,2,2,1],'SAME',name='dc_conv1',reuse=reuse)
disc_conv1 = lrelu(disc_conv1)
disc_conv2 = conv3d(disc_conv1,[4,4,4,64,128],[1,2,2,2,1],'SAME',name='dc_conv2',reuse=reuse)
disc_conv2 = batch_norm(disc_conv2,eps=1e-3,scope='dc_bn_conv2', is_training=IS_TRAIN,reuse=reuse)
disc_conv2 = lrelu(disc_conv2)
disc_conv3 = conv3d(disc_conv2,[4,4,4,128,256],[1,2,2,2,1],'SAME',name='dc_conv3',reuse=reuse)
disc_conv3 = batch_norm(disc_conv3,eps=1e-3,scope='dc_bn_conv3', is_training=IS_TRAIN,reuse=reuse)
disc_conv3 = lrelu(disc_conv3)
disc_conv4 = conv3d(disc_conv3,[4,4,4,256,512],[1,2,2,2,1],'SAME',name='dc_conv4',reuse=reuse)
disc_conv4 = batch_norm(disc_conv4,eps=1e-3,scope='dc_bn_conv4', is_training=IS_TRAIN,reuse=reuse)
disc_conv4 = lrelu(disc_conv4)
disc_conv5 = conv3d(disc_conv4,[2,4,4,512,1],[1,1,1,1,1],'VALID',name='dc_conv5',reuse=reuse)
final = tf.reshape(disc_conv5,[batch_size])
return tf.nn.sigmoid(final),final
def gen_video(z):
background = static_net(z)
background = tf.tile(tf.reshape(background,[batch_size,1,64,64,3]),[1,32,1,1,1])
foreground,mask,penalty = video_net_and_mask(z)
mask = tf.tile(mask,[1,1,1,1,3])
video = foreground*mask + background*(1-mask)
return video,penalty
z = tf.placeholder(tf.float32,[batch_size,z_size],name='z')
x = tf.placeholder(tf.float32,[batch_size,32,64,64,3],name='x_real')
gen_videos,penalty = gen_video(z)
gen_D, gen_D_logits = discriminator_net(gen_videos)
real_D, real_D_logits = discriminator_net(x,reuse=True)
d_loss_real = tf.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(logits=real_D_logits, labels=tf.ones_like(real_D_logits)))
d_loss_gen = tf.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(logits=gen_D_logits, labels=tf.zeros_like(gen_D_logits)))
g_loss = tf.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(logits=gen_D_logits, labels=tf.ones_like(gen_D_logits)))
d_loss = d_loss_gen + d_loss_real
g_loss = g_loss + penalty
t_vars = tf.trainable_variables()
d_vars = [var for var in t_vars if 'dc_' in var.name]
g_vars = [var for var in t_vars if var not in d_vars]
d_optim = tf.train.AdamOptimizer(lr, beta1=beta1).minimize(d_loss, var_list=d_vars)
g_optim = tf.train.AdamOptimizer(lr, beta1=beta1).minimize(g_loss, var_list=g_vars)
dataset_filename = 'dataset/path_name'
hf = h5py.File(dataset_filename,'r')
caps = hf.get('filenames')
total_count = caps.shape[0]
train_data = hf.get('video_data')[:total_count].reshape(-1,C, A*B*channel)
print "loaded"
fetches = []
fetches.extend([d_loss,g_loss,d_optim,g_optim])
d_lss = [0]*train_iters
g_lss = [0]*train_iters
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess=tf.InteractiveSession(config=config)
saver = tf.train.Saver()
tf.initialize_all_variables().run()
for i in range(train_iters):
xtrain = next_batch(train_data)
xtrain = xtrain.reshape(-1,C,64,64,channel)
z_sample = np.random.normal(size=(batch_size,z_size))
# print z_sample.shape
feed_dict={x:xtrain,z:z_sample}
results = sess.run(fetches,feed_dict)
d_lss[i],g_lss[i],_,_=results
if i%10==0:
print("iter=%d : D_Loss: %f G_Loss: %f" % (i, d_lss[i], g_lss[i]))
if (i+1)%100==0:
ckpt_file=model_file_name+".ckpt"
print("Model saved in file: %s" % saver.save(sess,ckpt_file))
gen_sample = sess.run(gen_videos,{z:z_sample})
gen_sample = np.array(gen_sample)
out_file=model_file_name+".npy"
np.save(out_file,[gen_sample,d_lss,g_lss])
print("Outputs saved in file: %s" % out_file)
ckpt_file=model_file_name+".ckpt"
print("Model saved in file: %s" % saver.save(sess,ckpt_file))
sess.close()
print('Done with V-GAN!')