-
Notifications
You must be signed in to change notification settings - Fork 1
/
net.py
291 lines (207 loc) · 10.8 KB
/
net.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
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
from __future__ import division, print_function, absolute_import
import tensorflow as tf
import numpy as np
import matplotlib.pyplot as plt
import dataset
from tensorflow.python import debug as tf_debug
import voxel
from random import shuffle
import matplotlib.pyplot as plt
import datetime
# Training Parameters
num_steps = 1000
batch_size = 2
display_step = 1000
examples_to_show = 10
# Network Parameters
n_convfilter = [96, 128, 256, 256, 256, 256]
n_deconvfilter = [128, 128, 128, 64, 32, 2]
n_gru_vox = 4
n_fc_filters = [1024]
NUM_OF_IMAGES = 24
X = tf.placeholder(tf.float32, shape=[None, 127, 127, 3],name = "X")
p_H = tf.placeholder(tf.float32, [n_gru_vox, n_gru_vox, n_gru_vox, 1, n_deconvfilter[0]], name="p_H")
Y = tf.placeholder(tf.float32, shape=[32,32,32,batch_size,2],name = "Y")
G = tf.placeholder(tf.float32, shape=[4,4,4,batch_size,128],name = "GRU_OUT")
initializer = tf.variance_scaling_initializer(scale=2.0)
weights = {
#Encoder Part
'conv1a': tf.Variable(initializer([7,7,3,n_convfilter[0]])),
'conv2a': tf.Variable(initializer([3,3,n_convfilter[0],n_convfilter[1]])),
'conv3a': tf.Variable(initializer([3,3,n_convfilter[1],n_convfilter[2]])),
'conv4a': tf.Variable(initializer([3,3,n_convfilter[2],n_convfilter[3]])),
'conv5a': tf.Variable(initializer([3,3,n_convfilter[3],n_convfilter[4]])),
'conv6a': tf.Variable(initializer([3,3,n_convfilter[4],n_convfilter[5]])),
#Gru Part
'w_update': tf.Variable(initializer([1024,8192])), #
'update_gate': tf.Variable(initializer([3,3,3,n_deconvfilter[0],n_deconvfilter[0]])),
'hidden_gate': tf.Variable(initializer([3,3,3,n_deconvfilter[0],n_deconvfilter[0]])),
'reset_gate': tf.Variable(initializer([3, 3, 3, n_deconvfilter[0], n_deconvfilter[0]])),
'tanh_reset': tf.Variable(initializer([3, 3, 3, n_deconvfilter[0], n_deconvfilter[0]])),
#Decoder Part
'conv7a': tf.Variable(initializer([3,3,3,n_deconvfilter[0],n_deconvfilter[1]])),
'conv8a': tf.Variable(initializer([3,3,3,n_deconvfilter[1],n_deconvfilter[2]])),
'conv9a': tf.Variable(initializer([3,3,3,n_deconvfilter[2],n_deconvfilter[3]])),
'conv10a': tf.Variable(initializer([3,3,3,n_deconvfilter[3],n_deconvfilter[4]])),
'conv11a': tf.Variable(initializer([3,3,3,n_deconvfilter[4],n_deconvfilter[5]]))
}
biases = {
#Encoder Part
'conv1a': tf.Variable(tf.zeros([1,1,1,n_convfilter[0]])),
'conv2a': tf.Variable(tf.zeros([1,1,1,n_convfilter[1]])),
'conv3a': tf.Variable(tf.zeros([1,1,1,n_convfilter[2]])),
'conv4a': tf.Variable(tf.zeros([1,1,1,n_convfilter[3]])),
'conv5a': tf.Variable(tf.zeros([1,1,1,n_convfilter[4]])),
'conv6a': tf.Variable(tf.zeros([1,1,1,n_convfilter[5]])),
'fc7': tf.Variable(tf.zeros([n_fc_filters[0]])),
#Gru Part
'w_update': tf.Variable(tf.zeros([8192])),
'update_gate': tf.Variable(tf.zeros([1,1,1,n_deconvfilter[0]])),
'hidden_gate': tf.Variable(tf.zeros([1,1,1,n_deconvfilter[0]])),
'reset_gate': tf.Variable(tf.zeros([1,1,1,n_deconvfilter[0]])),
'tanh_reset': tf.Variable(tf.zeros([1,1,1,n_deconvfilter[0]])),
#Decoder Part
'conv7a': tf.Variable(tf.zeros([1,1,1,n_deconvfilter[1]])),
'conv8a': tf.Variable(tf.zeros([1,1,1,n_deconvfilter[2]])),
'conv9a': tf.Variable(tf.zeros([1,1,1,n_deconvfilter[3]])),
'conv10a': tf.Variable(tf.zeros([1,1,1,n_deconvfilter[4]])),
'conv11a': tf.Variable(tf.zeros([1,1,1,n_deconvfilter[5]]))
}
def unpool(x):
x = tf.transpose(x,perm=[3,1,2,0,4])
x = tf.keras.layers.UpSampling3D(size=[2, 2, 2])(x)
x = tf.transpose(x,perm=[3,1,2,0,4])
return x
def gru():
with tf.name_scope("Encoder"):
# Convolutional Layer #1
conv1a = tf.nn.conv2d(input=X,filter=weights['conv1a'],strides=[1,1,1,1],padding="SAME")
conv1a = tf.add(conv1a,biases['conv1a'])
conv1a = tf.nn.max_pool(conv1a,ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding="SAME")
conv1a = tf.nn.leaky_relu(conv1a,alpha=0.01)
# Convolutional Layer #2
conv2a = tf.nn.conv2d(input=conv1a,filter=weights['conv2a'],strides=[1,1,1,1],padding="SAME")
conv2a = tf.add(conv2a,biases['conv2a'])
conv2a = tf.nn.max_pool(conv2a,ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding="SAME")
conv2a = tf.nn.leaky_relu(conv2a,alpha=0.01)
# Convolutional Layer #3
conv3a = tf.nn.conv2d(input=conv2a,filter=weights['conv3a'],strides=[1,1,1,1],padding="SAME")
conv3a = tf.add(conv3a,biases['conv3a'])
conv3a = tf.nn.max_pool(conv3a,ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding="SAME")
conv3a = tf.nn.leaky_relu(conv3a,alpha=0.01)
# Convolutional Layer #4
conv4a = tf.nn.conv2d(input=conv3a,filter=weights['conv4a'],strides=[1,1,1,1],padding="SAME")
conv4a = tf.add(conv4a,biases['conv4a'])
conv4a = tf.nn.max_pool(conv4a,ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding="SAME")
conv4a = tf.nn.leaky_relu(conv4a,alpha=0.01)
# Convolutional Layer #5
conv5a = tf.nn.conv2d(input=conv4a,filter=weights['conv5a'],strides=[1,1,1,1],padding="SAME")
conv5a = tf.add(conv5a,biases['conv5a'])
conv5a = tf.nn.max_pool(conv5a,ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding="SAME")
conv5a = tf.nn.leaky_relu(conv5a,alpha=0.01)
# Convolutional Layer #6
conv6a = tf.nn.conv2d(input=conv5a,filter=weights['conv6a'],strides=[1,1,1,1],padding="SAME")
conv6a = tf.add(conv6a,biases['conv6a'])
conv6a = tf.nn.max_pool(conv6a,ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding="SAME")
# Flatten Layer
flat7 = tf.layers.flatten(conv6a)
# FC Layer
fc7 = tf.layers.dense(flat7,1024,activation=tf.nn.leaky_relu,use_bias=True)
with tf.name_scope("GRU"):
prev_hidden = p_H
update_dense = tf.layers.dense(fc7,8192,activation=tf.nn.leaky_relu,use_bias=True)
update_dense = tf.reshape(update_dense, [4, 4, 4, -1, 128])
reset_dense = tf.layers.dense(fc7,8192,activation=tf.nn.leaky_relu,use_bias=True)
reset_dense = tf.reshape(reset_dense, [4, 4, 4, -1, 128])
hidden_dense = tf.layers.dense(fc7,8192,activation=tf.nn.leaky_relu,use_bias=True)
hidden_dense = tf.reshape(hidden_dense, [4, 4, 4, -1, 128])
t_x_s_update = tf.nn.conv3d(prev_hidden, weights['update_gate'], strides=[1, 1, 1, 1, 1], padding="SAME") + update_dense
t_x_s_update = tf.add(t_x_s_update, biases['update_gate'])
update_gate = tf.sigmoid(t_x_s_update)
t_x_s_reset = tf.nn.conv3d(prev_hidden, weights['reset_gate'], strides=[1, 1, 1, 1, 1], padding="SAME") + reset_dense
t_x_s_reset = tf.add(t_x_s_reset, biases['reset_gate'])
reset_gate = tf.sigmoid(t_x_s_reset)
hidden_gate = tf.nn.conv3d(reset_gate * prev_hidden, weights['hidden_gate'], strides=[1, 1, 1, 1, 1], padding="SAME") + hidden_dense
hidden_gate = tf.add(hidden_gate,biases['hidden_gate'])
gru_out = (1 - update_gate) * prev_hidden + update_gate * tf.tanh(hidden_gate)
return gru_out
# Building the decoder
def decoder():
with tf.name_scope("Decoder"):
unpool7 = unpool(G)
conv7a = tf.nn.conv3d(unpool7,weights['conv7a'],strides=[1,1,1,1,1],padding="SAME")
conv7a = tf.add(conv7a,biases['conv7a'])
conv7a = tf.nn.leaky_relu(conv7a,alpha=0.01)
unpool8 = unpool(conv7a)
conv8a = tf.nn.conv3d(unpool8,weights['conv8a'],strides=[1,1,1,1,1],padding="SAME")
conv8a = tf.add(conv8a,biases['conv8a'])
conv8a = tf.nn.leaky_relu(conv8a,alpha=0.01)
unpool9 = unpool(conv8a)
conv9a = tf.nn.conv3d(unpool9,weights['conv9a'],strides=[1,1,1,1,1],padding="SAME")
conv9a = tf.add(conv9a,biases['conv9a'])
conv9a = tf.nn.leaky_relu(conv9a,alpha=0.01)
conv10a = tf.nn.conv3d(conv9a,weights['conv10a'],strides=[1,1,1,1,1],padding="SAME")
conv10a = tf.add(conv10a,biases['conv10a'])
conv10a = tf.nn.leaky_relu(conv10a,alpha=0.01)
conv11a = tf.nn.conv3d(conv10a,weights['conv11a'],strides=[1,1,1,1,1],padding="SAME")
conv11a = tf.add(conv11a,biases['conv11a'])
loss_tmp = 0
exp_x = tf.exp(conv11a)
sum_exp_x = tf.reduce_sum(exp_x, reduction_indices=[4], keepdims=True)
for j in range(1,batch_size+1):
tmp = tf.reduce_mean(
tf.reduce_sum(-Y[:,:,:,j-1:j,:] * conv11a[:,:,:,j-1:j,:], reduction_indices=[4], keepdims=True) +
tf.log(sum_exp_x[:,:,:,j-1:j,:])
)
loss_tmp += tmp
loss = loss_tmp / batch_size
return conv11a, loss
# Construct model
gru_op = gru()
output, loss = decoder()
optimizer = tf.train.AdamOptimizer(5e-5).minimize(loss)
# Initialize the variables (i.e. assign their default value)
init = tf.global_variables_initializer()
with tf.Session() as sess:
sess.run(init)
merged = tf.summary.merge_all()
train_writer = tf.summary.FileWriter('./train',sess.graph)
x_train = dataset.train_data()
y_train = dataset.train_labels()
no = 0
while(len(x_train)!=0):
no += 1
i = 1
x_test = np.zeros([n_gru_vox, n_gru_vox, n_gru_vox, batch_size, n_deconvfilter[0]])
y_test = np.ones((32, 32, 32, batch_size, 2), dtype=float)
for image_hash in x_train.keys():
prev_state = np.zeros([n_gru_vox, n_gru_vox, n_gru_vox, 1, n_deconvfilter[0]])
images = x_train[image_hash]
shuffle(images) #Shuffle views
for image in images:
ims = tf.convert_to_tensor(image)
ims = tf.reshape(ims,[-1,127,127,3])
ims = ims.eval()
prev_s = sess.run([gru_op], feed_dict={X: ims, p_H: prev_state})
prev_s = np.array(prev_s)
prev_s = prev_s[0,:,:,:,:,:]
prev_state = prev_s
x_test[:, :, :, i-1:i, :] = prev_state
vox = tf.convert_to_tensor(y_train[image_hash])
vox = tf.cast(vox,tf.float64)
vox = vox.eval()
y_test[:, :, :, i-1, 1] = vox
y_test[:, :, :, i-1, 0] = (tf.ones_like(vox) - vox).eval()
i += 1
# Run optimization op (backprop) and cost op (to get loss value)
l, o, _ = sess.run([loss, output, optimizer], feed_dict={G: x_test, Y: y_test})
currentDT = datetime.datetime.now()
print(str(currentDT) + " Batch: " + str(no) + " Loss: " + str(l))
exp_x = tf.exp(o) # 32, 32, 32, 1 ,2
sum_exp_x = tf.reduce_sum(exp_x, reduction_indices=[4], keepdims=True) # 32, 32, 32, 1, 1
pred = exp_x / sum_exp_x
pred = pred.eval()
pred_name = "test_pred_" + str(no) + ".obj"
voxel.voxel2obj(pred_name, pred[:, :, :, 0, 0])
x_train = dataset.train_data()
y_train = dataset.train_labels()