-
Notifications
You must be signed in to change notification settings - Fork 4
/
models.py
269 lines (218 loc) · 13.5 KB
/
models.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
import tensorflow as tf
import numpy as np
from utils import tf_util
from tf_ops.nn_distance import tf_nndistance
BN_INIT_DECAY = 0.5
BN_DECAY_RATE = 0.5
BN_DECAY_CLIP = 0.99
def placeholder_inputs(batch_size, num_point):
point_clouds_ph = tf.placeholder(tf.float32, shape=(batch_size, num_point, 3))
gt_ph = tf.placeholder(tf.float32, shape=(batch_size, num_point, 3))
return point_clouds_ph, gt_ph
def build_parts_aes_graphs(num_parts, num_points, part_embedding_dim, base_learning_rate, batch_size, decay_step,
decay_rate, bn_decay_step):
point_clouds_phs = []
ae_ops = []
for i in xrange(num_parts):
with tf.variable_scope('part' + str(i)):
print 'Graph part ' + str(i)
point_clouds_ph, gt_ph = placeholder_inputs(batch_size, num_points)
point_clouds_phs.append(point_clouds_ph)
is_training_ph = tf.placeholder(tf.bool, shape=())
batch = tf.Variable(0)
bn_momentum = tf.train.exponential_decay(BN_INIT_DECAY, batch * batch_size, bn_decay_step, BN_DECAY_RATE,
staircase=True)
bn_decay = tf.minimum(BN_DECAY_CLIP, 1 - bn_momentum)
bn_decay_summary = tf.summary.scalar('bn_decay' + ' ' + str(i), bn_decay)
print "Get model and loss"
pred, end_points = get_model_ae(point_clouds_ph, is_training_ph, batch_size, num_points, bn_decay,
part_embedding_dim)
loss, end_points_tmp = get_loss_ae(pred, gt_ph, end_points)
loss_summary = tf.summary.scalar('loss', loss)
print "Get training operator"
learning_rate = tf.train.exponential_decay(base_learning_rate, batch * batch_size, decay_step,
decay_rate, staircase=True)
learning_rate_summary = tf.summary.scalar('learning_rate', learning_rate)
optimizer = tf.train.AdamOptimizer(learning_rate)
train_op = optimizer.minimize(loss, global_step=batch)
merged_summary = tf.summary.merge((bn_decay_summary, loss_summary, learning_rate_summary))
ae_ops.append({'point_clouds_ph': point_clouds_ph,
'gt_ph': gt_ph,
'is_training_ph': is_training_ph,
'pred': pred,
'loss': loss,
'train_op': train_op,
'merged_summary': merged_summary,
'step': batch,
'end_points': end_points})
return ae_ops, point_clouds_phs
def get_model_ae(point_cloud, is_training, batch_size, num_point, bn_decay=None, embedding_dim=64, reuse=False):
input_point_cloud = tf.expand_dims(point_cloud, -1)
net, end_points = ae_encoder(batch_size, num_point, 3, input_point_cloud, is_training, bn_decay=bn_decay,
embedding_dim=embedding_dim)
net = ae_decoder(batch_size, num_point, net, is_training, bn_decay=bn_decay, reuse=reuse)
return net, end_points
def ae_encoder(batch_size, num_point, point_dim, input_image, is_training, bn_decay=None, embedding_dim=128):
net = tf_util.conv2d(input_image, 64, [1, point_dim],
padding='VALID', stride=[1, 1],
bn=True, is_training=is_training,
scope='conv1', bn_decay=bn_decay)
net = tf_util.conv2d(net, 64, [1, 1],
padding='VALID', stride=[1, 1],
bn=True, is_training=is_training,
scope='conv2', bn_decay=bn_decay)
net = tf_util.conv2d(net, 64, [1, 1],
padding='VALID', stride=[1, 1],
bn=True, is_training=is_training,
scope='conv3', bn_decay=bn_decay)
net = tf_util.conv2d(net, 128, [1, 1],
padding='VALID', stride=[1, 1],
bn=True, is_training=is_training,
scope='conv4', bn_decay=bn_decay)
net = tf_util.conv2d(net, embedding_dim, [1, 1],
padding='VALID', stride=[1, 1],
bn=True, is_training=is_training,
scope='conv5', bn_decay=bn_decay)
global_feat = tf_util.max_pool2d(net, [num_point, 1],
padding='VALID', scope='maxpool')
net = tf.reshape(global_feat, [batch_size, -1])
end_points = {'embedding': net}
return net, end_points
def ae_decoder(batch_size, num_point, net, is_training, bn_decay=None, reuse=False):
net = tf_util.fully_connected(net, 256, bn=True, is_training=is_training, scope='fc1', bn_decay=bn_decay,
reuse=reuse)
net = tf_util.fully_connected(net, 256, bn=True, is_training=is_training, scope='fc2', bn_decay=bn_decay,
reuse=reuse)
net = tf_util.fully_connected(net, num_point * 3, activation_fn=None, scope='fc3', reuse=reuse)
net = tf.reshape(net, (batch_size, num_point, 3))
return net
def get_loss_ae(pred, gt, end_points):
dists_forward, _, dists_backward, _ = tf_nndistance.nn_distance(pred, gt)
loss = tf.reduce_mean(dists_forward + dists_backward)
end_points['pcloss'] = loss
loss = loss * 100
end_points['loss'] = loss
return loss, end_points
def build_parts_pcn_graph(ae_ops, point_clouds_ph, num_parts, num_points, noise_embedding_dim, base_learning_rate,
batch_size, decay_step, decay_rate, bn_decay_step):
print '\nGraph PCN'
with tf.variable_scope('pcn'):
pcn_is_training = tf.placeholder(tf.bool, shape=())
y = tf.placeholder(tf.float32, [None, num_parts, num_points, 3], name='y')
y_mask = tf.placeholder(tf.float32, [None, num_parts, num_points], name='y_mask')
noise = tf.placeholder(tf.float32, shape=[None, noise_embedding_dim])
pcn_enc = tf.concat([ae_ops[0]['end_points']['embedding'], ae_ops[1]['end_points']['embedding']], axis=-1)
for p in xrange(2, num_parts):
pcn_enc = tf.concat([pcn_enc, ae_ops[p]['end_points']['embedding']], axis=-1)
pcn_enc = tf.concat([pcn_enc, noise], axis=-1)
pcn_batch = tf.Variable(0)
bn_momentum = tf.train.exponential_decay(BN_INIT_DECAY, pcn_batch * batch_size, bn_decay_step, BN_DECAY_RATE,
staircase=True)
pcn_bn_decay = tf.minimum(BN_DECAY_CLIP, 1 - bn_momentum)
bn_decay_summary = tf.summary.scalar('bn_decay', pcn_bn_decay)
print "--- Get model and loss"
x = tf.stack(point_clouds_ph, axis=1)
y_hat = get_model_pcn(pcn_enc, x, num_parts, pcn_is_training, bn_decay=pcn_bn_decay)
pcn_loss = get_loss_pcn(y, y_hat, y_mask, num_parts, batch_size)
loss_summary = tf.summary.scalar('loss', pcn_loss)
print "--- Get training operator"
pcn_learning_rate = tf.train.exponential_decay(base_learning_rate, pcn_batch * batch_size, decay_step,
decay_rate, staircase=True)
learning_rate_summary = tf.summary.scalar('learning_rate', pcn_learning_rate)
pcn_param = [var for var in tf.trainable_variables() if any(x in var.name for x in ['pcn'])]
optimizer = tf.train.AdamOptimizer(pcn_learning_rate).minimize(pcn_loss, global_step=pcn_batch,
var_list=pcn_param)
merged_summary = tf.summary.merge((bn_decay_summary, loss_summary, learning_rate_summary))
pcn_ops = ({'y': y,
'y_mask': y_mask,
'noise': noise,
'is_training_ph': pcn_is_training,
'pcn_enc': pcn_enc,
'pred': y_hat,
'loss': pcn_loss,
'train_op': optimizer,
'merged_summary': merged_summary,
'step': pcn_batch})
return pcn_ops
def get_model_pcn(pcn_enc, x, num_parts, is_training, bn_decay=None, reuse=False):
with tf.variable_scope('trans', reuse=reuse):
bias_initializer = np.array([[1., 0, 1, 0, 1, 0] for _ in xrange(num_parts)])
bias_initializer = bias_initializer.astype('float32').flatten()
net = tf_util.fully_connected(pcn_enc, 256, bn=True, is_training=is_training, scope='fc1', bn_decay=bn_decay,
reuse=reuse)
net = tf_util.fully_connected(net, 128, bn=True, is_training=is_training, scope='fc2', bn_decay=bn_decay,
reuse=reuse)
trans = tf_util.fully_connected(net, num_parts * 6, activation_fn=None, scope='fc3',
weights_initializer=tf.zeros_initializer(),
biases_initializer=tf.constant_initializer(bias_initializer), reuse=reuse)
# Perform transformation
with tf.variable_scope('pcn', reuse=reuse):
zeros_dims = tf.stack([tf.shape(x)[0], 1])
zeros_col = tf.fill(zeros_dims, 0.0)
'''
sx 0 0 tx
0 sy 0 ty
0 0 sz tz
'''
trans_mat = tf.concat((tf.expand_dims(trans[:, 0], axis=1), zeros_col, zeros_col,
tf.expand_dims(trans[:, 1], axis=1), zeros_col, tf.expand_dims(trans[:, 2], axis=1),
zeros_col, tf.expand_dims(trans[:, 3], axis=1), zeros_col, zeros_col,
tf.expand_dims(trans[:, 4], axis=1), tf.expand_dims(trans[:, 5], axis=1),
tf.expand_dims(trans[:, 6], axis=1), zeros_col, zeros_col,
tf.expand_dims(trans[:, 7], axis=1), zeros_col, tf.expand_dims(trans[:, 8], axis=1),
zeros_col, tf.expand_dims(trans[:, 9], axis=1), zeros_col, zeros_col,
tf.expand_dims(trans[:, 10], axis=1), tf.expand_dims(trans[:, 11], axis=1)), axis=1)
for p in xrange(2, num_parts):
start_ind = 6 * p
trans_mat = tf.concat((trans_mat, tf.expand_dims(trans[:, start_ind], axis=1), zeros_col, zeros_col,
tf.expand_dims(trans[:, start_ind + 1], axis=1), zeros_col,
tf.expand_dims(trans[:, start_ind + 2], axis=1), zeros_col,
tf.expand_dims(trans[:, start_ind + 3], axis=1), zeros_col, zeros_col,
tf.expand_dims(trans[:, start_ind + 4], axis=1),
tf.expand_dims(trans[:, start_ind + 5], axis=1)), axis=1)
trans_mat = tf.reshape(trans_mat, (-1, num_parts, 3, 4))
# adding 1 (w coordinate) to every point (x,y,z,1)
w = tf.ones([tf.shape(x)[0], tf.shape(x)[1], tf.shape(x)[2], 1])
x = tf.concat((x, w), axis=-1)
x_t = tf.transpose(x, [0, 1, 3, 2])
y_hat_t = tf.matmul(trans_mat, x_t)
y_hat = tf.transpose(y_hat_t, [0, 1, 3, 2])
return y_hat
def get_loss_pcn(gt, pred, gt_mask, num_parts, batch_size):
dists_forward_total = tf.zeros(batch_size)
dists_backward_total = tf.zeros(batch_size)
for part in xrange(num_parts):
dists_forward, _, dists_backward, _ = tf_nndistance.nn_distance(pred[:, part], gt[:, part])
# zero out the non-existing parts
dists_forward = tf.reduce_sum(tf.multiply(dists_forward, gt_mask[:, part]), axis=-1)
dists_backward = tf.reduce_sum(tf.multiply(dists_backward, gt_mask[:, part]), axis=-1)
dists_forward_total += dists_forward
dists_backward_total += dists_backward
loss = dists_forward_total + dists_backward_total
# divide by the number of parts
div = tf.reduce_sum(tf.reduce_mean(gt_mask, axis=-1), axis=-1)
loss = tf.reduce_mean(tf.div(loss, div))
return loss * 100
def build_test_graph(num_parts, num_points, part_embedding_dim, base_learning_rate, batch_size, decay_step, decay_rate,
bn_decay_step, noise_embedding_dim):
ae_ops, point_clouds_ph = build_parts_aes_graphs(num_parts, num_points, part_embedding_dim, base_learning_rate,
batch_size, decay_step, decay_rate, bn_decay_step)
for i in xrange(num_parts):
with tf.variable_scope('part' + str(i)):
samples = (tf.placeholder(tf.float32, shape=(batch_size, part_embedding_dim)))
dec = ae_decoder(batch_size, num_points, samples, ae_ops[i]["is_training_ph"], reuse=True)
ae_ops[i]['samples'] = samples
ae_ops[i]['dec'] = dec
pcn_ops = build_parts_pcn_graph(ae_ops, point_clouds_ph, num_parts, num_points, noise_embedding_dim,
base_learning_rate, batch_size, decay_step, decay_rate, bn_decay_step)
with tf.variable_scope('pcn'):
x_full = tf.stack((ae_ops[0]['dec'], ae_ops[1]['dec']), axis=1)
for p in xrange(2, num_parts):
x_full = tf.concat((x_full, tf.expand_dims(ae_ops[p]['dec'], axis=1)), axis=1)
cpcn_enc_full = tf.concat([ae_ops[0]['samples'], ae_ops[1]['samples']], axis=-1)
for p in xrange(2, num_parts):
cpcn_enc_full = tf.concat([cpcn_enc_full, ae_ops[p]['samples']], axis=-1)
cpcn_enc_full = tf.concat([cpcn_enc_full, pcn_ops['noise']], axis=-1)
y_hat_full = get_model_pcn(cpcn_enc_full, x_full, num_parts, pcn_ops['is_training_ph'], reuse=True)
pcn_ops['pred_full'] = y_hat_full
return ae_ops, pcn_ops