Skip to content

Commit b395af5

Browse files
alex04072000alex04072000
authored
Add files via upload
1 parent 5f6f66a commit b395af5

8 files changed

+1345
-0
lines changed

‎CyclicGen_model.py

+135
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,135 @@
1+
"""Implements a voxel flow model."""
2+
from __future__ import absolute_import
3+
from __future__ import division
4+
from __future__ import print_function
5+
6+
import tensorflow as tf
7+
import tensorflow.contrib.slim as slim
8+
from utils.loss_utils import l1_loss, l2_loss, vae_loss
9+
from utils.geo_layer_utils import vae_gaussian_layer
10+
from utils.geo_layer_utils import bilinear_interp
11+
from utils.geo_layer_utils import meshgrid
12+
13+
FLAGS = tf.app.flags.FLAGS
14+
epsilon = 0.001
15+
16+
17+
class Voxel_flow_model(object):
18+
def __init__(self, is_train=True, is_extrapolation=False):
19+
self.is_train = is_train
20+
self.is_extrapolation = is_extrapolation
21+
22+
def inference(self, input_images):
23+
"""Inference on a set of input_images.
24+
Args:
25+
"""
26+
return self._build_model(input_images)
27+
28+
def total_var(self, images):
29+
pixel_dif1 = images[:, 1:, :, :] - images[:, :-1, :, :]
30+
pixel_dif2 = images[:, :, 1:, :] - images[:, :, :-1, :]
31+
tot_var = (tf.reduce_mean(tf.sqrt(tf.square(pixel_dif1) + epsilon**2)) + tf.reduce_mean(tf.sqrt(tf.square(pixel_dif2) + epsilon**2)))
32+
return tot_var
33+
34+
def loss(self, predictions, targets):
35+
"""Compute the necessary loss for training.
36+
Args:
37+
Returns:
38+
"""
39+
# self.reproduction_loss = l1_loss(predictions, targets)
40+
self.reproduction_loss = tf.reduce_mean(tf.sqrt(tf.square(predictions - targets) + epsilon**2))
41+
42+
self.motion_loss = self.total_var(self.flow)
43+
self.mask_loss = self.total_var(self.mask)
44+
45+
# return [self.reproduction_loss, self.prior_loss]
46+
return self.reproduction_loss + 0.01 * self.motion_loss + 0.005 * self.mask_loss
47+
48+
def l1loss(self, predictions, targets):
49+
self.reproduction_loss = l1_loss(predictions, targets)
50+
return self.reproduction_loss
51+
52+
def _build_model(self, input_images):
53+
with slim.arg_scope([slim.conv2d],
54+
activation_fn=tf.nn.relu,
55+
weights_initializer=tf.truncated_normal_initializer(0.0, 0.01),
56+
weights_regularizer=slim.l2_regularizer(0.0001)):
57+
# Define network
58+
batch_norm_params = {
59+
'decay': 0.9997,
60+
'epsilon': 0.001,
61+
'is_training': self.is_train,
62+
}
63+
with slim.arg_scope([slim.batch_norm], is_training=self.is_train, updates_collections=None):
64+
with slim.arg_scope([slim.conv2d], normalizer_fn=slim.batch_norm,
65+
normalizer_params=batch_norm_params):
66+
x0 = slim.conv2d(input_images, 64, [5, 5], stride=1, scope='conv1')
67+
# with tf.name_scope('conv1') as scope:
68+
# kernel = tf.Variable(tf.truncated_normal([5, 5, 8, 64], dtype=tf.float32, stddev=1e-1), name='weights')
69+
# conv = tf.nn.atrous_conv2d(input_images, kernel, 2, padding='SAME')
70+
# x0 = tf.nn.relu(conv, name=scope)
71+
72+
net = slim.max_pool2d(x0, [2, 2], scope='pool1')
73+
x1 = slim.conv2d(net, 128, [5, 5], stride=1, scope='conv2')
74+
75+
net = slim.max_pool2d(x1, [2, 2], scope='pool2')
76+
x2 = slim.conv2d(net, 256, [3, 3], stride=1, scope='conv3')
77+
78+
net = slim.max_pool2d(x2, [2, 2], scope='pool3')
79+
net = slim.conv2d(net, 256, [3, 3], stride=1, scope='conv4')
80+
81+
net = tf.image.resize_bilinear(net, [x2.get_shape().as_list()[1], x2.get_shape().as_list()[2]])
82+
net = slim.conv2d(tf.concat([net, x2], -1), 256, [3, 3], stride=1, scope='conv5')
83+
84+
net = tf.image.resize_bilinear(net, [x1.get_shape().as_list()[1], x1.get_shape().as_list()[2]])
85+
net = slim.conv2d(tf.concat([net, x1], -1), 128, [3, 3], stride=1, scope='conv6')
86+
87+
net = tf.image.resize_bilinear(net, [x0.get_shape().as_list()[1], x0.get_shape().as_list()[2]])
88+
y0 = slim.conv2d(tf.concat([net, x0], -1), 64, [5, 5], stride=1, scope='conv7')
89+
90+
91+
net = slim.conv2d(y0, 3, [5, 5], stride=1, activation_fn=tf.tanh,
92+
normalizer_fn=None, scope='conv8')
93+
net_copy = net
94+
95+
flow = net[:, :, :, 0:2]
96+
mask = tf.expand_dims(net[:, :, :, 2], 3)
97+
98+
self.flow = flow
99+
100+
101+
grid_x, grid_y = meshgrid(x0.get_shape().as_list()[1], x0.get_shape().as_list()[2])
102+
grid_x = tf.tile(grid_x, [FLAGS.batch_size, 1, 1])
103+
grid_y = tf.tile(grid_y, [FLAGS.batch_size, 1, 1])
104+
105+
flow = 0.5 * flow
106+
107+
flow_ratio = tf.constant([255.0 / (x0.get_shape().as_list()[2]-1), 255.0 / (x0.get_shape().as_list()[1]-1)])
108+
flow = flow * tf.expand_dims(tf.expand_dims(tf.expand_dims(flow_ratio, 0), 0), 0)
109+
110+
if self.is_extrapolation:
111+
coor_x_1 = grid_x + flow[:, :, :, 0] * 2
112+
coor_y_1 = grid_y + flow[:, :, :, 1] * 2
113+
coor_x_2 = grid_x + flow[:, :, :, 0]
114+
coor_y_2 = grid_y + flow[:, :, :, 1]
115+
else:
116+
coor_x_1 = grid_x + flow[:, :, :, 0]
117+
coor_y_1 = grid_y + flow[:, :, :, 1]
118+
coor_x_2 = grid_x - flow[:, :, :, 0]
119+
coor_y_2 = grid_y - flow[:, :, :, 1]
120+
121+
output_1 = bilinear_interp(input_images[:, :, :, 0:3], coor_x_1, coor_y_1, 'interpolate')
122+
output_2 = bilinear_interp(input_images[:, :, :, 3:6], coor_x_2, coor_y_2, 'interpolate')
123+
124+
self.warped_img1 = output_1
125+
self.warped_img2 = output_2
126+
127+
self.warped_flow1 = bilinear_interp(-flow[:, :, :, 0:3]*0.5, coor_x_1, coor_y_1, 'interpolate')
128+
self.warped_flow2 = bilinear_interp(flow[:, :, :, 0:3]*0.5, coor_x_2, coor_y_2, 'interpolate')
129+
130+
mask = 0.5 * (1.0 + mask)
131+
self.mask = mask
132+
mask = tf.tile(mask, [1, 1, 1, 3])
133+
net = tf.multiply(mask, output_1) + tf.multiply(1.0 - mask, output_2)
134+
135+
return [net, net_copy]

‎CyclicGen_model_large.py

+147
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,147 @@
1+
"""Implements a voxel flow model."""
2+
from __future__ import absolute_import
3+
from __future__ import division
4+
from __future__ import print_function
5+
6+
import tensorflow as tf
7+
import tensorflow.contrib.slim as slim
8+
from utils.loss_utils import l1_loss, l2_loss, vae_loss
9+
from utils.geo_layer_utils import vae_gaussian_layer
10+
from utils.geo_layer_utils import bilinear_interp
11+
from utils.geo_layer_utils import meshgrid
12+
13+
FLAGS = tf.app.flags.FLAGS
14+
epsilon = 0.001
15+
16+
17+
class Voxel_flow_model(object):
18+
def __init__(self, is_train=True):
19+
self.is_train = is_train
20+
21+
def inference(self, input_images):
22+
"""Inference on a set of input_images.
23+
Args:
24+
"""
25+
return self._build_model(input_images)
26+
27+
def total_var(self, images):
28+
pixel_dif1 = images[:, 1:, :, :] - images[:, :-1, :, :]
29+
pixel_dif2 = images[:, :, 1:, :] - images[:, :, :-1, :]
30+
tot_var = (tf.reduce_mean(tf.sqrt(tf.square(pixel_dif1) + epsilon**2)) + tf.reduce_mean(tf.sqrt(tf.square(pixel_dif2) + epsilon**2)))
31+
return tot_var
32+
33+
def loss(self, predictions, targets):
34+
"""Compute the necessary loss for training.
35+
Args:
36+
Returns:
37+
"""
38+
# self.reproduction_loss = l1_loss(predictions, targets)
39+
self.reproduction_loss = tf.reduce_mean(tf.sqrt(tf.square(predictions - targets) + epsilon**2))
40+
41+
self.motion_loss = self.total_var(self.flow)
42+
self.mask_loss = self.total_var(self.mask)
43+
44+
# return [self.reproduction_loss, self.prior_loss]
45+
return self.reproduction_loss + 0.01 * self.motion_loss + 0.005 * self.mask_loss
46+
47+
def l1loss(self, predictions, targets):
48+
self.reproduction_loss = l1_loss(predictions, targets)
49+
return self.reproduction_loss
50+
51+
def _build_model(self, input_images):
52+
with slim.arg_scope([slim.conv2d],
53+
activation_fn=tf.nn.leaky_relu,
54+
weights_initializer=tf.truncated_normal_initializer(0.0, 0.01),
55+
weights_regularizer=slim.l2_regularizer(0.0001)):
56+
# Define network
57+
batch_norm_params = {
58+
'decay': 0.9997,
59+
'epsilon': 0.001,
60+
'is_training': self.is_train,
61+
}
62+
with slim.arg_scope([slim.batch_norm], is_training=self.is_train, updates_collections=None):
63+
with slim.arg_scope([slim.conv2d], normalizer_fn=slim.batch_norm,
64+
normalizer_params=batch_norm_params):
65+
x0 = slim.conv2d(input_images, 32, [7, 7], stride=1, scope='conv1')
66+
x0_1 = slim.conv2d(x0, 32, [7, 7], stride=1, scope='conv1_1')
67+
68+
net = slim.avg_pool2d(x0_1, [2, 2], scope='pool1')
69+
x1 = slim.conv2d(net, 64, [5, 5], stride=1, scope='conv2')
70+
x1_1 = slim.conv2d(x1, 64, [5, 5], stride=1, scope='conv2_1')
71+
72+
net = slim.avg_pool2d(x1_1, [2, 2], scope='pool2')
73+
x2 = slim.conv2d(net, 128, [3, 3], stride=1, scope='conv3')
74+
x2_1 = slim.conv2d(x2, 128, [3, 3], stride=1, scope='conv3_1')
75+
76+
net = slim.avg_pool2d(x2_1, [2, 2], scope='pool3')
77+
x3 = slim.conv2d(net, 256, [3, 3], stride=1, scope='conv4')
78+
x3_1 = slim.conv2d(x3, 256, [3, 3], stride=1, scope='conv4_1')
79+
80+
net = slim.avg_pool2d(x3_1, [2, 2], scope='pool4')
81+
x4 = slim.conv2d(net, 512, [3, 3], stride=1, scope='conv5')
82+
x4_1 = slim.conv2d(x4, 512, [3, 3], stride=1, scope='conv5_1')
83+
84+
net = slim.avg_pool2d(x4_1, [2, 2], scope='pool5')
85+
net = slim.conv2d(net, 512, [3, 3], stride=1, scope='conv6')
86+
net = slim.conv2d(net, 512, [3, 3], stride=1, scope='conv6_1')
87+
88+
net = tf.image.resize_bilinear(net, [x4.get_shape().as_list()[1], x4.get_shape().as_list()[2]])
89+
net = slim.conv2d(tf.concat([net, x4_1], -1), 512, [3, 3], stride=1, scope='conv7')
90+
net = slim.conv2d(net, 512, [3, 3], stride=1, scope='conv7_1')
91+
92+
net = tf.image.resize_bilinear(net, [x3.get_shape().as_list()[1], x3.get_shape().as_list()[2]])
93+
net = slim.conv2d(tf.concat([net, x3_1], -1), 256, [3, 3], stride=1, scope='conv8')
94+
net = slim.conv2d(net, 256, [3, 3], stride=1, scope='conv8_1')
95+
96+
net = tf.image.resize_bilinear(net, [x2.get_shape().as_list()[1], x2.get_shape().as_list()[2]])
97+
net = slim.conv2d(tf.concat([net, x2_1], -1), 128, [3, 3], stride=1, scope='conv9')
98+
net = slim.conv2d(net, 128, [3, 3], stride=1, scope='conv9_1')
99+
100+
net = tf.image.resize_bilinear(net, [x1.get_shape().as_list()[1], x1.get_shape().as_list()[2]])
101+
net = slim.conv2d(tf.concat([net, x1_1], -1), 64, [3, 3], stride=1, scope='conv10')
102+
net = slim.conv2d(net, 64, [3, 3], stride=1, scope='conv10_1')
103+
104+
net = tf.image.resize_bilinear(net, [x0.get_shape().as_list()[1], x0.get_shape().as_list()[2]])
105+
net = slim.conv2d(tf.concat([net, x0_1], -1), 32, [3, 3], stride=1, scope='conv11')
106+
y0 = slim.conv2d(net, 32, [3, 3], stride=1, scope='conv11_1')
107+
108+
net = slim.conv2d(y0, 3, [5, 5], stride=1, activation_fn=tf.tanh,
109+
normalizer_fn=None, scope='conv12')
110+
net_copy = net
111+
112+
flow = net[:, :, :, 0:2]
113+
mask = tf.expand_dims(net[:, :, :, 2], 3)
114+
115+
self.flow = flow
116+
117+
118+
grid_x, grid_y = meshgrid(x0.get_shape().as_list()[1], x0.get_shape().as_list()[2])
119+
grid_x = tf.tile(grid_x, [FLAGS.batch_size, 1, 1])
120+
grid_y = tf.tile(grid_y, [FLAGS.batch_size, 1, 1])
121+
122+
flow = 0.5 * flow
123+
124+
flow_ratio = tf.constant([255.0 / (x0.get_shape().as_list()[2]-1), 255.0 / (x0.get_shape().as_list()[1]-1)])
125+
flow = flow * tf.expand_dims(tf.expand_dims(tf.expand_dims(flow_ratio, 0), 0), 0)
126+
127+
coor_x_1 = grid_x + flow[:, :, :, 0]
128+
coor_y_1 = grid_y + flow[:, :, :, 1]
129+
130+
coor_x_2 = grid_x - flow[:, :, :, 0]
131+
coor_y_2 = grid_y - flow[:, :, :, 1]
132+
133+
output_1 = bilinear_interp(input_images[:, :, :, 0:3], coor_x_1, coor_y_1, 'interpolate')
134+
output_2 = bilinear_interp(input_images[:, :, :, 3:6], coor_x_2, coor_y_2, 'interpolate')
135+
136+
self.warped_img1 = output_1
137+
self.warped_img2 = output_2
138+
139+
self.warped_flow1 = bilinear_interp(-flow[:, :, :, 0:3]*0.5, coor_x_1, coor_y_1, 'interpolate')
140+
self.warped_flow2 = bilinear_interp(flow[:, :, :, 0:3]*0.5, coor_x_2, coor_y_2, 'interpolate')
141+
142+
mask = 0.5 * (1.0 + mask)
143+
self.mask = mask
144+
mask = tf.tile(mask, [1, 1, 1, 3])
145+
net = tf.multiply(mask, output_1) + tf.multiply(1.0 - mask, output_2)
146+
147+
return [net, net_copy]

0 commit comments

Comments
 (0)