-
Notifications
You must be signed in to change notification settings - Fork 0
/
dan_model.py
180 lines (135 loc) · 8.4 KB
/
dan_model.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
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import numpy as np
import tensorflow as tf
_BATCH_NORM_DECAY = 0.997
_BATCH_NORM_EPSILON = 1e-5
DEFAULT_VERSION = 2
def batch_norm(inputs,training,data_format):
"""Performs a batch normalization using a standard set of parameters."""
# We set fused=True for a significant performance boost. See
# https://www.tensorflow.org/performance/performance_guide#common_fused_ops
return tf.layers.batch_normalization(inputs=inputs, axis=1 if data_format == 'channels_first' else -1,
momentum=_BATCH_NORM_DECAY, epsilon=_BATCH_NORM_EPSILON, center=True,
scale=True, training=training, fused=True)
def vgg_block(inputs,filters,num_convs,training,kernel_size,maxpool,data_format):
for i in range(num_convs):
inputs = batch_norm(tf.layers.conv2d(inputs, filters,kernel_size, 1,
padding='same',activation=tf.nn.relu, kernel_initializer=tf.glorot_uniform_initializer(),
data_format=data_format),training=training,data_format=data_format)
if maxpool:
inputs = tf.layers.max_pooling2d(inputs,2,2)
return inputs
class Model(object):
def __init__(self, num_lmark, img_size, filter_sizes, num_convs, kernel_size, data_format=None):
if not data_format:
data_format = ('channels_first' if tf.test.is_built_with_cuda() else 'channels_last')
self.data_format = data_format
self.filter_sizes = filter_sizes
self.num_convs = num_convs
self.num_lmark = num_lmark
self.kernel_size = kernel_size
self.img_size = img_size
self.__pixels__ = tf.constant([(x, y) for y in range(self.img_size) for x in range(self.img_size)], dtype=tf.float32,shape=[1,self.img_size,self.img_size,2])
#self.__pixels__ = tf.tile(self.__pixels__,[num_lmark,1,1,1])
def __calc_affine_params(self, from_shape, to_shape):
from_shape = tf.reshape(from_shape,[-1,self.num_lmark,2])
to_shape = tf.reshape(to_shape,[-1,self.num_lmark,2])
from_mean = tf.reduce_mean(from_shape, axis=1, keepdims=True)
to_mean = tf.reduce_mean(to_shape, axis=1, keepdims=True)
from_centralized = from_shape - from_mean
to_centralized = to_shape - to_mean
dot_result = tf.reduce_sum(tf.multiply(from_centralized, to_centralized), axis=[1, 2])
norm_pow_2 = tf.pow(tf.norm(from_centralized, axis=[1, 2]), 2)
a = dot_result / norm_pow_2
b = tf.reduce_sum(tf.multiply(from_centralized[:, :, 0], to_centralized[:, :, 1]) - tf.multiply(from_centralized[:, :, 1], to_centralized[:, :, 0]), 1) / norm_pow_2
r = tf.reshape(tf.stack([a, b, -b, a], axis=1), [-1, 2, 2])
t = to_mean - tf.matmul(from_mean, r)
return r, t
def __affine_image(self,imgs,r,t):
# The Tensor [imgs].format is [NHWC]
r = tf.matrix_inverse(r)
r = tf.matrix_transpose(r)
rm = tf.reshape(tf.pad(r, [[0, 0], [0, 0], [0, 1]], mode='CONSTANT'), [-1, 6])
rm = tf.pad(rm, [[0, 0], [0, 2]], mode='CONSTANT')
tm = tf.contrib.image.translations_to_projective_transforms(tf.reshape(t, [-1, 2]))
rtm = tf.contrib.image.compose_transforms(rm, tm)
return tf.contrib.image.transform(imgs, rtm, "BILINEAR")
def __affine_shape(self,shapes,r,t,isinv=False):
if isinv:
r = tf.matrix_inverse(r)
t = tf.matmul(-t,r)
shapes = tf.matmul(shapes,r) + t
return shapes
def __gen_heatmap(self,shapes):
shapes = shapes[:, :, tf.newaxis, tf.newaxis, :]
value = self.__pixels__ - shapes
value = tf.norm(value,axis=-1)
value = 1.0 / (tf.reduce_min(value,axis=1) + 1.0)
value = tf.expand_dims(value,axis=-1)
return value
def __call__(self,
inputs_imgs,
s1_training,
s2_training,
mean_shape,
imgs_mean,
imgs_std):
rd = {}
inputs_imgs = tf.reshape(inputs_imgs, [-1, self.img_size, self.img_size, 1])
tf.summary.image('image', inputs_imgs, max_outputs=6)
rd['img'] = inputs_imgs
mean_shape = tf.reshape(mean_shape,[self.num_lmark,2]) if mean_shape is not None else tf.zeros([self.num_lmark,2],tf.float32)
imgs_mean = tf.reshape(imgs_mean,[self.img_size,self.img_size,1]) if imgs_mean is not None else tf.zeros([self.img_size,self.img_size,1], tf.float32)
imgs_std = tf.reshape(imgs_std,[self.img_size,self.img_size,1]) if imgs_std is not None else tf.ones([self.img_size,self.img_size,1], tf.float32)
imgs_mean_tensor = tf.get_variable('imgs_mean',trainable=False,initializer=imgs_mean)
imgs_std_tensor = tf.get_variable('imgs_std',trainable=False,initializer=imgs_std)
shape_mean_tensor = tf.get_variable('shape_mean',trainable=False,initializer=mean_shape)
inputs_imgs = (inputs_imgs - imgs_mean_tensor) / imgs_std_tensor
# Convert the inputs from channels_last (NHWC) to channels_first
# (NCHW).
# This provides a large performance boost on GPU. See
# https://www.tensorflow.org/performance/performance_guide#data_formats
with tf.variable_scope('s1'):
inputs = inputs_imgs
if self.data_format == 'channels_first':
inputs = tf.transpose(inputs, [0, 3, 1, 2])
for i, num_filter in enumerate(self.filter_sizes):
inputs = vgg_block(inputs=inputs,filters=num_filter,num_convs=self.num_convs,
training=s1_training,kernel_size=self.kernel_size,maxpool=True,
data_format=self.data_format)
inputs = tf.contrib.layers.flatten(inputs)
inputs = tf.layers.dropout(inputs,0.5,training=s1_training)
s1_fc1 = tf.layers.dense(inputs,256,activation=tf.nn.relu, kernel_initializer=tf.glorot_uniform_initializer())
s1_fc1 = batch_norm(s1_fc1,s1_training,data_format=self.data_format)
s1_fc2 = tf.layers.dense(s1_fc1,self.num_lmark * 2,activation=None)
rd['s1_ret'] = tf.identity(tf.reshape(s1_fc2,[-1,self.num_lmark,2]) + shape_mean_tensor, name='output_landmark')
with tf.variable_scope('s2'):
r,t = self.__calc_affine_params(rd['s1_ret'], shape_mean_tensor)
inputs = self.__affine_image(inputs_imgs,r,t)
s2_lmark = self.__affine_shape(rd['s1_ret'],r,t)
s2_heatmap = self.__gen_heatmap(s2_lmark)
s2_feature = tf.layers.dense(s1_fc1,(self.img_size // 2) ** 2,activation=tf.nn.relu,kernel_initializer=tf.glorot_uniform_initializer())
s2_feature = tf.reshape(s2_feature,[-1,self.img_size // 2,self.img_size // 2,1])
s2_feature_upscale = tf.image.resize_images(s2_feature,[self.img_size,self.img_size])
tf.summary.image('heatmap', s2_heatmap, max_outputs=6)
tf.summary.image('feature', s2_feature, max_outputs=6)
tf.summary.image('image', inputs, max_outputs=6)
if self.data_format == 'channels_first':
inputs = tf.transpose(inputs, [0, 3, 1, 2])
s2_heatmap = tf.transpose(s2_heatmap,[0, 3, 1, 2])
s2_feature_upscale = tf.transpose(s2_feature_upscale, [0, 3, 1, 2])
inputs = tf.concat([inputs,s2_heatmap,s2_feature_upscale],axis= 1 if self.data_format == 'channels_first' else 3)
inputs = batch_norm(inputs,s2_training,self.data_format)
for i, num_filter in enumerate(self.filter_sizes):
inputs = vgg_block(inputs=inputs,filters=num_filter,num_convs=self.num_convs,
training=s2_training,kernel_size=self.kernel_size,maxpool=True, data_format=self.data_format)
inputs = tf.contrib.layers.flatten(inputs)
inputs = tf.layers.dropout(inputs,0.5,training=s2_training)
s2_fc1 = tf.layers.dense(inputs,256,activation=tf.nn.relu,kernel_initializer=tf.glorot_uniform_initializer())
s2_fc1 = batch_norm(s2_fc1,s2_training,data_format=self.data_format)
s2_fc2 = tf.layers.dense(s2_fc1,self.num_lmark * 2, activation=None)
s2_fc2 = tf.reshape(s2_fc2, [-1,self.num_lmark,2]) + s2_lmark
rd['s2_ret'] = tf.identity(self.__affine_shape(s2_fc2,r,t,isinv=True), name='output_landmark')
return rd