-
Notifications
You must be signed in to change notification settings - Fork 35
/
hourglass_net_normal.py
119 lines (88 loc) · 4.81 KB
/
hourglass_net_normal.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
# Borrowed from https://github.com/xuwenzhe/EECS442-Challenge-Surface-Normal-Estimation
import numpy as np
import tensorflow as tf
VARIABLE_COUNTER = 0
NUM_CH = [64,128,256,512,1024]
KER_SZ = 3
########## TABLE 1 ARCHITECTURE PARAMETERS #########
color_encode = "ABCDEFG"
block_in = [128,128,128,128,256,256,256]
block_out = [64,128,128,256,256,256,128]
block_inter = [64,32,64,32,32,64,32]
block_conv1 = [1,1,1,1,1,1,1]
block_conv2 = [3,3,3,3,3,3,3]
block_conv3 = [7,5,7,5,5,7,5]
block_conv4 = [11,7,11,7,7,11,7]
####################### TABLE 1 END#################
def variable(name, shape, initializer,regularizer=None):
global VARIABLE_COUNTER
with tf.device('/gpu:0'):
VARIABLE_COUNTER += np.prod(np.array(shape))
return tf.get_variable(name, shape, initializer=initializer, regularizer=regularizer, dtype=tf.float32, trainable=True)
def conv_layer(input_tensor,name,kernel_size,output_channels,initializer=tf.contrib.layers.variance_scaling_initializer(),stride=1,bn=False,training=False,relu=True):
input_channels = input_tensor.get_shape().as_list()[-1]
with tf.variable_scope(name) as scope:
kernel = variable('weights', [kernel_size, kernel_size, input_channels, output_channels], initializer, regularizer=tf.contrib.layers.l2_regularizer(0.0005))
conv = tf.nn.conv2d(input_tensor, kernel, [1, stride, stride, 1], padding='SAME')
biases = variable('biases', [output_channels], tf.constant_initializer(0.0))
conv_layer = tf.nn.bias_add(conv, biases)
if bn:
conv_layer = batch_norm_layer(conv_layer,scope,training)
if relu:
conv_layer = tf.nn.relu(conv_layer, name=scope.name)
# print('Conv layer {0} -> {1}'.format(input_tensor.get_shape().as_list(),conv_layer.get_shape().as_list()))
return conv_layer
def max_pooling(input_tensor,name,factor=2):
pool = tf.nn.max_pool(input_tensor, ksize=[1, factor, factor, 1], strides=[1, factor, factor, 1], padding='SAME', name=name)
# print('Pooling layer {0} -> {1}'.format(input_tensor.get_shape().as_list(),pool.get_shape().as_list()))
return pool
def batch_norm_layer(input_tensor,scope,training):
return tf.contrib.layers.batch_norm(input_tensor,scope=scope,is_training=training,decay=0.99)
def hourglass_normal_prediction(netIN,training):
print('-'*30)
print('Hourglass Architecture')
print('-'*30)
global VARIABLE_COUNTER
VARIABLE_COUNTER = 0;
layer_name_dict = {}
def layer_name(base_name):
if base_name not in layer_name_dict:
layer_name_dict[base_name] = 0
layer_name_dict[base_name] += 1
name = base_name + str(layer_name_dict[base_name])
return name
bn = True
def hourglass_stack_no_incep(stack_in):
c0 = conv_layer(stack_in,layer_name('conv'),KER_SZ,NUM_CH[0],bn=bn,training=training)
c1 = conv_layer(c0,layer_name('conv'),KER_SZ,NUM_CH[0],bn=bn,training=training)
c2 = conv_layer(c1,layer_name('conv'),KER_SZ,NUM_CH[0],bn=bn,training=training)
p0 = max_pooling(c2,layer_name('pool'))
c3 = conv_layer(p0,layer_name('conv'),KER_SZ,NUM_CH[1],bn=bn,training=training)
p1 = max_pooling(c3,layer_name('pool'))
c4 = conv_layer(p1,layer_name('conv'),KER_SZ,NUM_CH[2],bn=bn,training=training)
p2 = max_pooling(c4,layer_name('pool'))
c5 = conv_layer(p2,layer_name('conv'),KER_SZ,NUM_CH[3],bn=bn,training=training)
p3 = max_pooling(c5,layer_name('pool'))
c6 = conv_layer(p3,layer_name('conv'),KER_SZ,NUM_CH[4],bn=bn,training=training)
c7 = conv_layer(c6,layer_name('conv'),KER_SZ,NUM_CH[4],bn=bn,training=training)
c8 = conv_layer(c7,layer_name('conv'),KER_SZ,NUM_CH[3],bn=bn,training=training)
r0 = tf.image.resize_images(c8,[c8.get_shape().as_list()[1]*2, c8.get_shape().as_list()[2]*2])
cat0 = tf.concat([r0,c5],3)
c9 = conv_layer(cat0,layer_name('conv'),KER_SZ,NUM_CH[3],bn=bn,training=training)
c10 = conv_layer(c9,layer_name('conv'),KER_SZ,NUM_CH[2],bn=bn,training=training)
r1 = tf.image.resize_images(c10,[c10.get_shape().as_list()[1]*2, c10.get_shape().as_list()[2]*2])
cat1 = tf.concat([r1,c4],3)
c11 = conv_layer(cat1,layer_name('conv'),KER_SZ,NUM_CH[2],bn=bn,training=training)
c12 = conv_layer(c11,layer_name('conv'),KER_SZ,NUM_CH[1],bn=bn,training=training)
r2 = tf.image.resize_images(c12,[c12.get_shape().as_list()[1]*2, c12.get_shape().as_list()[2]*2])
cat2 = tf.concat([r2,c3],3)
c13 = conv_layer(cat2,layer_name('conv'),KER_SZ,NUM_CH[1],bn=bn,training=training)
c14 = conv_layer(c13,layer_name('conv'),KER_SZ,NUM_CH[0],bn=bn,training=training)
r3 = tf.image.resize_images(c14,[c14.get_shape().as_list()[1]*2, c14.get_shape().as_list()[2]*2])
cat3 = tf.concat([r3,c2],3)
c15 = conv_layer(cat3,layer_name('conv'),KER_SZ,NUM_CH[0],bn=bn,training=training)
c16 = conv_layer(c15,layer_name('conv'),KER_SZ,NUM_CH[0],bn=bn,training=training)
stack_out_d = conv_layer(c16, layer_name('conv'), 1, 3, bn=False, training=training, relu=False)
return stack_out_d
out0_n = hourglass_stack_no_incep(netIN)
return out0_n