forked from FFZhang1231/Facial-expression-recognition
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathops.py
150 lines (125 loc) · 5.76 KB
/
ops.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
from __future__ import division
import tensorflow as tf
import numpy as np
from scipy.misc import imread, imresize, imsave
import pdb
def conv2d(input_map, num_output_channels, size_kernel=5, stride=2, name='conv2d'):
with tf.variable_scope(name):
stddev = np.sqrt(2.0 / (np.sqrt(input_map.get_shape()[-1].value * num_output_channels) * size_kernel ** 2))
kernel = tf.get_variable(
name='w',
shape=[size_kernel, size_kernel, input_map.get_shape()[-1], num_output_channels],
dtype=tf.float32,
initializer=tf.truncated_normal_initializer(stddev=stddev)
)
biases = tf.get_variable(
name='b',
shape=[num_output_channels],
dtype=tf.float32,
initializer=tf.constant_initializer(0.0)
)
conv = tf.nn.conv2d(input_map, kernel, strides=[1, stride, stride, 1], padding='SAME')
return tf.nn.bias_add(conv, biases)
def conv2d2(input_map, num_output_channels, size_kernel=5, stride=2, name='conv2d2'):
with tf.variable_scope(name):
stddev = np.sqrt(2.0 / (np.sqrt(input_map.get_shape()[-1].value * num_output_channels) * size_kernel ** 2))
kernel = tf.get_variable(
name='w',
shape=[size_kernel, size_kernel, input_map.get_shape()[-1], num_output_channels],
dtype=tf.float32,
initializer=tf.truncated_normal_initializer(stddev=stddev)
)
biases = tf.get_variable(
name='b',
shape=[num_output_channels],
dtype=tf.float32,
initializer=tf.constant_initializer(0.0)
)
conv = tf.nn.conv2d(input_map, kernel, strides=[1, stride, stride, 1], padding='SAME')
return tf.nn.bias_add(conv, biases)
def max_pool(bottom):
return tf.nn.max_pool(bottom, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME', name='POOL')
def avg_pool(bottom):
return tf.nn.avg_pool(bottom, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME', name='POOL')
def fc(input_vector, num_output_length, name='fc'):
with tf.variable_scope(name):
stddev = np.sqrt(1.0 / (np.sqrt(input_vector.get_shape()[-1].value * num_output_length)))
w = tf.get_variable(
name='w',
shape=[input_vector.get_shape()[1], num_output_length],
dtype=tf.float32,
initializer=tf.random_normal_initializer(stddev=stddev)
)
b = tf.get_variable(
name='b',
shape=[num_output_length],
dtype=tf.float32,
initializer=tf.constant_initializer(0.0)
)
return tf.matmul(input_vector, w) + b
def deconv2d(input_map, output_shape, size_kernel=5, stride=2, stddev=0.02, name='deconv2d'):
with tf.variable_scope(name):
stddev = np.sqrt(1.0 / (np.sqrt(input_map.get_shape()[-1].value * output_shape[-1]) * size_kernel ** 2))
# filter : [height, width, output_channels, in_channels]
kernel = tf.get_variable(
name='w',
shape=[size_kernel, size_kernel, output_shape[-1], input_map.get_shape()[-1]],
dtype=tf.float32,
initializer=tf.random_normal_initializer(stddev=stddev)
)
biases = tf.get_variable(
name='b',
shape=[output_shape[-1]],
dtype=tf.float32,
initializer=tf.constant_initializer(0.0)
)
deconv = tf.nn.conv2d_transpose(input_map, kernel, strides=[1, stride, stride, 1], output_shape=output_shape)
return tf.nn.bias_add(deconv, biases)
def lrelu(logits, leak=0.2):
return tf.maximum(logits, leak*logits)
def concat_label(x, label, duplicate=1):
x_shape = x.get_shape().as_list()
if duplicate < 1:
return x
# duplicate the label to enhance its effect, does it really affect the result?
label = tf.tile(label, [1, duplicate])
label_shape = label.get_shape().as_list()
if len(x_shape) == 2:
return tf.concat(1, [x, label])
elif len(x_shape) == 4:
label = tf.reshape(label, [x_shape[0], 1, 1, label_shape[-1]])
return tf.concat(3, [x, label*tf.ones([x_shape[0], x_shape[1], x_shape[2], label_shape[-1]])])
def load_image(
image_path, # path of a image
image_size=64, # expected size of the image
image_value_range=(-1, 1), # expected pixel value range of the image
is_gray=False, # gray scale or color image
):
if is_gray:
image = imread(image_path, flatten=True).astype(np.float32)
else:
image = imread(image_path).astype(np.float32)
image = imresize(image, [image_size, image_size])
# pdb.set_trace()
image = image.astype(np.float32) * (image_value_range[-1] - image_value_range[0]) / 255.0 + image_value_range[0]
# array_img = np.array(image)
# image= (array_img - array_img.mean()) / array_img.std()
return image
def save_batch_images(
batch_images, # a batch of images
save_path, # path to save the images
image_value_range=(-1,1), # value range of the input batch images
size_frame=None # size of the image matrix, number of images in each row and column
):
# transform the pixcel value to 0~1
images = (batch_images - image_value_range[0]) / (image_value_range[-1] - image_value_range[0])
if size_frame is None:
auto_size = int(np.ceil(np.sqrt(images.shape[0])))
size_frame = [auto_size, auto_size]
img_h, img_w = batch_images.shape[1], batch_images.shape[2]
frame = np.zeros([img_h * size_frame[0], img_w * size_frame[1], 3])
for ind, image in enumerate(images):
ind_col = ind % size_frame[1]
ind_row = ind // size_frame[1]
frame[(ind_row * img_h):(ind_row * img_h + img_h), (ind_col * img_w):(ind_col * img_w + img_w), :] = image
imsave(save_path, frame)