-
Notifications
You must be signed in to change notification settings - Fork 4
/
ops.py
139 lines (123 loc) · 5.45 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
import tensorflow as tf
import numpy as np
import cv2 as cv
import math,os,h5py,argparse,sys
def conv(name,x,kers,s,outs,pad):
with tf.variable_scope(name):
ker = int(math.sqrt(kers))
shape = [i.value for i in x.get_shape()]
w = tf.get_variable('w',
[ker,ker,shape[-1],outs],
tf.float32,
tf.contrib.layers.xavier_initializer_conv2d())
b = tf.get_variable('b',[outs],tf.float32,tf.constant_initializer(0.))
padd = "SAME" if pad else "VALID"
x_conv = tf.nn.conv2d(x,w,[1,s,s,1],padd) + b
return x_conv
def res_block(name,x):
with tf.variable_scope(name):
shape = [i.value for i in x.get_shape()]
conv1 = conv(name+'_conv1',x,3*3,shape[-1],1,True)
bn1 = tf.nn.relu(tf.contrib.layers.batch_norm(conv1,scale=True,updates_collections=None))
conv2 = conv(name+'_conv2',bn1,3*3,shape[-1],1,True)
bn2 = tf.contrib.layers.batch_norm(conv2,scale=True,updates_collections=None)
return tf.nn.relu(bn2+x)
def conv_trans(name,x,sizes,s,outs,ifpad):
with tf.variable_scope(name):
ker = int(math.sqrt(sizes))
shape = [i.value for i in x.get_shape()]
ins = shape[-1]//4
w = tf.get_variable('w',[ker,ker,ins,outs],tf.float32,tf.contrib.layers.xavier_initializer_conv2d())
b = tf.get_variable('b',[outs],tf.float32,tf.constant_initializer(0.))
pad = "SAME" if ifpad else "VALID"
x_conv = tf.nn.conv2d(tf.depth_to_space(x,2),w,[1,s,s,1],pad)+b
return x_conv
def lrelu(name,x):
with tf.variable_scope(name):
return tf.nn.relu(x)
def tanh(name,x):
with tf.variable_scope(name):
return tf.nn.tanh(x)
def BN(name,x):
with tf.variable_scope(name):
return tf.contrib.layers.batch_norm(x,scale=True,updates_collections=None)
def FC_location(name,x,outs,im_size,hw_size):
with tf.variable_scope(name):
raw_shape = [i.value for i in x.get_shape()]
new_shape = int(raw_shape[1]*raw_shape[2]*raw_shape[3])
x_resh = tf.reshape(x,[-1,new_shape])
w = tf.get_variable('w',[new_shape,outs],tf.float32,tf.contrib.layers.xavier_initializer())
b = tf.get_variable('b',[1,outs],tf.float32,tf.constant_initializer(0.))
fc_ = tf.matmul(x_resh,w)+b
# Add a range limit to the position coordinates while making the gradient softer.
xy_constraint = tf.nn.sigmoid(fc_)*(im_size-hw_size)
return tf.cast(tf.round(xy_constraint),tf.int32)
def get_constant(im_size):
zero_x = np.zeros([im_size,im_size],np.float32)
zero_y = np.zeros([im_size,im_size],np.float32)
for i in range(im_size):
zero_x[:,i] = i+1
zero_y[i,:] = i+1
return zero_x,zero_y
def sigmoid_mask(x,k):
k = int(k)
X = tf.cast(x,tf.float32)
return 1/(1+tf.exp(-1*k*X))
def get_mask(xy,reigon_w,reigon_h,im_size,k,B):
# xy: [batch, 2]: coordinates
# reigon_w, reigon_h: size of the area
# im_size: Image size for generating the original X
# k: The growth factor of the sigmoid function
with tf.variable_scope('Mask'):
x_left = tf.expand_dims(tf.expand_dims(tf.expand_dims(xy[:,0],1),2),3)
x_right = tf.expand_dims(tf.expand_dims(tf.expand_dims(xy[:,0]+reigon_w,1),2),3)
y_top = tf.expand_dims(tf.expand_dims(tf.expand_dims(xy[:,1],1),2),3)
y_bottom = tf.expand_dims(tf.expand_dims(tf.expand_dims(xy[:,1]+reigon_h,1),2),3)
x_value,y_value = get_constant(im_size)
x_constant = np.tile(np.expand_dims(np.expand_dims(x_value,0),3),[B,1,1,1])
y_constant = np.tile(np.expand_dims(np.expand_dims(y_value,0),3),[B,1,1,1])
A = sigmoid_mask(x_constant-x_left,k)
C = sigmoid_mask(x_constant-x_right,k)
D = sigmoid_mask(y_constant-y_top,k)
E = sigmoid_mask(y_constant-y_bottom,k)
return (A-C)*(D-E)
def FC(name,x,outs):
with tf.variable_scope(name):
shape = [i.value for i in x.get_shape()]
size = int(shape[1]*shape[2]*shape[3])
x_reshape = tf.reshape(x,[-1,size])
w = tf.get_variable('w',[size,outs],tf.float32,tf.contrib.layers.xavier_initializer())
b = tf.get_variable('b',[1,outs],tf.float32,tf.constant_initializer(0.))
return tf.nn.sigmoid(tf.matmul(x_reshape,w)+b)
def resnet_classifer(name,x): # x: batch，4,4,512
with tf.variable_scope(name):
res1 = res_block('res1',x)
res2 = res_block('res2',res1)
res3 = res_block('res3',res2)
res4 = res_block('res4',res3)
res5 = res_block('res5',res4)
res6 = res_block('res6',res5)
res7 = res_block('res7',res6)
probs = FC('probs',res7,10)
return probs
def One_ep_Iter(im_dir,batch):
num_ims = min(len(os.listdir(im_dir+'positive')),len(os.listdir(im_dir+'negtive')))
return num_ims//batch
def get_shape(x):
L = [i.value for i in x.get_shape()]
return L
def save_im(sample_im,save_dir,cur_ep,cur_batch,batch):
for i in range(batch):
im_S = sample_im[i,:,:,:]
im_s = (im_S[:,:,::-1]+1)*127.5
s_name = 'ep'+str(cur_ep)+'_sample'+str(cur_batch)+'_batch'+str(i)+'.png'
cv.imwrite(save_dir+s_name,im_s)
return True
def Save_load(ims,num,save_dir):
b = np.shape(ims)[0]
for i in range(b):
im_S = ims[i,:,:,:]
im_s = (im_S[:,:,::-1]+1)*127.5
name = 'num_'+str(num)+'.png'
cv.imwrite(save_dir+name,im_s)
return True