-
Notifications
You must be signed in to change notification settings - Fork 8
/
Copy pathops.py
115 lines (102 loc) · 4.91 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
import tensorflow as tf
import tensorflow.contrib as contrib
from PatchMatch import *
def conv(inputs, w, b, strides=1):
return tf.nn.conv2d(inputs, w, [1, strides, strides, 1], "SAME") + b
def max_pooling(inputs, ksize=2, strides=2):
return tf.nn.max_pool(inputs, [1, ksize, ksize, 1], [1, strides, strides, 1], "SAME")
def relu(inputs):
return tf.nn.relu(inputs)
def vggnet(inputs):
inputs = tf.reverse(inputs, [-1]) - np.array([103.939, 116.779, 123.68])
para = np.load("./vgg19//vgg19.npy", encoding="latin1").item()
F = {}
inputs = relu(conv(inputs, para["conv1_1"][0], para["conv1_1"][1]))
F["conv1_1"] = inputs
inputs = relu(conv(inputs, para["conv1_2"][0], para["conv1_2"][1]))
inputs = max_pooling(inputs)
inputs = relu(conv(inputs, para["conv2_1"][0], para["conv2_1"][1]))
F["conv2_1"] = inputs
inputs = relu(conv(inputs, para["conv2_2"][0], para["conv2_2"][1]))
inputs = max_pooling(inputs)
inputs = relu(conv(inputs, para["conv3_1"][0], para["conv3_1"][1]))
F["conv3_1"] = inputs
inputs = relu(conv(inputs, para["conv3_2"][0], para["conv3_2"][1]))
inputs = relu(conv(inputs, para["conv3_3"][0], para["conv3_3"][1]))
inputs = relu(conv(inputs, para["conv3_4"][0], para["conv3_4"][1]))
inputs = max_pooling(inputs)
inputs = relu(conv(inputs, para["conv4_1"][0], para["conv4_1"][1]))
F["conv4_1"] = inputs
inputs = relu(conv(inputs, para["conv4_2"][0], para["conv4_2"][1]))
inputs = relu(conv(inputs, para["conv4_3"][0], para["conv4_3"][1]))
inputs = relu(conv(inputs, para["conv4_4"][0], para["conv4_4"][1]))
inputs = max_pooling(inputs)
inputs = relu(conv(inputs, para["conv5_1"][0], para["conv5_1"][1]))
F["conv5_1"] = inputs
return F
def vgg_block(layer, inputs):
para = np.load("./vgg19//vgg19.npy", encoding="latin1").item()
if layer == 1:
inputs = relu(conv(inputs, para["conv1_1"][0], para["conv1_1"][1]))
inputs_L = inputs * 1.0
inputs = relu(conv(inputs, para["conv1_2"][0], para["conv1_2"][1]))
inputs = max_pooling(inputs)
inputs = relu(conv(inputs, para["conv2_1"][0], para["conv2_1"][1]))
return inputs, inputs_L
if layer == 2:
inputs = relu(conv(inputs, para["conv2_1"][0], para["conv2_1"][1]))
inputs_L = inputs * 1.0
inputs = relu(conv(inputs, para["conv2_2"][0], para["conv2_2"][1]))
inputs = max_pooling(inputs)
inputs = relu(conv(inputs, para["conv3_1"][0], para["conv3_1"][1]))
return inputs, inputs_L
if layer == 3:
inputs = relu(conv(inputs, para["conv3_1"][0], para["conv3_1"][1]))
inputs_L = inputs * 1.0
inputs = relu(conv(inputs, para["conv3_2"][0], para["conv3_2"][1]))
inputs = relu(conv(inputs, para["conv3_3"][0], para["conv3_3"][1]))
inputs = relu(conv(inputs, para["conv3_4"][0], para["conv3_4"][1]))
inputs = max_pooling(inputs)
inputs = relu(conv(inputs, para["conv4_1"][0], para["conv4_1"][1]))
return inputs, inputs_L
if layer == 4:
inputs = relu(conv(inputs, para["conv4_1"][0], para["conv4_1"][1]))
inputs_L = inputs * 1.0
inputs = relu(conv(inputs, para["conv4_2"][0], para["conv4_2"][1]))
inputs = relu(conv(inputs, para["conv4_3"][0], para["conv4_3"][1]))
inputs = relu(conv(inputs, para["conv4_4"][0], para["conv4_4"][1]))
inputs = max_pooling(inputs)
inputs = relu(conv(inputs, para["conv5_1"][0], para["conv5_1"][1]))
return inputs, inputs_L
def preprocessing(A, B_prime):
F_A = vggnet(A)
F_B_prime = vggnet(B_prime)
return F_A, F_B_prime
def Deconvolve(sess, name, layer, warped_F_B_prime_L):
H = np.size(warped_F_B_prime_L, 0)
W = np.size(warped_F_B_prime_L, 1)
if layer == 1:
C = 3
elif layer == 2:
C = 64
elif layer == 3:
C = 128
else:
C = 256
var_name = name + "R_L_1"
temp = tf.get_variable(var_name, [1, 2 * H, 2 * W, C], initializer=tf.truncated_normal_initializer(stddev=0.02))
inputs, R_L_1 = vgg_block(layer, temp)
Loss = tf.reduce_sum(tf.square(inputs[0, :, :, :] - warped_F_B_prime_L))
sess.run(tf.global_variables_initializer())
print("Before L-BFGS, Loss: %f" % (sess.run(Loss)))
opt = contrib.opt.ScipyOptimizerInterface(Loss, method="L-BFGS-B", options={"maxiter": 50}, var_list=tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, var_name))
opt.minimize(sess)
print("After L-BFGS, Loss: %f" % (sess.run(Loss)))
R_L_1 = sess.run(R_L_1)
return R_L_1
def get_W_L_1(F_A_L_1, alpha_L_1, k=300, tau=0.05):
temp = np.sum(np.square(F_A_L_1), axis=-1, keepdims=True)
temp = (temp - np.min(temp)) / (np.max(temp) - np.min(temp))
W_L_1 = np.float32(temp > tau) * alpha_L_1
# W_L_1 = alpha_L_1 / (1 + np.exp(-k * (np.square(F_A_L_1) - tau)))
return W_L_1