-
Notifications
You must be signed in to change notification settings - Fork 0
/
model.py
145 lines (124 loc) · 7.21 KB
/
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
# -*- coding: UTF-8 -*-
import tensorflow as tf
class Model(object):
def __init__(self):
self.weight_init = tf.truncated_normal_initializer(mean=0.0, stddev=0.02)
def lrelu(self, bottom, a=0.2):
with tf.name_scope("lrelu"):
bottom = tf.identity(bottom)
return (0.5 * (1 + a)) * bottom + (0.5 * (1 - a)) * tf.abs(bottom)
def avg_pool(self, bottom, name):
return tf.nn.avg_pool(bottom, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME', name=name)
def conv(self, bottom, channels, kernel=3, stride=2, use_bias=True, sn=False, scope='conv_0'):
with tf.variable_scope(scope):
if sn:
w = tf.get_variable("kernel", shape=[kernel, kernel, bottom.get_shape()[-1], channels],
initializer=self.weight_init,
regularizer=tf.contrib.layers.l2_regularizer(0.003))
bias = tf.get_variable("bias", [channels], initializer=tf.constant_initializer(0.0))
w_sn = self.spectral_norm(w)
bottom = tf.nn.conv2d(input=bottom, filter=w_sn,
strides=[1, stride, stride, 1], padding='SAME')
if use_bias:
bottom = tf.nn.bias_add(bottom, bias)
else:
bottom = tf.layers.conv2d(inputs=bottom, filters=channels,
kernel_size=kernel, kernel_initializer=self.weight_init,
kernel_regularizer=tf.contrib.layers.l2_regularizer(0.003),
strides=(stride, stride), padding='SAME', use_bias=use_bias)
return bottom
def deconv(self, bottom, channels, kernel=3, stride=2, use_bias=True, sn=False, scope='deconv_0'):
with tf.variable_scope(scope):
bottom_shape = bottom.shape.as_list()
# None 无法由list转为Tensor
# output_shape = [x_shape[0], x_shape[1] * stride, x_shape[2] * stride, channels]
output_shape = [1, bottom_shape[1] * stride, bottom_shape[2] * stride, channels]
output_shape = tf.convert_to_tensor(output_shape[:])
if sn:
w = tf.get_variable("kernel", shape=[kernel, kernel, channels, bottom.get_shape()[-1]],
initializer=self.weight_init, regularizer=tf.contrib.layers.l2_regularizer(0.003))
w_sn = self.spectral_norm(w)
bottom = tf.nn.conv2d_transpose(value=bottom, filter=w_sn, output_shape=output_shape,
strides=[1, stride, stride, 1], padding='SAME')
if use_bias:
bias = tf.get_variable("bias", [channels], initializer=tf.constant_initializer(0.0))
bottom = tf.nn.bias_add(bottom, bias)
else:
bottom = tf.layers.conv2d_transpose(inputs=bottom, filters=channels,
kernel_size=kernel, kernel_initializer=self.weight_init,
kernel_regularizer=tf.contrib.layers.l2_regularizer(0.003),
strides=(stride, stride), padding='SAME', use_bias=use_bias)
return bottom
def dis_conv(self, bottom, channels, kernel=4, stride=2, use_bias=True, sn=False, scope='conv_0'):
with tf.variable_scope(scope):
if sn:
w = tf.get_variable("kernel", shape=[kernel, kernel, bottom.get_shape()[-1], channels],
initializer=self.weight_init,
regularizer=tf.contrib.layers.l2_regularizer(0.003))
bias = tf.get_variable("bias", [channels], initializer=tf.constant_initializer(0.0))
w_sn = self.spectral_norm(w)
# tf.add_to_collection(scope.split('_')[0] + "_loss", self.ortho_norm(w_sn))
bottom = tf.nn.conv2d(input=bottom, filter=w_sn, strides=[1, stride, stride, 1], padding='SAME')
if use_bias:
bottom = tf.nn.bias_add(bottom, bias)
else:
bottom = tf.layers.conv2d(inputs=bottom, filters=channels,
kernel_size=kernel, kernel_initializer=self.weight_init,
kernel_regularizer=tf.contrib.layers.l2_regularizer(0.003),
strides=(stride, stride), padding='SAME', use_bias=use_bias)
return bottom
def batchnorm(self, bottom, is_training, scope):
with tf.variable_scope(scope):
return tf.layers.batch_normalization(bottom, axis=3, epsilon=1e-5, momentum=0.1, training=is_training,
gamma_initializer=tf.random_normal_initializer(1.0, 0.02))
def spectral_norm(self, w, iteration=1):
# tensor 可以使用as_list？
w_shape = w.shape.as_list()
w = tf.reshape(w, [-1, w_shape[-1]])
# Gets an existing variable with these parameters or create a new one.
u = tf.get_variable("u", [1, w_shape[-1]], initializer=tf.truncated_normal_initializer(), trainable=False)
u_hat = u
v_hat = None
for i in range(iteration):
"""
power iteration
Usually iteration = 1 will be enough
"""
v_ = tf.matmul(u_hat, tf.transpose(w))
v_hat = self.l2_norm(v_)
u_ = tf.matmul(v_hat, w)
u_hat = self.l2_norm(u_)
sigma = tf.matmul(tf.matmul(v_hat, w), tf.transpose(u_hat))
w_norm = w / sigma
with tf.control_dependencies([u.assign(u_hat)]):
w_norm = tf.reshape(w_norm, w_shape)
return w_norm
def ortho_norm(self, w):
w_shape = w.shape.as_list()
w = tf.reshape(w, [-1, w_shape[-1]])
I = tf.eye(w_shape[-1])
x = tf.multiply(tf.matmul(tf.transpose(w), w), 1 - I)
return tf.reduce_mean(x)
def l2_norm(self, v, eps=1e-12):
return v / (tf.reduce_sum(v ** 2) ** 0.5 + eps)
def average_gradients(self, tower_grads):
average_grads = []
for grad_and_vars in zip(*tower_grads):
# Note that each grad_and_vars looks like the following:
# ((grad0_gpu0, var0_gpu0), ... , (grad0_gpuN, var0_gpuN))
grads = []
for g, _ in grad_and_vars:
# Add 0 dimension to the gradients to represent the tower.
expanded_g = tf.expand_dims(g, 0)
# Append on a 'tower' dimension which we will average over below.
grads.append(expanded_g)
# Average over the 'tower' dimension.
grad = tf.concat(grads, 0)
grad = tf.reduce_mean(grad, 0)
# Keep in mind that the Variables are redundant because they are shared
# across towers. So .. we will just return the first tower's pointer to
# the Variable.
v = grad_and_vars[0][1]
grad_and_var = (grad, v)
average_grads.append(grad_and_var)
return average_grads