-
Notifications
You must be signed in to change notification settings - Fork 11
/
model.py
161 lines (132 loc) · 5.05 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
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
import tensorflow as tf
"""
Convolutioanl Winner-Take-All Autoencoder TensorFlow implementation.
Usage :
ae = ConvWTA(sess)
# 1. to train an Autoencoder
loss = ae.loss(x)
train = optimizer.minimize(loss)
sess.run(train, feed_dict={...})
# 2. to get the sparse codes
h = ae.encoder(x)
sess.run(h, feed_dict={...})
# 3. to get the reconstructed results
y = ae.reconstruct(x)
sess.run(y, feed_dict={...})
# 4. to get the learned features
f = ae.features() # np.float32 array with shape [11, 11, 1, 16]
# 4-1. to train a different number of features
ae = ConvWTA(sess, num_features=32)
# 5. to save & restore the variables
ae.save(save_path)
ae.restore(save_path)
Reference:
[1] https://arxiv.org/pdf/1409.2752.pdf
"""
class ConvWTA(object):
"""
Args :
sess : TensorFlow session.
x : Input tensor.
"""
def __init__(self, sess, num_features=16, name="ConvWTA"):
self.sess = sess
self.name = name
self.size = [1, 128, 128, num_features] # ref [1]
self._set_variables()
self.t_vars = tf.get_collection(
tf.GraphKeys.TRAINABLE_VARIABLES, scope=self.name)
self.sess.run(tf.variables_initializer(self.t_vars))
self.saver = tf.train.Saver(self.t_vars)
def encoder(self, x):
with tf.variable_scope(self.name) as vs:
h = self._conv(x, self.size[1], 5, 5, 1, 1, "conv_1")
h = self._conv(h, self.size[2], 5, 5, 1, 1, "conv_2")
h = self._conv(h, self.size[3], 5, 5, 1, 1, "conv_3")
return h
def _decoder(self, h):
shape = tf.shape(h)
out_shape = tf.stack([shape[0], shape[1], shape[2], 1])
with tf.variable_scope(self.name) as vs:
y = self._deconv(h, out_shape, self.size[0],
11, 11, 1, 1, "deconv", end=True)
return y
def loss(self, x, lifetime_sparsity=0.20):
h = self.encoder(x)
h, winner = self._spatial_sparsity(h)
h = self._lifetime_sparsity(h, winner, lifetime_sparsity)
y = self._decoder(h)
return tf.reduce_sum(tf.square(y - x))
def reconstruct(self, x):
h = self.encoder(x)
h, _ = self._spatial_sparsity(h)
y = self._decoder(h)
return y
def _set_variables(self):
with tf.variable_scope(self.name) as vs:
self._conv_var(self.size[0], self.size[1], 5, 5, "conv_1")
self._conv_var(self.size[1], self.size[2], 5, 5, "conv_2")
self._conv_var(self.size[2], self.size[3], 5, 5, "conv_3")
self.f, _ = self._deconv_var(
self.size[-1], self.size[0], 11, 11, "deconv")
def _conv_var(self, in_dim, out_dim, k_h, k_w, name, stddev=0.1):
with tf.variable_scope(name) as vs:
k = tf.get_variable('filter',
[k_h, k_w, in_dim, out_dim],
initializer=tf.truncated_normal_initializer(stddev=stddev))
b = tf.get_variable('biases', [out_dim],
initializer=tf.constant_initializer(0.0001))
return k, b
def _deconv_var(self, in_dim, out_dim, k_h, k_w, name, stddev=0.1):
with tf.variable_scope(name) as vs:
k = tf.get_variable('filter',
[k_h, k_w, out_dim, in_dim],
initializer=tf.truncated_normal_initializer(stddev=stddev))
b = tf.get_variable('biases', [out_dim],
initializer=tf.constant_initializer(0.0001))
return k, b
def _conv(self, x, out_dim,
k_h, k_w, s_h, s_w, name, end=False):
with tf.variable_scope(name, reuse=True) as vs:
k = tf.get_variable('filter')
b = tf.get_variable('biases')
conv = tf.nn.conv2d(x, k, [1, s_h, s_w, 1], "SAME") + b
return conv if end else tf.nn.relu(conv)
def _deconv(self, x, out_shape, out_dim,
k_h, k_w, s_h, s_w, name, end=False):
with tf.variable_scope(name, reuse=True) as vs:
k = tf.get_variable('filter')
b = tf.get_variable('biases')
deconv = tf.nn.conv2d_transpose(
x, k, out_shape, [1, s_h, s_w, 1], "SAME") + b
return deconv if end else tf.nn.relu(deconv)
def _spatial_sparsity(self, h):
shape = tf.shape(h)
n = shape[0]
c = shape[3]
h_t = tf.transpose(h, [0, 3, 1, 2]) # n, c, h, w
h_r = tf.reshape(h_t, tf.stack([n, c, -1])) # n, c, h*w
th, _ = tf.nn.top_k(h_r, 1) # n, c, 1
th_r = tf.reshape(th, tf.stack([n, 1, 1, c])) # n, 1, 1, c
drop = tf.where(h < th_r,
tf.zeros(shape, tf.float32), tf.ones(shape, tf.float32))
# spatially dropped & winner
return h*drop, tf.reshape(th, tf.stack([n, c])) # n, c
def _lifetime_sparsity(self, h, winner, rate):
shape = tf.shape(winner)
n = shape[0]
c = shape[1]
k = tf.cast(rate * tf.cast(n, tf.float32), tf.int32)
winner = tf.transpose(winner) # c, n
th_k, _ = tf.nn.top_k(winner, k) # c, k
shape_t = tf.stack([c, n])
drop = tf.where(winner < th_k[:,k-1:k], # c, n
tf.zeros(shape_t, tf.float32), tf.ones(shape_t, tf.float32))
drop = tf.transpose(drop) # n, c
return h * tf.reshape(drop, tf.stack([n, 1, 1, c]))
def features(self):
return self.sess.run(self.f)
def save(self, ckpt_path):
self.saver.save(self.sess, ckpt_path)
def restore(self, ckpt_path):
self.saver.restore(self.sess, ckpt_path)