-
Notifications
You must be signed in to change notification settings - Fork 0
/
network.py
247 lines (203 loc) · 9.76 KB
/
network.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
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
import tensorflow as tf
import numpy as np
import aiutils.tftools.images as images
import aiutils.tftools.layers as layers
import aiutils.tftools.batch_normalizer as batch_normalizer
class BaseNetwork(object):
def __init__():
pass
""" Call this after setup, while in the correct scope """
def accumulate_variables(self):
scope_name = tf.get_variable_scope().name
assert scope_name != '', 'You almost certainly wanted to be in a variable scope when you created this network'
# Only care about trainable variables
self.vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope=scope_name)
class LegacyNetwork(BaseNetwork):
def __init__(self, input, params_path=None, trainable=False):
self.trainable = trainable
if params_path:
self.params = np.load(params_path).item()
self.layers = []
self.layerdict = {}
self.batch_size = int(input.get_shape()[0])
self.add_('input', input)
self.setup()
self.accumulate_variables()
def setup(self):
raise NotImplementedError('Must be subclassed.')
def get_unique_name_(self, prefix):
id = sum(t.startswith(prefix) for t,_ in self.layers)+1
return '%s_%d'%(prefix, id)
def add_(self, name, var):
self.layers.append((name, var))
self.layerdict[name] = var
def get_output(self):
return self.layers[-1][1]
def conv(self, h, w, c_i, c_o, stride=1, name=None):
name = name or self.get_unique_name_('conv')
with tf.variable_scope(name) as scope:
weights = tf.get_variable('W', initializer=tf.constant(self.params[name][0].astype(np.float32)), trainable=self.trainable, dtype=tf.float32)
conv = tf.nn.conv2d(self.get_output(), weights, [1,stride,stride,1], padding='SAME')
if len(self.params[name]) > 1:
biases = tf.get_variable('b', initializer=tf.constant(self.params[name][1].astype(np.float32)), trainable=self.trainable, dtype=tf.float32)
bias = tf.nn.bias_add(conv, biases)
relu = tf.nn.relu(bias, name=scope.name)
else:
relu = tf.nn.relu(conv, name=scope.name)
self.add_(name, relu)
return self
def pool(self, size=2, stride=2, name=None):
name = name or self.get_unique_name_('pool')
# pool = tf.nn.avg_pool(self.get_output(),
pool = tf.nn.max_pool(self.get_output(),
ksize=[1, size, size, 1],
strides=[1, stride, stride, 1],
padding='SAME',
name=name)
self.add_(name, pool)
return self
class Network(BaseNetwork):
def __init__(self, input, phase_train=None):
self.layers = []
self.layerdict = {}
self.batch_size = int(input.get_shape()[0])
self.add_('input', input)
if phase_train is not None:
self.phase_train = phase_train
else:
self.phase_train = tf.placeholder_with_default(True, (), 'PhaseTrain')
self.setup()
self.accumulate_variables()
def setup(self):
raise NotImplementedError('Must be subclassed.')
def get_unique_name_(self, prefix):
id = sum(t.startswith(prefix) for t,_ in self.layers)+1
return '%s_%d'%(prefix, id)
def add_(self, name, var):
self.layers.append((name, var))
self.layerdict[name] = var
def get_output(self):
return self.layers[-1][1]
def bnconv(self, out_dim, stride=1, filter_size=3, name=None, func=tf.nn.relu, inp_layer=None):
name = name or self.get_unique_name_('conv')
if inp_layer is None:
inp_layer = self.get_output()
new_layer = layers.conv2d(inp_layer, filter_size, out_dim, name, strides=[1,stride,stride,1], func=None)
new_layer = layers.batch_norm(new_layer, self.phase_train, name=name+'bn')
if func is not None:
new_layer = func(new_layer)
self.add_(name, new_layer)
return self
def conv(self, out_dim, stride=1, filter_size=3, name=None, func=tf.nn.relu, inp_layer=None):
name = name or self.get_unique_name_('conv')
if inp_layer is None:
inp_layer = self.get_output()
new_layer = layers.conv2d(inp_layer, filter_size, out_dim, name, strides=[1,stride,stride,1], func=func)
self.add_(name, new_layer)
return self
def atrous_conv2d(self, out_dim, rate=[1], filter_size=3, name=None, func=tf.nn.relu, inp_layer=None):
name = name or self.get_unique_name_('aconv')
if inp_layer is None:
inp_layer = self.get_output()
new_layers = []
for r in rate:
new_layers.append(layers.atrous_conv2d(inp_layer, filter_size, out_dim, rate=rate, func=func))
self.add_(name, tf.concat(3, new_layers))
return self
def reduce_mean_image(self, name=None, inp_layer=None):
name = name or self.get_unique_name_('reduce')
if inp_layer is None:
inp_layer = self.get_output()
new_layer = tf.reduce_mean(inp_layer, [1,2])
self.add_(name, new_layer)
return self
def bn_fully_connected(self, out_size, name=None, func=tf.nn.relu, inp_layer=None):
name = name or self.get_unique_name_('fc')
if inp_layer is None:
inp_layer = self.get_output()
new_layer = layers.full(inp_layer, out_size, name, func=None)
new_layer = layers.batch_norm(new_layer, self.phase_train, name=name+'bn')
if func is not None:
new_layer = func(new_layer)
self.add_(name, new_layer)
return self
def fully_connected(self, out_size, name=None, func=tf.nn.relu, inp_layer=None):
name = name or self.get_unique_name_('fc')
if inp_layer is None:
inp_layer = self.get_output()
new_layer = layers.full(inp_layer, out_size, name, func=func)
self.add_(name, new_layer)
return self
def softmax_layer(self, num_classes, name=None, inp_layer = None):
name = name or self.get_unique_name_('softmax')
if inp_layer is None:
inp_layer = self.get_output()
new_layer = layers.full(inp_layer, num_classes, name, func=tf.nn.softmax)
self.add_(name, new_layer)
return self
def batch_norm(self, name=None, inp_layer=None):
name = name or self.get_unique_name_('bn')
if inp_layer is None:
inp_layer = self.get_output()
new_layer = layers.batch_norm(inp_layer, self.phase_train, name=name)
self.add_(name, new_layer)
return self
def resize_like(self, target_layer, name=None, inp_layer=None):
name = name or self.get_unique_name_('resize_like')
if inp_layer is None:
inp_layer = self.get_output()
with tf.variable_scope(name):
new_layer = images.resize_images_like(inp_layer, self.layerdict[target_layer])
self.add_(name, new_layer)
def atrous_residual(self, outer_dim, inner_dim, name=None, inp_layer=None, rates=[1,4]):
name = name or self.get_unique_name_('residual')
if inp_layer is None:
inp_layer = self.get_output()
with tf.variable_scope(name):
if inp_layer.get_shape().as_list()[-1] != outer_dim:
input = layers.conv2d(inp_layer, 3, outer_dim, name+'_rch_conv', func=None)
input = layers.batch_norm(input, self.phase_train, name=name+'_bn')
input = tf.maximum(.01*input, input)
else:
input = inp_layer
with tf.variable_scope('atrous_conv'):
atrous_layers = []
for rate in rates:
atrous_layers.append(layers.atrous_conv2d(input, 5, outer_dim // len(rates), 'aconv_'+str(rate), rate=rate, func=None))
conv = tf.concat(3, atrous_layers)
bn = layers.batch_norm(conv, self.phase_train, name='bn1')
leaky_relu = tf.maximum(.01*bn, bn)
conv2 = layers.conv2d(leaky_relu, 3, outer_dim, 'conv_outer', func=None)
#bn2 = layers.batch_norm(conv, self.phase_train, name='bn2')
res = conv2 + input
bn2 = layers.batch_norm(res, self.phase_train, name='bn2')
out = tf.maximum(.01*bn2, bn2)
self.add_(name, out)
return self
def residual(self, outer_dim, inner_dim, name=None, inp_layer=None):
name = name or self.get_unique_name_('residual')
if inp_layer is None:
inp_layer = self.get_output()
with tf.variable_scope(name):
if inp_layer.get_shape().as_list()[-1] != outer_dim:
input = layers.conv2d(inp_layer, 3, outer_dim, name+'_rch_conv', func=None)
input = layers.batch_norm(input, self.phase_train, name=name+'_bn')
input = tf.maximum(.01*input, input)
else:
input = self.get_output()
conv = layers.conv2d(input, 5, inner_dim, 'conv_inner', func=None)
bn = layers.batch_norm(conv, self.phase_train, name='bn1')
leaky_relu = tf.maximum(.01*bn, bn)
conv2 = layers.conv2d(leaky_relu, 3, outer_dim, 'conv_outer', func=None)
#bn2 = layers.batch_norm(conv, self.phase_train, name='bn2')
res = conv2 + input
bn2 = layers.batch_norm(res, self.phase_train, name='bn2')
out = tf.maximum(.01*bn2, bn2)
self.add_(name, out)
def dropout(self, keep_prob=.8, name=None, inp_layer=None):
name = name or self.get_unique_name_('dropout')
if inp_layer is None:
inp_layer = self.get_output()
self.add_(name, tf.cond(self.phase_train,
lambda: tf.nn.dropout(inp_layer, keep_prob),
lambda: inp_layer))