-
Notifications
You must be signed in to change notification settings - Fork 24
/
erfnet.py
211 lines (175 loc) · 11.4 KB
/
erfnet.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
from __future__ import print_function, division, unicode_literals
import numpy as np
import tensorflow as tf
# USEFUL LAYERS
fc = tf.contrib.layers.fully_connected
conv = tf.contrib.layers.conv2d
# convsep = tf.contrib.layers.separable_conv2d
deconv = tf.contrib.layers.conv2d_transpose
relu = tf.nn.relu
maxpool = tf.contrib.layers.max_pool2d
dropout_layer = tf.layers.dropout
batchnorm = tf.contrib.layers.batch_norm
winit = tf.contrib.layers.xavier_initializer()
repeat = tf.contrib.layers.repeat
arg_scope = tf.contrib.framework.arg_scope
l2_regularizer = tf.contrib.layers.l2_regularizer
def get_conv_arg_scope(is_training, bn=True, reg=None, use_deconv=False, use_relu=True, bn_decay=0.9):
with arg_scope(
[deconv if use_deconv else conv],
padding = "SAME",
stride = 1,
activation_fn = relu if use_relu else None,
normalizer_fn = batchnorm if bn else None,
normalizer_params = {"is_training": is_training, "decay": bn_decay},
weights_regularizer = reg,
variables_collections = None,
) as scope:
return scope
def factorized_res_moduleOLD(x, is_training, dropout=0.3, dilation=1, name="fres"):
with arg_scope(get_conv_arg_scope(is_training=is_training, bn=True)):
with tf.variable_scope(name):
n_filters = x.shape.as_list()[-1]
y = conv(x, num_outputs=n_filters, kernel_size=[3,1], normalizer_fn=None, scope="conv_a_3x1")
y = conv(y, num_outputs=n_filters, kernel_size=[1,3], scope="conv_a_1x3")
y = conv(y, num_outputs=n_filters, kernel_size=[3,1], rate=dilation, normalizer_fn=None, scope="conv_b_3x1")
y = conv(y, num_outputs=n_filters, kernel_size=[1,3], rate=dilation, scope="conv_b_1x3")
y = dropout_layer(y, rate=dropout)
y = tf.add(x,y, name="add")
print("DEBUG: {} {}".format(name, y.shape.as_list()))
return y
def factorized_res_module(x, is_training, dropout=0.3, dilation=[1,1], l2=None, name="fres"):
reg = None if l2 is None else l2_regularizer(l2)
with arg_scope(get_conv_arg_scope(reg=reg, is_training=is_training, bn=True)):
with tf.variable_scope(name):
n_filters = x.shape.as_list()[-1]
y = conv(x, num_outputs=n_filters, kernel_size=[3,1], rate=dilation[0], normalizer_fn=None, scope="conv_a_3x1")
y = conv(y, num_outputs=n_filters, kernel_size=[1,3], rate=dilation[0], scope="conv_a_1x3")
y = conv(y, num_outputs=n_filters, kernel_size=[3,1], rate=dilation[1], normalizer_fn=None, scope="conv_b_3x1")
y = conv(y, num_outputs=n_filters, kernel_size=[1,3], rate=dilation[1], scope="conv_b_1x3")
y = dropout_layer(y, rate=dropout)
y = tf.add(x,y, name="add")
print("DEBUG: {} {}".format(name, y.shape.as_list()))
print("DEBUG: L2 in factorized res module {}".format(l2))
return y
def downsample(x, n_filters, is_training, bn=False, use_relu=False, l2=None, name="down"):
with tf.variable_scope(name):
reg = None if l2 is None else l2_regularizer(l2)
with arg_scope(get_conv_arg_scope(reg=reg, is_training=is_training, bn=bn, use_relu=use_relu)):
n_filters_in = x.shape.as_list()[-1]
n_filters_conv = n_filters - n_filters_in
branch_a = conv(x, num_outputs=n_filters_conv, kernel_size=3, stride=2, scope="conv")
branch_b = maxpool(x, kernel_size=2, stride=2, padding='VALID', scope="maxpool")
y = tf.concat([branch_a, branch_b], axis=-1, name="concat")
print("DEBUG: {} {}".format(name, y.shape.as_list()))
return y
def upsample(x, n_filters, is_training=False, use_relu=False, bn=False, l2=None, name="up"):
reg = None if l2 is None else l2_regularizer(l2)
with arg_scope(get_conv_arg_scope(reg=reg, is_training=is_training, bn=bn, use_deconv=True, use_relu=use_relu)):
y = deconv(x, num_outputs=n_filters, kernel_size=4, stride=2, scope=name)
print("DEBUG: {} {}".format(name, y.shape.as_list()))
return y
def romeranetB_logits(X, Y, n_classes, alpha=0.001, dropout=0.3, l2=None, is_training=False):
"""
"""
# TODO: Use TF repeat for some of these repeating layers
# TODO: register factorized_res_module and upsample and downsample with arg_scope
# and pass dropout, is_training, etc to it.
# TODO: Add weight decay.
with tf.name_scope("preprocess") as scope:
x = tf.div(X, 255., name="rescaled_inputs")
x = downsample(x, n_filters=16, is_training=is_training, name="d1")
x = downsample(x, n_filters=64, is_training=is_training, name="d2")
x = factorized_res_moduleOLD(x, is_training=is_training, dilation=1, name="fres3")
x = factorized_res_moduleOLD(x, is_training=is_training, dilation=1, name="fres4")
x = factorized_res_moduleOLD(x, is_training=is_training, dilation=1, name="fres5")
x = factorized_res_moduleOLD(x, is_training=is_training, dilation=1, name="fres6")
x = factorized_res_moduleOLD(x, is_training=is_training, dilation=1, name="fres7")
# TODO: Use dilated convolutions
x = downsample(x, n_filters=128, is_training=is_training, name="d8")
x = factorized_res_moduleOLD(x, is_training=is_training, dilation=2, name="fres9")
x = factorized_res_moduleOLD(x, is_training=is_training, dilation=4, name="fres10")
x = factorized_res_moduleOLD(x, is_training=is_training, dilation=8, name="fres11")
x = factorized_res_moduleOLD(x, is_training=is_training, dilation=16, name="fres12")
x = factorized_res_moduleOLD(x, is_training=is_training, dilation=2, name="fres13")
x = factorized_res_moduleOLD(x, is_training=is_training, dilation=4, name="fres14")
x = factorized_res_moduleOLD(x, is_training=is_training, dilation=8, name="fres15")
x = factorized_res_moduleOLD(x, is_training=is_training, dilation=16, name="fres16")
x = upsample(x, n_filters=64, is_training=is_training, name="up17")
x = factorized_res_moduleOLD(x, is_training=is_training, dilation=1, name="fres18")
x = factorized_res_moduleOLD(x, is_training=is_training, dilation=1, name="fres19")
x = upsample(x, n_filters=16, is_training=is_training, name="up20")
x = factorized_res_moduleOLD(x, is_training=is_training, dilation=1, name="fres21")
x = factorized_res_moduleOLD(x, is_training=is_training, dilation=1, name="fres22")
x = upsample(x, n_filters=n_classes, is_training=is_training, name="up23")
return x
def erfnetA(X, Y, n_classes, alpha=0.001, dropout=0.3, l2=None, is_training=False):
"""
"""
# TODO: Use TF repeat for some of these repeating layers
# TODO: register factorized_res_module and upsample and downsample with arg_scope
# and pass dropout, is_training, etc to it.
# TODO: Add weight decay.
with tf.name_scope("preprocess") as scope:
x = tf.div(X, 255., name="rescaled_inputs")
x = downsample(x, n_filters=16, is_training=is_training, name="d1")
x = downsample(x, n_filters=64, is_training=is_training, name="d2")
x = factorized_res_module(x, is_training=is_training, dilation=[1, 1], name="fres3")
x = factorized_res_module(x, is_training=is_training, dilation=[1, 1], name="fres4")
x = factorized_res_module(x, is_training=is_training, dilation=[1, 1], name="fres5")
x = factorized_res_module(x, is_training=is_training, dilation=[1, 1], name="fres6")
x = factorized_res_module(x, is_training=is_training, dilation=[1, 1], name="fres7")
# TODO: Use dilated convolutions
x = downsample(x, n_filters=128, is_training=is_training, name="d8")
x = factorized_res_module(x, is_training=is_training, dilation=[1, 2], name="fres9")
x = factorized_res_module(x, is_training=is_training, dilation=[1, 4], name="fres10")
x = factorized_res_module(x, is_training=is_training, dilation=[1, 8], name="fres11")
x = factorized_res_module(x, is_training=is_training, dilation=[1, 16], name="fres12")
x = factorized_res_module(x, is_training=is_training, dilation=[1, 2], name="fres13")
x = factorized_res_module(x, is_training=is_training, dilation=[1, 4], name="fres14")
x = factorized_res_module(x, is_training=is_training, dilation=[1, 8], name="fres15")
x = factorized_res_module(x, is_training=is_training, dilation=[1, 16], name="fres16")
x = upsample(x, n_filters=64, is_training=is_training, name="up17")
x = factorized_res_module(x, is_training=is_training, dilation=[1, 1], name="fres18")
x = factorized_res_module(x, is_training=is_training, dilation=[1, 1], name="fres19")
x = upsample(x, n_filters=16, is_training=is_training, name="up20")
x = factorized_res_module(x, is_training=is_training, dilation=[1, 1], name="fres21")
x = factorized_res_module(x, is_training=is_training, dilation=[1, 1], name="fres22")
x = upsample(x, n_filters=n_classes, is_training=is_training, name="up23")
return x
def erfnetB(X, Y, n_classes, alpha=0.001, dropout=0.3, l2=None, is_training=False):
"""
Uses L2 regularization.
"""
print("DEBUG: L2 passed on to ERFNETB{}".format(l2))
# TODO: Use TF repeat for some of these repeating layers
# TODO: register factorized_res_module and upsample and downsample with arg_scope
# and pass dropout, is_training, etc to it.
# TODO: Add weight decay.
with tf.name_scope("preprocess") as scope:
x = tf.div(X, 255., name="rescaled_inputs")
x = downsample(x, n_filters=16, is_training=is_training, l2=l2, name="d1")
x = downsample(x, n_filters=64, is_training=is_training, l2=l2, name="d2")
x = factorized_res_module(x, is_training=is_training, dilation=[1, 1], l2=l2, name="fres3")
x = factorized_res_module(x, is_training=is_training, dilation=[1, 1], l2=l2, name="fres4")
x = factorized_res_module(x, is_training=is_training, dilation=[1, 1], l2=l2, name="fres5")
x = factorized_res_module(x, is_training=is_training, dilation=[1, 1], l2=l2, name="fres6")
x = factorized_res_module(x, is_training=is_training, dilation=[1, 1], l2=l2, name="fres7")
# TODO: Use dilated convolutions
x = downsample(x, n_filters=128, is_training=is_training, l2=l2, name="d8")
x = factorized_res_module(x, is_training=is_training, dilation=[1, 2], l2=l2, name="fres9")
x = factorized_res_module(x, is_training=is_training, dilation=[1, 4], l2=l2, name="fres10")
x = factorized_res_module(x, is_training=is_training, dilation=[1, 8], l2=l2, name="fres11")
x = factorized_res_module(x, is_training=is_training, dilation=[1, 16], l2=l2, name="fres12")
x = factorized_res_module(x, is_training=is_training, dilation=[1, 2], l2=l2, name="fres13")
x = factorized_res_module(x, is_training=is_training, dilation=[1, 4], l2=l2, name="fres14")
x = factorized_res_module(x, is_training=is_training, dilation=[1, 8], l2=l2, name="fres15")
x = factorized_res_module(x, is_training=is_training, dilation=[1, 16], l2=l2, name="fres16")
x = upsample(x, n_filters=64, is_training=is_training, l2=l2, name="up17")
x = factorized_res_module(x, is_training=is_training, dilation=[1, 1], l2=l2, name="fres18")
x = factorized_res_module(x, is_training=is_training, dilation=[1, 1], l2=l2, name="fres19")
x = upsample(x, n_filters=16, is_training=is_training, l2=l2, name="up20")
x = factorized_res_module(x, is_training=is_training, dilation=[1, 1], l2=l2, name="fres21")
x = factorized_res_module(x, is_training=is_training, dilation=[1, 1], l2=l2, name="fres22")
x = upsample(x, n_filters=n_classes, is_training=is_training, l2=l2, name="up23")
return x