-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmodel.py
160 lines (139 loc) · 7.12 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
from keras import backend as K
import tensorflow as tf
layers = tf.keras.layers
def conv(inputs, filters, kernel_size, data_format='channels_last'):
return layers.Conv2D(filters, kernel_size,
padding='same',
data_format=data_format,
kernel_initializer="he_uniform",
use_bias=False)(inputs)
def Dense_Bottleneck(inputs, growth_rate, data_format='channels_first',
bottleneck=True,
dropout_rate=None,
training=True):
concat_axis = 1 if data_format == 'channels_first' else -1
x = tf.layers.batch_normalization(inputs,
axis=concat_axis,
epsilon=1.1e-5,
fused=True,
training=training)
x = layers.ReLU(max_value=6.0)(x)
if bottleneck:
inter_channel = growth_rate * 4
x = conv(x, inter_channel, (1, 1), data_format)
x = tf.layers.batch_normalization(inputs,
axis=concat_axis,
epsilon=1.1e-5,
fused=True,
training=training)
x = layers.ReLU(max_value=6.0)(x)
x = conv(x, growth_rate, (3, 3), data_format)
if dropout_rate:
x = tf.layers.dropout(x,
rate=dropout_rate,
training=training)
return x
def DenseLayer(inputs, growth_rate, num_layers,
data_format='channels_first',
bottleneck=True,
dropout_rate=None,
training=True):
concat_axis = 1 if data_format == 'channels_first' else -1
x = inputs
for i in range(num_layers):
y = Dense_Bottleneck(x, growth_rate, data_format, bottleneck, dropout_rate, training)
x = layers.concatenate([x, y], axis=concat_axis)
return x
def TransLayer(inputs, data_format='channels_first',
compression=1.0,
dropout_rate=None,
training=True):
concat_axis = 1 if data_format == 'channels_first' else -1
output_channels = int(K.int_shape(inputs)[concat_axis]*compression)
x = tf.layers.batch_normalization(inputs,
axis=concat_axis,
epsilon=1.1e-5,
fused=True,
training=training)
x = layers.ReLU(max_value=6.0)(x)
x = conv(x, output_channels, (1, 1), data_format)
if dropout_rate:
x = tf.layers.dropout(rate=dropout_rate,
training=training)(x)
x = layers.AvgPool2D((2, 2),
strides=(2, 2),
data_format=data_format)(x)
return x
def DenseNet(inputs, growth_rate, depth, num_dense_block,
num_init_filters,
sub_sample_image,
num_classes,
training=True,
num_layer_list=-1,
bottleneck=True,
dropout_rate=None,
compression=1.0,
data_format='channels_first',
all_config=None):
if type(num_layer_list) is list or type(num_layer_list) is tuple:
num_layers = list(num_layer_list)
assert len(num_layers) == num_dense_block, "num_dense_block or num_layer_list is wrong"
final_nb_layer = num_layers[-1]
num_layers = num_layers[:-1]
else:
assert (depth - 4) % 3 == 0, 'Depth must be 3 N + 4 if nb_layers_per_block == -1'
count = int((depth - 4) / 3)
if bottleneck:
count = count // 2
num_layers = [count for _ in range(num_dense_block-1)]
final_nb_layer = count
if sub_sample_image:
initial_kernel = (7, 7)
initial_strides = (2, 2)
else:
initial_kernel = (3, 3)
initial_strides = (1, 1)
########################################################################
# Summary the image for inputs
########################################################################
summary_inputs = inputs * tf.reshape(tf.convert_to_tensor(all_config.IMAGE_STD, tf.float32), (1, 3, 1, 1))
summary_inputs += tf.reshape(tf.convert_to_tensor(all_config.IMAGE_MEANS, tf.float32), (1, 3, 1, 1))
summary_inputs = tf.transpose(summary_inputs, [0, 2, 3, 1])
tf.summary.image('inputs', summary_inputs, max_outputs=1, family='features')
with tf.variable_scope("Stem_block"):
x = layers.Conv2D(num_init_filters,
initial_kernel,
initial_strides,
padding='same',
data_format=data_format,
kernel_initializer='he_uniform',
name="a",
use_bias=False)(inputs)
concat_axis = 1 if data_format == 'channels_first' else -1
if sub_sample_image:
x = tf.layers.batch_normalization(inputs,
axis=concat_axis,
epsilon=1.1e-5,
fused=True,
training=training)
x = layers.ReLU(max_value=6.0)(x)
x = layers.MaxPool2D((2, 2), (2, 2), data_format=data_format)(x)
for block_idx in range(num_dense_block-1):
with tf.variable_scope('DenseBlock_{}'.format(block_idx+1)):
x = DenseLayer(x, growth_rate, num_layers[block_idx], data_format, bottleneck, dropout_rate, training)
with tf.variable_scope('TransLayer_{}'.format(block_idx+1)):
x = TransLayer(x, data_format, compression, dropout_rate, training)
##############################################################################
# summary for last_layer images
##############################################################################
summary_last_x = x[0]
summary_last_x = tf.expand_dims(summary_last_x, axis=-1)
tf.summary.image('last_features', summary_last_x, max_outputs=10, family='features')
with tf.variable_scope('DenseBlock_{}'.format(num_dense_block)):
x = DenseLayer(x, growth_rate, final_nb_layer, data_format, bottleneck, dropout_rate, training)
with tf.variable_scope("Global_Avg"):
x = tf.layers.batch_normalization(x, axis=concat_axis, epsilon=1.1e-5, fused=True, training=training)
x = layers.ReLU(max_value=6.0)(x)
x = layers.GlobalAveragePooling2D(data_format)(x)
x = layers.Dense(num_classes, kernel_initializer='he_uniform')(x)
return x