forked from machrisaa/tensorflow-vgg
-
Notifications
You must be signed in to change notification settings - Fork 2
/
vgg16_trainable.py
182 lines (136 loc) · 6.71 KB
/
vgg16_trainable.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
import tensorflow as tf
import numpy as np
from functools import reduce
VGG_MEAN = [103.939, 116.779, 123.68]
class Vgg16:
"""
A trainable version VGG16.
"""
def __init__(self, vgg16_npy_path=None, trainable=True, dropout=0.5, output_dim=1000, retrain="semi"):
if vgg16_npy_path is not None:
self.data_dict = np.load(vgg16_npy_path, encoding='latin1').item()
else:
self.data_dict = None
self.var_dict = {}
self.trainable = trainable
self.dropout = dropout
self.output_dim=output_dim
self.retrain=retrain
def build(self, rgb, train_mode=None):
"""
load variable from npy to build the VGG
:param rgb: rgb image [batch, height, width, 3] values scaled [0, 1]
:param train_mode: a bool tensor, usually a placeholder: if True, dropout will be turned on
"""
rgb_scaled = rgb * 255.0
# Convert RGB to BGR
red, green, blue = tf.split(axis=3, num_or_size_splits=3, value=rgb_scaled)
assert red.get_shape().as_list()[1:] == [224, 224, 1]
assert green.get_shape().as_list()[1:] == [224, 224, 1]
assert blue.get_shape().as_list()[1:] == [224, 224, 1]
bgr = tf.concat(axis=3, values=[
blue - VGG_MEAN[0],
green - VGG_MEAN[1],
red - VGG_MEAN[2],
])
assert bgr.get_shape().as_list()[1:] == [224, 224, 3]
self.conv1_1 = self.conv_layer(bgr, 3, 64, "conv1_1")
self.conv1_2 = self.conv_layer(self.conv1_1, 64, 64, "conv1_2")
self.pool1 = self.max_pool(self.conv1_2, 'pool1')
self.conv2_1 = self.conv_layer(self.pool1, 64, 128, "conv2_1")
self.conv2_2 = self.conv_layer(self.conv2_1, 128, 128, "conv2_2")
self.pool2 = self.max_pool(self.conv2_2, 'pool2')
self.conv3_1 = self.conv_layer(self.pool2, 128, 256, "conv3_1")
self.conv3_2 = self.conv_layer(self.conv3_1, 256, 256, "conv3_2")
self.conv3_3 = self.conv_layer(self.conv3_2, 256, 256, "conv3_3")
self.pool3 = self.max_pool(self.conv3_3, 'pool3')
self.conv4_1 = self.conv_layer(self.pool3, 256, 512, "conv4_1")
self.conv4_2 = self.conv_layer(self.conv4_1, 512, 512, "conv4_2")
self.conv4_3 = self.conv_layer(self.conv4_2, 512, 512, "conv4_3")
self.pool4 = self.max_pool(self.conv4_3, 'pool4')
self.conv5_1 = self.conv_layer(self.pool4, 512, 512, "conv5_1")
self.conv5_2 = self.conv_layer(self.conv5_1, 512, 512, "conv5_2")
self.conv5_3 = self.conv_layer(self.conv5_2, 512, 512, "conv5_3")
self.pool5 = self.max_pool(self.conv5_3, 'pool5')
self.fc6 = self.fc_layer(self.pool5, 25088, 4096, "fc6-custom") # 25088 = ((224 // (2 ** 5)) ** 2) * 512
self.relu6 = tf.nn.relu(self.fc6)
if train_mode is not None:
self.relu6 = tf.cond(train_mode, lambda: tf.nn.dropout(self.relu6, self.dropout), lambda: self.relu6)
elif self.trainable:
self.relu6 = tf.nn.dropout(self.relu6, self.dropout)
self.fc7 = self.fc_layer(self.relu6, 4096, 4096, "fc7-custom")
self.relu7 = tf.nn.relu(self.fc7)
if train_mode is not None:
self.relu7 = tf.cond(train_mode, lambda: tf.nn.dropout(self.relu7, self.dropout), lambda: self.relu7)
elif self.trainable:
self.relu7 = tf.nn.dropout(self.relu7, self.dropout)
self.fc8 = self.fc_layer(self.relu7, 4096, self.output_dim, "fc8-custom")
self.prob = self.fc8
self.data_dict = None
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 max_pool(self, bottom, name):
return tf.nn.max_pool(bottom, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME', name=name)
def conv_layer(self, bottom, in_channels, out_channels, name):
with tf.variable_scope(name):
filt, conv_biases = self.get_conv_var(3, in_channels, out_channels, name)
conv = tf.nn.conv2d(bottom, filt, [1, 1, 1, 1], padding='SAME')
bias = tf.nn.bias_add(conv, conv_biases)
relu = tf.nn.relu(bias)
return relu
def fc_layer(self, bottom, in_size, out_size, name):
with tf.variable_scope(name):
weights, biases = self.get_fc_var(in_size, out_size, name)
x = tf.reshape(bottom, [-1, in_size])
fc = tf.nn.bias_add(tf.matmul(x, weights), biases)
return fc
def get_conv_var(self, filter_size, in_channels, out_channels, name):
initial_value = tf.truncated_normal([filter_size, filter_size, in_channels, out_channels], 0.0, 0.001)
if self.retrain == 'complete':
rt = True
elif self.retrain == 'semi':
if 'conv1' in name or 'conv2' in name:
rt = True
else:
rt = False
else:
rt = False
filters = self.get_var(initial_value, name, 0, name + "_filters", retrain=rt)
initial_value = tf.truncated_normal([out_channels], .0, .001)
biases = self.get_var(initial_value, name, 1, name + "_biases", retrain=rt)
return filters, biases
def get_fc_var(self, in_size, out_size, name):
initial_value = tf.truncated_normal([in_size, out_size], 0.0, 0.001)
weights = self.get_var(initial_value, name, 0, name + "_weights")
initial_value = tf.truncated_normal([out_size], .0, .001)
biases = self.get_var(initial_value, name, 1, name + "_biases")
return weights, biases
def get_var(self, initial_value, name, idx, var_name, retrain=True):
if self.data_dict is not None and name in self.data_dict:
value = self.data_dict[name][idx]
else:
value = initial_value
if self.trainable and retrain:
var = tf.Variable(value, name=var_name)
else:
var = tf.constant(value, dtype=tf.float32, name=var_name)
self.var_dict[(name, idx)] = var
# print var_name, var.get_shape().as_list()
assert var.get_shape() == initial_value.get_shape()
return var
def save_npy(self, sess, npy_path="./vgg16-save.npy"):
assert isinstance(sess, tf.Session)
data_dict = {}
for (name, idx), var in list(self.var_dict.items()):
var_out = sess.run(var)
if name not in data_dict:
data_dict[name] = {}
data_dict[name][idx] = var_out
np.save(npy_path, data_dict)
print(("file saved", npy_path))
return npy_path
def get_var_count(self):
count = 0
for v in list(self.var_dict.values()):
count += reduce(lambda x, y: x * y, v.get_shape().as_list())
return count