-
Notifications
You must be signed in to change notification settings - Fork 4
/
convnet.py
244 lines (226 loc) · 8.06 KB
/
convnet.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 h5py
import numpy as np
import tensorflow as tf
import tensorflow.contrib.slim as slim
from PIL import Image
from inception.inception_v3 import inception_v3
from inception.inception_v4 import inception_v4
from inception.inception_utils import inception_arg_scope
def build_inception(
name,
input_layer,
minibatch_size,
tf_session,
tf_graph,
pretrained_model_file_path=None,
reuse=False,
scope=None,
):
if name == 'inception_v3':
build_fn = inception_v3
elif name == 'inception_v4':
build_fn = inception_v4
with slim.arg_scope(inception_arg_scope()):
logits, endpoints = build_fn(
input_layer,
create_aux_logits=False,
is_training=False,
reuse=reuse,
)
if not reuse:
var_dict = {}
for var in tf_graph.get_collection('variables', scope=scope.name):
var_name_prefix = '{}/'.format(scope.name)
var_name_suffix = ':0'
saved_var_name = var.name[len(var_name_prefix)
:-len(var_name_suffix)]
var_dict[saved_var_name] = var
saver = tf.train.Saver(
var_list=var_dict,
)
saver.restore(
tf_session,
save_path=pretrained_model_file_path,
)
return endpoints
def build_vgg16(
input_layer,
minibatch_size,
pretrained_model_file_path='pretrained/vgg16_weights.h5'
):
network_config = [
('block1',
(
('conv1', {'W': [3, 3, 3, 64], 'b': [64]}),
('conv2', {'W': [3, 3, 64, 64], 'b': [64]}),
('pool', {'k': [2, 2], 's': [2, 2]}),
)
),
('block2',
(
('conv1', {'W': [3, 3, 64, 128], 'b': [128]}),
('conv2', {'W': [3, 3, 128, 128], 'b': [128]}),
('pool', {'k': [2, 2], 's': [2, 2]}),
)
),
('block3',
(
('conv1', {'W': [3, 3, 128, 256], 'b': [256]}),
('conv2', {'W': [3, 3, 256, 256], 'b': [256]}),
('conv3', {'W': [3, 3, 256, 256], 'b': [256]}),
('pool', {'k': [2, 2], 's': [2, 2]}),
)
),
('block4',
(
('conv1', {'W': [3, 3, 256, 512], 'b': [512]}),
('conv2', {'W': [3, 3, 512, 512], 'b': [512]}),
('conv3', {'W': [3, 3, 512, 512], 'b': [512]}),
('pool', {'k': [2, 2], 's': [2, 2]}),
)
),
('block5',
(
('conv1', {'W': [3, 3, 512, 512], 'b': [512]}),
('conv2', {'W': [3, 3, 512, 512], 'b': [512]}),
('conv3', {'W': [3, 3, 512, 512], 'b': [512]}),
('pool', {'k': [2, 2], 's': [2, 2]}),
)
),
('top',
(
('flatten', ()),
('fc1', (4096)),
('fc2', (4096)),
('predictions', (1000)),
)
),
]
weights_f = h5py.File(
pretrained_model_file_path,
mode='r',
)
prev_layer = input_layer
for block_name, block_conf in network_config:
with tf.variable_scope(block_name):
for layer_name, layer_conf in block_conf:
with tf.variable_scope(layer_name):
block_layer_name = block_name + '_' + layer_name
if 'conv' in layer_name:
conv_var = {}
for var_name, var_shape in layer_conf.items():
conv_var[var_name] = get_vgg16_weights(
weights_f,
block_layer_name,
var_name,
var_shape,
)
tensor = tf.nn.conv2d(
prev_layer,
conv_var['W'],
strides=[1, 1, 1, 1],
padding='SAME',
)
tensor = tf.nn.bias_add(
tensor,
conv_var['b']
)
new_layer = tf.nn.relu(
tensor,
)
elif 'pool' in layer_name:
new_layer = tf.nn.max_pool(
prev_layer,
ksize=([1] + layer_conf['k'] + [1]),
strides=([1] + layer_conf['s'] + [1]),
padding='SAME',
)
elif 'flatten' in layer_name:
new_layer = tf.reshape(
prev_layer,
[minibatch_size, -1],
)
elif (
'fc' in layer_name
or 'predictions' in layer_name
):
if 'fc' in layer_name:
f_layer = tf.nn.relu
elif 'predictions' in layer_name:
f_layer = tf.nn.softmax
input_dim = prev_layer.shape[-1].value
output_dim = layer_conf
layer_var = {}
for var_name, var_shape in (
('W', (input_dim, output_dim)),
('b', (output_dim)),
):
layer_var[var_name] = get_vgg16_weights(
weights_f,
layer_name,
var_name,
var_shape,
)
preactivation = tf.add(
tf.matmul(prev_layer, layer_var['W']),
layer_var['b'],
name='preactivation',
)
new_layer = f_layer(
preactivation,
name='activation',
)
else:
raise NotImplementedError
# End of building a layer.
prev_layer = new_layer
# End of building VGG16.
def get_vgg16_weights(weights_f, block_layer_name, var_name, var_shape):
dset_name = block_layer_name + '_' + var_name + '_1:0'
return tf.get_variable(
var_name,
shape=var_shape,
initializer=tf.constant_initializer(
weights_f
[block_layer_name]
[dset_name]
.value
),
trainable=False,
)
def resize_image(image, size, crop=True):
if crop:
width, height = image.size
min_size = min(width, height)
left = int((width - min_size) / 2.0)
upper = int((height - min_size) / 2.0)
image = image.crop((left, upper, left + min_size, upper + min_size))
image = image.resize((size, size))
return image
def preprocess_image(convnet_name, image, size=None):
if size is not None:
image = resize_image(image, size)
width, height = image.size
assert(width == height)
x = np.array(image, dtype=np.float32)
if len(x.shape) == 2:
rgbimg = Image.new("RGB", image.size)
rgbimg.paste(image)
x = np.array(rgbimg, dtype=np.float32)
if convnet_name == 'vgg16':
assert(width == 224)
# Substracting the mean, from Keras' imagenet_utils.preprocess_input.
# 'RGB'->'BGR'
x = x[:, :, ::-1]
# Zero-center by mean pixel
x[:, :, 0] -= 103.939
x[:, :, 1] -= 116.779
x[:, :, 2] -= 123.68
elif 'inception' in convnet_name:
assert(width == 299)
x /= (np.iinfo(np.uint8).max / 2.0)
x -= 1.0
else:
print('Unknown convolutional network: {}'.format(convnet_name))
raise NotImplementedError
return x