-
Notifications
You must be signed in to change notification settings - Fork 4
/
Inceptionv3.py
369 lines (339 loc) · 19.2 KB
/
Inceptionv3.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
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import tensorflow as tf
import numpy as np
class Inceptionv3:
def __init__(self, input_shape, num_classes, weight_decay, keep_prob, data_format, config=None):
assert(data_format in ['channels_last', 'channels_first'])
if data_format == 'channels_last':
assert((input_shape[0] >= 43) & (input_shape[1] >= 43))
self.input_shape = input_shape
self.output_shape = np.array([input_shape[0], input_shape[1]], dtype=np.int32)
else:
assert((input_shape[1] >= 43) & (input_shape[2] >= 43))
self.input_shape = input_shape
self.output_shape = np.array([input_shape[1], input_shape[2]], dtype=np.int32)
self.num_classes = num_classes
self.weight_decay = weight_decay
self.prob = 1. - keep_prob
self.data_format = data_format
if config is not None:
self.is_SENet = config['is_SENet']
if config['is_SENet']:
self.reduction = config['reduction']
else:
self.is_SENet = False
self.global_step = tf.train.get_or_create_global_step()
self.is_training = True
self._define_input()
self._build_graph()
self._init_session()
def _define_input(self):
shape = [None]
shape.extend(self.input_shape)
self.images = tf.placeholder(dtype=tf.float32, shape=shape, name='images')
self.labels = tf.placeholder(dtype=tf.int32, shape=[None, self.num_classes], name='labels')
self.lr = tf.placeholder(dtype=tf.float32, shape=[], name='lr')
def _build_graph(self):
with tf.variable_scope('stem'):
conv1_1 = self._conv_bn_activation(self.images, 32, 3, 2, 'valid')
self.output_shape = self._compute_output_shape(self.output_shape, 3, 'valid', 2)
conv1_2 = self._conv_bn_activation(conv1_1, 32, 3, 1, 'valid')
self.output_shape = self._compute_output_shape(self.output_shape, 3, 'valid', 1)
conv1_3 = self._conv_bn_activation(conv1_2, 64, 3, 1, 'same')
self.output_shape = self._compute_output_shape(self.output_shape, 3, 'same', 1)
pool1 = self._max_pooling(conv1_3, 3, 2, 'valid', 'pool1')
self.output_shape = self._compute_output_shape(self.output_shape, 3, 'valid', 2)
conv2_1 = self._conv_bn_activation(pool1, 80, 3, 1, 'valid')
self.output_shape = self._compute_output_shape(self.output_shape, 3, 'valid', 1)
conv2_2 = self._conv_bn_activation(conv2_1, 192, 3, 2, 'valid')
self.output_shape = self._compute_output_shape(self.output_shape, 3, 'valid', 2)
conv2_3 = self._conv_bn_activation(conv2_2, 288, 3, 1, 'valid')
self.output_shape = self._compute_output_shape(self.output_shape, 3, 'valid', 1)
with tf.variable_scope('inception_1'):
inception1_1 = self._inception_block1(conv2_3, [64, 48, 96, 32], 'inception1_1')
inception1_2 = self._inception_block1(inception1_1, [64, 48, 96, 64], 'inception1_2')
inception1_3 = self._inception_block1(inception1_2, [64, 48, 96, 64], 'inception1_3')
inception1_reduction = self._grid_size_reduction1(inception1_3, 'inception1_reduction')
self.output_shape = self._compute_output_shape(self.output_shape, 3, 'same', 2)
self.auxiliary_outshape = self.output_shape
with tf.variable_scope('inception_2'):
inception2_1 = self._inception_block2(inception1_reduction, [192, 128], 'inception2_1')
inception2_2 = self._inception_block2(inception2_1, [192, 160], 'inception2_2')
inception2_3 = self._inception_block2(inception2_2, [192, 160], 'inception2_3')
inception2_4 = self._inception_block2(inception2_3, [192, 192], 'inception2_4')
inception2_reduction = self._grid_size_reduction2(inception2_4, 'inception2_reduction')
self.output_shape = self._compute_output_shape(self.output_shape, 3, 'same', 2)
with tf.variable_scope('inception_3'):
inception3_1 = self._inception_block3(inception2_reduction, [320, 384, 448, 192], 'inception3_1')
inception3_2 = self._inception_block3(inception3_1, [320, 384, 448, 192], 'inception3_2')
with tf.variable_scope('classifier'):
global_pool = self._max_pooling(inception3_2, self.output_shape.astype(np.int32).tolist(), 1, 'valid', 'global_pool')
dropout = self._dropout(global_pool, 'dropout')
final_dense = self._conv_bn_activation(dropout, self.num_classes, 1, 1, 'same', None)
logit = tf.squeeze(final_dense, name='logit')
self.logit = tf.nn.softmax(logit, name='softmax')
with tf.variable_scope('auxilary'):
auxiliary_avgpool = self._avg_pooling(inception1_reduction, 5, 3, 'same', 'auxiliary_avgpool')
self.auxiliary_outshape = self._compute_output_shape(self.auxiliary_outshape, 5, 'same', 3)
auxiliary_conv1x1 = self._conv_bn_activation(auxiliary_avgpool, 128, 1, 1, 'same')
self.auxiliary_outshape = self._compute_output_shape(self.auxiliary_outshape, 1, 'same', 1)
new_channels = np.prod(self.auxiliary_outshape).astype(np.int32) * 128
auxiliary_conv1x1_reshape = tf.reshape(auxiliary_conv1x1, [-1, 1, 1, new_channels], name='auxiliary_conv1x1_reshape')
auxiliary_final_dense = self._conv_bn_activation(auxiliary_conv1x1_reshape, self.num_classes, 1, 1, 'valid', None)
auxiliary_logit = tf.squeeze(auxiliary_final_dense, name='logit')
auxiliary_loss = 0.4 * tf.losses.softmax_cross_entropy(self.labels, auxiliary_logit, label_smoothing=0.1, reduction=tf.losses.Reduction.MEAN)
with tf.variable_scope('optimizer'):
loss = tf.losses.softmax_cross_entropy(self.labels, logit, label_smoothing=0.1, reduction=tf.losses.Reduction.MEAN)
l2_loss = self.weight_decay * tf.add_n(
[tf.nn.l2_loss(var) for var in tf.trainable_variables()]
)
total_loss = loss + l2_loss + auxiliary_loss
lossavg = tf.train.ExponentialMovingAverage(0.9, name='loss_moveavg')
lossavg_op = lossavg.apply([loss, total_loss])
with tf.control_dependencies([lossavg_op]):
self.total_loss = tf.identity(total_loss)
var_list = tf.trainable_variables()
varavg = tf.train.ExponentialMovingAverage(0.9, name='var_moveavg')
varavg_op = varavg.apply(var_list)
optimizer = tf.train.RMSPropOptimizer(self.lr, momentum=0.9, decay=0.9, epsilon=1.)
grads = optimizer.compute_gradients(self.total_loss)
clipped_grads = [(tf.clip_by_value(grad, -2., 2.), var) for grad, var in grads]
train_op = optimizer.apply_gradients(clipped_grads, global_step=self.global_step)
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
self.train_op = tf.group([update_ops, lossavg_op, varavg_op, train_op])
self.accuracy = tf.reduce_mean(
tf.cast(
tf.equal(
tf.argmax(self.logit, 1), tf.argmax(self.labels, 1)
), tf.float32
), name='accuracy')
def _init_session(self):
self.sess = tf.InteractiveSession()
self.sess.run(tf.global_variables_initializer())
self.saver = tf.train.Saver()
self.best_saver = tf.train.Saver()
def train_one_batch(self, images, labels, lr, sess=None):
self.is_training = True
if sess is None:
sess_ = self.sess
else:
sess_ = sess
_, loss, acc = sess_.run([self.train_op, self.total_loss, self.accuracy],
feed_dict={
self.images: images,
self.labels: labels,
self.lr: lr
})
return loss, acc
def validate_one_batch(self, images, labels, sess=None):
self.is_training = False
if sess is None:
sess_ = self.sess
else:
sess_ = sess
logit, acc = sess_.run([self.logit, self.accuracy], feed_dict={
self.images: images,
self.labels: labels,
self.lr: 0.
})
return logit, acc
def test_one_batch(self, images, sess=None):
self.is_training = False
if sess is None:
sess_ = self.sess
else:
sess_ = sess
logit = sess_.run([self.logit], feed_dict={
self.images: images,
self.lr: 0.
})
return logit
def save_weight(self, mode, path, sess=None):
assert(mode in ['latest', 'best'])
if sess is None:
sess_ = self.sess
else:
sess_ = sess
saver = self.saver if mode == 'latest' else self.best_saver
saver.save(sess_, path, global_step=self.global_step)
print('save', mode, 'model in', path, 'successfully')
def load_weight(self, mode, path, sess=None):
assert(mode in ['latest', 'best'])
if sess is None:
sess_ = self.sess
else:
sess_ = sess
saver = self.saver if mode == 'latest' else self.best_saver
ckpt = tf.train.get_checkpoint_state(path)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess_, path)
print('load', mode, 'model in', path, 'successfully')
else:
raise FileNotFoundError('Not Found Model File!')
def _conv_bn_activation(self, bottom, filters, kernel_size, strides, padding, activation=tf.nn.relu):
assert(padding in ['same', 'valid'])
conv = tf.layers.conv2d(
inputs=bottom,
filters=filters,
kernel_size=kernel_size,
strides=strides,
data_format=self.data_format,
padding=padding
)
bn = tf.layers.batch_normalization(
inputs=conv,
axis=3 if self.data_format == 'channels_last' else 1,
training=self.is_training
)
if activation is not None:
return activation(bn)
else:
return bn
# squeeze-and-excitation block
def squeeze_and_excitation(self, bottom):
axes = [2, 3] if self.data_format == 'channels_first' else [1, 2]
channels = bottom.get_shape()[1] if self.data_format == 'channels_first' else bottom.get_shape()[3]
squeeze = tf.reduce_mean(
input_tensor=bottom,
axis=axes,
keepdims=False,
)
excitation = tf.layers.dense(
inputs=squeeze,
units=int(channels // self.reduction),
activation=tf.nn.relu,
)
excitation = tf.layers.dense(
inputs=excitation,
units=channels,
activation=tf.nn.sigmoid,
)
weight = tf.reshape(excitation, [-1, 1, 1, channels])
scaled = weight * bottom
return scaled
def _max_pooling(self, bottom, pool_size, strides, padding, name):
assert(padding in ['same', 'valid'])
return tf.layers.max_pooling2d(
inputs=bottom,
pool_size=pool_size,
strides=strides,
padding=padding,
data_format=self.data_format,
name=name
)
def _avg_pooling(self, bottom, pool_size, strides, padding, name):
assert(padding in ['same', 'valid'])
return tf.layers.average_pooling2d(
inputs=bottom,
pool_size=pool_size,
strides=strides,
padding=padding,
data_format=self.data_format,
name=name
)
def _dropout(self, bottom, name):
return tf.layers.dropout(
inputs=bottom,
rate=self.prob,
training=self.is_training,
name=name
)
def _inception_block1(self, bottom, filters, scope):
with tf.variable_scope(scope):
with tf.variable_scope('branch_pool1x1'):
branch_pool1x1 = self._avg_pooling(bottom, 3, 1, 'same', 'avg_pool')
branch_pool1x1 = self._conv_bn_activation(branch_pool1x1, filters[3], 1, 1, 'same')
with tf.variable_scope('branch_1x1'):
branch_1x1 = self._conv_bn_activation(bottom, filters[0], 1, 1, 'same')
with tf.variable_scope('branch_1x1x3x3'):
branch_1x1x3x3 = self._conv_bn_activation(bottom, filters[1], 1, 1, 'same')
branch_1x1x3x3 = self._conv_bn_activation(branch_1x1x3x3, filters[0], 3, 1, 'same')
with tf.variable_scope('branch_1x1x5x5'):
branch_1x1x5x5 = self._conv_bn_activation(bottom, filters[0], 1, 1, 'same')
branch_1x1x5x5 = self._conv_bn_activation(branch_1x1x5x5, filters[2], 3, 1, 'same')
branch_1x1x5x5 = self._conv_bn_activation(branch_1x1x5x5, filters[2], 3, 1, 'same')
axes = 3 if self.data_format == 'channels_last' else 1
if self.is_SENet:
return self.squeeze_and_excitation(tf.concat([branch_1x1, branch_pool1x1, branch_1x1x3x3, branch_1x1x5x5], axis=axes))
else:
return tf.concat([branch_1x1, branch_pool1x1, branch_1x1x3x3, branch_1x1x5x5], axis=axes)
def _grid_size_reduction1(self, bottom, scope):
with tf.variable_scope(scope):
with tf.variable_scope('branch_3x3'):
branch_3x3 = self._conv_bn_activation(bottom, 384, 3, 2, 'same')
with tf.variable_scope('branch_1x1x5x5'):
branch_1x1x5x5 = self._conv_bn_activation(bottom, 64, 1, 1, 'same')
branch_1x1x5x5 = self._conv_bn_activation(branch_1x1x5x5, 96, 3, 1, 'same')
branch_1x1x5x5 = self._conv_bn_activation(branch_1x1x5x5, 96, 3, 2, 'same')
with tf.variable_scope('max_pool'):
branch_pool = self._max_pooling(bottom, 3, 2, 'same', 'pool')
axes = 3 if self.data_format == 'channels_last' else 1
return tf.concat([branch_3x3, branch_1x1x5x5, branch_pool], axis=axes)
def _inception_block2(self, bottom, filters, scope):
with tf.variable_scope(scope):
with tf.variable_scope('branch_1x1'):
branch_1x1 = self._conv_bn_activation(bottom, filters[0], 1, 1, 'same')
with tf.variable_scope('branch_1x1x7x7'):
branch_1x1x7x7 = self._conv_bn_activation(bottom, filters[1], 1, 1, 'same')
branch_1x1x7x7 = self._conv_bn_activation(branch_1x1x7x7, filters[1], [1, 7], 1, 'same')
branch_1x1x7x7 = self._conv_bn_activation(branch_1x1x7x7, filters[0], [7, 1], 1, 'same')
with tf.variable_scope('branch_1x1x7x7x7x7'):
branch_1x1x7x7x7x7 = self._conv_bn_activation(bottom, filters[1], 1, 1, 'same')
branch_1x1x7x7x7x7 = self._conv_bn_activation(branch_1x1x7x7x7x7, filters[1], [7, 1], 1, 'same')
branch_1x1x7x7x7x7 = self._conv_bn_activation(branch_1x1x7x7x7x7, filters[1], [1, 7], 1, 'same')
branch_1x1x7x7x7x7 = self._conv_bn_activation(branch_1x1x7x7x7x7, filters[1], [7, 1], 1, 'same')
branch_1x1x7x7x7x7 = self._conv_bn_activation(branch_1x1x7x7x7x7, filters[0], [1, 7], 1, 'same')
with tf.variable_scope('branch_pool1x1'):
branch_pool1x1 = self._avg_pooling(bottom, 3, 1, 'same', 'pool')
branch_pool1x1 = self._conv_bn_activation(branch_pool1x1, filters[0], 1, 1, 'same')
axes = 3 if self.data_format == 'channels_last' else 1
if self.is_SENet:
return self.squeeze_and_excitation( tf.concat([branch_1x1, branch_1x1x7x7, branch_1x1x7x7x7x7, branch_pool1x1], axis=axes))
else:
return tf.concat([branch_1x1, branch_1x1x7x7, branch_1x1x7x7x7x7, branch_pool1x1], axis=axes)
def _grid_size_reduction2(self, bottom, scope):
with tf.variable_scope(scope):
with tf.variable_scope('branch_1x1x3x3'):
branch_1x1x3x3 = self._conv_bn_activation(bottom, 192, 1, 1, 'same')
branch_1x1x3x3 = self._conv_bn_activation(branch_1x1x3x3, 320, 3, 2, 'same')
with tf.variable_scope('branch_1x1x7x7x3x3'):
branch_1x1x7x7x3x3 = self._conv_bn_activation(bottom, 192, 1, 1, 'same')
branch_1x1x7x7x3x3 = self._conv_bn_activation(branch_1x1x7x7x3x3, 192, [1, 7], 1, 'same')
branch_1x1x7x7x3x3 = self._conv_bn_activation(branch_1x1x7x7x3x3, 192, [7, 1], 1, 'same')
branch_1x1x7x7x3x3 = self._conv_bn_activation(branch_1x1x7x7x3x3, 192, 3, 2, 'same')
with tf.variable_scope('branch_pool'):
branch_pool = self._max_pooling(bottom, 3, 2, 'same', 'pool')
axes = 3 if self.data_format == 'channels_last' else 1
return tf.concat([branch_1x1x3x3, branch_1x1x7x7x3x3, branch_pool], axis=axes)
def _inception_block3(self, bottom, filters, scope):
with tf.variable_scope(scope):
with tf.variable_scope('branch_1x1'):
branch_1x1 = self._conv_bn_activation(bottom, filters[0], 1, 1, 'same')
with tf.variable_scope('branch_1x1x3x3'):
branch_1x1x3x3 = self._conv_bn_activation(bottom, filters[1], 1, 1, 'same')
branch_1x1x3x3x1x3 = self._conv_bn_activation(branch_1x1x3x3, filters[1], [1, 3], 1, 'same')
branch_1x1x3x3x3x1 = self._conv_bn_activation(branch_1x1x3x3, filters[1], [3, 1], 1, 'same')
axes = 3 if self.data_format == 'channels_last' else 1
branch_1x1x3x3 = tf.concat([branch_1x1x3x3x1x3, branch_1x1x3x3x3x1], axis=axes)
with tf.variable_scope('branch_1x1x3x3x3x3'):
branch_1x1x3x3x3x3 = self._conv_bn_activation(bottom, filters[2], 1, 1, 'same')
branch_1x1x3x3x3x3 = self._conv_bn_activation(branch_1x1x3x3x3x3, filters[1], 3, 1, 'same')
branch_1x1x3x3x3x3x1x3 = self._conv_bn_activation(branch_1x1x3x3x3x3, filters[1], [1, 3], 1, 'same')
branch_1x1x3x3x3x3x3x1 = self._conv_bn_activation(branch_1x1x3x3x3x3, filters[1], [3, 1], 1, 'same')
branch_1x1x3x3x3x3 = tf.concat([branch_1x1x3x3x3x3x1x3, branch_1x1x3x3x3x3x3x1], axis=axes)
with tf.variable_scope('branch_pool1x1'):
branch_pool1x1 = self._avg_pooling(bottom, 3, 1, 'same', 'pool')
branch_pool1x1 = self._conv_bn_activation(branch_pool1x1, filters[3], 1, 1, 'same')
if self.is_SENet:
return self.squeeze_and_excitation(tf.concat([branch_1x1, branch_1x1x3x3, branch_1x1x3x3x3x3, branch_pool1x1], axis=axes))
else:
return tf.concat([branch_1x1, branch_1x1x3x3, branch_1x1x3x3x3x3, branch_pool1x1], axis=axes)
def _compute_output_shape(self, shape, kernel, padding, strides):
assert(padding in ['same', 'valid'])
if padding == 'valid':
padded = 0
else:
padded = kernel // 2
return (shape - kernel + 2 * padded) // strides + 1