-
Notifications
You must be signed in to change notification settings - Fork 2
/
capsnet.py
267 lines (242 loc) · 12.9 KB
/
capsnet.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
"""
License: Apache-2.0
Author: Fajie Ye
E-mail: yefajie@live.cn
"""
import tensorflow as tf
import os
import time
from tensorflow.python.ops import array_ops
class CapsModel(object):
def __init__(self, data_provider, params):
self.data_provider = data_provider
self.params = params
self.init_learning_rate = params.init_learning_rate
self._define_inputs()
self._build_graph()
self._initialize_session()
self._count_trainable_params()
def _define_inputs(self):
self.tensor_input = tf.placeholder(dtype=tf.float32, shape=[None, 28, 28, 1],
name='tensor_input')
self.tensor_output = tf.placeholder(dtype=tf.float32, shape=[None, 10],
name='tensor_output')
def _build_graph(self):
conv1 = self._build_relu_conv1(self.tensor_input) # shape=[batch_size, 20, 20, 256]
primary_caps = self._build_primary_caps(conv1) # shape=[batch_size, 32*6*6, 8]
digit_caps = self._build_digit_caps(primary_caps) # shape=[batch_size, 10, 16]
self.accuracy = self._compute_accuracy(digit_caps)
tf.summary.scalar('accuracy', self.accuracy)
self.loss = self._compute_loss(digit_caps)
global_step = tf.Variable(0, trainable=False)
learning_rate = tf.train.exponential_decay(self.init_learning_rate, global_step=global_step,
decay_rate=0.5, decay_steps=2000, staircase=True)
tf.summary.scalar('learning_rate', learning_rate)
self.train_op = tf.train.AdamOptimizer(learning_rate).minimize(self.loss, global_step=global_step)
self.summary_all = tf.summary.merge_all()
print('build model graph done...')
def _initialize_session(self):
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
self.sess = tf.Session(config=config)
self.sess.run(tf.global_variables_initializer())
self.saver = tf.train.Saver()
self.train_writer = tf.summary.FileWriter(os.path.join(self.params.log_dir, 'train'), graph=self.sess.graph)
self.test_writer = tf.summary.FileWriter(os.path.join(self.params.log_dir, 'test'), graph=self.sess.graph)
def _count_trainable_params(self):
total_parameters = 0
for variable in tf.trainable_variables():
shape = variable.get_shape()
variable_parameters = 1
for dim in shape:
variable_parameters *= dim.value
total_parameters += variable_parameters
print("Total training params: %.1fM" % (total_parameters / 1e6))
def _build_relu_conv1(self, _input):
"""
Build relu conv1
:param _input: => [batch_size, 28, 28, 1]
:return: conv1 => [batch_size, 20, 20, 256]
"""
with tf.variable_scope('relu_conv1'):
conv1 = tf.contrib.layers.conv2d(_input, 256, 9, stride=1, padding='VALID')
conv1 = tf.nn.relu(conv1)
return conv1
def _build_primary_caps(self, _input):
"""
Build primary caps
:param _input: => [batch_size, 20, 20, 256]
:return: primary_caps => [batch_size, 32x6x6, 8]
"""
with tf.variable_scope('primary_caps'):
conv = tf.contrib.layers.conv2d(_input, 32*8, 9, stride=2, padding='VALID',
activation_fn=None)
primary_caps = tf.reshape(conv, [-1, 32*6*6, 8])
primary_caps = self._squashing(primary_caps)
return primary_caps
def _build_digit_caps(self, _input):
"""
Build digit caps
:param _input: => [batch_size, 32*6*6, 8]
:return: digit_caps => [batch_size, 10, 16]
"""
with tf.variable_scope('digit_caps'):
batch_size = array_ops.shape(_input)[0]
w = tf.get_variable(name='weights', shape=[1, 32*6*6, 10, 8, 16],
initializer=tf.truncated_normal_initializer(stddev=self.params.stddev))
w = tf.tile(w, multiples=[batch_size, 1, 1, 1, 1]) # shape=[batch_size, 32*6*6, 10, 8, 16]
input_caps = tf.reshape(_input, [-1, 32*6*6, 1, 1, 8])
input_caps = tf.tile(input_caps, multiples=[1, 1, 10, 1, 1]) # shape=[batch_size, 32*6*6, 10, 1, 8]
u = tf.matmul(input_caps, w) # shape=[batch_size, 32*6*6, 10, 1, 16]
u = tf.reshape(u, shape=[-1, 32*6*6, 10, 16])
digit_caps = self._routing(u)
return digit_caps
def _compute_loss(self, _input):
"""
Compute margin loss and reconstruction loss(if used)
:param _input: digit_prob => [batch_size, 10, 16]
:return: loss = margin_loss + reconstruction loss
"""
digit_prob = tf.sqrt(tf.reduce_sum(tf.square(_input), axis=2)) # shape=[batch_size, 10]
margin_loss1 = tf.reduce_sum(tf.square(tf.maximum(0.0, 0.9-digit_prob)) * self.tensor_output, axis=1)
margin_loss2 = 0.5 * tf.reduce_sum(tf.square(tf.maximum(0.0, digit_prob-0.1)) * (1.0-self.tensor_output), axis=1)
self.margin_loss = tf.reduce_mean(margin_loss1 + margin_loss2)
tf.summary.scalar('margin_loss', self.margin_loss)
loss = self.margin_loss
if self.params.reconstruction:
self.reconstruction_output, self.reconstruction_loss = self._build_reconstruction_graph(_input)
tf.summary.scalar('reconstruction_loss', self.reconstruction_loss)
loss += self.params.recon_factor * self.reconstruction_loss
tf.summary.scalar('total_loss', loss)
return loss
def _compute_accuracy(self, _input):
"""
Compute accuracy
:param _input: digit_caps => [batch_size, 10, 16]
:return: accuracy => scalar
"""
digit_prob = tf.sqrt(tf.reduce_sum(tf.square(_input), axis=2)) # shape=[batch_size, 10]
correct_prediction = tf.equal(tf.argmax(digit_prob, 1), tf.argmax(self.tensor_output, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
return accuracy
def _build_reconstruction_graph(self, _input):
"""
Use an additional reconstruction loss to encourage the digit capsules to encode the instantiation
parameters of the input digit. During training, we mask out all but the activity vector of the correct
digit capsule. Then we use this activity vector to reconstruct the input image
:param _input: digit_caps => [batch_size, 10, 16]
:return: reconstruction_output => [batch_size, 784]
reconstruction_loss => scalar
"""
def fc_layer(input, units, activation_fn):
return tf.contrib.layers.fully_connected(input, units, activation_fn=activation_fn)
with tf.variable_scope('reconstruction_graph'):
mask = tf.reshape(self.tensor_output, [-1, 10, 1])
digit_caps_masked = _input * mask
digit_caps_masked = tf.reshape(digit_caps_masked, [-1, 10*16])
fc_relu1 = fc_layer(digit_caps_masked, 512, tf.nn.relu)
fc_relu2 = fc_layer(fc_relu1, 1024, tf.nn.relu)
reconstruction_output = fc_layer(fc_relu2, 784, tf.nn.sigmoid)
tensor_input_reshape = tf.reshape(self.tensor_input, [-1, 784])
reconstruction_loss = tf.reduce_mean(tf.reduce_sum(tf.square(reconstruction_output-tensor_input_reshape),
axis=1))
return reconstruction_output, reconstruction_loss
def _squashing(self, _input):
"""
A non-linear "squashing" function to ensure that short vectors get shrunk
to almost zero length and long vectors get shrunk to a length slightly below 1.
:param _input: => [batch_size, num_caps, cap_len]
:return: squashed_caps=> [batch_size, num_caps, cap_len]
"""
squared_norm = tf.reduce_sum(tf.square(_input), axis=2, keep_dims=True)
squash_factor = squared_norm / (1 + squared_norm) / tf.sqrt(squared_norm)
squashed_caps = _input * squash_factor
return squashed_caps
def _routing(self, _input):
"""
Routing algorithm
:param _input: prediction vectors => [batch_size, 32*6*6, 10, 16]
:return: digit_caps => [batch_size, 10, 16]
"""
u = _input
u_temp = tf.stop_gradient(u)
batch_size = array_ops.shape(_input)[0]
b = tf.zeros(shape=[batch_size, 32*6*6, 10, 1], dtype=tf.float32)
for i in range(self.params.iter_routing):
c = tf.nn.softmax(b)
if i < self.params.iter_routing-1:
s = tf.reduce_sum(u_temp*c, axis=1) # shape=[batch_size, 10, 16]
v = self._squashing(s)
# v_temp => [batch_size, 32*6*6, 10, 16]
v_temp = tf.tile(tf.reshape(v, [-1, 1, 10, 16]), multiples=[1, 32*6*6, 1, 1])
# u_multi_v => [batch_size, 32*6*6, 10, 1]
u_multi_v = tf.reduce_sum(u_temp * v_temp, axis=-1, keep_dims=True)
b = b + u_multi_v
else:
s = tf.reduce_sum(u * c, axis=1) # shape=[batch_size, 10, 16]
v = self._squashing(s)
digit_caps = v
return digit_caps
def train(self):
"""
Train capsnet model
"""
print('training start...')
batch_size = self.params.batch_size
train_steps = self.params.train_epoch * self.data_provider.train.num_examples // batch_size
for step in range(train_steps):
data, targets = self.data_provider.train.next_batch(batch_size)
time1 = time.time()
self.sess.run(self.train_op, feed_dict={self.tensor_input: data, self.tensor_output: targets})
time_per_batch = time.time() - time1
if step % 100 == 0:
fetches = [self.loss, self.accuracy, self.summary_all]
feed_dict = {self.tensor_input: data, self.tensor_output: targets}
train_loss, accuracy, merged = self.sess.run(fetches, feed_dict=feed_dict)
print('train step: {:d}/{:d}, time_per_batch: {:.4f}s, train loss: {:.4f}, accuracy: {:.4f}, '
'complete after {:s}'.format(step, train_steps, time_per_batch, train_loss, accuracy,
get_time_left(time_per_batch, train_steps-step)))
self.train_writer.add_summary(merged, global_step=step)
if step % 1000 == 0:
fetches = [self.loss, self.accuracy, self.summary_all]
test_data, test_targets = self.data_provider.test.next_batch(1000)
feed_dict = {self.tensor_input: test_data, self.tensor_output: test_targets}
test_loss, accuracy, merged = self.sess.run(fetches, feed_dict=feed_dict)
print('=========== test loss: {:.4f}\taccuracy: {:.4f}============='.format(
test_loss, accuracy))
self.test_writer.add_summary(merged, global_step=step)
self.saver.save(self.sess, os.path.join(self.params.model_dir, 'model.ckpt'), global_step=step)
print('training over...')
test_accuracy = self.sess.run(self.accuracy, feed_dict={self.tensor_input: self.data_provider.test.images,
self.tensor_output: self.data_provider.test.labels})
print('final test accuracy: {:.5f}'.format(test_accuracy))
def test(self):
"""
Test capsnet model
"""
batch_size = 100
total_accuracy = []
time1 = time.time()
for i in range(self.data_provider.test.num_examples // batch_size):
data, targets = self.data_provider.test.next_batch(batch_size)
accuracy = self.sess.run(self.accuracy,
feed_dict={self.tensor_input: data, self.tensor_output: targets})
total_accuracy.append(accuracy)
print('{:d}, acc: {:.4f}'.format(i, accuracy))
print('test accuracy: {:.5f}, use time: {:.3f}s'.format(sum(total_accuracy)/len(total_accuracy),
time.time()-time1))
def load_model(self):
"""
Load model from trained model files
"""
last_ckpt = tf.train.latest_checkpoint(self.params.model_dir)
if not last_ckpt:
return False
self.saver.restore(self.sess, last_ckpt)
print("Successfully load model from save path: {:s}".format(last_ckpt))
return True
def get_time_left(time_per_batch, batch_num_left):
seconds = int(time_per_batch * batch_num_left)
m, s = divmod(seconds, 60)
h, m = divmod(m, 60)
return '{:d}:{:d}:{:d}'.format(h, m, s)