-
Notifications
You must be signed in to change notification settings - Fork 0
/
Adaptive-diverse-capsule-network.py
316 lines (243 loc) · 14.1 KB
/
Adaptive-diverse-capsule-network.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
# -*- coding: utf-8 -*
"""
This implement is an improved version of real-valued capsule network from our paper《Cv-CapsNet:complex-valued capsule network》,
We introduce an attentional mechanism for fusing of three levels of features by weights so as to eliminate the manual setting
of capsule dimensions in the coding stage.
Some key layers used for constructing a Capsule Network. These layers can used to construct CapsNet on other dataset,
not just on CIFAR10.
AUTHOR Jiangnan He E-mail: [email protected], github:https://github.com/Johnnan002/Adaptive-diverse-capsule-network
We refer to the implementation of the capsule network Github: `https://github.com/XifengGuo/CapsNet-Keras`
Usage:
python Adaptive-diverse-capsule-network.py
... ...
Result:
Validation accuracy > 88.5% after 25 epochs.
About 600 seconds per epoch on a single tesla k80 GPU card
"""
from keras.layers import Lambda
import numpy as np
from keras import layers, models, optimizers
from keras import backend as K
from keras.utils import to_categorical
import matplotlib.pyplot as plt
from capsulelayers import CapsuleLayer, PrimaryCap, Length,bottleneck
from keras.layers.merge import concatenate
from capsulelayers import squash
K.set_image_data_format('channels_last')
def CapsNet(input_shape, n_class, routings):
"""
Adaptive diverse Capsule Network on CIFAR10.
:param input_shape: data shape, 3d, [width, height, channels]
:param n_class: number of classes
:param routings: number of routing iterations
:return: Two Keras Models, the first one used for training, and the second one for evaluation.
`eval_model` can also be used for training.
"""
concat_axis = 1 if K.image_data_format() == 'channels_first' else -1
x = layers.Input(shape=input_shape)
l_1 = bottleneck(x, 32, (3, 3), e=1, s=1, squeeze=False, nl='RE')
m_1 = concatenate([x, l_1], axis=concat_axis)
l_2 = bottleneck(m_1, 32, (3, 3), e=1, s=1, squeeze=False, nl='RE')
m_2 = concatenate([m_1, l_2], axis=concat_axis)
l_3 = bottleneck(m_2, 32, (3, 3), e=1, s=1, squeeze=False, nl='RE')
m_3 = concatenate([m_2, l_3], axis=concat_axis)
l_4 = bottleneck(m_3, 32, (3, 3), e=1, s=1, squeeze=False, nl='RE')
m_4 = concatenate([m_3, l_4], axis=concat_axis)
l_5 = bottleneck(m_4, 32, (3, 3), e=1, s=1, squeeze=False, nl='RE')
m_5 = concatenate([m_4, l_5], axis=concat_axis)
l_6 = bottleneck(m_5, 32, (3, 3), e=1, s=1, squeeze=False, nl='RE')
m_6 = concatenate([m_5, l_6], axis=concat_axis)
l_7 = bottleneck(m_6, 32, (3, 3), e=1, s=1, squeeze=False, nl='RE')
m_7 = concatenate([m_6, l_7], axis=concat_axis)
l_8 = bottleneck(m_7, 32, (3, 3), e=1, s=1, squeeze=False, nl='HS')
m_8 = concatenate([m_7, l_8], axis=concat_axis)
input1,primarycaps1,a1 = PrimaryCap(m_8,dim_capsule=12, n_channels=10, kernel_size=5, strides=(2,2), padding='valid')
#################################################level1 低层特征提取分支###############################################
l2_1= bottleneck(input1, 32, (3, 3), e=1, s=1, squeeze=False, nl='RE')
m2_1 = concatenate([input1, l2_1], axis=concat_axis)
l2_2 = bottleneck(m2_1, 32, (3, 3), e=1, s=1, squeeze=False, nl='RE')
m2_2 = concatenate([m2_1, l2_2], axis=concat_axis)
l2_3 = bottleneck(m2_2, 32, (3, 3), e=1, s=1, squeeze=False, nl='RE')
m2_3 = concatenate([m2_2, l2_3], axis=concat_axis)
l2_4 = bottleneck(m2_3, 32, (3, 3), e=1, s=1, squeeze=False, nl='RE')
m2_4 = concatenate([m2_3,l2_4], axis=concat_axis)
l2_5 = bottleneck(m2_4, 32, (3, 3), e=1, s=1, squeeze=False, nl='RE')
m2_5 = concatenate([m2_4, l2_5], axis=concat_axis)
l2_6= bottleneck(m2_5, 32, (3, 3), e=1, s=1, squeeze=False, nl='RE')
m2_6 = concatenate([m2_5, l2_6], axis=concat_axis)
l2_7 = bottleneck(m2_6, 32, (3, 3), e=1, s=1, squeeze=False, nl='RE')
m2_7 = concatenate([m2_6, l2_7], axis=concat_axis)
l2_8 = bottleneck(m2_7, 32, (3, 3), e=1, s=1, squeeze=False, nl='HS')
m2_8 = concatenate([m2_7, l2_8], axis=concat_axis)
input2, primarycaps2 ,a2= PrimaryCap(m2_8, dim_capsule=12, n_channels=10, kernel_size=5, strides=(2, 2), padding='valid')
###########################################level2 中层特征提取分支##################################################
l3_1 = bottleneck(input2, 32, (3, 3), e=1, s=1, squeeze=False, nl='RE')
m3_1 = concatenate([input2, l3_1], axis=concat_axis)
l3_2 = bottleneck( m3_1 , 32, (3, 3), e=1, s=1, squeeze=False, nl='RE')
m3_2 = concatenate([m3_1,l3_2], axis=concat_axis)
l3_3 = bottleneck( m3_2 , 32, (3, 3), e=1, s=1, squeeze=False, nl='RE')
m3_3 = concatenate([m3_2,l3_3], axis=concat_axis)
l3_4= bottleneck( m3_3 , 32, (3, 3), e=1, s=1, squeeze=False, nl='RE')
m3_4 = concatenate([m3_3, l3_4], axis=concat_axis)
l3_5 = bottleneck( m3_4 , 32, (3, 3), e=1, s=1, squeeze=False, nl='RE')
m3_5=concatenate([ m3_4, l3_5], axis=concat_axis)
l3_6= bottleneck( m3_5 , 32, (3, 3), e=1, s=1, squeeze=False, nl='RE')
m3_6=concatenate([m3_5, l3_6], axis=concat_axis)
l3_7= bottleneck( m3_6 , 32, (3, 3), e=1, s=1, squeeze=False, nl='RE')
m3_7=concatenate([m3_6, l3_7], axis=concat_axis)
l3_8 = bottleneck( m3_7 , 32, (3, 3), e=1, s=1, squeeze=False, nl='RE')
m3_8 = concatenate([m3_7,l3_8], axis=concat_axis)
input3, primarycaps3 ,a3= PrimaryCap(m3_8, dim_capsule=12, n_channels=10, kernel_size=3, strides=(1, 1), padding='valid')
##################################################level3 高层特征提取分支##############################################
primarycaps1=layers.Reshape(target_shape=(-1, 12), name='primarycaps11')(primarycaps1)
primarycaps2=layers.Reshape(target_shape=(-1, 12), name='primarycaps21')(primarycaps2)
primarycaps3 = layers.Reshape(target_shape=(-1,12), name='primarycaps31')(primarycaps3)
#低层 中层 高层 特征编码到相同维度12D 经过路由层降维到6D 通过SE模块学到的参数，将三层特征进行加权融合，
#避免了低层 中层 高层手动设置不同维度胶囊来平衡各层级特征。这里通过SE模块学的参数来进行平衡。
digitcaps2= CapsuleLayer(num_capsule=n_class, dim_capsule=6, routings=routings,
name='digitcaps2')(primarycaps1)
digitcaps3= CapsuleLayer(num_capsule=n_class, dim_capsule=6, routings=routings,
name='digitcaps3')(primarycaps2)
digitcaps4= CapsuleLayer(num_capsule=n_class, dim_capsule=6, routings=routings,
name='digitcaps4')(primarycaps3)
digitcaps2 = layers.Reshape(target_shape=(-1, 6), name='digitcaps21')(digitcaps2)
a1 = K.tile(a1, [1,10, 6])#加权系数a1
weight_1 = Lambda(lambda x: x * a1)
digitcaps2 = weight_1(digitcaps2)
digitcaps3 = layers.Reshape(target_shape=(-1, 6), name='digitcaps31') (digitcaps3)
a2= K.tile(a2, [1,10, 6])
weight_2 = Lambda(lambda x: x * a2)
digitcaps3=weight_2(digitcaps3)
digitcaps4 = layers.Reshape(target_shape=(-1, 6), name='digitcaps41')(digitcaps4)
a3 = K.tile(a3, [1,10, 6])
weight_3 = Lambda(lambda x: x * a3)
digitcaps4 = weight_3(digitcaps4)
digitcaps = concatenate([ digitcaps2, digitcaps3], axis=-1)#加权后拼接成最终的胶囊
digitcaps = concatenate([ digitcaps, digitcaps4], axis=-1)
digitcaps = layers.Lambda(squash)(digitcaps)
out_caps = Length(name='capsnet')(digitcaps)
y = layers.Input(shape=(n_class,))
train_model = models.Model([x, y], out_caps)
eval_model = models.Model(x, out_caps )
return train_model, eval_model
####################################################capsnet################################
def margin_loss(y_true, y_pred):
"""
Margin loss for Eq.(4). When y_true[i, :] contains not just one `1`, this loss should work too. Not test it.
:param y_true: [None, n_classes]
:param y_pred: [None, num_capsule]
:return: a scalar loss value.
"""
#如果标签为0 则预测概率不能离0.1 太远 否则loss大 如果标签为1 则预测概率不能离0.9太远 否则loss大
L = y_true * K.square(K.maximum(0., 0.9 - y_pred)) + \
0.5 * (1 - y_true) * K.square(K.maximum(0., y_pred - 0.1))
return K.mean(K.sum(L, 1))
def train(model, data, args):
"""
Training a CapsuleNet
:param model: the CapsuleNet model
:param data: a tuple containing training and testing data, like `((x_train, y_train), (x_test, y_test))`
:param args: arguments
:return: The trained model
"""
# unpacking the data
(x_train, y_train), (x_test, y_test) = data
# callbacks
log = callbacks.CSVLogger(args.save_dir + '/log.csv')
tb = callbacks.TensorBoard(log_dir=args.save_dir + '/tensorboard-logs',
batch_size=args.batch_size, histogram_freq=int(args.debug))
checkpoint = callbacks.ModelCheckpoint(args.save_dir + '/weights-{epoch:02d}.h5', monitor='val_capsnet_acc',
save_best_only=True, save_weights_only=True, verbose=2)
lr_decay = callbacks.LearningRateScheduler(schedule=lambda epoch: args.lr * (args.lr_decay ** epoch))
model.compile(optimizer=optimizers.Adam(lr=args.lr),
loss=margin_loss,
loss_weights=[1.],
metrics={'capsnet': 'accuracy'})
# Begin: Training with data augmentation ---------------------------------------------------------------------#
def train_generator(x, y, batch_size, shift_fraction=0.):
train_datagen = ImageDataGenerator(width_shift_range=shift_fraction,
height_shift_range=shift_fraction)
generator = train_datagen.flow(x, y, batch_size=batch_size)
while 1:
x_batch, y_batch = generator.next()
yield ([x_batch, y_batch], y_batch)
# Training with data augmentation. If shift_fraction=0., also no augmentation.
model.fit_generator(generator=train_generator(x_train, y_train, args.batch_size, args.shift_fraction),
steps_per_epoch=int(y_train.shape[0] / args.batch_size),
epochs=args.epochs,
validation_data=[[x_test, y_test], y_test],
callbacks=[log, tb, checkpoint, lr_decay])
# callbacks=[log, tb, checkpoint])
# End: Training with data augmentation -----------------------------------------------------------------------#
model.save_weights(args.save_dir + '/trained_model.h5')
print('Trained model saved to \'%s/trained_model.h5\'' % args.save_dir)
from utils1 import plot_log
plot_log(args.save_dir + '/log.csv', show=True)
return model
def test(model, data, args):
x_test, y_test = data
y_pred= model.predict(x_test, batch_size=128)
print(y_pred)
print('-'*30 + 'Begin: test' + '-'*30)
print('Test acc:', np.sum(np.argmax(y_pred, 1) == np.argmax(y_test, 1))/y_test.shape[0])
print()
print('Reconstructed images are saved to %s/real_and_recon.png' % args.save_dir)
print('-' * 30 + 'End: test' + '-' * 30)
plt.imshow(plt.imread(args.save_dir + "/real_and_recon.png"))
plt.show()
def load_mnist():
# the data, shuffled and split between train and test sets
from keras.datasets import cifar10
(x_train, y_train), (x_test, y_test) = cifar10.load_data()
x_train = x_train.reshape(-1, 32, 32, 3).astype('float32') / 255.
x_test = x_test.reshape(-1, 32, 32, 3).astype('float32') / 255.
y_train = to_categorical(y_train,10)
y_test = to_categorical(y_test,10)
y_train = y_train.reshape(-1, 10)
y_test = y_test.reshape(-1, 10)
return (x_train, y_train), (x_test, y_test)
if __name__ == "__main__":
import os
import argparse
from keras.preprocessing.image import ImageDataGenerator
from keras import callbacks
# setting the hyper parameters
parser = argparse.ArgumentParser(description="Adaptive Diverse Capsule Network on CIFAR10.")
parser.add_argument('--epochs', default=25, type=int)
parser.add_argument('--batch_size', default=128, type=int)
parser.add_argument('--lr', default=0.003, type=float,
help="Initial learning rate")
parser.add_argument('--lr_decay', default=0.90, type=float,
help="The value multiplied by lr at each epoch. Set a larger value for larger epochs")
parser.add_argument('-r', '--routings', default=3, type=int,
help="Number of iterations used in routing algorithm. should > 0")
parser.add_argument('--shift_fraction', default=0.1, type=float,
help="Fraction of pixels to shift at most in each direction.")
parser.add_argument('--debug', action='store_true',
help="Save weights by TensorBoard")
parser.add_argument('--save_dir', default='./result')
parser.add_argument('-t', '--testing', action='store_true',
help="Test the trained model on testing dataset")
parser.add_argument('--digit', default=5, type=int,
help="Digit to manipulate")
parser.add_argument('-w', '--weights', default=None,
help="The path of the saved weights. Should be specified when testing")
args = parser.parse_args()
print(args)
if not os.path.exists(args.save_dir):
os.makedirs(args.save_dir)
(x_train, y_train), (x_test, y_test) = load_mnist()
x_train = np.reshape(x_train, (-1, 32, 32, 3))
x_test = np.reshape(x_test, (-1, 32, 32, 3))
model, eval_model= CapsNet(input_shape=x_train.shape[1:],
n_class=10,routings=args.routings)
model.summary()
# train or test
if args.weights is not None: # init the model weights with provided one
model.load_weights(args.weights)
if not args.testing:
train(model=model, data=((x_train, y_train), (x_test, y_test)), args=args)
else: # as long as weights are given, will run testing
if args.weights is None:
print('No weights are provided. Will test using random initialized weights.')
test(model=eval_model, data=(x_test, y_test), args=args)