├── README.md
└── cnn_imdb_vib.py


/README.md:
--------------------------------------------------------------------------------
1 | # Variational Information Bottleneck
2 | 


--------------------------------------------------------------------------------
/cnn_imdb_vib.py:
--------------------------------------------------------------------------------
 1 | #! -*- coding: utf-8 -*-
 2 | 
 3 | from keras.preprocessing import sequence
 4 | from keras.layers import *
 5 | from keras.models import Model
 6 | import keras.backend as K
 7 | from keras.datasets import imdb
 8 | 
 9 | 
10 | class VIB(Layer):
11 |     """变分信息瓶颈层
12 |     """
13 |     def __init__(self, lamb, **kwargs):
14 |         self.lamb = lamb
15 |         super(VIB, self).__init__(**kwargs)
16 |     def call(self, inputs):
17 |         z_mean, z_log_var = inputs
18 |         u = K.random_normal(shape=K.shape(z_mean))
19 |         kl_loss = - 0.5 * K.sum(K.mean(1 + z_log_var - K.square(z_mean) - K.exp(z_log_var), 0))
20 |         self.add_loss(self.lamb * kl_loss)
21 |         u = K.in_train_phase(u, 0.)
22 |         return z_mean + K.exp(z_log_var / 2) * u
23 |     def compute_output_shape(self, input_shape):
24 |         return input_shape[0]
25 | 
26 | 
27 | max_features = 5000
28 | maxlen = 400
29 | batch_size = 32
30 | embedding_dims = 50
31 | filters = 250
32 | kernel_size = 3
33 | epochs = 5
34 | 
35 | 
36 | (x_train, y_train), (x_test, y_test) = imdb.load_data(num_words=max_features)
37 | x_train = sequence.pad_sequences(x_train, maxlen=maxlen)
38 | x_test = sequence.pad_sequences(x_test, maxlen=maxlen)
39 | 
40 | 
41 | seq_in = Input(shape=(maxlen,))
42 | seq = seq_in
43 | 
44 | seq = Embedding(max_features, embedding_dims)(seq)
45 | seq = Conv1D(filters,
46 |              kernel_size,
47 |              padding='valid',
48 |              activation='relu',
49 |              strides=1)(seq)
50 | seq = GlobalMaxPooling1D()(seq)
51 | 
52 | z_mean = Dense(128)(seq)
53 | z_log_var = Dense(128)(seq)
54 | seq = VIB(0.1)([z_mean, z_log_var])
55 | seq = Dense(1, activation='sigmoid')(seq)
56 | 
57 | model = Model(seq_in, seq)
58 | model.compile(loss='binary_crossentropy',
59 |               optimizer='adam',
60 |               metrics=['accuracy'])
61 | 
62 | model.fit(x_train, y_train,
63 |           batch_size=batch_size,
64 |           epochs=epochs,
65 |           validation_data=(x_test, y_test))
66 | 
67 | 
68 | """
69 | # 去掉VIB后的模型
70 | 
71 | seq_in = Input(shape=(maxlen,))
72 | seq = seq_in
73 | 
74 | seq = Embedding(max_features, embedding_dims)(seq)
75 | seq = Conv1D(filters,
76 |              kernel_size,
77 |              padding='valid',
78 |              activation='relu',
79 |              strides=1)(seq)
80 | seq = GlobalMaxPooling1D()(seq)
81 | seq = Dense(1, activation='sigmoid')(seq)
82 | 
83 | model = Model(seq_in, seq)
84 | model.compile(loss='binary_crossentropy',
85 |               optimizer='adam',
86 |               metrics=['accuracy'])
87 | 
88 | model.fit(x_train, y_train,
89 |           batch_size=batch_size,
90 |           epochs=epochs,
91 |           validation_data=(x_test, y_test))
92 | """
93 | 


--------------------------------------------------------------------------------