-
Notifications
You must be signed in to change notification settings - Fork 8
/
cnn_model.py
231 lines (192 loc) · 8.96 KB
/
cnn_model.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
import data_utils
import tensorflow as tf
import os
class CNN(object):
def __init__(
self,
ntags=2,
word_num=100,
word_dim=50,
filter_sizes='2,3,4',
kernel_num=100,
learning_rate_base=0.001,
batch_size=5,
sent_len=100,
l2_alpha=1e-4,
word_embedding=None,
input_type='CNN-static',
vocab=None,
dropout_prob=0.5,
epoch=20,
model_path=''
):
self.ntags = ntags
self.word_num = word_num
self.word_dim = word_dim
self.filter_sizes = [int(e) for e in filter_sizes.split(',')]
self.kernel_num = kernel_num
self.learning_rate_base = learning_rate_base
self.batch_size = batch_size
self.sent_len = sent_len
self.l2_alpha = l2_alpha
self.pre_embedding = None
self.input_type = input_type
self.vocab = vocab
self.dropout_prob = dropout_prob
self.epoch = epoch
self.model_path = model_path
if self.input_type == 'CNN-rand':
self.pre_embedding = None
self.train_able = True
elif self.input_type == 'CNN-static':
self.pre_embedding = word_embedding
self.train_able = False
elif self.input_type == 'CNN-non-static':
self.pre_embedding = word_embedding
self.train_able = True
elif self.input_type == 'CNN-multichannel':
self.pre_embedding = word_embedding
self.train_able = True
self.init_graph()
def place_holder(self):
self.input_x = tf.placeholder(tf.int32, shape=[None, self.sent_len], name="input_x")
self.input_y = tf.placeholder(tf.int64, shape=[None], name="input_y")
self.dropout = tf.placeholder(dtype=tf.float32, shape=[], name="dropout")
def embedding(self):
with tf.variable_scope('word_embedding'):
if self.pre_embedding is None:
self._W_emb = tf.get_variable(name='embedding', shape=[self.word_num, self.word_dim],
initializer=tf.random_uniform_initializer(-1.0, 1.0), trainable=self.train_able)
else:
self._W_emb = tf.Variable(self.pre_embedding, name='_word_embeddings', dtype=tf.float32,
trainable=self.train_able)
word_embedding = tf.nn.embedding_lookup(self._W_emb, self.input_x)
# 3 channel
self.word_embedding = tf.expand_dims(word_embedding, -1)
if self.input_type == 'CNN-multichannel':
self._W_emb_2 = tf.Variable(self.pre_embedding, name='_word_embeddings_f', dtype=tf.float32, trainable=False)
word_embedding2 = tf.nn.embedding_lookup(self._W_emb_2, self.input_x)
word_embedding2 = tf.expand_dims(word_embedding2, -1)
self.word_embedding = tf.concat([self.word_embedding, word_embedding2], axis=3)
self.input_dim = self.word_dim
def conv_layer(self):
pooled_outputs = []
for i, filter_size in enumerate(self.filter_sizes):
with tf.variable_scope('conv-maxpool-%s' % filter_size, reuse=tf.AUTO_REUSE):
filter_shape = [filter_size, self.input_dim, 1, self.kernel_num]
W = tf.get_variable(name='conv_w', shape=filter_shape,
initializer=tf.truncated_normal_initializer(-0.1, 0.1))
b = tf.get_variable(name='conv_b', shape=[self.kernel_num], initializer=tf.constant_initializer(0.1))
conv = tf.nn.conv2d(
self.word_embedding,
W,
strides=[1, 1, 1, 1],
padding="VALID",
name="conv"
)
activation = tf.nn.relu(tf.nn.bias_add(conv, b), name="relu")
conv_len = activation.get_shape()[1]
pooled = tf.nn.max_pool(
activation,
ksize=[1, conv_len, 1, 1],
strides=[1, 1, 1, 1],
padding='VALID',
name='pool'
)
pooled_outputs.append(pooled)
self.num_filters_total = self.kernel_num * len(self.filter_sizes)
self.h_pool = tf.concat(pooled_outputs, 1)
self.final_feature = tf.reshape(self.h_pool, [-1, self.num_filters_total])
with tf.name_scope("dropout"):
self.final_feature = tf.nn.dropout(self.final_feature, self.dropout)
def fc_layer(self):
with tf.variable_scope("proj"):
self.fc_w = tf.get_variable(name='fc_w', shape=[self.num_filters_total, self.ntags],
initializer=tf.truncated_normal_initializer(stddev=0.1))
fc_b = tf.get_variable(name='fc_b', shape=[self.ntags], initializer=tf.constant_initializer(0.1))
self.fc = tf.matmul(self.final_feature, self.fc_w) + fc_b
self.logits = self.fc
def train_op(self):
losses = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=self.logits, labels=self.input_y)
self.l2_loss = tf.contrib.layers.apply_regularization(
regularizer=tf.contrib.layers.l2_regularizer(self.l2_alpha),
weights_list=tf.trainable_variables())
self.loss = tf.reduce_mean(losses) + self.l2_loss
global_step = tf.Variable(0, trainable=False)
self.learning_rate = tf.train.exponential_decay(self.learning_rate_base, global_step, 200, 0.99, staircase=True)
self.train_step = tf.train.AdamOptimizer(self.learning_rate).minimize(self.loss, global_step=global_step)
self.pred = tf.nn.softmax(self.logits)
self.pred_index = tf.argmax(self.pred, 1)
self.correct_prediction = tf.equal(tf.argmax(self.pred, 1), self.input_y)
self.accuracy = tf.reduce_mean(tf.cast(self.correct_prediction, "float"))
def init_graph(self):
tf.reset_default_graph()
self.place_holder()
self.embedding()
self.conv_layer()
self.fc_layer()
self.train_op()
self.sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True, log_device_placement=False))
self.sess.run(tf.global_variables_initializer())
self.saver = tf.train.Saver()
def save_session(self, dir_model):
if not os.path.exists(dir_model):
os.makedirs(dir_model)
self.saver.save(self.sess, dir_model)
def restore_session(self, dir_model):
self.saver.restore(self.sess, dir_model)
def close_session(self):
self.sess.close()
def train(self, train, valid):
accuracy = 0.0
for e in range(self.epoch):
for i, (data, y) in enumerate(data_utils.minibatches(train, self.batch_size)):
logits, _, lr, loss = self.sess.run([self.logits,self.train_step, self.learning_rate, self.loss], feed_dict=
{
self.input_x: data,
self.input_y: y,
self.dropout: self.dropout_prob
})
if i % 100 == 0:
acc_test = self.evaluate(valid)
if acc_test > accuracy:
self.save_session(self.model_path)
print('This is the ' + str(e) + ' epoch training, the ' + str(i) + ' batch data，learning rate = ' + str(round(lr, 5)) +
', loss = ' + str(round(loss, 2)) + ', accuracy = ' + str(acc_test))
def predict(self, sentences):
pred, pred_index = self.sess.run([self.pred, self.pred_index], feed_dict=
{
self.input_x: sentences,
self.dropout: 1.0
})
return pred_index
def evaluate(self, valid):
acc_total, loss_total, cnt = 0, 0, 0
for i, (data, y) in enumerate(data_utils.minibatches(valid, self.batch_size)):
cnt += 1
acc = self.sess.run(self.accuracy, feed_dict={
self.input_x: data,
self.input_y: y,
self.dropout: 1.0
})
acc_total += self.batch_size * acc
acc_valid = round(acc_total * 1.0 / len(valid), 3)
return acc_valid
if __name__ == '__main__':
vocabulary_path = './input/data/vocabulary.txt'
vocab, rev_vocab = data_utils.initialize_vocabulary(vocabulary_path)
embed_path = './input/data/embed/glove.6B.300d.npz'
embeddings = data_utils.get_trimmed_glove_vectors(embed_path)
model = CNN(
batch_size=10,
word_embedding=embeddings,
sent_len=100,
input_type='CNN-static',
word_num=len(rev_vocab),
word_dim=300,
vocab=vocab
)
train_data = data_utils.text_dataset('./input/data/train_data.ids', 100)
valid_data = data_utils.text_dataset('./input/data/valid_data.ids', 100)
print('train set={a},valid set={b}'.format(a=train_data.__len__(), b=valid_data.__len__()))
model.train(train_data, valid_data)