forked from jimmyyfeng/TD-LSTM
-
Notifications
You must be signed in to change notification settings - Fork 0
/
at_lstm.py
376 lines (330 loc) · 17 KB
/
at_lstm.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
370
371
372
373
374
375
376
#!/usr/bin/env python
# encoding: utf-8
# @author: newbie
# email: zhengshiliang0@gmail.com
import tensorflow as tf
from utils import load_w2v, batch_index, load_word_embedding, load_aspect2id, load_inputs_twitter_at
FLAGS = tf.app.flags.FLAGS
tf.app.flags.DEFINE_integer('embedding_dim', 300, 'dimension of word embedding')
tf.app.flags.DEFINE_integer('batch_size', 25, 'number of example per batch')
tf.app.flags.DEFINE_integer('n_hidden', 300, 'number of hidden unit')
tf.app.flags.DEFINE_float('learning_rate', 0.01, 'learning rate')
tf.app.flags.DEFINE_integer('n_class', 3, 'number of distinct class')
tf.app.flags.DEFINE_integer('max_sentence_len', 80, 'max number of tokens per sentence')
tf.app.flags.DEFINE_float('l2_reg', 0.001, 'l2 regularization')
tf.app.flags.DEFINE_integer('display_step', 4, 'number of test display step')
tf.app.flags.DEFINE_integer('n_iter', 20, 'number of train iter')
tf.app.flags.DEFINE_float('keep_prob1', 1.0, 'dropout keep prob')
tf.app.flags.DEFINE_float('keep_prob2', 1.0, 'dropout keep prob')
tf.app.flags.DEFINE_string('train_file_path', 'data/restaurant/rest_2014_lstm_train_new.txt', 'training file')
tf.app.flags.DEFINE_string('validate_file_path', 'data/restaurant/rest_2014_lstm_test_new.txt', 'validating file')
tf.app.flags.DEFINE_string('test_file_path', 'data/restaurant/rest_2014_lstm_test_new.txt', 'testing file')
tf.app.flags.DEFINE_string('embedding_file_path', 'data/restaurant/rest_2014_word_embedding_300_new.txt', 'embedding file')
tf.app.flags.DEFINE_string('word_id_file_path', 'data/restaurant/word_id_new.txt', 'word-id mapping file')
tf.app.flags.DEFINE_string('aspect_id_file_path', 'data/restaurant/aspect_id_new.txt', 'word-id mapping file')
tf.app.flags.DEFINE_string('method', 'AT', 'model type: AE, AT or AEAT')
tf.app.flags.DEFINE_string('t', 'last', 'model type: ')
class LSTM(object):
def __init__(self, embedding_dim=100, batch_size=64, n_hidden=100, learning_rate=0.01,
n_class=3, max_sentence_len=50, l2_reg=0., display_step=4, n_iter=100, type_=''):
self.embedding_dim = embedding_dim
self.batch_size = batch_size
self.n_hidden = n_hidden
self.learning_rate = learning_rate
self.n_class = n_class
self.max_sentence_len = max_sentence_len
self.l2_reg = l2_reg
self.display_step = display_step
self.n_iter = n_iter
self.type_ = type_
self.word_id_mapping, self.w2v = load_word_embedding(FLAGS.word_id_file_path, FLAGS.embedding_file_path, self.embedding_dim)
# self.word_embedding = tf.constant(self.w2v, dtype=tf.float32, name='word_embedding')
self.word_embedding = tf.Variable(self.w2v, dtype=tf.float32, name='word_embedding')
# self.word_id_mapping = load_word_id_mapping(FLAGS.word_id_file_path)
# self.word_embedding = tf.Variable(
# tf.random_uniform([len(self.word_id_mapping), self.embedding_dim], -0.1, 0.1), name='word_embedding')
self.aspect_id_mapping, self.aspect_embed = load_aspect2id(FLAGS.aspect_id_file_path, self.word_id_mapping, self.w2v, self.embedding_dim)
self.aspect_embedding = tf.Variable(self.aspect_embed, dtype=tf.float32, name='aspect_embedding')
self.keep_prob1 = tf.placeholder(tf.float32)
self.keep_prob2 = tf.placeholder(tf.float32)
with tf.name_scope('inputs'):
self.x = tf.placeholder(tf.int32, [None, self.max_sentence_len], name='x')
self.y = tf.placeholder(tf.int32, [None, self.n_class], name='y')
self.sen_len = tf.placeholder(tf.int32, None, name='sen_len')
self.aspect_id = tf.placeholder(tf.int32, None, name='aspect_id')
with tf.name_scope('weights'):
self.weights = {
'softmax': tf.get_variable(
name='softmax_w',
shape=[self.n_hidden, self.n_class],
initializer=tf.random_uniform_initializer(-0.01, 0.01),
regularizer=tf.contrib.layers.l2_regularizer(self.l2_reg)
)
}
with tf.name_scope('biases'):
self.biases = {
'softmax': tf.get_variable(
name='softmax_b',
shape=[self.n_class],
initializer=tf.random_uniform_initializer(-0.01, 0.01),
regularizer=tf.contrib.layers.l2_regularizer(self.l2_reg)
)
}
self.W = tf.get_variable(
name='W',
shape=[self.n_hidden + self.embedding_dim, self.n_hidden + self.embedding_dim],
initializer=tf.random_uniform_initializer(-0.01, 0.01),
regularizer=tf.contrib.layers.l2_regularizer(self.l2_reg)
)
self.w = tf.get_variable(
name='w',
shape=[self.n_hidden + self.embedding_dim, 1],
initializer=tf.random_uniform_initializer(-0.01, 0.01),
regularizer=tf.contrib.layers.l2_regularizer(self.l2_reg)
)
self.Wp = tf.get_variable(
name='Wp',
shape=[self.n_hidden, self.n_hidden],
initializer=tf.random_uniform_initializer(-0.01, 0.01),
regularizer=tf.contrib.layers.l2_regularizer(self.l2_reg)
)
self.Wx = tf.get_variable(
name='Wx',
shape=[self.n_hidden, self.n_hidden],
initializer=tf.random_uniform_initializer(-0.01, 0.01),
regularizer=tf.contrib.layers.l2_regularizer(self.l2_reg)
)
def dynamic_rnn(self, cell, inputs, length, max_len, scope_name, out_type='all'):
outputs, state = tf.nn.dynamic_rnn(
cell(self.n_hidden),
inputs=inputs,
sequence_length=length,
dtype=tf.float32,
scope=scope_name
) # outputs -> batch_size * max_len * n_hidden
batch_size = tf.shape(outputs)[0]
if out_type == 'last':
index = tf.range(0, batch_size) * max_len + (length - 1)
outputs = tf.gather(tf.reshape(outputs, [-1, self.n_hidden]), index) # batch_size * n_hidden
elif out_type == 'all_avg':
outputs = LSTM.reduce_mean(outputs, length)
return outputs
def bi_dynamic_rnn(self, cell, inputs, length, max_len, scope_name, out_type='all'):
outputs, state = tf.nn.bidirectional_dynamic_rnn(
cell_fw=cell(self.n_hidden),
cell_bw=cell(self.n_hidden),
inputs=inputs,
sequence_length=length,
dtype=tf.float32,
scope=scope_name
)
if out_type == 'last':
outputs_fw, outputs_bw = outputs
outputs_bw = tf.reverse_sequence(outputs_bw, tf.cast(length, tf.int64), seq_dim=1)
outputs = tf.concat([outputs_fw, outputs_bw], 2)
else:
outputs = tf.concat(outputs, 2) # batch_size * max_len * 2n_hidden
batch_size = tf.shape(outputs)[0]
if out_type == 'last':
index = tf.range(0, batch_size) * max_len + (length - 1)
outputs = tf.gather(tf.reshape(outputs, [-1, 2 * self.n_hidden]), index) # batch_size * 2n_hidden
elif out_type == 'all_avg':
outputs = LSTM.reduce_mean(outputs, length) # batch_size * 2n_hidden
return outputs
def AE(self, inputs, target, type_='last'):
"""
:params: self.x, self.seq_len, self.weights['softmax_lstm'], self.biases['sof
:return: non-norm prediction values
"""
print('I am AE.')
batch_size = tf.shape(inputs)[0]
target = tf.reshape(target, [-1, 1, self.embedding_dim])
target = tf.ones([batch_size, self.max_sentence_len, self.embedding_dim], dtype=tf.float32) * target
inputs = tf.concat([inputs, target], 2)
inputs = tf.nn.dropout(inputs, keep_prob=self.keep_prob1)
cell = tf.nn.rnn_cell.LSTMCell
outputs = self.dynamic_rnn(cell, inputs, self.sen_len, self.max_sentence_len, 'AE', FLAGS.t)
return LSTM.softmax_layer(outputs, self.weights['softmax'], self.biases['softmax'], self.keep_prob2)
def AT(self, inputs, target, type_=''):
print('I am AT.')
batch_size = tf.shape(inputs)[0]
target = tf.reshape(target, [-1, 1, self.embedding_dim])
target = tf.ones([batch_size, self.max_sentence_len, self.embedding_dim], dtype=tf.float32) * target
in_t = tf.concat([inputs, target], 2)
in_t = tf.nn.dropout(in_t, keep_prob=self.keep_prob1)
cell = tf.nn.rnn_cell.LSTMCell
hiddens = self.dynamic_rnn(cell, in_t, self.sen_len, self.max_sentence_len, 'AT', 'all')
h_t = tf.reshape(tf.concat([hiddens, target], 2), [-1, self.n_hidden + self.embedding_dim])
M = tf.matmul(tf.tanh(tf.matmul(h_t, self.W)), self.w)
alpha = LSTM.softmax(tf.reshape(M, [-1, 1, self.max_sentence_len]), self.sen_len, self.max_sentence_len)
self.alpha = tf.reshape(alpha, [-1, self.max_sentence_len])
r = tf.reshape(tf.matmul(alpha, hiddens), [-1, self.n_hidden])
index = tf.range(0, batch_size) * self.max_sentence_len + (self.sen_len - 1)
hn = tf.gather(tf.reshape(hiddens, [-1, self.n_hidden]), index) # batch_size * n_hidden
h = tf.tanh(tf.matmul(r, self.Wp) + tf.matmul(hn, self.Wx))
return LSTM.softmax_layer(h, self.weights['softmax'], self.biases['softmax'], self.keep_prob2)
@staticmethod
def softmax_layer(inputs, weights, biases, keep_prob):
with tf.name_scope('softmax'):
outputs = tf.nn.dropout(inputs, keep_prob=keep_prob)
predict = tf.matmul(outputs, weights) + biases
predict = tf.nn.softmax(predict)
return predict
@staticmethod
def reduce_mean(inputs, length):
"""
:param inputs: 3-D tensor
:param length: the length of dim [1]
:return: 2-D tensor
"""
length = tf.cast(tf.reshape(length, [-1, 1]), tf.float32) + 1e-9
inputs = tf.reduce_sum(inputs, 1, keep_dims=False) / length
return inputs
@staticmethod
def softmax(inputs, length, max_length):
inputs = tf.cast(inputs, tf.float32)
max_axis = tf.reduce_max(inputs, 2, keep_dims=True)
inputs = tf.exp(inputs - max_axis)
length = tf.reshape(length, [-1])
mask = tf.reshape(tf.cast(tf.sequence_mask(length, max_length), tf.float32), tf.shape(inputs))
inputs *= mask
_sum = tf.reduce_sum(inputs, reduction_indices=2, keep_dims=True) + 1e-9
return inputs / _sum
def run(self):
inputs = tf.nn.embedding_lookup(self.word_embedding, self.x)
aspect = tf.nn.embedding_lookup(self.aspect_embedding, self.aspect_id)
if FLAGS.method == 'AE':
prob = self.AE(inputs, aspect, FLAGS.t)
elif FLAGS.method == 'AT':
prob = self.AT(inputs, aspect, FLAGS.t)
with tf.name_scope('loss'):
reg_loss = tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES)
# cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(prob, self.y))
cost = - tf.reduce_mean(tf.cast(self.y, tf.float32) * tf.log(prob)) + sum(reg_loss)
with tf.name_scope('train'):
global_step = tf.Variable(0, name="tr_global_step", trainable=False)
optimizer = tf.train.AdamOptimizer(learning_rate=self.learning_rate).minimize(cost, global_step=global_step)
# optimizer = tf.train.AdagradOptimizer(learning_rate=self.learning_rate).minimize(cost, global_step=global_step)
with tf.name_scope('predict'):
correct_pred = tf.equal(tf.argmax(prob, 1), tf.argmax(self.y, 1))
true_y = tf.argmax(self.y, 1)
pred_y = tf.argmax(prob, 1)
accuracy = tf.reduce_sum(tf.cast(correct_pred, tf.int32))
_acc = tf.reduce_mean(tf.cast(correct_pred, tf.float32))
with tf.Session() as sess:
title = '-d1-{}d2-{}b-{}r-{}l2-{}sen-{}dim-{}h-{}c-{}'.format(
FLAGS.keep_prob1,
FLAGS.keep_prob2,
FLAGS.batch_size,
FLAGS.learning_rate,
FLAGS.l2_reg,
FLAGS.max_sentence_len,
FLAGS.embedding_dim,
FLAGS.n_hidden,
FLAGS.n_class
)
summary_loss = tf.summary.scalar('loss' + title, cost)
summary_acc = tf.summary.scalar('acc' + title, _acc)
train_summary_op = tf.summary.merge([summary_loss, summary_acc])
validate_summary_op = tf.summary.merge([summary_loss, summary_acc])
test_summary_op = tf.summary.merge([summary_loss, summary_acc])
import time
timestamp = str(int(time.time()))
_dir = 'logs/' + str(timestamp) + '_' + title
train_summary_writer = tf.summary.FileWriter(_dir + '/train', sess.graph)
test_summary_writer = tf.summary.FileWriter(_dir + '/test', sess.graph)
validate_summary_writer = tf.summary.FileWriter(_dir + '/validate', sess.graph)
saver = tf.train.Saver(write_version=tf.train.SaverDef.V2)
init = tf.global_variables_initializer()
sess.run(init)
# saver.restore(sess, 'models/logs/1481529975__r0.005_b2000_l0.05self.softmax/-1072')
save_dir = 'models/' + _dir + '/'
import os
if not os.path.exists(save_dir):
os.makedirs(save_dir)
tr_x, tr_sen_len, tr_target_word, tr_y = load_inputs_twitter_at(
FLAGS.train_file_path,
self.word_id_mapping,
self.aspect_id_mapping,
self.max_sentence_len,
self.type_
)
te_x, te_sen_len, te_target_word, te_y = load_inputs_twitter_at(
FLAGS.test_file_path,
self.word_id_mapping,
self.aspect_id_mapping,
self.max_sentence_len,
self.type_
)
max_acc = 0.
max_alpha = None
max_ty, max_py = None, None
for i in range(self.n_iter):
for train, _ in self.get_batch_data(tr_x, tr_sen_len, tr_y, tr_target_word, self.batch_size, FLAGS.keep_prob1, FLAGS.keep_prob2):
_, step, summary = sess.run([optimizer, global_step, train_summary_op], feed_dict=train)
train_summary_writer.add_summary(summary, step)
acc, loss, cnt = 0., 0., 0
flag = True
summary, step = None, None
alpha = None
ty, py = None, None
for test, num in self.get_batch_data(te_x, te_sen_len, te_y, te_target_word, 2000, 1.0, 1.0, False):
_loss, _acc, _summary, _step, alpha, ty, py = sess.run([cost, accuracy, validate_summary_op, global_step, self.alpha, true_y, pred_y],
feed_dict=test)
acc += _acc
loss += _loss * num
cnt += num
if flag:
summary = _summary
step = _step
flag = False
alpha = alpha
ty = ty
py = py
print('all samples={}, correct prediction={}'.format(cnt, acc))
test_summary_writer.add_summary(summary, step)
saver.save(sess, save_dir, global_step=step)
print('Iter {}: mini-batch loss={:.6f}, test acc={:.6f}'.format(i, loss / cnt, acc / cnt))
if acc / cnt > max_acc:
max_acc = acc / cnt
max_alpha = alpha
max_ty = ty
max_py = py
print('Optimization Finished! Max acc={}'.format(max_acc))
fp = open('weight.txt', 'w')
for y1, y2, ws in zip(max_ty, max_py, max_alpha):
fp.write(str(y1) + ' ' + str(y2) + ' ' + ' '.join([str(w) for w in ws]) + '\n')
print('Learning_rate={}, iter_num={}, batch_size={}, hidden_num={}, l2={}'.format(
self.learning_rate,
self.n_iter,
self.batch_size,
self.n_hidden,
self.l2_reg
))
def get_batch_data(self, x, sen_len, y, target_words, batch_size, keep_prob1, keep_prob2, is_shuffle=True):
for index in batch_index(len(y), batch_size, 1, is_shuffle):
feed_dict = {
self.x: x[index],
self.y: y[index],
self.sen_len: sen_len[index],
self.aspect_id: target_words[index],
self.keep_prob1: keep_prob1,
self.keep_prob2: keep_prob2,
}
yield feed_dict, len(index)
def main(_):
lstm = LSTM(
embedding_dim=FLAGS.embedding_dim,
batch_size=FLAGS.batch_size,
n_hidden=FLAGS.n_hidden,
learning_rate=FLAGS.learning_rate,
n_class=FLAGS.n_class,
max_sentence_len=FLAGS.max_sentence_len,
l2_reg=FLAGS.l2_reg,
display_step=FLAGS.display_step,
n_iter=FLAGS.n_iter,
type_=FLAGS.method
)
lstm.run()
if __name__ == '__main__':
tf.app.run()