-
Notifications
You must be signed in to change notification settings - Fork 7
/
Copy pathps_SC_MSA_R.py
209 lines (192 loc) · 12.7 KB
/
ps_SC_MSA_R.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
import sys
sys.path.append('..')
import tensorflow as tf
import pickle
import random
from numpy import linalg
import numpy as np
import math
import os
import utils
if __name__ == '__main__':
# ################ Prepare Data ###################
# basic config
dim = 200 # dimension of embedding
lr_decay_rate = 0.99 # learning rate decay rate
# Note that there is a mistake about lr_decay_rate, we correct it here.
# To achieve the reported results, actually we do not need to decay the learning rate.
batch_size = 1 # batch size, set to 1 because we use SGD
learning_rate = 0.01 # initial learning rate
# learning_rate = float(sys.argv[1]) # initial learning rate
labda = 0.0001 # regularization term labmda
# labda = float(sys.argv[2]) # regularization term labmda
total_epoch = 50 # total training epoches
# total_epoch = int(sys.argv[3]) # total training epoches
trunc_num = 5 # dimention for decomposed sparse matrix
# trunc_num = int(sys.argv[4])
hownet_filename = 'dataset/hownet.txt'
comp_filename = 'dataset/all.bin'
train_filename = 'dataset/train.bin'
test_filename = 'dataset/test.bin'
dev_filename = 'dataset/dev.bin'
embedding_filename = 'dataset/word_embedding.txt'
sem_embed_filename = 'dataset/sememe_vector.txt'
logdir_name = 'phrase_sim/SCMSApos_'+'trunc'+str(trunc_num)
# load hownet，并把hownet.comp分成test_set和train_set
hownet = utils.Hownet(hownet_file=hownet_filename, comp_file=comp_filename)
hownet.build_hownet()
hownet.token2id()
hownet.load_split_dataset(train_filename=train_filename, test_filename=test_filename, dev_filename=dev_filename)
word_embedding_np, hownet = utils.load_word_embedding(embedding_filename, hownet, scale=False) # load word embedding
sememe_embedding_np = utils.load_sememe_embedding(sem_embed_filename, hownet, scale=True) # load sememe embedding
hownet, wordsim_words = utils.fliter_wordsim_all(hownet) # remove MWEs in testset
train_num = len(hownet.comp_train)
pos_dict, word_remove = utils.load_hownet_pos()
hownet, cls_dict = utils.divide_data_with_pos(pos_dict, hownet)
print("number of dataset in training set:{}".format(len(hownet.comp_train)))
print("number of dataset in test set:{}".format(len(hownet.comp_test)))
print("number of dataset in dev set:{}".format(len(hownet.comp_dev)))
if not os.path.exists(logdir_name):
os.makedirs(logdir_name)
os.makedirs(os.path.join(logdir_name, 'print_files'))
os.makedirs(os.path.join(logdir_name, 'model_file'))
os.makedirs(os.path.join(logdir_name, 'tensorboard_logs'))
os.makedirs(os.path.join(logdir_name, 'example_files'))
# ################ Prepare Data ###################
# ################ Model and Run ###################
input_word_l = tf.placeholder(tf.int32, shape=[1], name='word_left')
input_word_r = tf.placeholder(tf.int32, shape=[1], name='word_right')
input_word_t = tf.placeholder(tf.int32, shape=[1], name='word_truth')
input_pos = tf.placeholder(tf.int32, shape=[1], name='pos_tag')
input_sememes_l = tf.placeholder(tf.int32, shape=[None], name='sememes_left')
input_sememes_r = tf.placeholder(tf.int32, shape=[None], name='sememes_right')
word_embedding = tf.Variable(tf.constant(0.0, shape=[word_embedding_np.shape[0], word_embedding_np.shape[1]]),trainable=False,name='word_embed')
embedding_placeholder = tf.placeholder(tf.float32, [word_embedding_np.shape[0], word_embedding_np.shape[1]])
embedding_init = word_embedding.assign(embedding_placeholder)
W_a = tf.Variable(tf.truncated_normal([dim, dim], stddev=0.5), tf.float32, name='W_a')
b_a = tf.Variable(tf.zeros([1, dim]), tf.float32, name='b_a')
U = tf.Variable(tf.truncated_normal([4, 2 * dim, trunc_num], stddev=0.5), tf.float32, name='U')
V = tf.Variable(tf.truncated_normal([4, dim, trunc_num], stddev=0.5), tf.float32, name='V')
U_i = tf.reshape(tf.nn.embedding_lookup(U, input_pos), [2 * dim, trunc_num])
V_i = tf.reshape(tf.nn.embedding_lookup(V, input_pos), [dim, trunc_num])
W_c_i = tf.matmul(U_i, V_i, transpose_b=True)
W_c_base = tf.Variable(tf.truncated_normal([2 * dim, dim], stddev=1.0), tf.float32, name='W_c_base')
b_c = tf.Variable(tf.zeros([1, dim]), tf.float32, name='b_c')
global_step = tf.Variable(0, trainable=False)
# regularizer
tf.add_to_collection('losses', tf.contrib.layers.l2_regularizer(labda)(W_a))
tf.add_to_collection('losses', tf.contrib.layers.l2_regularizer(labda)(U))
tf.add_to_collection('losses', tf.contrib.layers.l2_regularizer(labda)(V))
tf.add_to_collection('losses', tf.contrib.layers.l2_regularizer(labda)(W_c_base))
with tf.name_scope('word_embedding'):
embed_word_r = tf.nn.embedding_lookup(word_embedding, input_word_r)
embed_word_l = tf.nn.embedding_lookup(word_embedding, input_word_l)
embed_truth = tf.nn.embedding_lookup(word_embedding, input_word_t)
with tf.name_scope('sememe_embedding'):
sememe_embedding = tf.Variable(tf.constant(0.0, shape=[hownet.sem_num, dim]),trainable=True,name='Sememe_embeddings')
sememe_placeholder = tf.placeholder(tf.float32, [hownet.sem_num, dim])
sememe_init = sememe_embedding.assign(sememe_placeholder)
embed_sememe_l = tf.nn.embedding_lookup(sememe_embedding, input_sememes_l)
embed_sememe_r = tf.nn.embedding_lookup(sememe_embedding, input_sememes_r)
embed_sememe_r = utils.norm(embed_sememe_r)
embed_sememe_l = utils.norm(embed_sememe_l)
# 在att_no_wordvec基础上，把att_no_wordvec的聚合得到的词向量变成预训练的词向量，同时与word embedding一起使用
with tf.name_scope('attention_left'):
embed_word_align_r = tf.nn.tanh(tf.matmul(embed_word_r, W_a)+b_a, name='embed_align_r') # 1 * dim
att_l = tf.nn.softmax(tf.matmul(embed_sememe_l, tf.transpose(embed_word_align_r)), axis=0)
embed_sememe_l = att_l * embed_sememe_l
embed_aggre_word_l_pure = tf.reduce_sum(embed_sememe_l, axis=0, keepdims=True, name='embed_aggre_word_l_pure')
with tf.name_scope('attention_right'):
embed_word_align_l = tf.nn.tanh(tf.matmul(embed_word_l, W_a)+b_a, name='embed_align_l')
att_r = tf.nn.softmax(tf.matmul(embed_sememe_r, tf.transpose(embed_word_align_l)), axis=0)
embed_sememe_r = att_r * embed_sememe_r
embed_aggre_word_r_pure = tf.reduce_sum(embed_sememe_r, axis=0, keepdims=True, name='embed_aggre_word_r_pure')
with tf.name_scope('phrase_embedding'):
embed_word_whole = embed_word_r + embed_word_l
embed_sememe_whole = embed_aggre_word_r_pure + embed_aggre_word_l_pure
phrase_vec = tf.nn.tanh(tf.matmul(tf.concat([embed_word_whole, embed_sememe_whole], 1), W_c_i+W_c_base)+b_c, name="phrase_vec")
with tf.name_scope('loss'):
loss_pure = tf.reduce_mean((phrase_vec - embed_truth) ** 2)
tf.summary.scalar("loss pure", loss_pure)
tf.add_to_collection("losses", loss_pure)
cross_entropy_mean = tf.add_n(tf.get_collection("losses")) # 只是变量名叫cross_entropy_mean, 实际上是loss_all
tf.summary.scalar('loss_all', cross_entropy_mean)
lr = tf.train.exponential_decay(learning_rate, global_step=global_step, decay_steps=train_num/batch_size,decay_rate=lr_decay_rate, staircase=True)
tf.summary.scalar("learning_rate", lr)
opt = tf.train.GradientDescentOptimizer(learning_rate=lr)
grads_vars = opt.compute_gradients(cross_entropy_mean)
capped_grads_vars = [(tf.clip_by_value(grad, -5., 5.), var) for grad, var in grads_vars]
train_one_example = opt.apply_gradients(capped_grads_vars)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
sess.run(embedding_init, feed_dict={embedding_placeholder: word_embedding_np}) # initialize word embedding
sess.run(sememe_init, feed_dict={sememe_placeholder: sememe_embedding_np}) # initialize sememe embedding
merged = tf.summary.merge_all()
train_writer = tf.summary.FileWriter(logdir_name + '/tensorboard_logs', sess.graph) # writer of tensorboard
print_writer_filename = logdir_name + '/print_files/print.txt' # saver for printing
saver = tf.train.Saver(max_to_keep=1) # saver for saving model
# training
random.shuffle(hownet.comp_train)
for epoch in range(total_epoch):
example_writer_filename = logdir_name + '/example_files/epoch' + str(epoch + 1) + '.txt' # file for writing examples
loss_train = 0
loss_this_epoch = 0
print("\nEpoch:" + str(epoch + 1))
for current_num, train_tup in enumerate(hownet.comp_train):
total_num = epoch * train_num + current_num
batch_dict = utils.generate_one_example(hownet, train_tup)
_, summary_i, step_num, loss_i = sess.run(
[train_one_example, merged, global_step, cross_entropy_mean],
feed_dict={input_word_l: [batch_dict['wl']],
input_word_r: [batch_dict['wr']],
input_word_t: [batch_dict['lb']],
input_pos:[batch_dict['pos']],
input_sememes_l: batch_dict['sl'],
input_sememes_r: batch_dict['sr'], })
loss_this_epoch += loss_i
if current_num % 100 == 0:
sys.stdout.flush()
sys.stdout.write('\rTraining num: ' + str(current_num) + ' of ' + str(train_num) + ' loss:' + str(loss_this_epoch / (0.1 + current_num)))
train_writer.add_summary(summary=summary_i, global_step=step_num)
with open(print_writer_filename, 'a', encoding='utf-8') as fprint:
fprint.write('epoch: '+str(epoch+1)+' loss:'+str(loss_this_epoch/(0.1+len(hownet.comp_train)))+'\n')
saver.save(sess, logdir_name + '/model_file/model_ckpt', global_step=epoch + 1)
# dev
loss_dev = 0
for current_num, dev_tup in enumerate(hownet.comp_dev):
batch_dict = utils.generate_one_example(hownet, dev_tup)
loss_i = sess.run(cross_entropy_mean,
feed_dict={input_word_l: [batch_dict['wl']],
input_word_r: [batch_dict['wr']],
input_word_t: [batch_dict['lb']],
input_pos: [batch_dict['pos']],
input_sememes_l: batch_dict['sl'],
input_sememes_r: batch_dict['sr'], })
loss_dev += loss_i
sys.stdout.flush()
sys.stdout.write('\nDev set loss:' + str(loss_dev / (0.1 + len(hownet.comp_dev))) + '\n')
with open(print_writer_filename, 'a', encoding='utf-8') as fprint:
fprint.write('Dev set loss:' + str(loss_dev / (0.1 + len(hownet.comp_dev))) + '\n')
# test MWE similarity: write embedding to phrase_vec_file
number = 0
phrase_vec_file = os.path.join(logdir_name, 'example_files', 'phrase_vector_epoch.txt')
for current_num, test_tup in enumerate(hownet.comp_test):
if test_tup[4] in wordsim_words:
batch_test = utils.generate_one_example(hownet, test_tup)
phrase_vector = sess.run(phrase_vec, feed_dict={input_word_l: [batch_test['wl']],
input_word_r: [batch_test['wr']],
input_word_t: [batch_test['lb']],
input_pos: [batch_test['pos']],
input_sememes_l: batch_test['sl'],
input_sememes_r: batch_test['sr'], })
with open(phrase_vec_file, 'a', encoding='utf-8') as f_phrase_embed:
f_phrase_embed.write(test_tup[4] + ' ')
phrase_vector = phrase_vector.tolist()[0]
phrase_vector = [str(vec) for vec in phrase_vector]
f_phrase_embed.write(' '.join(phrase_vector))
f_phrase_embed.write('\n')
number += 1
print('Have written {} words to phrase_vector.txt'.format(number))
with open(print_writer_filename, 'a', encoding='utf-8') as fprint:
fprint.write('Have written {} words to phrase_vector.txt'.format(number))
train_writer.close()