-
Notifications
You must be signed in to change notification settings - Fork 7
/
train_MLBiNet.py
337 lines (288 loc) · 17.7 KB
/
train_MLBiNet.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
#!/usr/bin/env python
#-*- coding: utf-8 -*-
import os
import time
import json
import random
import numpy as np
import tensorflow as tf
tf.flags.DEFINE_integer('encode_h', 100, 'dim of encoding layer')
tf.flags.DEFINE_integer('decode_h', 200, 'dim of decoding layer')
tf.flags.DEFINE_integer('tag_dim', 100, 'dimension of tags')
tf.flags.DEFINE_integer('event_info_h', 100, 'hidden size of sentence level information aggregation layer')
tf.flags.DEFINE_integer('batch_size', 64, 'batch size')
tf.flags.DEFINE_integer('max_doc_len', 8, 'max number of sentences in a document')
tf.flags.DEFINE_integer('max_seq_len', 50, 'maximum length of sequence')
tf.flags.DEFINE_integer('num_tag_layers', 2, 'number of tagging layers')
tf.flags.DEFINE_integer('reverse_seq', 1, 'decoder mechanism')
tf.flags.DEFINE_string('tagging_mechanism', "backward_decoder", 'decoder mechanism')
tf.flags.DEFINE_integer('ner_dim_1', 20, 'embedding size of level-1 NER')
tf.flags.DEFINE_integer('ner_dim_2', 20, 'embedding size of level-2 NER')
tf.flags.DEFINE_integer('self_att_not', 1, 'self attention or not')
tf.flags.DEFINE_integer('context_info', 1,
'0: single sentence information, 1: information of two neighbor sentences')
tf.flags.DEFINE_float('penalty_coef', 2e-5, 'penalty coefficient')
tf.flags.DEFINE_float('event_vector_trans', 1, 'event_vector_trans')
tf.flags.DEFINE_integer('num_epochs', 50, 'Number of epoches')
tf.flags.DEFINE_integer('eval_every_steps', 100, 'Number of epoches')
tf.flags.DEFINE_integer('num_epochs_warm', 0, 'Number of epoches of warm start')
tf.flags.DEFINE_integer('nconsect_epoch', 3, 'early stopping epoches')
tf.flags.DEFINE_float('weight_decay', 1, 'truncation of event attention weights')
tf.flags.DEFINE_float('warm_learning_rate', 1e-5, 'warm-up learning rate')
tf.flags.DEFINE_float('learning_rate', 5e-4, 'learning rate')
tf.flags.DEFINE_float('decay_rate', 0.99, 'decay rate')
tf.flags.DEFINE_float('dropout_rate', 0.5, 'dropout rate')
tf.flags.DEFINE_float('grad_clip', 10, 'grad clip to prevent gradient exlode')
tf.flags.DEFINE_float('positive_weights', 1, 'weights for positive sample')
tf.flags.DEFINE_string('train_file', './data-ACE/example_new.train', 'train file')
tf.flags.DEFINE_string('dev_file', './data-ACE/example_new.dev', 'dev file')
tf.flags.DEFINE_string('test_file', './data-ACE/example_new.test', 'test file')
tf.flags.DEFINE_string('embedding_file','./embedding/embeddings.txt','pretrained embedding file')
tf.flags.DEFINE_integer('word_emb_dim', 100, 'word embedding size')
tf.flags.DEFINE_string('NER_dict_file', './dict/event_types.txt', 'ner dict file')
tf.flags.DEFINE_string('ner_1_dict_file', './dict/ner_1.txt', 'level-1 ner dict file')
tf.flags.DEFINE_string('ner_2_dict_file', './dict/ner_2.txt', 'level-2 ner dict file')
FLAGS = tf.flags.FLAGS
lower_case = False
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
config_gpu = tf.ConfigProto()
config_gpu.gpu_options.per_process_gpu_memory_fraction = 0.6
def train(seed_id=1):
# set seed
tf.set_random_seed(seed_id)
from MLBiNet import MLBiNet
from utils_init import load_ED_data
from utils_init import data_transformation_doc
from utils_init import batch_generation_doc
from utils_init import load_vocab
from utils_init import load_pretrain
from ace_model_evaluation import write_2_file, ace_pred_result_stat
with tf.Graph().as_default() as g:
# loading the embedding matrix
embedding_matrix, vocab_words, vocab_2_id, id_2_vocab = load_pretrain(FLAGS.embedding_file,
FLAGS.word_emb_dim)
print('shape of embedding_matrix is:', np.asmatrix(embedding_matrix).shape)
# load train, dev, test data
sents_train, ners_train, ner_vocab, ner_1_train, ner_2_train, doc_file_to_sents_train = \
load_ED_data(FLAGS.train_file, lower_case=lower_case)
# load the vocab of event type
_, ED_2_id = load_vocab(FLAGS.NER_dict_file)
print("ner_2_id is:\t", ED_2_id)
sents_dev, ners_dev, _, ner_1_dev, ner_2_dev, doc_file_to_sents_dev = \
load_ED_data(FLAGS.dev_file, lower_case=lower_case)
sents_test, ners_test, _, ner_1_test, ner_2_test, doc_file_to_sents_test = \
load_ED_data(FLAGS.test_file, lower_case=lower_case)
print("load_ner_data finished!")
print("doc_file_to_sents_test:\t", doc_file_to_sents_test)
# load NER label
ner_vocab_1, ner_to_id_1 = load_vocab(FLAGS.ner_1_dict_file)
ner_vocab_2, ner_to_id_2 = load_vocab(FLAGS.ner_2_dict_file)
print("NER vocab loaded!")
# encoding the train, dev, test data
encode_train = data_transformation_doc(sents_train, ner_1_train, ner_2_train, ners_train,
vocab_2_id, ED_2_id, vocab_2_id['<UNK>'], ner_to_id_1, ner_to_id_2)
encode_dev = data_transformation_doc(sents_dev, ner_1_dev, ner_2_dev, ners_dev,
vocab_2_id, ED_2_id, vocab_2_id['<UNK>'], ner_to_id_1, ner_to_id_2)
encode_test = data_transformation_doc(sents_test, ner_1_test, ner_2_test, ners_test,
vocab_2_id, ED_2_id, vocab_2_id['<UNK>'], ner_to_id_1, ner_to_id_2)
print("Document data transformation finished!")
# batch generating
train_batches = batch_generation_doc(doc_file_to_sents_train, encode_train, FLAGS.batch_size, FLAGS.max_doc_len,
FLAGS.max_seq_len, vocab_2_id, ED_2_id, num_epoches=FLAGS.num_epochs)
dev_batches = batch_generation_doc(doc_file_to_sents_dev, encode_dev, FLAGS.batch_size, FLAGS.max_doc_len,
FLAGS.max_seq_len, vocab_2_id, ED_2_id, num_epoches=1)
test_batches = batch_generation_doc(doc_file_to_sents_test, encode_test, FLAGS.batch_size, FLAGS.max_doc_len,
FLAGS.max_seq_len, vocab_2_id, ED_2_id, num_epoches=1)
print("batch_generation_doc finished!")
print('Begin model initialization!')
with tf.Session(config=config_gpu) as sess:
model = MLBiNet(
encode_h = FLAGS.encode_h,
decode_h = FLAGS.decode_h,
tag_dim = FLAGS.tag_dim,
event_info_h = FLAGS.event_info_h,
word_emb_mat = np.array(embedding_matrix),
batch_size = FLAGS.batch_size,
max_doc_len = FLAGS.max_doc_len,
max_seq_len = FLAGS.max_seq_len,
id_O = ED_2_id['O'],
num_tag_layers = FLAGS.num_tag_layers,
weight_decay = FLAGS.weight_decay,
reverse_seq = FLAGS.reverse_seq,
class_size = len(ED_2_id),
tagging_mechanism = FLAGS.tagging_mechanism,
ner_size_1 = len(ner_to_id_1),
ner_dim_1 = FLAGS.ner_dim_1,
ner_size_2 = len(ner_to_id_2),
ner_dim_2 = FLAGS.ner_dim_2,
self_att_not = FLAGS.self_att_not,
context_info = FLAGS.context_info,
event_vector_trans = FLAGS.event_vector_trans
)
print('encoder-decoder model initialized!')
loss_ed = model.loss
for tvarsi in tf.trainable_variables():
if tvarsi.name != 'word_emb_mat:0':
loss_ed += FLAGS.penalty_coef * tf.reduce_sum(tvarsi ** 2)
else:
print("\n\n{} is not penalied!\n\n".format(tvarsi))
with tf.name_scope('accuracy'):
label_pred_naive = model.label_pred
label_pred = model.label_pred
label_true = model.label_true
acc_cnt_naive = tf.reduce_sum(tf.cast(tf.equal(label_pred_naive,label_true),dtype=tf.float32))
acc_cnt = tf.reduce_sum(tf.cast(tf.equal(label_pred,label_true),dtype=tf.float32))
cnt_all = tf.reduce_sum(tf.cast(tf.greater(label_true,-1),dtype=tf.float32))
acc_rate = acc_cnt / cnt_all
valid_len_final = model.valid_len_list
timestamp = str(int(time.time()))
out_dir = os.path.join('./runs', timestamp)
checkpoint_dir = os.path.join(out_dir, "checkpoints")
if not os.path.exists(checkpoint_dir):
os.makedirs(checkpoint_dir)
param_dict = FLAGS.flag_values_dict()
param_dict['lower_case'] = lower_case
with open(os.path.join(checkpoint_dir,'config.json'), "w") as f:
f.write(json.dumps(param_dict, indent=2, ensure_ascii=False))
tvars = tf.trainable_variables()
for kk, tvarsi in enumerate(tvars):
print('The %d-th tvars is %s' % (kk, tvarsi))
global_step = tf.Variable(0, trainable=False)
learning_rate = tf.train.exponential_decay(learning_rate=FLAGS.learning_rate,
global_step=global_step,
decay_steps=len(train_batches) // int(FLAGS.num_epochs),
decay_rate=FLAGS.decay_rate)
tvars_no_emb = [x for x in tvars if 'word_emb_mat' not in x.name]
opt_ed_NO_emb_sent = tf.train.AdamOptimizer(learning_rate)
#
grads_trig_sent_NO_EMB, _ = tf.clip_by_global_norm(tf.gradients(loss_ed, tvars_no_emb), FLAGS.grad_clip)
grads_and_vars_trig_sent_NO_EMB = tuple(zip(grads_trig_sent_NO_EMB, tvars_no_emb))
train_ed_NO_emb = opt_ed_NO_emb_sent.apply_gradients(grads_and_vars_trig_sent_NO_EMB, global_step=global_step)
sess.run(tf.global_variables_initializer())
def train_step(train_batch,epoch):
positive_weights = FLAGS.positive_weights
feed_dict = {
model.dropout_rate: FLAGS.dropout_rate,
model.input_docs: np.array(train_batch[0]),
model.ner_docs_1: np.array(train_batch[1]),
model.ner_docs_2: np.array(train_batch[2]),
model.input_label_docs: np.array(train_batch[3]),
model.valid_batch: train_batch[4],
model.valid_sent_len: np.array(train_batch[5]),
model.valid_words_len: np.array(train_batch[6]),
model.positive_weights: positive_weights
}
_, loss_trigger_tmp, acc_rate_tmp, step_curr = sess.run([train_ed_NO_emb, loss_ed, acc_rate, global_step],
feed_dict)
return loss_trigger_tmp,step_curr,acc_rate_tmp
def dev_test_step(dev_batches):
def dev_ont_batch(dev_batch):
feed_dict = {
model.dropout_rate: 0,
model.input_docs: np.array(dev_batch[0]),
model.ner_docs_1: np.array(dev_batch[1]),
model.ner_docs_2: np.array(dev_batch[2]),
model.input_label_docs: np.array(dev_batch[3]),
model.valid_batch: dev_batch[4],
model.valid_sent_len: np.array(dev_batch[5]),
model.valid_words_len: np.array(dev_batch[6]),
model.positive_weights: 1.0
}
acc_cnt_tmp, cnt_all_tmp,acc_cnt_naive_tmp,valid_len_tmp,\
label_pred_tmp, label_pred_naive_tmp, label_true_tmp, final_words_id_tmp, loss_tmp \
= sess.run([acc_cnt,cnt_all,acc_cnt_naive,valid_len_final,label_pred, label_pred_naive,
label_true,model.final_words_id,loss_ed], feed_dict)
return acc_cnt_tmp, cnt_all_tmp,acc_cnt_naive_tmp,valid_len_tmp,label_pred_tmp, \
label_pred_naive_tmp, label_true_tmp,final_words_id_tmp, loss_tmp
acc_cnt_list, cnt_all_list = [], []
acc_cnt_naive_list, cnt_all_naive_list = [], []
label_pred_list, label_pred_naive_list, label_true_list = [],[],[]
valid_len_list = []
words_sents = []
loss_dev_test = 0
len_seq_all = 0
for dev_batchi in dev_batches:
acc_cnt_tmp, cnt_all_tmp,acc_cnt_naive_tmp,valid_len_tmp,\
label_pred_tmp, label_pred_naive_tmp, label_true_tmp,final_words_id_tmp, loss_tmp_i\
= dev_ont_batch(dev_batchi)
acc_cnt_list.append(acc_cnt_tmp)
cnt_all_list.append(cnt_all_tmp)
acc_cnt_naive_list.append(acc_cnt_naive_tmp)
label_pred_list.extend(label_pred_tmp)
label_pred_naive_list.extend(label_pred_naive_tmp)
label_true_list.extend(label_true_tmp)
valid_len_list.extend(valid_len_tmp)
words_sents.extend(final_words_id_tmp)
loss_dev_test += loss_tmp_i * len(label_pred_naive_tmp)
len_seq_all += len(label_pred_naive_tmp)
loss_dev_test = loss_dev_test / (len_seq_all + 1e-8)
prec_dev = sum(acc_cnt_list) / sum(cnt_all_list)
prec_dev_naive = sum(acc_cnt_naive_list) / sum(cnt_all_list)
return prec_dev,prec_dev_naive,words_sents,label_pred_list,\
label_true_list,valid_len_list,loss_dev_test
print('Total train batch is:\t',len(train_batches),flush=True)
prec_test_best = 0
loss_dev_best = 10000
loss_dev_second = 10000
loss_dev_list = []
nconsect = 0
print("total train steps:\t", len(train_batches))
for i, train_batchi in enumerate(train_batches):
epoch = i // FLAGS.eval_every_steps
loss_trigger_tmp, step_curr, acc_rate_tmp = train_step(train_batchi,0)
if i % 1e1 == 0:
print('epoch {}, step: {},loss: {},acc_rate: {}'.format(
epoch,step_curr,loss_trigger_tmp,acc_rate_tmp), flush=True)
if i % FLAGS.eval_every_steps == 0 or i == len(train_batches) - 1:
prec_dev,prec_dev_naive,words_sents,label_pred_list,\
label_true_list,valid_len_list,loss_dev_ = dev_test_step(dev_batches)
print('epoch {} prec_dev is: \n'.format(epoch), prec_dev, flush=True)
if epoch == 0:
os.makedirs(os.path.join(checkpoint_dir, 'dev'))
filename_dev = os.path.join(checkpoint_dir, 'dev/test_result_{}.txt').format(step_curr)
write_2_file(filename_dev, ED_2_id, label_true_list,valid_len_list,
words_sents, label_pred_list, id_2_vocab)
prec_event_dev, recall_event_dev, f1_event_dev = ace_pred_result_stat(filename_dev)
print('epoch: {}, loss_dev_: {}'.format(epoch, loss_dev_), flush=True)
print('epoch: {}, prec_event_dev: {}, recall_event_dev: {}, f1_event_dev: {}'.format(
epoch, prec_event_dev, recall_event_dev, f1_event_dev), flush=True)
loss_dev_list.append(loss_dev_)
loss_dev_list = sorted(loss_dev_list,key=lambda x: x, reverse=False)
if len(loss_dev_list) > 2:
loss_dev_second = loss_dev_list[2]
if loss_dev_ > loss_dev_best:
if loss_dev_ > loss_dev_second:
nconsect += 1
else:
nconsect = 0
else:
nconsect = 0
loss_dev_best = loss_dev_
print('\n')
prec_test,prec_test_naive,words_sents,label_pred_list,\
label_true_list,valid_len_list, loss_test_ = dev_test_step(test_batches)
print('epoch {} prec_test is: \n'.format(epoch), prec_test, flush=True)
print('\n')
# write to file
if epoch == 0:
os.makedirs(os.path.join(checkpoint_dir, 'test'))
filename_test = os.path.join(checkpoint_dir, 'test/test_result_{}.txt').format(step_curr)
write_2_file(filename_test, ED_2_id, label_true_list,valid_len_list,
words_sents, label_pred_list, id_2_vocab)
prec_event_test, recall_event_test, f1_event_test = ace_pred_result_stat(filename_test)
print('epoch: {}, prec_event_test: {}, recall_event_test: {}, f1_event_test:{}'.format(
epoch, prec_event_test, recall_event_test, f1_event_test), flush=True)
if prec_test_best < f1_event_test:
prec_test_best = f1_event_test
print('The best dev loss value is:\t', [loss_dev_best,nconsect])
# print('The best dev f1 value is:\t', [prec_dev_best,nconsect])
print('The best test f1 value is:\t', prec_test_best)
with open(os.path.join(checkpoint_dir, 'test_result.txt'), encoding='utf-8', mode='a') as f:
f.write('\t'.join([str(epoch), str(prec_event_test),str(recall_event_test),
str(f1_event_test), str(loss_dev_best), str(loss_dev_second), str(nconsect)]) + '\n')
if nconsect >= FLAGS.nconsect_epoch:
break
tf.reset_default_graph()
if __name__ == "__main__":
# train()
pass