forked from wavewangyue/ner
-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_multitask_lstm.py
190 lines (155 loc) · 7.33 KB
/
train_multitask_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
import logging
import tensorflow as tf
import numpy as np
import os
from model_multitask_lstm import MyModel
from utils import DataProcessor_MTL_LSTM as DataProcessor
from utils import load_vocabulary
from utils import extract_kvpairs_in_bioes
from utils import cal_f1_score
# set logging
log_file_path = "./ckpt/run.log"
if os.path.exists(log_file_path): os.remove(log_file_path)
logger = logging.getLogger()
logger.setLevel(logging.INFO)
formatter = logging.Formatter("%(asctime)s | %(message)s", "%Y-%m-%d %H:%M:%S")
chlr = logging.StreamHandler()
chlr.setFormatter(formatter)
fhlr = logging.FileHandler(log_file_path)
fhlr.setFormatter(formatter)
logger.addHandler(chlr)
logger.addHandler(fhlr)
logger.info("loading vocab...")
w2i_char, i2w_char = load_vocabulary("./data/vocab_char.txt")
w2i_bio, i2w_bio = load_vocabulary("./data/vocab_bio.txt")
w2i_attr, i2w_attr = load_vocabulary("./data/vocab_attr.txt")
logger.info("loading data...")
data_processor_train = DataProcessor(
"./data/train/input.seq.char",
"./data/train/output.seq.bio",
"./data/train/output.seq.attr",
w2i_char,
w2i_bio,
w2i_attr,
shuffling=True
)
data_processor_valid = DataProcessor(
"./data/test/input.seq.char",
"./data/test/output.seq.bio",
"./data/test/output.seq.attr",
w2i_char,
w2i_bio,
w2i_attr,
shuffling=True
)
logger.info("building model...")
model = MyModel(embedding_dim=300,
hidden_dim=300,
vocab_size_char=len(w2i_char),
vocab_size_bio=len(w2i_bio),
vocab_size_attr=len(w2i_attr),
O_tag_index=w2i_bio["O"],
use_crf=False)
logger.info("model params:")
params_num_all = 0
for variable in tf.trainable_variables():
params_num = 1
for dim in variable.shape:
params_num *= dim
params_num_all += params_num
logger.info("\t {} {} {}".format(variable.name, variable.shape, params_num))
logger.info("all params num: " + str(params_num_all))
logger.info("start training...")
tf_config = tf.ConfigProto(allow_soft_placement=True)
tf_config.gpu_options.allow_growth = True
with tf.Session(config=tf_config) as sess:
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver(max_to_keep=50)
epoches = 0
losses = []
batches = 0
best_f1 = 0
batch_size = 32
while epoches < 20:
(inputs_seq_batch,
inputs_seq_len_batch,
outputs_seq_bio_batch,
outputs_seq_attr_batch) = data_processor_train.get_batch(batch_size)
feed_dict = {
model.inputs_seq: inputs_seq_batch,
model.inputs_seq_len: inputs_seq_len_batch,
model.outputs_seq_bio: outputs_seq_bio_batch,
model.outputs_seq_attr: outputs_seq_attr_batch
}
if batches == 0:
logger.info("###### shape of a batch #######")
logger.info("input_seq: " + str(inputs_seq_batch.shape))
logger.info("input_seq_len: " + str(inputs_seq_len_batch.shape))
logger.info("output_seq_bio: " + str(outputs_seq_bio_batch.shape))
logger.info("output_seq_attr: " + str(outputs_seq_attr_batch.shape))
logger.info("###### preview a sample #######")
logger.info("input_seq:" + " ".join([i2w_char[i] for i in inputs_seq_batch[0]]))
logger.info("input_seq_len :" + str(inputs_seq_len_batch[0]))
logger.info("output_seq_bio: " + " ".join([i2w_bio[i] for i in outputs_seq_bio_batch[0]]))
logger.info("output_seq_attr: " + " ".join([i2w_attr[i] for i in outputs_seq_attr_batch[0]]))
logger.info("###############################")
loss, _ = sess.run([model.loss, model.train_op], feed_dict)
losses.append(loss)
batches += 1
if data_processor_train.end_flag:
data_processor_train.refresh()
epoches += 1
def valid(data_processor, max_batches=None, batch_size=1024):
preds_kvpair = []
golds_kvpair = []
batches_sample = 0
while True:
(inputs_seq_batch,
inputs_seq_len_batch,
outputs_seq_bio_batch,
outputs_seq_attr_batch) = data_processor.get_batch(batch_size)
feed_dict = {
model.inputs_seq: inputs_seq_batch,
model.inputs_seq_len: inputs_seq_len_batch,
model.outputs_seq_bio: outputs_seq_bio_batch,
model.outputs_seq_attr: outputs_seq_attr_batch
}
preds_seq_bio_batch, preds_seq_attr_batch = sess.run(model.outputs, feed_dict)
for pred_seq_bio, gold_seq_bio, pred_seq_attr, gold_seq_attr, input_seq, l in zip(preds_seq_bio_batch,
outputs_seq_bio_batch,
preds_seq_attr_batch,
outputs_seq_attr_batch,
inputs_seq_batch,
inputs_seq_len_batch):
pred_seq_bio = [i2w_bio[i] for i in pred_seq_bio[:l]]
gold_seq_bio = [i2w_bio[i] for i in gold_seq_bio[:l]]
char_seq = [i2w_char[i] for i in input_seq[:l]]
pred_seq_attr = [i2w_attr[i] for i in pred_seq_attr[:l]]
gold_seq_attr = [i2w_attr[i] for i in gold_seq_attr[:l]]
pred_kvpair = extract_kvpairs_in_bioes(pred_seq_bio, char_seq, pred_seq_attr)
gold_kvpair = extract_kvpairs_in_bioes(gold_seq_bio, char_seq, gold_seq_attr)
preds_kvpair.append(pred_kvpair)
golds_kvpair.append(gold_kvpair)
if data_processor.end_flag:
data_processor.refresh()
break
batches_sample += 1
if (max_batches is not None) and (batches_sample >= max_batches):
break
p, r, f1 = cal_f1_score(preds_kvpair, golds_kvpair)
logger.info("Valid Samples: {}".format(len(preds_kvpair)))
logger.info("Valid P/R/F1: {} / {} / {}".format(round(p*100, 2), round(r*100, 2), round(f1*100, 2)))
return (p, r, f1)
if batches % 100 == 0:
logger.info("")
logger.info("Epoches: {}".format(epoches))
logger.info("Batches: {}".format(batches))
logger.info("Loss: {}".format(sum(losses) / len(losses)))
losses = []
ckpt_save_path = "./ckpt/model.ckpt.batch{}".format(batches)
logger.info("Path of ckpt: {}".format(ckpt_save_path))
saver.save(sess, ckpt_save_path)
p, r, f1 = valid(data_processor_valid, max_batches=10)
if f1 > best_f1:
best_f1 = f1
logger.info("############# best performance now here ###############")