-
Notifications
You must be signed in to change notification settings - Fork 29
/
train.py
174 lines (151 loc) · 6.82 KB
/
train.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
#!/usr/bin/env python
# -*- encoding:utf-8 -*-
"""
Training part of the Deep QA model
"""
from __future__ import print_function
from __future__ import division
__copyright__ = "Copyright (c) 2017 Xuming Lin. All Rights Reserved"
__author__ = "Xuming Lin, Hai Liang Wang<[email protected]>"
__date__ = "2017-08-21:23:42:17"
import os
curdir = os.path.dirname(os.path.abspath(__file__))
from qacnn import QACNN
from sklearn import metrics
from tqdm import tqdm
import tensorflow as tf
import datetime
import operator
import data
flags, FLAGS = tf.app.flags, tf.app.flags.FLAGS
flags.DEFINE_integer('sequence_length', 100, 'sequence length') # noqa: skipped autopep8 checking
flags.DEFINE_integer('evaluate_every', 1, 'evaluate every N steps') # noqa: skipped autopep8 checking
flags.DEFINE_integer('num_epochs', 300, 'epochs') # noqa: skipped autopep8 checking
flags.DEFINE_integer('batch_size', 100, 'min batch size') # noqa: skipped autopep8 checking
flags.DEFINE_integer('embedding_size', 50, 'embedding size') # noqa: skipped autopep8 checking
flags.DEFINE_integer('hidden_size', 80, 'hidden size') # noqa: skipped autopep8 checking
flags.DEFINE_integer('num_filters', 512, 'number of filters') # noqa: skipped autopep8 checking
flags.DEFINE_float('l2_reg_lambda', 0., 'L2 regularization factor') # noqa: skipped autopep8 checking
flags.DEFINE_float('keep_prob', 1.0, 'Dropout keep rate') # noqa: skipped autopep8 checking
flags.DEFINE_float('lr', 0.001, 'learning rate') # noqa: skipped autopep8 checking
flags.DEFINE_float('margin', 0.05, 'margin for computing loss') # noqa: skipped autopep8 checking
# Config函数
class Config(object):
def __init__(self, vocab_size):
# 输入序列(句子)长度
self.sequence_length = FLAGS.sequence_length
# 循环数
self.num_epochs = FLAGS.num_epochs
# batch大小
self.batch_size = FLAGS.batch_size
# 词表大小
self.vocab_size = vocab_size
# 词向量大小
self.embedding_size = FLAGS.embedding_size
# 不同类型的filter,相当于1-gram,2-gram,3-gram和5-gram
self.filter_sizes = [1, 2, 3, 5]
# 隐层大小
self.hidden_size = FLAGS.hidden_size
# 每种filter的数量
self.num_filters = FLAGS.num_filters
# 论文里给的是0.0001
self.l2_reg_lambda = FLAGS.l2_reg_lambda
# dropout
self.keep_prob = FLAGS.keep_prob
# 学习率
# 论文里给的是0.01
self.lr = FLAGS.lr
# margin
# 论文里给的是0.009
self.m = FLAGS.margin
# 设定GPU的性质,允许将不能在GPU上处理的部分放到CPU
# 设置log打印
self.cf = tf.ConfigProto(allow_soft_placement=True, log_device_placement=False)
'''
GPU内存使用策略
'''
# 自动增长
self.cf.gpu_options.allow_growth=True
# 只占用20%的GPU内存
# self.cf.gpu_options.per_process_gpu_memory_fraction = 0.2
self.test_data = data.load_test(self.sequence_length, self.sequence_length)
print('Loading Data...')
# 词映射ID
vocab = data.vocab_data
# 配置文件
config = Config(len(vocab['word2id']))
def main(unused_argv):
'''
开始训练和测试
'''
with tf.device('/gpu:0'), tf.Session(config=config.cf) as sess:
# 建立CNN网络
cnn = QACNN(config, sess)
# 保存Metrics数据
tf_writer = tf.summary.FileWriter(logdir=os.path.join(curdir, 'sdist/'), graph=sess.graph)
# Summaries for loss and accuracy during training
summary_loss = tf.summary.scalar("train/loss", cnn.loss)
summary_accu = tf.summary.scalar("train/accuracy", cnn.accu)
summary_op = tf.summary.merge([summary_loss, summary_accu])
# 训练函数
def train_step(x_batch_1, x_batch_2, x_batch_3):
feed_dict = {
cnn.q: x_batch_1,
cnn.aplus: x_batch_2,
cnn.aminus: x_batch_3,
cnn.keep_prob: config.keep_prob
}
_, step, loss, accuracy, summaries = sess.run(
[cnn.train_op, cnn.global_step, cnn.loss, cnn.accu, summary_op],
feed_dict)
tf_writer.add_summary(summaries, step)
time_str = datetime.datetime.now().isoformat()
print("{}: step {}, loss {:g}, acc {:g}".format(time_str, step, loss, accuracy))
return time_str, step, loss, accuracy
# 测试函数
def dev_step(step):
# 混淆矩阵建立评估
# http://www.uta.fi/sis/tie/tl/index/Rates.pdf
quality = {'tp': 0, 'tn': 0, 'fp': 0, 'fn': 0}
losses = []
labels = []
scores = []
pbar = tqdm(config.test_data)
pbar.set_description("evaluate step %s" % step)
for x in pbar:
_, loss, score = cnn.predict(dict({
'question': x[1],
'utterance': x[2]
}), x[3])
scores.append(score)
losses.append(loss)
labels.append(x[3])
# 使用Roc Curve生成Threshold
# http://alexkong.net/2013/06/introduction-to-auc-and-roc/
fpr, tpr, th = metrics.roc_curve(labels, scores)
threshold = round(metrics.auc(fpr, tpr), 5)
if score >= threshold and x[3]==1:
quality['tp'] += 1
elif score >= threshold and x[3]==0:
quality['fp'] += 1
elif score < threshold and x[3]==1:
quality['fn'] += 1
else:
quality['tn'] += 1
accuracy = float(quality['tp'] + quality['tn'] )/(quality['tp'] + quality['tn'] + quality['fp'] + quality['fn'])
loss = tf.reduce_mean(losses).eval()
tf_writer.add_summary(tf.Summary(value=[
tf.Summary.Value(tag="evaluate/loss", simple_value=loss),
tf.Summary.Value(tag="evaluate/accuracy", simple_value=accuracy)]), step)
print('evaluation @ step %d: 准确率: %d, 损失函数: %s, threshold: %d' % (step, accuracy, loss, threshold))
# 每500步测试一下
# 开始训练和测试
sess.run(tf.global_variables_initializer())
for i in range(config.num_epochs):
for (_, x_question, x_utterance, y) in data.load_train(config.batch_size, config.sequence_length, config.sequence_length):
if len(_) == config.batch_size: # 在epoch的最后一个mini batch中，数据条数可能不等于 batch_size
_, global_step, _, _ = train_step(x_question, x_utterance, y)
if global_step % FLAGS.evaluate_every == 0:
dev_step(global_step)
if __name__ == '__main__':
tf.app.run()