-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain.py
163 lines (134 loc) · 7.77 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
# -*- coding: utf-8 -*-
import os
import sys
import logging
from keras.backend.tensorflow_backend import set_session
logging.basicConfig(level=logging.INFO,
format='%(asctime)s %(filename)s[line:%(lineno)d] %(levelname)s %(message)s',
datefmt='%a, %d %b %Y %H:%M:%S')
from model import *
from dataset.data_load import *
def init_dataset(meta_files):
"""
初始化dataset
:param meta_files:[list],The list contains the meta file
"""
train_dataset, train_num_batch, dev_dataset, dev_num_batch = get_dataset(meta_files)
train_iterator = train_dataset.make_initializable_iterator()
train_next_element = train_iterator.get_next()
dev_iterator = dev_dataset.make_initializable_iterator()
dev_next_element = dev_iterator.get_next()
return train_next_element, train_num_batch, train_iterator, dev_next_element, dev_num_batch, dev_iterator
if __name__ == '__main__':
if not 'win' in sys.platform:
import setproctitle
setproctitle.setproctitle(pm.process_name)
# 限定使用的gpu和使用gpu的内存
os.environ["CUDA_VISIBLE_DEVICES"] = pm.gpu_devices
tf_config = tf.ConfigProto(allow_soft_placement=True, log_device_placement=pm.log_device_placement)
tf_config.gpu_options.per_process_gpu_memory_fraction = pm.gpu_memory_rate
with tf.device('/cpu:0'):
with tf.Graph().as_default():
# 创建多GPU的图
start_time = time.time()
svfObj = SvfOjb(is_training=True)
svfObj.build_model()
logging.info("Build model done! Duration time:{}s".format(time.time() - start_time))
# 创建session
sess = tf.Session(config=tf_config)
set_session(session=sess)
# 获取训练，验证数据集
logging.info("Get dataset....")
start_time = time.time()
train_next_element, train_num_batch, train_iterator, dev_next_element, dev_num_batch, dev_iterator = init_dataset(
pm.train_meta_files.split(","))
# 模型保存
train_summary_writer = tf.summary.FileWriter(pm.log_dir + os.sep + 'train', sess.graph)
# 保存dev的模型
dev_model_save_dir = pm.checkpoint + os.sep + 'dev'
Path(dev_model_save_dir).mkdir(exist_ok=True)
dev_summary_writer = tf.summary.FileWriter(pm.log_dir + os.sep + 'dev', sess.graph)
# 保存效果最好的dev模型
best_dev_model_save_dir = pm.checkpoint + os.sep + 'min_dev'
Path(best_dev_model_save_dir).mkdir(exist_ok=True)
# 保存模型
saver = tf.train.Saver(tf.global_variables(), max_to_keep=pm.max_keep_model)
try:
model_file = tf.train.latest_checkpoint(best_dev_model_save_dir)
saver.restore(sess, model_file)
logging.info('已加载最近训练模型!')
except Exception as e:
logging.info("无法加载旧模型,训练过程会覆盖旧模型!")
pass
if not os.path.exists(pm.checkpoint):
os.mkdir(pm.checkpoint)
sess.run(tf.global_variables_initializer())
sess.run(train_iterator.initializer)
sess.run(dev_iterator.initializer)
train_loss_cl = []
lr_cl = []
min_dev_eer = 1e8
for epoch in range(1, pm.epochs + 1):
# 训练
loss_cl = []
for step in range(train_num_batch):
feature = sess.run(train_next_element)
sim_matrix, loss, st, lr, _, summary = sess.run([svfObj.sim_matrix,
svfObj.loss,
svfObj.global_step,
svfObj.lr,
svfObj.train_op,
svfObj.summary_op],
feed_dict={svfObj.inpt: feature})
loss_cl.append(loss)
logging.info(
'train--step:{}({}) global_step:{} loss:{}'.format(step, train_num_batch, int(st), loss))
train_loss_cl.append(sum(loss_cl) / len(loss_cl))
lr_cl.append(lr)
eer = calculate_eer(sim_matrix)
train_summary_writer.add_summary(summary, epoch)
logging.info('------------------------------------------------------------------------------')
logging.info("train epoch:{} gs:{}".format(epoch, int(st)))
logging.info("lr:{}".format(lr))
logging.info("avg_loss:{}".format(sum(loss_cl) / len(loss_cl)))
logging.info("last batch eer:{}".format(eer))
logging.info('------------------------------------------------------------------------------')
logging.info("------------------------------------------------------------------------------")
if pm.display_loss:
for ls, lr in zip(train_loss_cl, lr_cl):
logging.info("lr:{} loss:{}".format(lr, ls))
logging.info("------------------------------------------------------------------------------")
loss_cl.clear()
if epoch > 0 and epoch % 1 == 0:
# 验证集
eer_cl = []
for step in range(dev_num_batch):
feature = sess.run(dev_next_element)
sim_matrix, loss, st, lr, summary = sess.run([svfObj.sim_matrix,
svfObj.loss,
svfObj.global_step,
svfObj.lr,
svfObj.summary_op],
feed_dict={svfObj.inpt: feature})
eer = calculate_eer(sim_matrix)
loss_cl.append(loss)
eer_cl.append(eer)
logging.info(
'dev--step:{}({}) global_step:{} loss:{} eer:{}'.format(step, dev_num_batch, int(st), loss,
eer))
dev_summary_writer.add_summary(summary, epoch)
logging.info('------------------------------------------------------------------------------')
logging.info("dev epoch:{} gs:{}".format(epoch, int(st)))
logging.info("lr:{}".format(lr))
logging.info("avg_loss:{}".format(sum(loss_cl) / len(loss_cl)))
logging.info("avg_eer:{}".format(sum(eer_cl) / len(eer_cl)))
logging.info('------------------------------------------------------------------------------')
cur_eer = np.mean(np.array(eer_cl))
saver.save(sess,
dev_model_save_dir + "/dev_mode_epoch_{}_gs_{}_eer_{:.6f}".format(epoch, st, cur_eer))
if cur_eer < min_dev_eer:
saver.save(sess,
best_dev_model_save_dir + "/dev_mode_epoch_{}_gs_min_{}_eer_{:.6f}".format(epoch, st,
cur_eer))
min_dev_eer = cur_eer
loss_cl.clear()