forked from LiuDongjing/BuildingChangeDetector
/
end2end_train.py
219 lines (193 loc) · 7.85 KB
/
end2end_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
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
# -*- coding: utf-8 -*-
"""
训练网络。
在本地运行，当前目录应有local_config_end2end.json配置文件；在PAI上运行，应制定
配置文件，格式参考unet-end2end-rgbn.txt。
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import os
import re
import sys
import time
import json
import argparse
import tensorflow as tf
from keras.optimizers import Adam
from keras.callbacks import (Callback, ModelCheckpoint, CSVLogger,
LearningRateScheduler)
import utils
import configs
from datasets import Datasets
from end2end_predict import test
def get_initial_epoch(log_path):
"""
从log文件中获取最近训练结束时的epoch，重新运行代码后会接着这个epoch继续训练。
"""
initial_epoch = 0
if tf.gfile.Exists(log_path):
with open(log_path) as log_file:
line_ind = -1
for _, line in enumerate(log_file):
line = line.strip()
if line == '':
continue
t = re.split(r',', line)[0]
if not t.isdigit():
continue
line_ind = 1+int(t)
initial_epoch = line_ind
return initial_epoch
class ModelBackup(Callback):
"""
保存训练好的网络权重。
PAI上运行代码时才会使用该功能。每个epoch结束后，都会将得到的最佳网络权重保存到
自己的oss目录里。
"""
def __init__(self, output_path, model_names):
"""
Args:
output_path: 保存网络权重的oss目录
model_namse: 权重文件的名称
"""
self.output_path = output_path
self.model_names = model_names
if sys.version_info > (3, 0):
super().__init__()
else:
super(ModelBackup, self).__init__()
def on_epoch_end(self, batch, logs=None):
for e in self.model_names:
if tf.gfile.Exists(e):
path = os.path.join(self.output_path, e)
print('Backup %s to %s.'%(e, path))
tf.gfile.Copy(e, path, overwrite=True)
else:
print('%s not found!'%e, file=sys.stderr)
class TrainModel():
"""
训练模型。
"""
def __init__(self, options):
"""
Args:
options: configs.Config类的实例
"""
self.options = options
input_shape = (options.target_size[0], options.target_size[1],
len(options.use_chans))
print('Making %s...'%options.model_name)
if utils.is_py3(): #本地运行
best_model = os.path.join(options.output_path, options.weight_path)
self.model = utils.make_model(self.options.model_name, input_shape)
if tf.gfile.Exists(best_model):
print('Load weights from %s.'%best_model)
self.model.load_weights(best_model)
else: #PAI上运行
path = os.path.join(options.output_path, options.weight_path)
self.model = utils.make_model(self.options.model_name, input_shape)
if tf.gfile.Exists(path):
utils.copy_from_oss([path], using_cache=False)
best_model = options.weight_path
print('Load weights from %s.'%best_model)
self.model.load_weights(best_model)
self.log_path = '%s-log.txt'%options.run_name
if utils.is_py3():
self.log_path = os.path.join(options.output_path, self.log_path)
else:
utils.copy_from_oss(
[os.path.join(options.output_path, self.log_path)],
using_cache=False)
if self.options.init_epoch < 0:
self.options.init_epoch = get_initial_epoch(self.log_path)
self.metrics = [utils.f1_score]
self.loss_function = utils.dice_coef_loss
def make_callbacks(self):
tmp_path = self.options.output_path
if not utils.is_py3():
tmp_path = ''
best_model_checkpoint = ModelCheckpoint(monitor='val_f1_score', mode='max',
filepath=os.path.join(tmp_path, options.weight_path), save_best_only=True,
save_weights_only=True)
logger = CSVLogger(self.log_path, append=True)
callbacks = [best_model_checkpoint, logger]
if not utils.is_py3():
callbacks.append(ModelBackup(self.options.output_path,
[self.options.weight_path,
self.log_path]))
if self.options.lr_epoch_decay: # 每个epoch都会降低lr
def get_lr(epoch):
w = epoch // 10
lr = self.options.init_lr / (options.lr_epoch_decay ** w)
if lr < 1e-10:
lr = 1e-10
return lr
callback = LearningRateScheduler(get_lr)
callbacks.append(callback)
return callbacks
def train_model(self):
t0 = time.time()
print(self.model.summary())
dataset = Datasets(self.options)
train_gen = dataset.make_split_generator('train')
validation_gen = dataset.make_split_generator('validation')
optimizer = Adam(lr=self.options.init_lr, decay=0)
self.model.compile(
optimizer, loss=self.loss_function, metrics=self.metrics)
callbacks = self.make_callbacks()
self.model.fit_generator(
train_gen,
initial_epoch=self.options.init_epoch,
steps_per_epoch=self.options.steps_per_epoch,
epochs=self.options.epochs,
validation_data=validation_gen,
validation_steps=self.options.validation_steps,
callbacks=callbacks)
print('Training time cost: %0.2f(min).'%((time.time()-t0)/60))
if __name__ == '__main__':
options = None
t0 = time.time()
train_files = []
val_files = []
test_files = []
if utils.is_py3():
# Python3
options = configs.get_config('local_config_end2end.json')
train_val = json.load(open(options.train_val_test_config))
for e in train_val['train']:
train_files.append(os.path.join(options.input_path,
options.data_path, e))
for e in train_val['validation']:
val_files.append(os.path.join(options.input_path,
options.data_path, e))
for e in train_val['test']:
test_files.append(os.path.join(options.input_path,
options.data_path, e))
else:
# Python2
options = configs.get_config(argparse.ArgumentParser())
train_val = json.load(open(options.train_val_test_config))
# 将data_path里所有的npy文件拷贝到当前目录
data = tf.gfile.Glob(os.path.join(
options.input_path, options.data_path, '*.npy'))
utils.copy_from_oss(data, using_cache=False)
for e in train_val['train']:
train_files.append(e)
for e in train_val['validation']:
val_files.append(e)
for e in train_val['test']:
test_files.append(e)
counter = utils.SampleCounter(train_files+val_files,
options.target_size, options.batch_size,
1)
options.train_files = train_files
options.validation_files = val_files
options.test_files = test_files
options.steps_per_epoch = counter.count_sample(options.train_files)
options.validation_steps = counter.count_sample(options.validation_files)
options.show()
model = TrainModel(options=options)
model.train_model()
test(model.model, options)
print('Total time cost: %0.2f(min).'%((time.time()-t0)/60))