-
Notifications
You must be signed in to change notification settings - Fork 5
/
train.py
110 lines (88 loc) · 4.3 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
# -*- coding: utf-8 -*-
# @File : train.py
# @Author: Runist
# @Time : 2021/7/2 15:04
# @Software: PyCharm
# @Brief: 训练脚本
import os
import tensorflow as tf
from tensorflow.keras import callbacks, optimizers, losses
from core.config import args
from core.dataset import Dataset
from core.models import get_model
from core.callback import CosineAnnealingLRScheduler, print_lr
if __name__ == '__main__':
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu
gpus = tf.config.experimental.list_physical_devices("GPU")
if gpus:
for gpu in gpus:
tf.config.experimental.set_memory_growth(gpu, True)
if not os.path.exists("./weights/{}".format(args.network)):
os.makedirs("./weights/{}".format(args.network))
print("Preparing train {}.".format(args.network))
model = get_model(args.network, args.input_shape, args.num_classes, include_top=args.include_top)
cbk = [
callbacks.ModelCheckpoint(
'./weights/{}/'.format(args.network)
+ 'epoch={epoch:02d}_val_loss={val_loss:.04f}_val_acc={val_accuracy:.04f}.h5',
save_weights_only=True)
]
# 第一阶段
if args.first_stage_epochs > 0:
train_dataset = Dataset(args.train_image_dir, args.label_name, batch_size=args.batch_size, aug=True)
test_dataset = Dataset(args.test_image_dir, args.label_name, batch_size=1)
train_steps = len(train_dataset) // args.batch_size
test_steps = len(test_dataset)
train_gen = train_dataset.tf_dataset()
test_gen = test_dataset.tf_dataset()
learning_rate = CosineAnnealingLRScheduler(args.first_stage_epochs * train_steps,
args.learn_rate_init, args.learn_rate_end,
warmth_rate=0.05)
optimizer = optimizers.Adam(learning_rate)
if args.include_top:
freeze_layers = len(model.layers) - 1
else:
freeze_layers = len(model.layers) - 4
for i in range(freeze_layers):
model.layers[i].trainable = False
print('Freeze the first {} layers of total {} layers. Train {} epoch.'.format(freeze_layers,
len(model.layers),
args.first_stage_epochs))
model.compile(optimizer=optimizer,
loss=losses.CategoricalCrossentropy(),
metrics=['accuracy'])
model.fit(train_gen,
steps_per_epoch=train_steps,
validation_data=test_gen,
validation_steps=test_steps,
epochs=args.first_stage_epochs,
shuffle=True,
callbacks=cbk)
# 第二阶段
if args.second_stage_epochs > 0:
train_dataset = Dataset(args.train_image_dir, args.label_name, batch_size=args.batch_size, aug=True)
test_dataset = Dataset(args.test_image_dir, args.label_name, batch_size=1)
train_steps = len(train_dataset) // args.batch_size
test_steps = len(test_dataset)
train_gen = train_dataset.tf_dataset()
test_gen = test_dataset.tf_dataset()
learning_rate = CosineAnnealingLRScheduler(args.second_stage_epochs * train_steps,
args.learn_rate_init/10, args.learn_rate_end/10,
warmth_rate=0.05)
optimizer = optimizers.Adam(learning_rate)
for i in range(len(model.layers)):
if 'BatchNormalization' in str(model.layers[i].name_scope):
continue
model.layers[i].trainable = True
print('Unfreeze all layers. Train {} epoch.'.format(args.second_stage_epochs))
model.compile(optimizer=optimizer,
loss=losses.CategoricalCrossentropy(),
metrics=['accuracy'])
model.fit(train_gen,
steps_per_epoch=train_steps,
validation_data=test_gen,
validation_steps=test_steps,
epochs=args.first_stage_epochs + args.second_stage_epochs,
initial_epoch=args.first_stage_epochs,
shuffle=True,
callbacks=cbk)