-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
66 lines (51 loc) · 2.31 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
from utils.Dataset import DatasetLoader
from loss import YOLOLoss
from models import network
from tensorflow import keras
import argparse
parser = argparse.ArgumentParser(description='Train detector file')
parser.add_argument('--img_size', default=448, type=int, help="image input size")
parser.add_argument('--dataset_dir', required=True, type=str, help="train tfrecord dir")
parser.add_argument('--s', default=7, type=int, help="output grid num")
parser.add_argument('--num_class', default=20, type=int, help="the number of class")
parser.add_argument('--num_epoch', default=105, type=int, help='train_epoch')
parser.add_argument('--batch_size', default=32, type=int, help='batch_size')
args = parser.parse_args()
if __name__ == '__main__':
img_size = args.img_size
s = args.s
num_class = args.num_class
num_epochs = args.num_epoch
batch_size = args.batch_size
yolo = network.build_model((img_size, img_size, 3))
lr_schedule = keras.optimizers.schedules.ExponentialDecay(
initial_learning_rate=1e-4,
decay_steps=40000,
decay_rate=0.5,
staircase=True,
)
optimizer = keras.optimizers.Adam(learning_rate=lr_schedule)
loss = YOLOLoss.get_yolo_loss(img_size, s)
yolo.compile(loss=loss, optimizer=optimizer)
loader = DatasetLoader(args.dataset_dir, img_size, s, num_class)
train_ds, val_ds = loader.get_dataset(batch_size)
callbacks_list = [keras.callbacks.ModelCheckpoint(
filepath='./ckpt/valid_best_yolo',
monitor='val_loss',
mode='min',
save_weights_only=True,
save_best_only=True,
verbose=1),
keras.callbacks.TerminateOnNaN()]
hist = yolo.fit(train_ds, validation_data=val_ds, epochs=num_epochs, callbacks=callbacks_list, verbose=1)
yolo.save_weights('./ckpt/yolo_final')
# model save
decoder = network.OutputDecoder()
out = decoder(yolo.output)
final_model = keras.models.Model(yolo.input, out)
final_model.save("./model_asset/yolo_final")
# valid best model
yolo.load_weights("./ckpt/valid_best_yolo")
out = decoder(yolo.output)
final_model = keras.models.Model(yolo.input, out)
final_model.save("./model_asset/valid_best_yolo")