-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
156 lines (136 loc) · 7.04 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
import numpy as np
import torch
import torch.backends.cudnn as cudnn
import torch.optim as optim
from torch.utils.data import DataLoader
from nets.yolo import YoloBody
from nets.yolo_training import YOLOLoss, weights_init
from utils.callbacks import LossHistory
from utils.dataloader import YoloDataset, yolo_dataset_collate
from utils.utils import get_classes
from utils.utils_fit import fit_one_epoch
if __name__ == "__main__":
Cuda = True
classes_path = 'model_data/voc_classes.txt'
model_path = 'model_data/yolox_s.pth'
input_shape = [640, 640]
phi = 's'
# mosaic 马赛克数据增强
# YOLOX作者强调要在训练结束前的N个epoch关掉Mosaic。因为Mosaic生成的训练图片，远远脱离自然图片的真实分布。
# 并且Mosaic大量的crop操作会带来很多不准确的标注框，本代码自动会在前90%个epoch使用mosaic，后面不使用。
# 马赛克数据增强将4张训练图像按一定比例组合成1张。Mosaic是一个新的数据增强技术。这使得模型能够学习如何识别比正常尺寸小的物体。
mosaic = False
# 余弦退火学习率,按一个公式确定学习率的衰减
Cosine_scheduler = False
# 训练分为两个阶段，分别是冻结阶段和解冻阶段
# 冻结阶段模型的主干backbone被冻结了，特征提取网络不发生改变，训练neck和head
Init_Epoch = 0
Freeze_Epoch = 50
Freeze_batch_size = 8
Freeze_lr = 1e-3
# 解冻阶段所有的参数一起训练
UnFreeze_Epoch = 100
Unfreeze_batch_size = 4
Unfreeze_lr = 1e-4
Freeze_Train = True
# 线程设置
num_workers = 4
# 得到图片路径和标签
train_annotation_path = '2007_train.txt'
val_annotation_path = '2007_val.txt'
# 得到类的名称和长度
class_names, num_classes = get_classes(classes_path)
# 模型设计需要种类数来确定cls尺寸
model = YoloBody(num_classes, phi)
# 初始化网络参数
weights_init(model)
if model_path != '':
print('Load weights {}.'.format(model_path))
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
#验证模型格式一致
model_dict = model.state_dict()
pretrained_dict = torch.load(model_path, map_location=device)
pretrained_dict = {k: v for k, v in pretrained_dict.items() if np.shape(model_dict[k]) == np.shape(v)}
model_dict.update(pretrained_dict)
model.load_state_dict(model_dict)
model_train = model.train()
#是否启用GPU
if Cuda:
model_train = torch.nn.DataParallel(model)
cudnn.benchmark = True
model_train = model_train.cuda()
yolo_loss = YOLOLoss(num_classes)
loss_history = LossHistory("logs/")
with open(train_annotation_path) as f:
train_lines = f.readlines()
with open(val_annotation_path) as f:
val_lines = f.readlines()
num_train = len(train_lines)
num_val = len(val_lines)
#设定冻结阶段的训练
if True:
batch_size = Freeze_batch_size
lr = Freeze_lr
start_epoch = Init_Epoch
end_epoch = Freeze_Epoch
epoch_step = num_train // batch_size #设置一个epoch下要几个batch
epoch_step_val = num_val // batch_size
if epoch_step == 0 or epoch_step_val == 0: #至少训练集数量要比一个batch大
raise ValueError("数据集过小，无法进行训练，请扩充数据集。")
optimizer = optim.Adam(model_train.parameters(), lr, weight_decay=5e-4) #设置优化器
if Cosine_scheduler:
lr_scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=5, eta_min=1e-5) #用不用余弦退火学习率
else:
lr_scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=1, gamma=0.92) #使用steplr的学习率调整方法
train_dataset = YoloDataset(train_lines, input_shape, num_classes, end_epoch - start_epoch, mosaic=mosaic,
train=True)
val_dataset = YoloDataset(val_lines, input_shape, num_classes, end_epoch - start_epoch, mosaic=False,
train=False)
gen = DataLoader(train_dataset, shuffle=True, batch_size=batch_size, num_workers=num_workers, pin_memory=True,
drop_last=True, collate_fn=yolo_dataset_collate)
gen_val = DataLoader(val_dataset, shuffle=True, batch_size=batch_size, num_workers=num_workers, pin_memory=True,
drop_last=True, collate_fn=yolo_dataset_collate)
# ------------------------------------#
# 冻结一定部分训练
# ------------------------------------#
if Freeze_Train:
for param in model.backbone.parameters():
param.requires_grad = False
for epoch in range(start_epoch, end_epoch):
#逐步优化
fit_one_epoch(model_train, model, yolo_loss, loss_history, optimizer, epoch,
epoch_step, epoch_step_val, gen, gen_val, end_epoch, Cuda)
#学习率调整，一个epoch*0.92
lr_scheduler.step()
if True:
batch_size = Unfreeze_batch_size
lr = Unfreeze_lr
start_epoch = Freeze_Epoch
end_epoch = UnFreeze_Epoch
epoch_step = num_train // batch_size
epoch_step_val = num_val // batch_size
if epoch_step == 0 or epoch_step_val == 0:
raise ValueError("数据集过小，无法进行训练，请扩充数据集。")
optimizer = optim.Adam(model_train.parameters(), lr, weight_decay=5e-4)
if Cosine_scheduler:
lr_scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=5, eta_min=1e-5)
else:
lr_scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=1, gamma=0.92)
train_dataset = YoloDataset(train_lines, input_shape, num_classes, end_epoch - start_epoch, mosaic=mosaic,
train=True)
val_dataset = YoloDataset(val_lines, input_shape, num_classes, end_epoch - start_epoch, mosaic=False,
train=False)
gen = DataLoader(train_dataset, shuffle=True, batch_size=batch_size, num_workers=num_workers, pin_memory=True,
drop_last=True, collate_fn=yolo_dataset_collate)
gen_val = DataLoader(val_dataset, shuffle=True, batch_size=batch_size, num_workers=num_workers, pin_memory=True,
drop_last=True, collate_fn=yolo_dataset_collate)
# ------------------------------------#
# 解冻后训练，这一部分与前面不同，其它基本相似。
# ------------------------------------#
if Freeze_Train:
for param in model.backbone.parameters():
param.requires_grad = True
for epoch in range(start_epoch, end_epoch):
fit_one_epoch(model_train, model, yolo_loss, loss_history, optimizer, epoch,
epoch_step, epoch_step_val, gen, gen_val, end_epoch, Cuda)
lr_scheduler.step()