-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathtrain.py
172 lines (156 loc) · 8.53 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
164
165
166
167
168
169
170
171
172
import torch
from torch.optim.lr_scheduler import ReduceLROnPlateau
from tools.data_loader import Dataset_CBCT, Dataset_PCA
from torch.utils.data import DataLoader
import math
from tools.loss_tool import PCA_loss, Log_cosh, PCA_smoothL1Loss
from tools.config import get_args
import yaml
from functools import partial
from tools.instanceExam import InstanceExam
from tools.tool_functions import *
from torch.utils.tensorboard import SummaryWriter
# 加载各类模型
from Model import *
def val(Dataset_loader, net, loss_function, device):
val_loss = 0
with torch.no_grad():
for i, (imgs, target) in enumerate(Dataset_loader):
imgs = imgs.to(device)
target = target.to(device)
prediction = net(imgs)
loss = loss_function(prediction, target)
val_loss += loss
return (val_loss / (i + 1))
def load_cfg(yaml_path):
with open(yaml_path, 'r', encoding='utf-8') as f:
cfg = yaml.load(f, Loader=yaml.FullLoader)
return cfg
def train(args, cfg):
# 初始化
exam_process_dict = cfg['EXAM_PROCESS']
total_exam_num = len(exam_process_dict)
cur_exam_num = 0
# 模型方式
model_methods = {
"CNN": CNN_model.CNN_net,
"Unet": partial(Unet_model.UNet_net, n_classes=3),
"Resnet": partial(Resnet_attention.resnet, layers=[2, 2, 2, 2]),
"Resnet_outCBAM": partial(Resnet_attention.resnet, layers=[2, 2, 2, 2], is_outAttention="CBAM"),
"Resnet_outSPA": partial(Resnet_attention.resnet, layers=[2, 2, 2, 2], is_outAttention="SPA"),
"Resnet_inCBAM": partial(Resnet_attention.resnet, layers=[2, 2, 2, 2], is_inlineAttention="CBAM"),
"Resnet_outSE": partial(Resnet_attention.resnet, layers=[2, 2, 2, 2], is_outAttention="SE"),
"Resnet_inSPA": partial(Resnet_attention.resnet, layers=[2, 2, 2, 2], is_inlineAttention="SPA"),
"Resnet_inSE": partial(Resnet_attention.resnet, layers=[2, 2, 2, 2], is_inlineAttention="SE"),
"Resnet_allSPA": partial(Resnet_attention.resnet, layers=[2, 2, 2, 2], is_inlineAttention="SPA",
is_outAttention="SPA"),
"Resnet_allCBAM": partial(Resnet_attention.resnet, layers=[2, 2, 2, 2], is_inlineAttention="CBAM",
is_outAttention="CBAM"),
"Resnet_allSE": partial(Resnet_attention.resnet, layers=[2, 2, 2, 2], is_inlineAttention="SE",
is_outAttention="SE"),
"Resnet_inSPA_outCBAM": partial(Resnet_attention.resnet, layers=[2, 2, 2, 2], is_inlineAttention="SPA",
is_outAttention="CBAM"),
"Resnet_inCBAM_outSPA": partial(Resnet_attention.resnet, layers=[2, 2, 2, 2], is_inlineAttention="CBAM",
is_outAttention="SPA"),
"Resnet_inCBAM_outSE": partial(Resnet_attention.resnet, layers=[2, 2, 2, 2], is_inlineAttention="CBAM",
is_outAttention="SE"),
"Resnet_inSPA_outSE": partial(Resnet_attention.resnet, layers=[2, 2, 2, 2], is_inlineAttention="SPA",
is_outAttention="SE"),
"Resnet_inSE_outSPA": partial(Resnet_attention.resnet, layers=[2, 2, 2, 2], is_inlineAttention="SE",
is_outAttention="SPA"),
"Resnet_inSE_outCBAM": partial(Resnet_attention.resnet, layers=[2, 2, 2, 2], is_inlineAttention="SE",
is_outAttention="CBAM"),
}
# 损失函数方式
lossfunction_methods = {
"MSE": PCA_loss,
"Smooth_MSE": PCA_smoothL1Loss,
"log_cosh": Log_cosh
}
# 数据加载方式
dataset_process_methods = {
"pca": Dataset_PCA,
"cbct": Dataset_CBCT,
"dvf": Dataset_CBCT
}
setup_seed(12)
# 超参数设定
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
batch_size = args.batch_size
loss_wcoeff = torch.FloatTensor([2 / math.sqrt(6), 1 / math.sqrt(6), 1 / math.sqrt(6)]).to(device)
img_folder = os.path.join(args.root_path, args.img_folder)
target_folder = os.path.join(args.root_path, args.CBCT_folder) if cfg["PREDICTION_MODE"] == "cbct" \
else os.path.join(args.root_path, args.PCA_folder)
# 进行各个exam
for exam_cfg in exam_process_dict:
exam_instance = InstanceExam(args, cfg, exam_cfg)
# 生成log文件
logger = get_logger(filename=os.path.join(exam_instance.log_dir, exam_instance.work_fileName + "_train.log"),
verbosity=1,
name=exam_instance.work_fileName)
# 生成run文件
writer = SummaryWriter(os.path.join(exam_instance.tensorboard_dir, exam_instance.work_fileName))
model = model_methods[exam_instance.model_method](exam_instance.inChannel_num).to(device)
loss_fn = lossfunction_methods[exam_instance.lossFunction_method](loss_wcoeff)
opt = torch.optim.Adam(model.parameters(), lr=args.lr)
scheduler = ReduceLROnPlateau(opt, mode='min', verbose=True, patience=3)
# 数据加载
dataset = dataset_process_methods[exam_instance.prediction_mode](img_folder, target_folder,
exam_instance.input_mode,
exam_instance.model_type)
test_size = int(len(dataset) * args.val_ratio)
train_size = int(len(dataset) - test_size)
train_dataset, test_dataset = torch.utils.data.random_split(dataset=dataset,
lengths=[train_size, test_size],
generator=torch.Generator().manual_seed(12))
train_data_loader = DataLoader(train_dataset, shuffle=True, batch_size=batch_size)
test_data_loader = DataLoader(test_dataset, shuffle=True, batch_size=batch_size)
# 训练参数设定
logger.info(
"DATASET:" + str(exam_instance.dataset)
+ "\tMODEL_TYPE:" + str(exam_instance.model_type)
+ "\tPREDICTION_MODE:" + str(exam_instance.prediction_mode)
+ '\tCOMPARE_MODE:' + str(exam_instance.compare_mode)
+ '\tINPUT_MODE:' + str(exam_instance.input_mode)
+ "\tPREIMG_NUM:" + str(exam_instance.preImg_num)
+ "\tMODEL:" + str(exam_instance.model_method)
+ "\tLOSSFUNCTION:" + str(exam_instance.lossFunction_method)
)
logger.info("Epoch:" + str(args.EPOCH) + "\ttrain_dataset_num:" + str(train_size) + "\ttest_dataset_num:" + str(
test_size))
logger.info("---" * 100)
loss_epoch = []
logger.info('start training!')
for epoch in range(args.EPOCH):
loss_mse = 0
for i, (imgs, target) in enumerate(train_data_loader):
imgs = imgs.to(device)
target = target.to(device)
prediction = model(imgs)
loss_item = loss_fn(target, prediction)
loss_mse += loss_item
opt.zero_grad() # 清空上一步残余更新参数值
loss_item.backward() # 误差反向传播，计算参数更新值
opt.step()
print("(epoch:%d--step:%d)------->loss:%.3f" % (epoch, i, loss_item.item()))
loss_mse = loss_mse / (i + 1)
loss_epoch.append(loss_mse.cpu().detach().numpy())
val_loss = val(test_data_loader, model, loss_fn, device)
print('epoch:%d train_loss:%.3f test_loss:%.3f' % (epoch, loss_mse.item(), val_loss.item()))
scheduler.step(loss_mse)
logger.info(
'Epoch:[{}/{}]\t train_loss={:.3f}\t test_loss={:.3f}'.format((epoch + 1), args.EPOCH, loss_mse.item(),
val_loss.item()))
writer.add_scalars("train_progress", {"train_loss": loss_mse.item(), "val_loss": val_loss.item()})
if (epoch + 1) % args.EPOCH == 0:
cur_ckpt_dir = os.path.join(exam_instance.ckpt_dir,exam_instance.work_fileName)
os.makedirs(cur_ckpt_dir, exist_ok=True)
cur_ckpt_file_name = os.path.join(cur_ckpt_dir, str(epoch + 1) + ".pth")
torch.save(model.state_dict(), cur_ckpt_file_name)
logger.info('finish training!')
logging.shutdown()
writer.close()
if __name__ == '__main__':
args = get_args()
cfg = load_cfg(yaml_path="./tools/cfg/pca_space.yaml")
train(args, cfg)