forked from ainazHjm/LandslidePrediction
-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
145 lines (129 loc) · 6.49 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
import model
import torch as th
import torch.optim as to
import torch.nn as nn
import os
from time import ctime
from tensorboardX import SummaryWriter
from torch.optim.lr_scheduler import ReduceLROnPlateau
from utils.plot import save_config
from unet import UNet
# pylint: disable=E1101,E0401,E1123
def create_dir(dir_name):
model_dir = dir_name+'/model/'
res_dir = dir_name+'/result/'
if not os.path.exists(model_dir):
os.mkdir(model_dir)
if not os.path.exists(res_dir):
os.mkdir(res_dir)
return model_dir, res_dir
def validate(model, val_loader, data_param, train_param, _log):
with th.no_grad():
criterion = nn.BCEWithLogitsLoss(pos_weight=th.Tensor([train_param['pos_weight']]).cuda())
val_iter = iter(val_loader)
running_loss = 0
prune = data_param['prune']
for _ in range(len(val_iter)):
batch_sample = val_iter.next()
data = batch_sample['data'].cuda()
gt = batch_sample['gt'].cuda()
prds = model.forward(data)[:, :, prune:-prune, prune:-prune]
indices = gt>=0
loss = criterion(prds[indices], gt[indices])
running_loss += loss.item()
del data, gt, prds, indices
return running_loss/len(val_iter)
def train(train_loader, val_loader, train_param, data_param, loc_param, _log, _run):
writer = SummaryWriter()
model_dir, _ = create_dir(writer.file_writer.get_logdir())
sig = nn.Sigmoid()
if train_param['model'] == "FCN":
train_model = model.FCN(data_param['feature_num']).cuda()
elif train_param['model'] == 'FCNwPool':
train_model = model.FCNwPool(data_param['feature_num'], data_param['pix_res']).cuda()
elif train_param['model'] == 'UNet':
train_model = UNet(data_param['feature_num'], 1).cuda()
elif train_param['model'] == 'FCNwBottleneck':
train_model = model.FCNwBottleneck(data_param['feature_num'], data_param['pix_res']).cuda()
elif train_param['model'] == 'SimplerFCNwBottleneck':
train_model = model.SimplerFCNwBottleneck(data_param['feature_num']).cuda()
elif train_param['model'] == 'Logistic':
train_model = model.Logistic(data_param['feature_num']).cuda()
elif train_param['model'] == 'PolyLogistic':
train_model = model.PolyLogistic(data_param['feature_num']).cuda()
if th.cuda.device_count() > 1:
train_model = nn.DataParallel(train_model)
if loc_param['load_model']:
train_model.load_state_dict(th.load(loc_param['load_model']))
_log.info('[{}] model is initialized ...'.format(ctime()))
if train_param['optim'] == 'Adam':
optimizer = to.Adam(train_model.parameters(), lr=train_param['lr'], weight_decay=train_param['decay'])
else:
optimizer = to.SGD(train_model.parameters(), lr=train_param['lr'], weight_decay=train_param['decay'])
scheduler = ReduceLROnPlateau(optimizer, mode='min', patience=train_param['patience'], verbose=True, factor=0.5)
criterion = nn.BCEWithLogitsLoss(pos_weight=th.Tensor([train_param['pos_weight']]).cuda())
valatZero = validate(train_model, val_loader, data_param, train_param, _log)
_log.info('[{}] validation loss before training: {}'.format(ctime(), valatZero))
_run.log_scalar('training.val_loss', valatZero, 0)
trainatZero = validate(train_model, train_loader, data_param, train_param, _log)
_log.info('[{}] train loss before training: {}'.format(ctime(), trainatZero))
_run.log_scalar('training.loss_epoch', trainatZero, 0)
loss_ = 0
prune = data_param['prune']
for epoch in range(train_param['n_epochs']):
running_loss = 0
train_iter = iter(train_loader)
for iter_ in range(len(train_iter)):
optimizer.zero_grad()
batch_sample = train_iter.next()
data, gt = batch_sample['data'].cuda(), batch_sample['gt'].cuda()
if train_param['model'] == 'UNET':
prds = train_model(data)[:, :, prune:-prune, prune:-prune]
else:
prds = train_model.forward(data)[:, :, prune:-prune, prune:-prune]
indices = gt>=0
loss = criterion(prds[indices], gt[indices])
running_loss += loss.item()
loss_ += loss.item()
loss.backward()
optimizer.step()
_run.log_scalar("training.loss_iter", loss.item(), epoch*len(train_iter)+iter_+1)
_run.log_scalar("training.max_prob", th.max(sig(prds)).item(), epoch*len(train_iter)+iter_+1)
_run.log_scalar("training.min_prob", th.min(sig(prds)).item(), epoch*len(train_iter)+iter_+1)
writer.add_scalar("loss/train_iter", loss.item(), epoch*len(train_iter)+iter_+1)
writer.add_scalars(
"probRange",
{'min': th.min(sig(prds)), 'max': th.max(sig(prds))},
epoch*len(train_iter)+iter_+1
)
if (epoch*len(train_iter)+iter_+1) % 20 == 0:
_run.log_scalar("training.loss_20", loss_/20, epoch*len(train_iter)+iter_+1)
writer.add_scalar("loss/train_20", loss_/20, epoch*len(train_iter)+iter_+1)
_log.info(
'[{}] loss at [{}/{}]: {}'.format(
ctime(),
epoch*len(train_iter)+iter_+1,
train_param['n_epochs']*len(train_iter),
loss_/20
)
)
loss_ = 0
v_loss = validate(train_model, val_loader, data_param, train_param, _log)
scheduler.step(v_loss)
_log.info('[{}] validation loss at [{}/{}]: {}'.format(ctime(), epoch+1, train_param['n_epochs'], v_loss))
_run.log_scalar('training.val_loss', v_loss, epoch+1)
_run.log_scalar('training.loss_epoch', running_loss/len(train_iter), epoch+1)
writer.add_scalars(
"loss/grouped",
{'test': v_loss, 'train': running_loss/len(train_iter)},
epoch+1
)
del data, gt, prds, indices
if (epoch+1) % loc_param['save'] == 0:
th.save(train_model.cpu().state_dict(), model_dir+'model_{}.pt'.format(str(epoch+1)))
train_model = train_model.cuda()
writer.export_scalars_to_json(model_dir+'loss.json')
th.save(train_model.cpu().state_dict(), model_dir+'trained_model.pt')
save_config(writer.file_writer.get_logdir()+'/config.txt', train_param, data_param)
_log.info('[{}] model has been trained and config file has been saved.'.format(ctime()))
return v_loss