-
Notifications
You must be signed in to change notification settings - Fork 3
/
main.py
315 lines (263 loc) · 14.8 KB
/
main.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
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
import itertools
import os
import torch
import numpy as np
from torch.utils.data import DataLoader
from torch.optim.lr_scheduler import StepLR
from tqdm import tqdm
from dataset.dataset import HiUCDDataset, LoveDADataset, LoveDADataset_for_SCD, NanjingDataset
from loss.losses import BCDLoss, AdditionalBackgroundSupervision
from utils.callbacks import AverageMeter
from utils.evaluator import BCDEvaluator, SEGEvaluator
from utils.helper import get_lr, seed_torch, get_model
from utils.parser import get_parser_with_args_from_json
from utils.saver import Saver
from utils.logger import Logger as Log
def split_sample(sample, seg_pretrain=False):
if not seg_pretrain:
img_A = sample['img_A'].cuda(non_blocking=True)
img_B = sample['img_B'].cuda(non_blocking=True)
label_BCD = sample['label_BCD'].cuda(non_blocking=True)
label_SGA = sample['label_SGA'].cuda(non_blocking=True)
label_SGB = sample['label_SGB'].cuda(non_blocking=True)
return img_A, img_B, label_BCD, label_SGA.long(), label_SGB.long()
else:
imgs = sample['img_A'].cuda(non_blocking=True)
labels = sample['label_SGA'].cuda(non_blocking=True)
batch_size = int(imgs.shape[0] / 2)
img_A = imgs[0:batch_size, :]
img_B = imgs[batch_size::, :]
label_SGA = labels[0:batch_size, :]
label_SGB = labels[batch_size::, :]
return img_A, img_B, None, label_SGA.long(), label_SGB.long()
def get_dataset(args):
if args.dataset == 'HiUCD':
train_dataset = HiUCDDataset(args, split='train')
val_dataset = HiUCDDataset(args, split='val')
elif args.dataset =='LoveDA':
train_dataset = LoveDADataset(args, split='train')
val_dataset = LoveDADataset(args, split='val')
elif args.dataset == 'LoveDAforSCD':
train_dataset = LoveDADataset_for_SCD(args, split='train')
val_dataset = LoveDADataset_for_SCD(args, split='val')
elif args.dataset == 'Nanjing':
train_dataset = NanjingDataset(args, split='train')
val_dataset = NanjingDataset(args, split='val')
return train_dataset, val_dataset
def main(args):
seed_torch()
model = get_model(args).cuda()
train_dataset, val_dataset = get_dataset(args)
drop_last = True
if args.dataset == 'Nanjing':
drop_last = False
train_loader = DataLoader(train_dataset, batch_size=args.batch_size, shuffle=True,
persistent_workers=True, pin_memory=True, num_workers=args.num_workers, drop_last=drop_last)
val_loader = DataLoader(val_dataset, batch_size=args.val_batch_size, shuffle=False,
persistent_workers=True, pin_memory=True, num_workers=args.val_num_workers, drop_last=drop_last)
loss_bcd = BCDLoss()
loss_seg = torch.nn.CrossEntropyLoss(ignore_index=args.num_segclass)
loss_abs = AdditionalBackgroundSupervision(ignore_index=args.num_segclass)
optimizer = torch.optim.Adam(itertools.chain(model.parameters()), lr=args.learning_rate, weight_decay=args.weight_decay)
lr_scheduler = StepLR(optimizer, step_size=10, gamma=0.9)
saver = Saver(args)
evaluator_bcd = BCDEvaluator()
evaluator_seg_A = SEGEvaluator(args.num_segclass)
evaluator_seg_B = SEGEvaluator(args.num_segclass)
evaluator_seg_total = SEGEvaluator(args.num_segclass)
metric_best = -1
metric_best_dict = {}
start_epoch = 1
Log.init(logfile_level="info", log_file=saver.experiment_dir + '/log.log')
if isinstance(args.resume, str):
checkpoint = torch.load(args.resume)
checkpoint['epoch'] = 1
model.load_state_dict(checkpoint['state_dict'])
Log.info("=> loaded checkpoint '{}' (epoch {})".format(args.resume, checkpoint['epoch']))
del checkpoint
epoch_best = start_epoch
for epoch in tqdm(range(start_epoch, args.epochs + 1), desc=args.congfig_name):
if epoch > args.warmup_epoch:
warmup = False
else:
warmup = True
# traning
losses_seg_A = AverageMeter()
losses_seg_B = AverageMeter()
losses_bcd = AverageMeter()
losses_bscc = AverageMeter()
losses_cont = AverageMeter()
losses_total = AverageMeter()
model.train()
for batch_idx, sample in enumerate(train_loader):
# wheter seg pretraining
img_A, img_B, label_BCD, label_SGA, label_SGB = split_sample(sample, seg_pretrain=args.seg_pretrain)
imgs = torch.cat([img_A, img_B], dim=1)
# whether contrastive learning
if args.with_sacl:
outputs = model(imgs, label_SGA, label_SGB, label_BCD, test=False, warmup=warmup)
elif args.with_bscc:
outputs = model(imgs, label_BCD=label_BCD, test=False)
else:
outputs = model(imgs)
# whether only segmentation
if not args.only_seg:
loss_cd = loss_bcd(outputs['BCD'], label_BCD)
# whether seg pretraining
elif args.only_seg and args.seg_pretrain:
# whether using additional background supervision
if args.with_abs:
loss_cd = loss_abs(torch.cat([outputs['seg_A'], outputs['seg_B']], dim=0), torch.cat([label_SGA, label_SGB], dim=0))
else:
loss_cd = torch.tensor(0)
else:
loss_cd = torch.tensor(0)
# whether only binary change detection or seg_pretrain
if not args.only_bcd and (not args.seg_pretrain):
loss_seg_A = loss_seg(outputs['seg_A'], label_SGA)
loss_seg_B = loss_seg(outputs['seg_B'], label_SGB)
elif not args.only_bcd and args.seg_pretrain:
loss_seg_A = loss_seg(torch.cat([outputs['seg_A'], outputs['seg_B']], dim=0), torch.cat([label_SGA, label_SGB], dim=0))
loss_seg_B = torch.tensor(0)
elif args.only_bcd:
loss_seg_A = torch.tensor(0)
loss_seg_B = torch.tensor(0)
# only global: must not warmup; only local: anyway
flag = (not warmup) or args.local_contrast
if args.with_sacl and flag:
loss_contrast = outputs['contrast_loss']
losses_cont.update(loss_contrast.item())
else:
loss_contrast = torch.tensor(0)
# bscc loss
if args.with_bscc:
loss_bscc = outputs['bscc_loss']
else:
loss_bscc = torch.tensor(0)
# total loss
loss = loss_cd + loss_seg_A + loss_seg_B + loss_contrast + loss_bscc
optimizer.zero_grad()
loss.backward()
optimizer.step()
losses_bcd.update(loss_cd.item())
losses_bscc.update(loss_bscc.item())
losses_seg_A.update(loss_seg_A.item())
losses_seg_B.update(loss_seg_B.item())
losses_total.update(loss.item())
if batch_idx % args.print_step == 0 or batch_idx == len(train_loader)-1:
print('[Epoch:%3d/%3d | Batch:%4d/%4d] loss_total: %.4f loss_bcd: %.4f loss_segA: %.4f loss_segB: %.4f loss_contrast: %.4f loss_bscc: %.4f lr: %5f' %
(epoch, args.epochs, batch_idx+1, train_loader.__len__(), losses_total.avg,
losses_bcd.avg, losses_seg_A.avg, losses_seg_B.avg, losses_cont.avg, losses_bscc.avg, get_lr(optimizer))
)
lr_scheduler.step()
Log.info('[Training Epoch:%3d/%3d] loss_total: %.4f loss_bcd: %.4f loss_segA: %.4f loss_segB: %.4f loss_contrast: %.4f loss_bscc: %.4f lr: %f' %
(epoch, args.epochs, losses_total.avg, losses_bcd.avg, losses_seg_A.avg, losses_seg_B.avg, losses_cont.avg, losses_bscc.avg, get_lr(optimizer)))
# validation
evaluator_bcd.reset()
evaluator_seg_A.reset()
evaluator_seg_B.reset()
evaluator_seg_total.reset()
valosses_seg_A = AverageMeter()
valosses_seg_B = AverageMeter()
valosses_bcd = AverageMeter()
valosses_total = AverageMeter()
model.eval()
with torch.no_grad():
for batch_idx, sample in enumerate(val_loader):
img_A, img_B, label_BCD, label_SGA, label_SGB = split_sample(sample, seg_pretrain=args.seg_pretrain)
imgs = torch.cat([img_A, img_B], dim=1)
if args.with_sacl or args.with_bscc:
outputs = model(imgs, test=True)
else:
outputs = model(imgs)
if not args.only_seg:
loss_cd = loss_bcd(outputs['BCD'], label_BCD)
else:
loss_cd = torch.tensor(0)
if not args.only_bcd and (not args.seg_pretrain):
loss_seg_A = loss_seg(outputs['seg_A'], label_SGA)
loss_seg_B = loss_seg(outputs['seg_B'], label_SGB)
elif not args.only_bcd and args.seg_pretrain:
loss_seg_A = loss_seg(torch.cat([outputs['seg_A'], outputs['seg_B']], dim=0), torch.cat([label_SGA, label_SGB], dim=0))
loss_seg_B = torch.tensor(0)
else:
loss_seg_A = torch.tensor(0)
loss_seg_B = torch.tensor(0)
loss = loss_cd + loss_seg_A + loss_seg_B
valosses_bcd.update(loss_cd.item())
valosses_seg_A.update(loss_seg_A.item())
valosses_seg_B.update(loss_seg_B.item())
valosses_total.update(loss.item())
if not args.only_seg:
pred_bcd = outputs['BCD'].sigmoid().squeeze().cpu().detach().numpy().round().astype('int')
evaluator_bcd.add_batch(label_BCD.cpu().numpy().astype('int').squeeze(), pred_bcd)
if not args.only_bcd and args.separate_val_seg:
pred_seg_A = torch.argmax(outputs['seg_A'], 1).cpu().detach().numpy().astype('int')
evaluator_seg_A.add_batch(label_SGA.cpu().numpy().astype('int'), pred_seg_A)
pred_seg_B = torch.argmax(outputs['seg_B'], 1).cpu().detach().numpy().astype('int')
evaluator_seg_B.add_batch(label_SGB.cpu().numpy().astype('int'), pred_seg_B)
elif not args.only_bcd and (not args.separate_val_seg):
pred_seg_A = torch.argmax(outputs['seg_A'], 1).cpu().detach().numpy().astype('int')
pred_seg_B = torch.argmax(outputs['seg_B'], 1).cpu().detach().numpy().astype('int')
pred_seg = np.concatenate([pred_seg_A, pred_seg_B], axis=0)
label_seg = torch.cat([label_SGA, label_SGB], dim=0)
evaluator_seg_total.add_batch(label_seg.cpu().numpy().astype('int'), pred_seg)
if batch_idx % args.print_step == 0 or batch_idx == len(val_loader)-1:
print('[Epoch:%3d/%3d | Batch:%4d/%4d] loss_total: %.4f loss_bcd: %.4f loss_segA: %.4f loss_segB: %.4f' %
(epoch, args.epochs, batch_idx+1, val_loader.__len__(), valosses_total.avg,
valosses_bcd.avg, valosses_seg_A.avg, valosses_seg_B.avg)
)
Log.info('[Validation Epoch:%3d/%3d] loss_total: %.4f loss_bcd: %.4f loss_segA: %.4f loss_segB: %.4f' %
(epoch, args.epochs, valosses_total.avg, valosses_bcd.avg, valosses_seg_A.avg, valosses_seg_B.avg))
if not args.only_seg:
OA_bcd = evaluator_bcd.Overall_Accuracy()
IoU_bcd = evaluator_bcd.Intersection_over_Union()
F1_bcd = evaluator_bcd.F1_score()
else:
OA_bcd = IoU_bcd = F1_bcd = 0
if not args.only_bcd:
OA_seg_A = evaluator_seg_A.Overall_Accuracy()
mIoU_seg_A = evaluator_seg_A.Mean_Intersection_over_Union()
F1_seg_A = evaluator_seg_A.F1_score().mean()
OA_seg_B = evaluator_seg_B.Overall_Accuracy()
mIoU_seg_B = evaluator_seg_B.Mean_Intersection_over_Union()
F1_seg_B = evaluator_seg_B.F1_score().mean()
OA_seg_total = evaluator_seg_total.Overall_Accuracy()
mIoU_seg_total = evaluator_seg_total.Mean_Intersection_over_Union()
F1_seg_total = evaluator_seg_total.F1_score().mean()
else:
OA_seg_A = mIoU_seg_A = F1_seg_A = OA_seg_B = mIoU_seg_B = F1_seg_B = OA_seg_total = mIoU_seg_total = F1_seg_total = 0
Log.info('[Validation Epoch:%3d/%3d] OA_BCD: %.4f IoU_BCD: %.4f F1_BCD: %.4f OA_SEG_A: %.4f mIoU_SEG_A: %.4f F1_SEG_A: %.4f OA_SEG_B: %.4f mIoU_SEG_B: %.4f F1_SEG_B: %.4f OA_SEG_total: %.4f mIoU_SEG_total: %.4f F1_SEG_total: %.4f' %
(epoch, args.epochs, OA_bcd, IoU_bcd, F1_bcd, OA_seg_A, mIoU_seg_A, F1_seg_A,
OA_seg_B, mIoU_seg_B, F1_seg_B, OA_seg_total, mIoU_seg_total, F1_seg_total))
if args.separate_val_seg:
metric_current = IoU_bcd + mIoU_seg_A + mIoU_seg_B
else:
metric_current = IoU_bcd + mIoU_seg_total
if (metric_current > metric_best) or (epoch == 1):
metric_best_dict = {}
metric_best_dict["IoU_BCD"] = IoU_bcd
metric_best_dict["mIoU_SEG_A"] = mIoU_seg_A
metric_best_dict["mIoU_SEG_B"] = mIoU_seg_B
metric_best_dict["mIoU_SEG_total"] = mIoU_seg_total
epoch_best = epoch
metric_best = metric_current
if args.epochs > 0:
if epoch > 0:
saver.save_checkpoint({
'state_dict': model.state_dict(),
}, epoch, metric_current)
print('=> Current metric %.4f Best metric %.4f' % (metric_current, metric_best))
Log.info('=> Best epoch {} Best metric {}'.format(epoch_best, metric_best_dict))
if __name__ == '__main__':
floder_list = [
"./configs/hiucd_mini",
"./configs/hiucd"
]
for config_floder in floder_list:
configs_list = os.listdir(config_floder)
configs_list = [cf for cf in configs_list if cf.endswith('json')]
for configs_name in configs_list:
configs_file = os.path.join(config_floder, configs_name)
args = get_parser_with_args_from_json(configs_file)
main(args)