-
Notifications
You must be signed in to change notification settings - Fork 6
/
engine.py
321 lines (261 loc) · 15.6 KB
/
engine.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
316
317
318
319
320
321
import sys
import time
import logging
import numpy as np
from tqdm import tqdm
from dataset import get_data
import torch
import torch.nn as nn
import torch.backends.cudnn as cudnn
from tensorboardX import SummaryWriter
from utils.optimizer import get_optim
from utils.prepare import clean, backup, get_month_day
from utils.metrics import compute_sdf, dice_loss
from utils.learning_rate import adjust_learning_rate, get_current_consistency_weight
from utils.runtime import random_resize, get_cls_label, name_list_to_cls_label
from utils.validate import val_mode_seg, val_mode_seg_multi_scale
from utils.loss import MyLabelSmoothingLoss
# get loss function
mse_loss = nn.MSELoss()
l1_loss = nn.L1Loss()
ce_loss = nn.CrossEntropyLoss()
weight_ce_loss_for_class_m = nn.CrossEntropyLoss(weight=torch.Tensor([1.0, 1.0, 1.0])).cuda() # 第1类的权重调高
label_smooth_ce_loss = MyLabelSmoothingLoss(classes=2, smoothing=0.0)
def init_program(args):
clean(args.exp_name)
backup(args.exp_name)
writer = SummaryWriter('./history/{}/'.format(get_month_day()) + args.exp_name)
logging.basicConfig(filename="./history/{}/{}/log.txt".format(get_month_day(), args.exp_name), level=logging.INFO,
format='[%(asctime)s.%(msecs)03d] %(message)s', datefmt='%H:%M:%S')
logging.getLogger().addHandler(logging.StreamHandler(sys.stdout))
logging.info(str(args))
cudnn.enabled = True
cudnn.benchmark = True
return writer
# a function that trains the model
def trainer(args, model):
writer = init_program(args)
if args.pretrain != '':
pretrained_dict = torch.load(args.pretrain)
model.load_state_dict(pretrained_dict)
print('Load Pre-trained Model Finish!')
optimizer = get_optim(args, model.parameters())
dataloader_group = get_data(args)
trainloader, valloader, testloader = dataloader_group['trainloader'], dataloader_group['valloader'], \
dataloader_group['testloader']
label_dic = get_cls_label(args)
best_val_jac = 0
best_cls = 0
iter_num = 0
for epoch in range(args.epoch):
for i_iter, batch in tqdm(enumerate(trainloader)):
test_flag = False
if test_flag:
print('*** Only Test ***')
break
test_flag = False
model.train()
images, labels, name = batch
images, labels = random_resize(args, images, labels)
images = images.cuda()
labels = labels.cuda().squeeze(1)
optimizer.zero_grad()
lr = adjust_learning_rate(args, optimizer, iter_num)
preds, dt_preds, cls_logits, attention_maps, ds_mask_logits = model(images)
cls_label = name_list_to_cls_label(name, label_dic)
### attention loss
last_att_maps = attention_maps[0]
last_att_maps = last_att_maps[:int(args.batch_size / 2), ...]
last_att_maps = torch.mean(last_att_maps, 1)[:, 0, 1:].view(int(args.batch_size / 2), 14, 14) # 4 196
labels_resize = torch.nn.functional.interpolate(labels.unsqueeze(1), size=(14, 14), mode='nearest')[
:int(args.batch_size / 2), ...]
labels_resize = labels_resize.squeeze(1)
attention_loss = ((1 - labels_resize) * last_att_maps).sum()
### cls seg consistency loss
last_att_maps = attention_maps[-1]
last_att_maps = torch.mean(last_att_maps, 1)[:, 0, 1:] # 8 196
last_att_maps = last_att_maps * (1 / (last_att_maps.sum() + 1e-3))
preds_resize = torch.nn.functional.interpolate(preds[:, 1, ...].unsqueeze(1), size=(14, 14),
mode='bilinear') # 8 14 14
preds_resize = preds_resize.squeeze(1).view(args.batch_size, -1) # 8 196
preds_resize = torch.softmax(preds_resize, 1)
cs_loss = ((last_att_maps - preds_resize) * (last_att_maps - preds_resize)).sum()
### activate consistent loss
preds_softmax = torch.softmax(preds, 1)
preds_resize = torch.nn.functional.interpolate(preds_softmax[:, 1, ...].unsqueeze(1), size=(14, 14),
mode='bilinear') # 8 14 14
preds_resize = preds_resize.squeeze(1).view(args.batch_size, -1) # 8 196
# preds_resize = torch.softmax(preds_resize, 1)
ac_loss = - (last_att_maps * preds_resize).sum()
### deep supervision loss
label_resize = torch.nn.functional.interpolate(labels.unsqueeze(1), size=(14, 14),
mode='nearest') # 8 14 14
label_resize = label_resize.squeeze(1)
deep_loss_seg = ce_loss(
ds_mask_logits[:int(args.batch_size / 2), ...], label_resize[:int(args.batch_size / 2), ...].long())
### sdf seg dice loss
with torch.no_grad():
gt_dis = compute_sdf(labels[:int(args.batch_size / 2)].cpu(
).numpy(), dt_preds[:int(args.batch_size / 2), 0, ...].shape)
gt_dis = torch.from_numpy(gt_dis).float().cuda()
loss_sdf = mse_loss(dt_preds[:int(args.batch_size / 2), 0, ...], gt_dis)
loss_seg = ce_loss(
preds[:int(args.batch_size / 2), ...], labels[:int(args.batch_size / 2), ...].long())
loss_label_smooth_ce_seg = label_smooth_ce_loss(preds[:int(args.batch_size / 2), ...],
labels[:int(args.batch_size / 2), ...].long())
loss_seg_dice = dice_loss(
preds[:int(args.batch_size / 2), 0, :, :], labels[:int(args.batch_size / 2)] == 0)
dis_to_mask = torch.sigmoid(-1500 * dt_preds)
### cls loss
cls_loss = ce_loss(cls_logits, cls_label)
### consistency loss
consistency_loss = torch.mean(
(torch.cat((1 - dis_to_mask, dis_to_mask), 1) - preds) ** 2)
consistency_weight = get_current_consistency_weight(epoch)
### calculate the total loss function
if args.train_mode == 'seg_only':
loss = loss_label_smooth_ce_seg
elif args.train_mode == 'cls_only':
loss = cls_loss
elif args.train_mode == 'seg+cls':
if iter_num <= args.cls_late:
loss = loss_label_smooth_ce_seg
else:
loss = loss_label_smooth_ce_seg + args.cls_weight * cls_loss
elif args.train_mode == 'seg+dual':
loss = loss_label_smooth_ce_seg + consistency_weight * consistency_loss + args.consis_weight * loss_sdf
elif args.train_mode == 'seg+cls+dual':
if iter_num <= args.cls_late:
loss = loss_label_smooth_ce_seg + consistency_weight * consistency_loss + args.consis_weight * loss_sdf
else:
loss = loss_label_smooth_ce_seg + consistency_weight * consistency_loss + args.consis_weight * loss_sdf + args.cls_weight * cls_loss
if args.att_loss != 0:
loss += args.att_loss * attention_loss
if args.cs_loss != 0:
loss += args.cs_loss * cs_loss
if args.ds != 0:
loss += args.ds * deep_loss_seg
if args.ac_loss != 0:
loss += args.ac_loss * ac_loss
if torch.isnan(loss):
continue
iter_num += 1
loss.backward()
optimizer.step()
writer.add_scalar('lr', lr, iter_num)
writer.add_scalar('loss/loss', loss, iter_num)
writer.add_scalar('loss/loss_seg', loss_seg, iter_num)
writer.add_scalar('loss/loss_dice', loss_seg_dice, iter_num)
writer.add_scalar('loss/loss_label_smooth_ce_seg', loss_label_smooth_ce_seg, iter_num)
writer.add_scalar('loss/loss_hausdorff', loss_sdf, iter_num)
writer.add_scalar('loss/consistency_weight', consistency_weight, iter_num)
writer.add_scalar('loss/consistency_loss', consistency_loss, iter_num)
writer.add_scalar('loss/cls_loss', cls_loss, iter_num)
writer.add_scalar('loss/att_loss', attention_loss, iter_num)
writer.add_scalar('loss/cs_loss', cs_loss, iter_num)
writer.add_scalar('loss/ds_loss', deep_loss_seg, iter_num)
writer.add_scalar('loss/ac_loss', deep_loss_seg, iter_num)
logging.info(
'iteration %d : loss : %f, loss_consis: %f, loss_seg_ce: %f, loss_haus: %f, loss_dice: %f, loss_cls: %f, att_loss: %f, cs_loss: %f, ds_loss: %f, ac_loss: %f' %
(iter_num, loss.item(), consistency_loss.item(), loss_seg.item(), loss_sdf.item(), loss_seg_dice.item(),
cls_loss.item(),
attention_loss.item(), cs_loss.item(), deep_loss_seg.item(), ac_loss.item()))
if iter_num % 20 == 0:
image = images[0, :, :, :]
image = (image - image.min()) / (image.max() - image.min())
writer.add_image('train/image', image, iter_num)
outputs = torch.argmax(torch.softmax(
preds, dim=1), dim=1, keepdim=True)
writer.add_image('train/mask_pred',
outputs[0, ...] * 200, iter_num)
writer.add_image('train/dt_to_mask',
dis_to_mask[0, ...] * 200, iter_num)
# print(labels[0,...].shape)
labs = labels[0, ...].unsqueeze(0) * 200
writer.add_image('train/mask_gt', labs, iter_num)
dis = gt_dis.unsqueeze(1)
dis = dis[0, :, :, :] * 200
dis = (dis - dis.min()) / (dis.max() - dis.min())
writer.add_image('train/dt_gt', dis, iter_num)
dt_pred = dt_preds[0, :, :, :] * 200
dt_pred = (dt_pred - dt_pred.min()) / (dt_pred.max() - dt_pred.min())
writer.add_image('train/dt_pred', dt_pred, iter_num)
##### Unlabel
image = images[4, :, :, :]
image = (image - image.min()) / (image.max() - image.min())
writer.add_image('train/Unlabel_image', image, iter_num)
outputs = torch.argmax(torch.softmax(
preds, dim=1), dim=1, keepdim=True)
writer.add_image('train/Unlabel_mask_pred',
outputs[4, ...] * 200, iter_num)
dt_pred = dt_preds[4, :, :, :] * 200
dt_pred = (dt_pred - dt_pred.min()) / (dt_pred.max() - dt_pred.min())
writer.add_image('train/Unlabel_dt_pred', dt_pred, iter_num)
############# Start the validation
if not test_flag:
with torch.no_grad():
print('start val!')
[vacc, vdice, vsen, vspe, vjac_score] = val_mode_seg(valloader, model, './history/{}/'.format(
get_month_day()) + args.exp_name, epoch)
logging.info("val%d: vacc=%f, vdice=%f, vsensitivity=%f, vspecifity=%f, vjac=%f \n" % \
(epoch, np.nanmean(vacc), np.nanmean(vdice), np.nanmean(vsen), np.nanmean(vspe),
np.nanmean(vjac_score)))
writer.add_scalar('val/vacc', np.nanmean(vacc), epoch)
writer.add_scalar('val/vdice', np.nanmean(vdice), epoch)
writer.add_scalar('val/vsen', np.nanmean(vsen), epoch)
writer.add_scalar('val/vspe', np.nanmean(vspe), epoch)
writer.add_scalar('val/vjac_score', np.nanmean(vjac_score), epoch)
with torch.no_grad():
print('start test!')
[vacc, vdice, vsen, vspe, vjac_score, total_acc, m_acc, s_acc, dic] = val_mode_seg_multi_scale(args,
testloader,
model,
'./history/{}/'.format(
get_month_day()) + args.exp_name,
test=True,
ph2=args.ph2_test,
logging=logging,
cls_dic=label_dic)
logging.info("test%d: tacc=%f, tdice=%f, tsensitivity=%f, tspecifity=%f, tjac=%f \n" % \
(epoch, np.nanmean(vacc), np.nanmean(vdice), np.nanmean(vsen), np.nanmean(vspe),
np.nanmean(vjac_score)))
logging.info('cls_acc=%f, m_acc=%f, s_acc=%f' % (total_acc, m_acc, s_acc))
# ############# Plot val curve
# val_jac.append(np.nanmean(vjac_score))
become_best_flag = False
if best_val_jac < np.nanmean(vjac_score):
best_val_jac = np.nanmean(vjac_score)
become_best_flag = True
become_cls_best = False
if m_acc + s_acc >= best_cls:
best_cls = m_acc + s_acc
become_cls_best = True
# if epoch % 5 == 0:
if become_best_flag:
# torch.save(model.state_dict(), path + 'CoarseSN_e' + str(epoch) + '.pth')
torch.save(model.state_dict(),
'./history/{}/'.format(get_month_day()) + args.exp_name + '/best_model.pth')
if become_cls_best:
# torch.save(model.state_dict(), path + 'CoarseSN_e' + str(epoch) + '.pth')
torch.save(model.state_dict(),
'./history/{}/'.format(get_month_day()) + args.exp_name + '/best_cls_model.pth')
if (epoch + 1) % 20 == 0:
torch.save(model.state_dict(),
'./history/{}/'.format(get_month_day()) + args.exp_name + '/last_epoch.pth')
writer.add_scalar('test/tacc', np.nanmean(vacc), epoch)
writer.add_scalar('test/tdice', np.nanmean(vdice), epoch)
writer.add_scalar('test/tsen', np.nanmean(vsen), epoch)
writer.add_scalar('test/tspe', np.nanmean(vspe), epoch)
writer.add_scalar('test/tjac_score', np.nanmean(vjac_score), epoch)
writer.add_scalar('test/cls_acc', total_acc, epoch)
writer.add_scalar('test/cls_macc', m_acc, epoch)
writer.add_scalar('test/cls_sacc', s_acc, epoch)
writer.add_scalar('test/cls_mauc', dic['mauc'], epoch)
writer.add_scalar('test/cls_msen', dic['msens'], epoch)
writer.add_scalar('test/cls_mspec', dic['mspec'], epoch)
writer.add_scalar('test/cls_sauc', dic['sauc'], epoch)
writer.add_scalar('test/cls_ssen', dic['ssens'], epoch)
writer.add_scalar('test/cls_sspec', dic['sspec'], epoch)
if test_flag:
print('Finish Testing')
exit(0)