-
Notifications
You must be signed in to change notification settings - Fork 4
Expand file tree
/
Copy pathLA_train.py
More file actions
492 lines (382 loc) · 20.1 KB
/
LA_train.py
File metadata and controls
492 lines (382 loc) · 20.1 KB
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
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
from asyncore import write
import imp
import os
from sre_parse import SPECIAL_CHARS
import sys
from xml.etree.ElementInclude import default_loader
from tqdm import tqdm
from tensorboardX import SummaryWriter
import shutil
import argparse
import logging
import random
import numpy as np
from medpy import metric
import torch
import torch.optim as optim
from torchvision import transforms
import torch.nn.functional as F
import torch.backends.cudnn as cudnn
import torch.nn as nn
import pdb
from yaml import parse
from skimage.measure import label
from torch.utils.data import DataLoader
from torch.autograd import Variable
from utils import losses, ramps, feature_memory, contrastive_losses, test_3d_patch
# from dataloaders.dataset import *
from dataloaders.LADataset import LAHeart
from utils.BCP_utils import context_mask, mix_loss, parameter_sharing, update_ema_variables
from utils.LA_utils import to_cuda
from utils.BCP_utils import *
from pancreas.losses import *
from pancreas.Vnet import VNet
from networks.ResVNet import ResVNet
parser = argparse.ArgumentParser()
parser.add_argument('--root_path', type=str, default='./Datasets/la/data', help='Name of Dataset')
parser.add_argument('--exp', type=str, default='SDCL', help='exp_name')
parser.add_argument('--model', type=str, default='VNet', help='model_name')
parser.add_argument('--pre_max_iteration', type=int, default=2000, help='maximum pre-train iteration to train')
parser.add_argument('--self_max_iteration', type=int, default=15000, help='maximum self-train iteration to train')
parser.add_argument('--max_samples', type=int, default=80, help='maximum samples to train')
parser.add_argument('--labeled_bs', type=int, default=4, help='batch_size of labeled data per gpu')
parser.add_argument('--batch_size', type=int, default=8, help='batch_size per gpu')
parser.add_argument('--base_lr', type=float, default=1e-3, help='maximum epoch number to train')
parser.add_argument('--deterministic', type=int, default=1, help='whether use deterministic training')
parser.add_argument('--labelnum', type=int, default=8, help='trained samples')
parser.add_argument('--gpu', type=str, default='0', help='GPU to use')
parser.add_argument('--seed', type=int, default=1345, help='random seed')
parser.add_argument('--consistency', type=float, default=1.0, help='consistency')
parser.add_argument('--consistency_rampup', type=float, default=40.0, help='consistency_rampup')
parser.add_argument('--magnitude', type=float, default='10.0', help='magnitude')
# -- setting of BCP
parser.add_argument('--u_weight', type=float, default=0.5, help='weight of unlabeled pixels')
parser.add_argument('--mask_ratio', type=float, default=2 / 3, help='ratio of mask/image')
# -- setting of mixup
parser.add_argument('--u_alpha', type=float, default=2.0, help='unlabeled image ratio of mixuped image')
parser.add_argument('--loss_weight', type=float, default=0.5, help='loss weight of unimage term')
args = parser.parse_args()
def create_Vnet(ema=False):
net = VNet(n_channels=1, n_classes=2, normalization='instancenorm', has_dropout=True)
net = nn.DataParallel(net)
model = net.cuda()
if ema:
for param in model.parameters():
param.detach_()
return model
def create_ResVnet(ema=False):
net = ResVNet(n_channels=1, n_classes=2, normalization='instancenorm', has_dropout=True)
net = nn.DataParallel(net)
model = net.cuda()
if ema:
for param in model.parameters():
param.detach_()
return model
def get_cut_mask(out, thres=0.5, nms=0):
probs = F.softmax(out, 1)
masks = (probs >= thres).type(torch.int64)
masks = masks[:, 1, :, :].contiguous()
if nms == 1:
masks = LargestCC_pancreas(masks)
return masks
def get_cut_mask_two(out1, out2, thres=0.5, nms=0):
probs1 = F.softmax(out1, 1)
probs2 = F.softmax(out2, 1)
probs = (probs1 + probs2) / 2
masks = (probs >= thres).type(torch.int64)
masks = masks[:, 1, :, :].contiguous()
if nms == 1:
masks = LargestCC_pancreas(masks)
return masks
def LargestCC_pancreas(segmentation):
N = segmentation.shape[0]
batch_list = []
for n in range(N):
n_prob = segmentation[n].detach().cpu().numpy()
labels = label(n_prob)
if labels.max() != 0:
largestCC = labels == np.argmax(np.bincount(labels.flat)[1:]) + 1
else:
largestCC = n_prob
batch_list.append(largestCC)
return torch.Tensor(batch_list).cuda()
def save_net_opt(net, optimizer, path):
state = {
'net': net.state_dict(),
'opt': optimizer.state_dict(),
}
torch.save(state, str(path))
def load_net_opt(net, optimizer, path):
state = torch.load(str(path))
net.load_state_dict(state['net'])
optimizer.load_state_dict(state['opt'])
def load_net(net, path):
state = torch.load(str(path))
net.load_state_dict(state['net'])
def get_current_consistency_weight(epoch):
# Consistency ramp-up from https://arxiv.org/abs/1610.02242
return args.consistency * ramps.sigmoid_rampup(epoch, args.consistency_rampup)
train_data_path = args.root_path
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
pre_max_iterations = args.pre_max_iteration
self_max_iterations = args.self_max_iteration
base_lr = args.base_lr
CE = nn.CrossEntropyLoss(reduction='none')
if args.deterministic:
cudnn.benchmark = False
cudnn.deterministic = True
torch.manual_seed(args.seed)
torch.cuda.manual_seed(args.seed)
random.seed(args.seed)
np.random.seed(args.seed)
patch_size = (112, 112, 80)
num_classes = 2
def load_net_opt(net, optimizer, path):
state = torch.load(str(path))
net.load_state_dict(state['net'])
optimizer.load_state_dict(state['opt'])
def save_net_opt(net, optimizer, path, epoch):
state = {
'net': net.state_dict(),
'opt': optimizer.state_dict(),
'epoch': epoch,
}
torch.save(state, str(path))
def get_XOR_region(mixout1, mixout2):
s1 = torch.softmax(mixout1, dim=1)
l1 = torch.argmax(s1, dim=1)
s2 = torch.softmax(mixout2, dim=1)
l2 = torch.argmax(s2, dim=1)
diff_mask = (l1 != l2)
return diff_mask
def cmp_dice_loss(score, target):
target = target.float()
smooth = 1e-5
intersect = torch.sum(score * target)
y_sum = torch.sum(target * target)
z_sum = torch.sum(score * score)
loss = (2 * intersect + smooth) / (z_sum + y_sum + smooth)
loss = 1 - loss
return loss
def pre_train(args, snapshot_path):
model = create_Vnet()
model2 = create_ResVnet()
c_batch_size = 2
trainset_lab_a = LAHeart(train_data_path, "./Datasets/la/data_split", split='train_lab', logging=logging)
lab_loader_a = DataLoader(trainset_lab_a, batch_size=c_batch_size, shuffle=False, num_workers=0, drop_last=True)
trainset_lab_b = LAHeart(train_data_path, "./Datasets/la/data_split", split='train_lab', reverse=True, logging=logging)
lab_loader_b = DataLoader(trainset_lab_b, batch_size=c_batch_size, shuffle=False, num_workers=0, drop_last=True)
optimizer = optim.Adam(model.parameters(), lr=1e-3)
optimizer2 = optim.Adam(model2.parameters(), lr=1e-3)
DICE = losses.mask_DiceLoss(nclass=2)
model.train()
model2.train()
logging.info("{} iterations per epoch".format(len(lab_loader_a)))
iter_num = 0
best_dice = 0
best_dice2 = 0
max_epoch = 81
iterator = tqdm(range(1, max_epoch), ncols=70)
for epoch_num in iterator:
logging.info("\n")
for step, ((img_a, lab_a), (img_b, lab_b)) in enumerate(zip(lab_loader_a, lab_loader_b)):
img_a, img_b, lab_a, lab_b = img_a.cuda(), img_b.cuda(), lab_a.cuda(), lab_b.cuda()
with torch.no_grad():
img_mask, loss_mask = context_mask(img_a, args.mask_ratio)
"""Mix Input"""
volume_batch = img_a * img_mask + img_b * (1 - img_mask)
label_batch = lab_a * img_mask + lab_b * (1 - img_mask)
outputs, _ = model(volume_batch)
loss_ce = F.cross_entropy(outputs, label_batch)
loss_dice = DICE(outputs, label_batch)
loss = (loss_ce + loss_dice) / 2
outputs2, _ = model2(volume_batch)
loss_ce2 = F.cross_entropy(outputs2, label_batch)
loss_dice2 = DICE(outputs2, label_batch)
loss2 = (loss_ce2 + loss_dice2) / 2
iter_num += 1
optimizer.zero_grad()
loss.backward()
optimizer.step()
optimizer2.zero_grad()
loss2.backward()
optimizer2.step()
logging.info(
'iteration %d : loss: %03f, loss_dice: %03f, loss_ce: %03f' % (iter_num, loss, loss_dice, loss_ce))
if epoch_num % 5 == 0:
model.eval()
dice_sample = test_3d_patch.var_all_case_LA(model, num_classes=num_classes, patch_size=patch_size,
stride_xy=18, stride_z=4)
if dice_sample > best_dice:
best_dice = round(dice_sample, 4)
save_mode_path = os.path.join(snapshot_path, 'iter_{}_dice_{}.pth'.format(iter_num, best_dice))
save_best_path = os.path.join(snapshot_path, 'best_model.pth'.format(args.model))
save_net_opt(model, optimizer, save_mode_path, epoch_num)
save_net_opt(model, optimizer, save_best_path, epoch_num)
logging.info("save best model to {}".format(save_mode_path))
model.train()
model2.eval()
dice_sample2 = test_3d_patch.var_all_case_LA(model2, num_classes=num_classes, patch_size=patch_size,
stride_xy=18, stride_z=4)
if dice_sample2 > best_dice2:
best_dice2 = round(dice_sample2, 4)
save_mode_path = os.path.join(snapshot_path, 'iter_{}_dice_{}_resnet.pth'.format(iter_num, best_dice2))
save_best_path = os.path.join(snapshot_path, 'best_model_resnet.pth'.format(args.model))
save_net_opt(model2, optimizer2, save_mode_path, epoch_num)
save_net_opt(model2, optimizer2, save_best_path, epoch_num)
logging.info("save best resnet model to {}".format(save_mode_path))
model2.train()
def self_train(args, pre_snapshot_path, self_snapshot_path):
model1 = create_Vnet()
model2 = create_ResVnet()
ema_model1 = create_Vnet(ema=True).cuda()
c_batch_size = 2
trainset_lab_a = LAHeart(train_data_path, "./Datasets/la/data_split", split='train_lab', logging=logging)
lab_loader_a = DataLoader(trainset_lab_a, batch_size=c_batch_size, shuffle=False, num_workers=0, drop_last=True)
trainset_lab_b = LAHeart(train_data_path, "./Datasets/la/data_split", split='train_lab', reverse=True, logging=logging)
lab_loader_b = DataLoader(trainset_lab_b, batch_size=c_batch_size, shuffle=False, num_workers=0, drop_last=True)
trainset_unlab_a = LAHeart(train_data_path, "./Datasets/la/data_split", split='train_unlab', logging=logging)
unlab_loader_a = DataLoader(trainset_unlab_a, batch_size=c_batch_size, shuffle=False, num_workers=0, drop_last=True)
trainset_unlab_b = LAHeart(train_data_path, "./Datasets/la/data_split", split='train_unlab', reverse=True, logging=logging)
unlab_loader_b = DataLoader(trainset_unlab_b, batch_size=c_batch_size, shuffle=False, num_workers=0, drop_last=True)
optimizer = optim.Adam(model1.parameters(), lr=1e-3)
optimizer2 = optim.Adam(model2.parameters(), lr=1e-3)
pretrained_model = os.path.join(pre_snapshot_path, 'best_model.pth')
pretrained_model2 = os.path.join(pre_snapshot_path, 'best_model_resnet.pth')
load_net_opt(model1, optimizer, pretrained_model)
load_net_opt(model2, optimizer2, pretrained_model2)
load_net_opt(ema_model1, optimizer, pretrained_model)
model1.train()
model2.train()
ema_model1.train()
logging.info("{} iterations per epoch".format(len(lab_loader_a)))
iter_num = 0
best_dice = 0
best_dice2 = 0
mean_best_dice = 0
max_epoch = 276
iterator = tqdm(range(1, max_epoch), ncols=70)
for epoch in iterator:
logging.info("\n")
for step, ((img_a, lab_a), (img_b, lab_b), (unimg_a, unlab_a), (unimg_b, unlab_b)) in enumerate(
zip(lab_loader_a, lab_loader_b, unlab_loader_a, unlab_loader_b)):
img_a, lab_a, img_b, lab_b, unimg_a, unlab_a, unimg_b, unlab_b = to_cuda(
[img_a, lab_a, img_b, lab_b, unimg_a, unlab_a, unimg_b, unlab_b])
with torch.no_grad():
unoutput_a_1, _ = ema_model1(unimg_a)
unoutput_b_1, _ = ema_model1(unimg_b)
plab_a = get_cut_mask(unoutput_a_1, nms=1)
plab_b = get_cut_mask(unoutput_b_1, nms=1)
img_mask, loss_mask = context_mask(img_a, args.mask_ratio)
mixl_img = unimg_a * img_mask + img_b * (1 - img_mask)
mixu_img = img_a * img_mask + unimg_b * (1 - img_mask)
outputs_l, _ = model1(mixl_img)
outputs_u, _ = model1(mixu_img)
loss_l = mix_loss(outputs_l, plab_a.long(), lab_b, loss_mask, u_weight=args.u_weight, unlab=True)
loss_u = mix_loss(outputs_u, lab_a, plab_b.long(), loss_mask, u_weight=args.u_weight)
outputs_l_2, _ = model2(mixl_img)
outputs_u_2, _ = model2(mixu_img)
loss_l_2 = mix_loss(outputs_l_2, plab_a.long(), lab_b, loss_mask, u_weight=args.u_weight, unlab=True)
loss_u_2 = mix_loss(outputs_u_2, lab_a, plab_b.long(), loss_mask, u_weight=args.u_weight)
with torch.no_grad():
diff_mask1 = get_XOR_region(outputs_l, outputs_l_2)
diff_mask2 = get_XOR_region(outputs_u, outputs_u_2)
net1_mse_loss_lab = mix_mse_loss(outputs_l, plab_a.long(), lab_b, loss_mask, unlab=True,
diff_mask=diff_mask1)
net1_kl_loss_lab = mix_max_kl_loss(outputs_l, plab_a.long(), lab_b, loss_mask, unlab=True,
diff_mask=diff_mask1)
net1_mse_loss_unlab = mix_mse_loss(outputs_u, lab_a, plab_b.long(), loss_mask, diff_mask=diff_mask2)
net1_kl_loss_unlab = mix_max_kl_loss(outputs_u, lab_a, plab_b.long(), loss_mask, diff_mask=diff_mask2)
net2_mse_loss_lab = mix_mse_loss(outputs_l_2, plab_a.long(), lab_b, loss_mask, unlab=True,
diff_mask=diff_mask1)
net2_kl_loss_lab = mix_max_kl_loss(outputs_l_2, plab_a.long(), lab_b, loss_mask, unlab=True,
diff_mask=diff_mask1)
net2_mse_loss_unlab = mix_mse_loss(outputs_u_2, lab_a, plab_b.long(), loss_mask, diff_mask=diff_mask2)
net2_kl_loss_unlab = mix_max_kl_loss(outputs_u_2, lab_a, plab_b.long(), loss_mask, diff_mask=diff_mask2)
loss = (loss_l + loss_u) + 0.5 * (net1_mse_loss_lab + net1_mse_loss_unlab) + 0.05 * (
net1_kl_loss_lab + net1_kl_loss_unlab)
loss_2 = (loss_l_2 + loss_u_2) + 0.5 * (net2_mse_loss_lab + net2_mse_loss_unlab) + 0.05 * (
net2_kl_loss_lab + net2_kl_loss_unlab)
iter_num += 1
optimizer.zero_grad()
loss.backward()
optimizer.step()
optimizer2.zero_grad()
loss_2.backward()
optimizer2.step()
logging.info('epoch %d iteration %d : loss: %03f, loss_l: %03f, loss_u: %03f \
net1_mse_loss_lab: %.4f, net1_mse_loss_unlab: %.4f, net1_kl_loss_lab: %.4f, net1_kl_loss_unlab: %.4f \
' % (epoch, iter_num, loss, loss_l, loss_u, net1_mse_loss_lab.item(), net1_mse_loss_unlab.item(),
net1_kl_loss_lab.item(), net1_kl_loss_unlab.item()))
update_ema_variables(model1, ema_model1, 0.99)
if epoch % 5 == 0:
model1.eval()
model2.eval()
dice_sample = test_3d_patch.var_all_case_LA(model1, num_classes=num_classes, patch_size=patch_size,
stride_xy=18, stride_z=4)
dice_sample2 = test_3d_patch.var_all_case_LA(model2, num_classes=num_classes, patch_size=patch_size,
stride_xy=18, stride_z=4)
mean_dice_sample = test_3d_patch.var_all_case_LA_mean(model1, model2, num_classes=num_classes,
patch_size=patch_size, stride_xy=18, stride_z=4)
if dice_sample > best_dice:
best_dice = round(dice_sample, 4)
save_mode_path = os.path.join(self_snapshot_path, 'iter_{}_dice_{}.pth'.format(iter_num, best_dice))
save_best_path = os.path.join(self_snapshot_path, 'best_model.pth')
torch.save(model1.state_dict(), save_mode_path)
torch.save(model1.state_dict(), save_best_path)
logging.info("save best model to {}".format(save_mode_path))
logging.info("cur dice %.4f, max dice %.4f" % (dice_sample, best_dice))
if dice_sample2 > best_dice2:
best_dice2 = round(dice_sample2, 4)
save_mode_path = os.path.join(self_snapshot_path,
'iter_{}_dice_{}_res.pth'.format(iter_num, best_dice2))
save_best_path = os.path.join(self_snapshot_path, 'best_model_res.pth')
torch.save(model2.state_dict(), save_mode_path)
torch.save(model2.state_dict(), save_best_path)
logging.info("resnet cur dice %.4f, max dice %.4f" % (dice_sample2, best_dice2))
if mean_dice_sample > mean_best_dice:
mean_best_dice = round(mean_dice_sample, 4)
save_mode_path1 = os.path.join(self_snapshot_path,
'iter_{}_dice_{}_v.pth'.format(iter_num, mean_best_dice))
save_best_path1 = os.path.join(self_snapshot_path, 'best_model_v.pth')
save_mode_path2 = os.path.join(self_snapshot_path,
'iter_{}_dice_{}_r.pth'.format(iter_num, mean_best_dice))
save_best_path2 = os.path.join(self_snapshot_path, 'best_model_r.pth')
torch.save(model1.state_dict(), save_mode_path1)
torch.save(model1.state_dict(), save_best_path1)
torch.save(model2.state_dict(), save_mode_path2)
torch.save(model2.state_dict(), save_best_path2)
logging.info("mean save best model to {}".format(save_mode_path1))
logging.info("mean cur dice %.4f, max dice %.4f" % (mean_dice_sample, mean_best_dice))
model1.train()
model2.train()
if __name__ == "__main__":
## make logger file
pre_snapshot_path = "./model/SDCL/LA_{}_{}_labeled/pre_train".format(args.exp, args.labelnum)
self_snapshot_path = "./model/SDCL/LA_{}_{}_labeled/self_train".format(args.exp, args.labelnum)
print("Starting SDCL training.")
for snapshot_path in [pre_snapshot_path, self_snapshot_path]:
if not os.path.exists(snapshot_path):
os.makedirs(snapshot_path)
if os.path.exists(snapshot_path + '/code'):
shutil.rmtree(snapshot_path + '/code')
shutil.copy('../code/LA_train.py', self_snapshot_path)
# -- Pre-Training
logging.basicConfig(filename=pre_snapshot_path + "/log.txt", 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))
pre_train(args, pre_snapshot_path)
# -- Self-training
logging.basicConfig(filename=self_snapshot_path + "/log.txt", 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))
if args.deterministic:
cudnn.benchmark = False
cudnn.deterministic = True
torch.manual_seed(args.seed)
torch.cuda.manual_seed(args.seed)
random.seed(args.seed)
np.random.seed(args.seed)
self_train(args, pre_snapshot_path, self_snapshot_path)