-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
390 lines (312 loc) · 17.4 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
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
import os
import time
import argparse
import numpy as np
import logging
import random
from scipy.ndimage import zoom
import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader
from utils import Segloss, ContrastiveLoss
from dataset import MOODDataset
from networks import *
logging.basicConfig(
format='%(asctime)s : %(levelname)s : %(message)s',
level=logging.INFO,
datefmt="%Y-%m-%d %H:%M:%S"
)
def pretask(model, device, scatter_patch, orig_patch, optimizer, scheduler, patch_size, writer):
model.train()
optimizer.zero_grad()
correct = 0.0
scatter_patch, orig_patch = scatter_patch.to(device), orig_patch.to(device)
pred_seg = model(scatter_patch.float())
# calculate the loss
BCELoss = nn.BCELoss()
reconLoss = BCELoss(pred_seg, orig_patch.float())
total_loss = reconLoss
logging.info(f"Recon Loss: {total_loss.item():.6f}")
# optimize the parameters
total_loss.backward()
optimizer.step()
scheduler.step()
return total_loss.item()
def train(model, pre_model, device, img_patch, imgGT_patch, img_cls, aug_patch, augGT_patch, aug_cls, optimizer, scheduler, patch_size, lamda, pre_optimizer, pre_scheduler, rotImage, rotAug, pre_augGT, lamda2, preaugGT_patch):
model.train()
pre_model.train()
optimizer.zero_grad()
pre_optimizer.zero_grad()
correct = 0.0
aug_index = np.transpose(np.nonzero(augGT_patch[0, :, :, :, :]))
img_patch, imgGT_patch, img_cls, rotImage = img_patch.to(device), imgGT_patch.to(device), img_cls.to(device), rotImage.to(device)
aug_patch, augGT_patch, aug_cls, rotAug = aug_patch.to(device), augGT_patch.to(device), aug_cls.to(device), rotAug.to(device)
activation = {}
def get_activation(name):
def hook(model, input, output):
activation[name + str(input[0].device)] = output
return hook
def merge_activation(name, device):
activation_list = []
for key in activation.keys():
if name in key:
activation_list.append(activation[key].to(device))
activation[name] = torch.cat(activation_list, 0)
latent_size = 4
batch_size = img_patch.shape[0]
#Normal Image: 3D Network
pred_normal_seg, pred_normal_cls, mood_latent_img = model(img_patch.float())
#Normal Image: 2D Network
sliceImgs = torch.reshape(rotImage, (batch_size * patch_size * 2, rotImage.shape[1], patch_size*2, patch_size*2))
outImg, latent_img = pre_model(sliceImgs.float())
latent_img = torch.reshape(latent_img, (batch_size, patch_size * 2, latent_img.shape[1], latent_size, latent_size))
latent_img = torch.transpose(latent_img, 1,2)
latent_img = latent_img.reshape(-1, 512, latent_img.shape[1]//512, 4, latent_img.shape[2]//4, latent_size, latent_size).permute(0,1,3,2,4,5,6)
latent_img = latent_img.reshape(-1, latent_img.shape[3] * latent_img.shape[4], latent_size * latent_size).permute(0,2,1)
latent_img = latent_img.reshape(-1, latent_img.shape[2])
reshape_latent_img = torch.amax(latent_img, dim=1).reshape(-1, 512, 4, latent_size, latent_size)
#Abnormal Image: 3D Network
pred_aug_seg, pred_aug_cls, mood_latent_aug = model(aug_patch.float())
#Abnormal Image: 2D Network
sliceAugs = torch.reshape(rotAug, (batch_size * patch_size * 2, rotAug.shape[1], patch_size*2, patch_size*2))
outAug, latent_aug = pre_model(sliceAugs.float())
latent_aug = torch.reshape(latent_aug, (batch_size, patch_size * 2, latent_aug.shape[1], latent_size, latent_size))
latent_aug = torch.transpose(latent_aug, 1,2)
latent_aug = latent_aug.reshape(-1, 512, latent_aug.shape[1]//512, 4, latent_aug.shape[2]//4, latent_size, latent_size).permute(0,1,3,2,4,5,6)
latent_aug = latent_aug.reshape(-1, latent_aug.shape[3] * latent_aug.shape[4], latent_size * latent_size).permute(0,2,1)
latent_aug = latent_aug.reshape(-1, latent_aug.shape[2])
reshape_latent_aug = torch.amax(latent_aug, dim=1).reshape(-1, 512, 4, latent_size, latent_size)
#Similarity Loss
rep_2d = torch.cat([reshape_latent_img, reshape_latent_aug, reshape_latent_img, reshape_latent_aug], dim = 0)
rep_3d = torch.cat([mood_latent_img, mood_latent_aug, mood_latent_aug, mood_latent_img], dim = 0)
rep_2d, rep_3d = torch.reshape(rep_2d, (batch_size*4, -1)), torch.reshape(rep_3d, (batch_size*4, -1))
SimLoss_ = ContrastiveLoss(rep_2d.to(device), rep_3d.to(device), batch_size*4)
SimLoss = SimLoss_ * (100 - lamda2) / 8000.0
# calculate the loss
bceLoss = nn.BCELoss()
pred_normal_cls = torch.flatten(pred_normal_cls)
pred_aug_cls = torch.flatten(pred_aug_cls)
pred_2d_normal_cls = torch.flatten(outImg)
pred_2d_aug_cls = torch.flatten(outAug)
pre_augGT = torch.flatten(pre_augGT.float())
pre_imgGT = img_cls.repeat(patch_size)
normal_seg_loss = bceLoss(pred_normal_seg, imgGT_patch.float())
aug_seg_loss = bceLoss(pred_aug_seg, augGT_patch.float())
seg_loss = normal_seg_loss + aug_seg_loss
normal_cls_loss = bceLoss(pred_normal_cls, img_cls.float())
aug_cls_loss = bceLoss(pred_aug_cls, aug_cls.float())
cls_loss = normal_cls_loss + aug_cls_loss
countAug = np.count_nonzero(pre_augGT)
pos_weight = torch.ones([1]) * ((pre_augGT.shape[0] - countAug) / (countAug))
bceLossW = nn.BCEWithLogitsLoss(pos_weight=pos_weight.to('cuda:0'))
aug_cls_loss_2d = bceLossW(pred_2d_aug_cls.to('cuda:0'), pre_augGT.to('cuda:0'))
cls_loss_2d = aug_cls_loss_2d
# aviod the trivial solution
_, FN = Segloss(pred_aug_seg, augGT_patch.float(), patch_size)
total_loss = seg_loss + FN * lamda + cls_loss + SimLoss
logging.info(f"Cls BCE Loss: {cls_loss.item():.6f} | Seg BCE Loss: {seg_loss.item():.6f} | FN loss {FN.item() * lamda:.6f} | Sim loss {SimLoss.item():.6f} | 2d Cls loss {cls_loss_2d.item() :.6f} | Pos weight {pos_weight.item() :.4f")
for i in range (pred_normal_cls.shape[0]):
if pred_normal_cls[i].round() == img_cls[i]:
correct += 0.5
if pred_aug_cls[i].round() == aug_cls[i]:
correct += 0.5
# optimize the parameters
total_loss.backward()
optimizer.step()
pre_optimizer.step()
scheduler.step()
pre_scheduler.step()
return total_loss.item(), correct
def test(model, device, img_patch, imgGT_patch, img_cls, aug_patch, augGT_patch, aug_cls):
torch.cuda.empty_cache()
model.eval()
correct = 0.0
img_patch, imgGT_patch, img_cls = img_patch.to(device), imgGT_patch.to(device), img_cls.to(device)
aug_patch, augGT_patch, aug_cls = aug_patch.to(device), augGT_patch.to(device), aug_cls.to(device)
with torch.no_grad():
pred_normal_seg, pred_normal_cls = model(img_patch.float())
pred_aug_seg, pred_aug_cls = model(aug_patch.float())
pred_normal_cls = torch.flatten(pred_normal_cls)
pred_aug_cls = torch.flatten(pred_aug_cls)
# calculate the loss
bceLoss = nn.BCELoss()
normal_seg_loss = bceLoss(pred_normal_seg, imgGT_patch.float())
aug_seg_loss = bceLoss(pred_aug_seg, augGT_patch.float())
seg_loss = normal_seg_loss + aug_seg_loss
normal_cls_loss = bceLoss(pred_normal_cls, img_cls.float())
aug_cls_loss = bceLoss(pred_aug_cls, aug_cls.float())
cls_loss = normal_cls_loss + aug_cls_loss
total_loss = seg_loss + cls_loss
logging.info(f"*TEST* Cls BCE Loss: {cls_loss.item():.6f} | Seg BCE Loss: {seg_loss.item():.6f}")
for i in range (pred_normal_cls.shape[0]):
if pred_normal_cls[i].round() == img_cls[i]:
correct += 0.5
if pred_aug_cls[i].round() == aug_cls[i]:
correct += 0.5
torch.cuda.empty_cache()
return seg_loss.item(), cls_loss.item(), total_loss.item(), correct
if __name__ == "__main__":
# Version of Pytorch
logging.info("Pytorch Version:%s" % torch.__version__)
# Training args
"""##############################################################################################################################
Brain = Put the resized into 128 x 128 x 128 images (2D Network: 128 x 128 images, 3D Network: 64 x 64 x 64 with patch size 64)
Abdom = Put the resized into 256 x 256 x 256 images (2D Network: 256 x 256 images, 3D Network: 128 x 128 x 128 volumes with patch size 64)
###################################################################################################################################################"""
parser = argparse.ArgumentParser(description='MOOD training')
parser.add_argument('--dataset', type=str, default='/root/MOOD2021/dataset/brain_train',
help='path of processed dataset')
parser.add_argument('--weight', type=str, default='./weights',
help='path of the weights folder')
parser.add_argument('--checkpoints', type=str, default='./checkpoints',
help='path of training snapshot')
parser.add_argument('--batch-size', type=int, default=1, metavar='N',
help='input batch size for training (default: 1)')
parser.add_argument('--epoches', type=int, default=100, metavar='N',
help='number of epoches to train (default: 1000)')
parser.add_argument('--log-interval', type=int, default=1, metavar='N',
help='number of iterations to log (default: 1000)')
parser.add_argument('--lr', type=float, default=1e-3, metavar='LR',
help='learning rate (default: 0.001)')
parser.add_argument('--seed', type=int, default=1, metavar='S',
help='random seed (default: 0)')
parser.add_argument('--gpu', type=str, default='0', metavar='N',
help='Select the GPU (defualt 0)')
parser.add_argument('--category', type=str, default='brain', metavar='N',
help='Select the category brain or abdom (defualt brain)')
parser.add_argument('--patch-size', type=int, default=128, metavar='N',
help='patch-size (default=128')
parser.add_argument('--pretrain', action='store_true',
help='pretraining step')
args = parser.parse_args()
os.environ["CUDA_DEVICE_ORDER"]="PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu
# set random seed for reproducibility
torch.cuda.manual_seed_all(args.seed)
torch.manual_seed(args.seed)
np.random.seed(args.seed)
random.seed(args.seed)
# Use GPU if it is available
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
logging.info("Torch use the device: %s" % device)
if args.category == 'brain':
model = Unet_brain(num_channels=64).to(device)
model_2d = resnet50().to(device)
patch_size = 64
elif args.category == 'abdom':
model = Unet_abdom(num_channels=64).to(device)
model_2d = resnet50().to(device)
patch_size = 64
else:
logging.info("Choose the Correct Category")
model = nn.DataParallel(model)
model_2d = nn.DataParallel(model_2d)
batch_size = args.batch_size
if args.pretrain:
train_dataset = MOODDataset(args.dataset, subset='train', category = args.category, pretrain = args.pretrain)
test_dataset = MOODDataset(args.dataset, subset='valid', category = args.category, pretrain = args.pretrain)
else:
train_dataset = MOODDataset(args.dataset, subset='train', category = args.category)
test_dataset = MOODDataset(args.dataset, subset='valid', category = args.category)
generator = torch.Generator()
generator.manual_seed(args.seed)
train_loader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True, num_workers=0, pin_memory=True, generator=generator)
test_loader = DataLoader(test_dataset, batch_size=batch_size, shuffle=False, num_workers=0, pin_memory=True, generator=generator)
train_iterator = iter(train_loader)
total_iteration = args.epoches * len(train_loader)
train_interval = args.log_interval * len(train_loader)
logging.info(f"total iter: {total_iteration}")
# optimizer
optimizer = optim.Adam(model.parameters(), lr=args.lr)
optimizer_2d = optim.Adam(model_2d.parameters(), lr=args.lr)
iteration = 1
best_train_loss, best_test_loss = float('inf'), float('inf')
scheduler = optim.lr_scheduler.OneCycleLR(optimizer, max_lr=args.lr*10, steps_per_epoch=len(train_loader), epochs=args.epoches, anneal_strategy='linear')
scheduler_2d = optim.lr_scheduler.OneCycleLR(optimizer_2d, max_lr=args.lr*10, steps_per_epoch=len(train_loader), epochs=args.epoches, anneal_strategy='linear')
epoch_train_loss = []
epoch_test_loss = []
epoch_test_segloss = []
epoch_test_clsloss = []
correct_train_count = 0
correct_test_count = 0
epoch_train_accuracy = 0.
epoch_test_accuracy = 0.
v_clsloss = 0.
start_time = time.time()
# Seed initializaiton for patch reproducibility
torch.cuda.manual_seed_all(args.seed)
torch.manual_seed(args.seed)
np.random.seed(args.seed)
random.seed(args.seed)
while iteration <= total_iteration:
# 0 epoch: 1.0 ~ max epoch* 0.3: 0.05 (linear)
lamda = 100 * (1.0 - iteration /(args.epoches * 0.3 * len(train_loader)))
lamda = 0.05 if lamda < 0 else lamda
try:
if args.pretrain:
img_patch, imgGT_patch, orig_patch = next(train_iterator)
else:
# Samples the batch
img_patch, img_2d, imgGT_patch, img_cls, aug_patch, aug_2d, augGT_patch, aug_cls, augGT_2d_cls, augGT_2d = next(train_iterator)
except StopIteration:
# restart the generator if the previous generator is exhausted.
train_iterator = iter(train_loader)
if args. pretrain:
img_patch, imgGT_patch, orig_patch = next(train_iterator)
else:
img_patch, img_2d, imgGT_patch, img_cls, aug_patch, aug_2d, augGT_patch, aug_cls, augGT_2d_cls, augGT_2d = next(train_iterator)
if args.pretrain:
t_total_loss = pretask(model, device, img_patch, orig_patch, optimizer, scheduler, patch_size)
else:
t_total_loss, t_correct = train(model, model_2d, device, img_patch, imgGT_patch, img_cls, aug_patch, augGT_patch, aug_cls, optimizer, scheduler, patch_size, lamda, batch_size, optimizer_2d, scheduler_2d, img_2d, aug_2d, augGT_2d_cls, augGT_2d)
if (iteration % train_interval == 0):
avg_train_loss = sum(epoch_train_loss) / len(epoch_train_loss)
epoch_train_accuracy = (correct_train_count / (args.log_interval * len(train_dataset))) * 100
logging.info(f'Iter {iteration / train_interval}-{total_iteration / train_interval}: \t Loss: {avg_train_loss:.6f}\t')
if avg_train_loss < best_train_loss:
best_train_loss = avg_train_loss
logging.info(f'--- Saving model at Avg Train Loss:{avg_train_loss:.6f} ---')
torch.save(model.state_dict(), os.path.join(args.weight, './mood_best_train_' + args.identifier +'.pth'))
# validation process
test_iterator = iter(test_loader)
for i in range(len(test_loader)):
if args.pretrain:
v_total_loss = pretask(model, device, img_patch, orig_patch, optimizer, scheduler, patch_size)
epoch_test_loss.append(v_total_loss)
else:
img_patch, img_2d, imgGT_patch, img_cls, aug_patch, aug_2d, augGT_patch, aug_cls, augGT_2d_cls, augGT_2d = next(test_iterator)
v_segloss, v_clsloss, v_total_loss, v_correct = test(model, device, img_patch, imgGT_patch, img_cls, aug_patch, augGT_patch, aug_cls)
epoch_test_loss.append(v_total_loss)
epoch_test_segloss.append(v_segloss)
epoch_test_clsloss.append(v_clsloss)
correct_test_count += v_correct
avg_test_loss = sum(epoch_test_loss) / len(epoch_test_loss)
avg_test_segloss = sum(epoch_test_segloss) / len(epoch_test_segloss)
avg_test_clsloss = sum(epoch_test_clsloss) / len(epoch_test_clsloss)
epoch_test_accuracy = (correct_test_count / len(test_dataset)) * 100
logging.info(f'Iter {iteration / train_interval}-{total_iteration / train_interval} eval: \t Loss: {avg_test_loss:.6f}\t')
if avg_test_loss < best_test_loss:
best_test_loss = avg_test_loss
logging.info(f'--- Saving model at Avg Valid Loss:{avg_test_loss:.6f} ---')
torch.save(model.state_dict(), os.path.join(args.weight, './mood_best_valid_' + args.identifier +'.pth'))
# save snapshot for resume training
logging.info('--- Saving snapshot ---')
torch.save({
'iteration': iteration+1,
'model_state_dict': model.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
'scheduler_state_dict': scheduler.state_dict(),
'best_train_loss': best_train_loss,
'best_test_loss': best_test_loss,
},
os.path.join(args.checkpoints, 'latest_checkpoints_' + args.identifier +'.pth'))
logging.info(f"--- {time.time() - start_time} seconds ---")
epoch_train_loss = []
epoch_test_loss = []
epoch_test_segloss = []
epoch_test_clsloss = []
correct_train_count = 0
correct_test_count = 0
start_time = time.time()
iteration += 1