-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathtrainer_Synapse.py
166 lines (142 loc) · 7.56 KB
/
trainer_Synapse.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
import logging
import os
import pdb
import random
import sys
import numpy as np
import torch
import torch.nn as nn
import torch.optim as optim
from tensorboardX import SummaryWriter
from torch.nn.modules.loss import CrossEntropyLoss
from torch.utils.data import DataLoader
from tqdm import tqdm
from utils import DiceLoss, save_parameters, test_single_volume
from torchvision import transforms
def inference(model, testloader, args, test_save_path=None):
model.eval()
metric_list = 0.0
for i_batch, sampled_batch in tqdm(enumerate(testloader)):
h, w = sampled_batch["image"].size()[2:]
image, label, case_name = sampled_batch["image"], sampled_batch["label"], sampled_batch['case_name'][0]
metric_i = test_single_volume(image, label, model, classes=args.num_classes,
patch_size=[args.img_size, args.img_size],
input_size=[args.input_size, args.input_size],
test_save_path=test_save_path, case=case_name)
metric_list += np.array(metric_i)
logging.info(' idx %d case %s mean_dice %f mean_hd95 %f' % (
i_batch, case_name, np.mean(metric_i, axis=0)[0], np.mean(metric_i, axis=0)[1]))
metric_list = metric_list / len(testloader.dataset)
for i in range(1, args.num_classes):
logging.info('Mean class %d mean_dice %f mean_hd95 %f' % (i, metric_list[i - 1][0], metric_list[i - 1][1]))
performance = np.mean(metric_list, axis=0)[0]
mean_hd95 = np.mean(metric_list, axis=0)[1]
logging.info('Testing performance in best val model: mean_dice : %f mean_hd95 : %f' % (performance, mean_hd95))
return performance, mean_hd95
def calc_loss(outputs, low_res_label_batch, ce_loss, dice_loss, dice_weight:float=0.8):
low_res_logits = outputs
loss_ce = ce_loss(low_res_logits, low_res_label_batch[:].long())
loss_dice = dice_loss(low_res_logits, low_res_label_batch, softmax=True)
loss = (1 - dice_weight) * loss_ce + dice_weight * loss_dice
return loss, loss_ce, loss_dice
def trainer_synapse(args, model, snapshot_path, split="train"):
from datasets.dataset_synapse import Synapse_dataset, RandomGenerator
logging.basicConfig(filename=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))
base_lr = args.base_lr
num_classes = args.num_classes
batch_size = args.batch_size * args.n_gpu
db_train = Synapse_dataset(base_dir=args.root_path, list_dir=args.list_dir, split=split,
transform=transforms.Compose(
[RandomGenerator(output_size=[args.img_size, args.img_size])]))
db_test = Synapse_dataset(base_dir=args.test_path, split="test_vol", list_dir=args.list_dir)
print("The length of train set is: {}".format(len(db_train)))
def worker_init_fn(worker_id):
random.seed(args.seed + worker_id)
trainloader = DataLoader(db_train, batch_size=batch_size, shuffle=True, num_workers=8, pin_memory=True,
worker_init_fn=worker_init_fn)
testloader = DataLoader(db_test, batch_size=1, shuffle=False, num_workers=1)
if args.n_gpu > 1:
model = nn.DataParallel(model)
model.train()
ce_loss = CrossEntropyLoss()
dice_loss = DiceLoss(num_classes)
if args.warmup:
b_lr = base_lr / args.warmup_period
else:
b_lr = base_lr
if args.AdamW:
optimizer = optim.AdamW(filter(lambda p: p.requires_grad, model.parameters()), lr=b_lr, betas=(0.9, 0.999), weight_decay=0.1)
else:
optimizer = optim.SGD(filter(lambda p: p.requires_grad, model.parameters()), lr=b_lr, momentum=0.9, weight_decay=0.0001) # Even pass the model.parameters(), the `requires_grad=False` layers will not update
writer = SummaryWriter(snapshot_path + '/log')
iter_num = 0
max_epoch = args.max_epochs
max_iterations = args.max_epochs * len(trainloader)
logging.info("{} iterations per epoch. {} max iterations ".format(len(trainloader), max_iterations))
best_performance = 0.0
iterator = tqdm(range(max_epoch), ncols=70)
for epoch_num in iterator:
for i_batch, sampled_batch in enumerate(trainloader):
image_batch, label_batch = sampled_batch['image'], sampled_batch['label']
image_batch, label_batch = image_batch.cuda(), label_batch.cuda()
assert image_batch.max() <= 3, f'image_batch max: {image_batch.max()}'
outputs = model(image_batch)
loss, loss_ce, loss_dice = calc_loss(outputs, label_batch, ce_loss, dice_loss, args.dice_param)
optimizer.zero_grad()
loss.backward()
optimizer.step()
if args.warmup and iter_num < args.warmup_period:
lr_ = base_lr * ((iter_num + 1) / args.warmup_period)
for param_group in optimizer.param_groups:
param_group['lr'] = lr_
else:
if args.warmup:
shift_iter = iter_num - args.warmup_period
assert shift_iter >= 0, f'Shift iter is {shift_iter}, smaller than zero'
else:
shift_iter = iter_num
lr_ = base_lr * (1.0 - shift_iter / max_iterations) ** 0.9 # learning rate adjustment depends on the max iterations
for param_group in optimizer.param_groups:
param_group['lr'] = lr_
iter_num = iter_num + 1
writer.add_scalar('info/lr', lr_, iter_num)
writer.add_scalar('info/total_loss', loss, iter_num)
writer.add_scalar('info/loss_ce', loss_ce, iter_num)
writer.add_scalar('info/loss_dice', loss_dice, iter_num)
logging.info('iteration %d : loss : %f, loss_ce: %f, loss_dice: %f' % (iter_num, loss.item(), loss_ce.item(), loss_dice.item()))
eval_interval = args.eval_interval
if epoch_num >= int(max_epoch / 2) and (epoch_num + 1) % eval_interval == 0:
filename = f'epoch_{epoch_num}.pth'
save_mode_path = os.path.join(snapshot_path, filename)
torch.save(model.state_dict(), save_mode_path)
logging.info("save model to {}".format(save_mode_path))
logging.info("*" * 20)
logging.info(f"Running Inference after epoch {epoch_num}")
print(f"Epoch {epoch_num}")
mean_dice, mean_hd95 = inference(model, testloader, args)
model.train()
if mean_dice > best_performance:
best_performance = mean_dice
save_mode_path = os.path.join(snapshot_path, 'epoch_{}_dice_{}.pth'.format(
epoch_num, round(best_performance, 4)))
save_best = os.path.join(snapshot_path, 'best_model.pth')
try:
save_parameters(model, save_mode_path)
save_parameters(model, save_best)
except:
save_parameters(model.module, save_mode_path)
save_parameters(model.module, save_best)
if epoch_num >= max_epoch - 1:
save_mode_path = os.path.join(snapshot_path, 'epoch_' + str(epoch_num) + '.pth')
try:
save_parameters(model, save_mode_path)
except:
save_parameters(model.module, save_mode_path)
logging.info("save model to {}".format(save_mode_path))
iterator.close()
break
writer.close()
return "Training Finished!"