-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain.py
167 lines (143 loc) · 7.67 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
# -*- coding: utf-8 -*-
# @Describe :
# @Time : 2022/07/04 15:15
# @Author : zpx
# @File : train.py
import os
import random
import hydra
import numpy as np
import torch
import torch.nn.functional as F
from ignite.contrib.handlers import TensorboardLogger, ProgressBar
from ignite.contrib.metrics import GpuInfo
from ignite.engine import Events, create_supervised_trainer, create_supervised_evaluator, convert_tensor
from ignite.handlers import CosineAnnealingScheduler, LRScheduler, EarlyStopping, global_step_from_engine, \
ModelCheckpoint
from ignite.metrics import Loss, DiceCoefficient, ConfusionMatrix
from torch import nn
from torch.optim.lr_scheduler import StepLR
from torch.utils.data import DataLoader
from appendix_dataset import Appendix_Dataset_3slice, RandomGenerator
from network import LSSED
from utils import logValidAndTest, HD95Metric, DiceLoss, LSS_LOSS
@hydra.main(config_path='config', config_name='config.yaml')
def run(config):
os.environ["CUDA_VISIBLE_DEVICES"] = "{}".format(config.gpu_num)
os.environ['PYTHONHASHSEED'] = str(config.seed)
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
random.seed(config.seed)
np.random.seed(config.seed)
torch.manual_seed(config.seed)
torch.cuda.manual_seed(config.seed)
device = torch.device(config.device)
net = LSSED(pretrained=config.pretrained, num_classes=config.num_classes, patch_size=config.patch_size).to(device)
logger_path = '{}_{}_{}'.format(config.log_dir, config.model_name, config.gamma)
tb_logger = TensorboardLogger(log_dir=logger_path)
# load data
train_loader = DataLoader(
Appendix_Dataset_3slice(base_dir=config.root_path, split="train", transform=RandomGenerator(
output_size=config.img_size, random_convert=True), wl=config.wl, ww=config.ww),
batch_size=config.batch_size, num_workers=config.num_workers, pin_memory=True, shuffle=True,
worker_init_fn=lambda worker_id: random.seed(config.seed + worker_id))
valid_loader = DataLoader(
Appendix_Dataset_3slice(base_dir=config.root_path, split="valid", transform=RandomGenerator(
output_size=config.img_size, random_convert=False), wl=config.wl, ww=config.ww),
batch_size=config.batch_size, num_workers=config.num_workers, pin_memory=True, shuffle=False,
worker_init_fn=lambda worker_id: random.seed(config.seed + worker_id))
test_loader = DataLoader(
Appendix_Dataset_3slice(base_dir=config.root_path, split="test", transform=RandomGenerator(
output_size=config.img_size, random_convert=False), wl=config.wl, ww=config.ww),
batch_size=config.batch_size, num_workers=config.num_workers, pin_memory=True, shuffle=False,
worker_init_fn=lambda worker_id: random.seed(config.seed + worker_id))
dice_loss = DiceLoss(config.num_classes)
lgem_loss = LSS_LOSS(H=config.H, Q=config.Q, net=net.seg_head, gamma=config.gamma)
if config.num_gpus > 1:
net = nn.DataParallel(net)
# define optimizer
if config.optimizer == 'AdamW':
optimizer = torch.optim.AdamW(net.parameters(), lr=config.lr)
elif config.optimizer == 'Adam':
optimizer = torch.optim.Adam(net.parameters(), lr=config.lr)
elif config.optimizer == 'RMSprop':
optimizer = torch.optim.RMSprop(net.parameters(), lr=config.lr, weight_decay=1e-4, momentum=0.9)
else:
optimizer = torch.optim.SGD(net.parameters(), lr=config.lr, momentum=0.9, weight_decay=1e-4)
# define learning scheduler
if config.scheduler == 'cosine':
scheduler = CosineAnnealingScheduler(optimizer, 'lr', 1e-3, 1e-8, config.epochs)
elif config.scheduler == 'step_lr':
scheduler = LRScheduler(StepLR(optimizer, step_size=5, gamma=0.5))
def prepare_data(sample, device, non_blocking):
return convert_tensor(sample['image'], device, non_blocking), \
convert_tensor(sample['label'].long(), device, non_blocking)
# define loss function
def loss_fun(p, y):
return F.cross_entropy(p[0], y) + dice_loss(p[0], y, softmax=True) + lgem_loss((p[1], p[0]), y)
train_engine = create_supervised_trainer(net, optimizer, loss_fn=loss_fun, device=device,
prepare_batch=prepare_data)
train_engine.add_event_handler(Events.EPOCH_STARTED, scheduler)
cm = ConfusionMatrix(config.num_classes, output_transform=lambda x: (x[0][0], x[1]))
train_metric = {
"loss_bce": Loss(F.cross_entropy, output_transform=lambda x: (x[0][0], x[1])),
"loss_dice": Loss(dice_loss, output_transform=lambda x: (x[0][0], x[1])),
"loss_lgem": Loss(lgem_loss, output_transform=lambda x: ((x[0][1], x[0][0]), x[1])),
"Dice": DiceCoefficient(cm, ignore_index=0),
"HD95": HD95Metric(output_transform=lambda x: (x[0][0], x[1]))
}
evaluate_metric = {
"Dice": DiceCoefficient(cm, ignore_index=0),
"HD95": HD95Metric(output_transform=lambda x: (x[0][0], x[1]))
}
train_evaluator = create_supervised_evaluator(net, train_metric, device=device, prepare_batch=prepare_data)
valid_evaluator = create_supervised_evaluator(net, evaluate_metric, device=device, prepare_batch=prepare_data)
test_evaluator = create_supervised_evaluator(net, evaluate_metric, device=device, prepare_batch=prepare_data)
for name, metric in train_metric.items():
metric.attach(train_evaluator, name)
for name, metric in evaluate_metric.items():
metric.attach(valid_evaluator, name)
metric.attach(test_evaluator, name)
GpuInfo().attach(train_engine, name='gpu')
progressBar = ProgressBar()
progressBar.attach(train_engine, metric_names='all')
# log setting
@train_engine.on(Events.EPOCH_COMPLETED(every=1))
def log_training_result(engine):
train_evaluator.run(train_loader)
metrics = train_evaluator.state.metrics
tb_logger.writer.add_scalar('training/lr', optimizer.state_dict()['param_groups'][0]['lr'], engine.state.epoch)
logValidAndTest(engine, "training", train_evaluator, train_loader, tb_logger, progressBar)
@train_engine.on(Events.EPOCH_COMPLETED(every=1))
def log_valid_result(engine):
logValidAndTest(engine, "Valid", valid_evaluator, valid_loader, tb_logger, progressBar)
@train_engine.on(Events.EPOCH_COMPLETED(every=1))
def log_test_result(engine):
logValidAndTest(engine, "Test", test_evaluator, test_loader, tb_logger, progressBar)
# save best model critical
def score_function(engine):
return engine.state.metrics['Dice'].item()
early_stop = EarlyStopping(patience=config.early_stop_patience, score_function=score_function, trainer=train_engine)
valid_evaluator.add_event_handler(Events.COMPLETED, early_stop)
for tag, evaluator in [('training', train_evaluator), ('validation', valid_evaluator), ("test", test_evaluator)]:
tb_logger.attach_output_handler(
evaluator,
event_name=Events.EPOCH_COMPLETED,
tag=tag,
metric_names="all",
global_step_transform=global_step_from_engine(train_engine)
)
model_checkpoint = ModelCheckpoint(
'checkpoint',
n_saved=config.checkpoint_epoch,
filename_prefix='best',
score_function=score_function,
score_name='Dice',
global_step_transform=global_step_from_engine(train_engine)
)
valid_evaluator.add_event_handler(Events.EPOCH_COMPLETED, model_checkpoint, {'model': net})
train_engine.run(train_loader, config.epochs)
torch.save('{}/final.pt'.format(logger_path), net.state_dict())
progressBar.log_message("===================Save final model and finished!!!=======================")
if __name__ == '__main__':
run()