-
Notifications
You must be signed in to change notification settings - Fork 4
/
Copy pathmain.py
204 lines (182 loc) · 10.5 KB
/
main.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
import argparse
from utils.data_utils import get_loader
from medical.trainer import Trainer, Validator
from monai.inferers import SlidingWindowInferer
import torch
import torch.nn.parallel
import torch.distributed as dist
import torch.multiprocessing as mp
import torch.utils.data.distributed
import numpy as np
from monai.metrics import DiceMetric
from monai.utils.enums import MetricReduction
from medical.optimizers.lr_scheduler import LinearWarmupCosineAnnealingLR
from monai.losses.dice import DiceLoss
from medical.model.nested_former import NestedFormer
parser = argparse.ArgumentParser(description='Swin UNETR segmentation pipeline for BRATS Challenge')
parser.add_argument('--model_name', default="swinunetr", help='the model will be trained')
parser.add_argument('--checkpoint', default=None, help='start training from saved checkpoint')
parser.add_argument('--logdir', default='test', type=str, help='directory to save the tensorboard logs')
parser.add_argument('--fold', default=0, type=int, help='data fold')
parser.add_argument('--pretrain_model_path', default='./model.pt', type=str, help='pretrained model name')
parser.add_argument('--load_pretrain', action="store_true", help='pretrained model name')
parser.add_argument('--data_dir', default='/mnt/datasets/brats2020/MICCAI_BraTS2020_TrainingData', type=str, help='dataset directory')
parser.add_argument('--json_list', default='./brats2020_datajson.json', type=str, help='dataset json file')
parser.add_argument('--max_epochs', default=300, type=int, help='max number of training epochs')
parser.add_argument('--batch_size', default=2, type=int, help='number of batch size')
parser.add_argument('--sw_batch_size', default=4, type=int, help='number of sliding window batch size')
parser.add_argument('--optim_lr', default=1e-4, type=float, help='optimization learning rate')
parser.add_argument('--optim_name', default='adamw', type=str, help='optimization algorithm')
parser.add_argument('--reg_weight', default=1e-5, type=float, help='regularization weight')
parser.add_argument('--momentum', default=0.99, type=float, help='momentum')
parser.add_argument('--val_every', default=10, type=int, help='validation frequency')
parser.add_argument('--distributed', action='store_true', help='start distributed training')
parser.add_argument('--world_size', default=1, type=int, help='number of nodes for distributed training')
parser.add_argument('--rank', default=0, type=int, help='node rank for distributed training')
parser.add_argument('--dist-url', default='tcp://127.0.0.1:23456', type=str, help='distributed url')
parser.add_argument('--dist-backend', default='nccl', type=str, help='distributed backend')
parser.add_argument('--norm_name', default='instance', type=str, help='normalization name')
parser.add_argument('--workers', default=8, type=int, help='number of workers')
parser.add_argument('--feature_size', default=24, type=int, help='feature size')
parser.add_argument('--in_channels', default=4, type=int, help='number of input channels')
parser.add_argument('--out_channels', default=3, type=int, help='number of output channels')
parser.add_argument('--cache_dataset', action='store_true', help='use monai Dataset class')
parser.add_argument('--a_min', default=-175.0, type=float, help='a_min in ScaleIntensityRanged')
parser.add_argument('--a_max', default=250.0, type=float, help='a_max in ScaleIntensityRanged')
parser.add_argument('--b_min', default=0.0, type=float, help='b_min in ScaleIntensityRanged')
parser.add_argument('--b_max', default=1.0, type=float, help='b_max in ScaleIntensityRanged')
parser.add_argument('--space_x', default=1.0, type=float, help='spacing in x direction')
parser.add_argument('--space_y', default=1.0, type=float, help='spacing in y direction')
parser.add_argument('--space_z', default=1.0, type=float, help='spacing in z direction')
parser.add_argument('--roi_x', default=128, type=int, help='roi size in x direction')
parser.add_argument('--roi_y', default=128, type=int, help='roi size in y direction')
parser.add_argument('--roi_z', default=128, type=int, help='roi size in z direction')
parser.add_argument('--dropout_rate', default=0.0, type=float, help='dropout rate')
parser.add_argument('--dropout_path_rate', default=0.0, type=float, help='drop path rate')
parser.add_argument('--RandFlipd_prob', default=0.2, type=float, help='RandFlipd aug probability')
parser.add_argument('--RandRotate90d_prob', default=0.2, type=float, help='RandRotate90d aug probability')
parser.add_argument('--RandScaleIntensityd_prob', default=0.1, type=float, help='RandScaleIntensityd aug probability')
parser.add_argument('--RandShiftIntensityd_prob', default=0.1, type=float, help='RandShiftIntensityd aug probability')
parser.add_argument('--infer_overlap', default=0.25, type=float, help='sliding window inference overlap')
parser.add_argument('--lrschedule', default='warmup_cosine', type=str, help='type of learning rate scheduler')
parser.add_argument('--warmup_epochs', default=50, type=int, help='number of warmup epochs')
parser.add_argument('--resume_ckpt', action='store_true', help='resume training from pretrained checkpoint')
def post_pred_func(pred):
pred = torch.sigmoid(pred)
pred = (pred > 0.5).float()
return pred
def main():
args = parser.parse_args()
args.logdir = './runs/' + args.logdir
if args.distributed:
args.ngpus_per_node = torch.cuda.device_count()
print('Found total gpus', args.ngpus_per_node)
args.world_size = args.ngpus_per_node * args.world_size
mp.spawn(main_worker,
nprocs=args.ngpus_per_node,
args=(args,))
else:
main_worker(gpu=0, args=args)
def main_worker(gpu, args):
if args.distributed:
torch.multiprocessing.set_start_method('fork', force=True)
np.set_printoptions(formatter={'float': '{: 0.3f}'.format}, suppress=True)
args.gpu = gpu
if args.distributed:
args.rank = args.rank * args.ngpus_per_node + gpu
dist.init_process_group(backend=args.dist_backend,
init_method=args.dist_url,
world_size=args.world_size,
rank=args.rank)
torch.cuda.set_device(args.gpu)
torch.backends.cudnn.benchmark = True
args.test_mode = False
train_loader, val_loader = get_loader(args)
print(args.rank, ' gpu', args.gpu)
if args.rank == 0:
print('Batch size is:', args.batch_size, 'epochs', args.max_epochs)
inf_size = [args.roi_x, args.roi_y, args.roi_z]
model = NestedFormer(model_num=args.in_channels,
out_channels=args.out_channels,
image_size=inf_size,
window_size=(4, 4, 4),
)
window_infer = SlidingWindowInferer(roi_size=inf_size,
sw_batch_size=args.sw_batch_size,
overlap=args.infer_overlap)
pytorch_total_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
print('Total parameters count', pytorch_total_params)
best_acc = 0
start_epoch = 0
if args.checkpoint is not None:
checkpoint = torch.load(args.checkpoint, map_location='cpu')
from collections import OrderedDict
new_state_dict = OrderedDict()
for k, v in checkpoint['state_dict'].items():
new_state_dict[k.replace('backbone.','')] = v
model.load_state_dict(new_state_dict, strict=False)
if 'epoch' in checkpoint:
start_epoch = checkpoint['epoch']
if 'best_acc' in checkpoint:
best_acc = checkpoint['best_acc']
print("=> loaded checkpoint '{}' (epoch {}) (bestacc {})".format(args.checkpoint, start_epoch, best_acc))
model.cuda(args.gpu)
if args.distributed:
torch.cuda.set_device(args.gpu)
if args.norm_name == 'batch':
model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model)
model.cuda(args.gpu)
model = torch.nn.parallel.DistributedDataParallel(model,
device_ids=[args.gpu],
output_device=args.gpu,
)
if args.optim_name == 'adam':
optimizer = torch.optim.Adam(model.parameters(),
lr=args.optim_lr,
weight_decay=args.reg_weight)
elif args.optim_name == 'adamw':
optimizer = torch.optim.AdamW(model.parameters(),
lr=args.optim_lr,
weight_decay=args.reg_weight)
elif args.optim_name == 'sgd':
optimizer = torch.optim.SGD(model.parameters(),
lr=args.optim_lr,
momentum=args.momentum,
nesterov=True,
weight_decay=args.reg_weight)
else:
raise ValueError('Unsupported Optimization Procedure: ' + str(args.optim_name))
if args.lrschedule == 'warmup_cosine':
scheduler = LinearWarmupCosineAnnealingLR(optimizer,
warmup_epochs=args.warmup_epochs,
max_epochs=args.max_epochs)
elif args.lrschedule == 'cosine_anneal':
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer,
T_max=args.max_epochs)
if args.checkpoint is not None:
scheduler.step(epoch=start_epoch)
else:
scheduler = None
dice_metric = DiceMetric(include_background=True,
reduction=MetricReduction.MEAN_BATCH,
get_not_nans=True)
validator = Validator(args,
model,
val_loader,
class_list=("TC", "WT", "ET"),
metric_functions=[["dice", dice_metric]],
sliding_window_infer=window_infer,
post_label=None,
post_pred=post_pred_func)
dice_loss = DiceLoss(to_onehot_y=False, sigmoid=True)
trainer = Trainer(args,
train_loader,
validator=validator,
loss_func=dice_loss,
)
best_acc = trainer.train(model,
optimizer=optimizer,
scheduler=scheduler)
return best_acc
if __name__ == '__main__':
main()