-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
213 lines (191 loc) · 6.48 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
import argparse
import os
import shutil
from functools import partial
from typing import List, Tuple, Optional, Sequence
from omegaconf import OmegaConf
import torch
import torch.nn.parallel
from torch.optim.lr_scheduler import LinearLR
from dotenv import load_dotenv
from trainer import run_training
from utils.model_utils import get_model
from utils.data_utils import get_loader
from optimizers.lr_schedule import LinearWarmupCosineAnnealingLR
from monai.inferers import sliding_window_inference
from monai.losses import DiceCELoss
from monai.metrics import DiceMetric
from monai.transforms import AsDiscrete
from monai.utils.enums import MetricReduction
def main(config: dict, debug: bool = False):
load_dotenv()
config = parse_congig(config, debug)
run(config=config, debug=debug, **config)
def parse_congig(config: dict, debug: bool = False) -> dict:
config['data_dir'] = os.path.expanduser(config['data_dir'])
if debug:
print('Using debug mode!!!')
config['batch_size'] = 2
config['num_workers'] = 0
config['val_every'] = 1
config['cache_num'] = 2
return config
def run(
log_dir: str,
batch_size: int,
model_name: str,
inf_size: int,
in_channels: int,
out_channels: int,
feature_size: int,
hidden_size: int,
mlp_dim: int,
num_heads: int,
pos_embed: str,
norm_name: Tuple or str,
dropout_rate: float,
pretrained: bool,
path_to_pretrain: Optional[str],
path_to_checkpoint: Optional[str],
smooth_nr: float,
smooth_dr: float,
sw_batch_size: int,
infer_overlap: float,
optim_name: str,
optim_lr: float,
optim_weight_decay: float,
momentum: Optional[float],
lrschedule_name: Optional[str],
warmup_epochs: Optional[int],
max_epochs: Optional[int],
val_every: int,
save_every: int,
data_dir: str,
spacing: Sequence[float],
modality: int or Sequence,
a_min: float,
a_max: float,
b_min: float,
b_max: float,
RandFlipd_prob: float,
RandRotate90d_prob: float,
RandScaleIntensityd_prob: float,
RandShiftIntensityd_prob: float,
gauss_noise_prob: float,
gauss_noise_std: float,
gauss_smooth_prob: float,
gauss_smooth_std: float or Tuple[float],
n_workers: int,
cache_num: int,
device: str,
config: dict,
debug: bool,
**kwargs,
):
device = torch.device(device if torch.cuda.is_available() else 'cpu')
print(f'device: {device.type}')
loader = get_loader(
data_dir=data_dir,
batch_size=batch_size,
spacing=spacing,
modality=modality,
a_min=a_min,
a_max=a_max,
b_min=b_min,
b_max=b_max,
roi_size=inf_size,
RandFlipd_prob=RandFlipd_prob,
RandRotate90d_prob=RandRotate90d_prob,
RandScaleIntensityd_prob=RandScaleIntensityd_prob,
RandShiftIntensityd_prob=RandShiftIntensityd_prob,
gauss_noise_prob=gauss_noise_prob,
gauss_noise_std=gauss_noise_std,
gauss_smooth_prob=gauss_smooth_prob,
gauss_smooth_std=gauss_smooth_std,
n_workers=n_workers,
cache_num=cache_num,
device=device,
debug=debug,
)
print(f"Batch size is: {batch_size}")
model = get_model(
model_name=model_name,
in_channels=in_channels,
out_channels=out_channels,
inf_size=inf_size,
feature_size=feature_size,
hidden_size=hidden_size,
mlp_dim=mlp_dim,
num_heads=num_heads,
pos_embed=pos_embed,
norm_name=norm_name,
dropout_rate=dropout_rate,
pretrained=pretrained,
path_to_pretrain=path_to_pretrain,
path_to_checkpoint=path_to_checkpoint,
device=device,
)
dice_loss = DiceCELoss(
to_onehot_y=True, softmax=True, squared_pred=True, smooth_nr=smooth_nr, smooth_dr=smooth_dr
)
post_label = AsDiscrete(to_onehot=out_channels)
post_pred = AsDiscrete(argmax=True, to_onehot=out_channels)
dice_acc = DiceMetric(include_background=True, reduction=MetricReduction.NONE, get_not_nans=True)
model_inferer = partial(
sliding_window_inference,
roi_size=(inf_size, inf_size, inf_size),
sw_batch_size=sw_batch_size,
predictor=model,
overlap=infer_overlap,
)
if optim_name == "adam":
optimizer = torch.optim.Adam(model.parameters(), lr=optim_lr, weight_decay=optim_weight_decay)
elif optim_name == "adamw":
optimizer = torch.optim.AdamW(model.parameters(), lr=optim_lr, weight_decay=optim_weight_decay)
elif optim_name == "sgd":
optimizer = torch.optim.SGD(
model.parameters(), lr=optim_lr, momentum=momentum, nesterov=True,
weight_decay=optim_weight_decay
)
else:
raise ValueError("Unsupported Optimization Procedure: " + str(optim_name))
if lrschedule_name == "warmup_cosine":
scheduler = LinearWarmupCosineAnnealingLR(
optimizer, warmup_epochs=warmup_epochs, max_epochs=max_epochs
)
elif lrschedule_name == "warmup_linear":
scheduler = LinearLR(optimizer, start_factor=1e-12, end_factor=1.0, total_iters=warmup_epochs)
elif lrschedule_name == "cosine_anneal":
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=max_epochs)
else:
scheduler = None
run_training(
log_dir=log_dir,
model=model,
train_loader=loader[0],
val_loader=loader[1],
optimizer=optimizer,
loss_func=dice_loss,
acc_func=dice_acc,
val_every=val_every,
save_every=save_every,
device=device,
config=config,
model_inferer=model_inferer,
scheduler=scheduler,
post_label=post_label,
post_pred=post_pred,
)
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--config", type=str, default="train_unetr_config.yaml")
parser.add_argument("--debug", type=bool, default=False)
args = parser.parse_args()
config_name = args.config
print(f'Using config {config_name}')
config_folder = "configs"
config_path = os.path.join(config_folder, config_name)
cfg = OmegaConf.load(config_path)
cfg = OmegaConf.to_container(cfg, resolve=True)
cfg["config_path"] = config_path
main(cfg, debug=args.debug)