-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathtrain_cls.py
193 lines (159 loc) · 7.18 KB
/
train_cls.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
import os
os.environ["MKL_NUM_THREADS"] = "1"
os.environ["NUMEXPR_NUM_THREADS"] = "1"
os.environ["OMP_NUM_THREADS"] = "1"
from sklearn.metrics import classification_report
from torch.nn import BCELoss
import cv2
cv2.ocl.setUseOpenCL(False)
cv2.setNumThreads(0)
import warnings
from training import augmentation_sets, dataset_cls
from training.augmentation_sets import FractureAugmentations
from training.utils import all_gather
import torch
from tqdm import tqdm
import torch.distributed as dist
import numpy as np
from training.config import load_config
warnings.filterwarnings("ignore")
import argparse
from typing import Dict
from training.trainer import TrainConfiguration, PytorchTrainer, Evaluator
from torch.utils.data import DataLoader
import torch.distributed
class SegFractureEvaluator(Evaluator):
def __init__(self, args: argparse.Namespace) -> None:
super().__init__()
self.args = args
self.competition_weights = {
'negative': torch.from_numpy(np.array([7, 1, 1, 1, 1, 1, 1, 1])).float(),
'positive': torch.from_numpy(np.array([14, 2, 2, 2, 2, 2, 2, 2])).float()
}
def init_metrics(self) -> Dict:
return {"loss": 10}
def validate(self, dataloader: DataLoader, model: torch.nn.Module, distributed: bool = False, local_rank: int = 0,
snapshot_name: str = "") -> Dict:
os.makedirs(self.args.pred_dir, exist_ok=True)
all_preds = []
all_targets = []
for sample in tqdm(dataloader):
imgs = sample["image"].cuda().float()[0]
label = sample["label"].cpu().float()[0]
all_targets.append(label.numpy())
with torch.no_grad():
with torch.cuda.amp.autocast():
output = model(imgs)["cls"]
preds = torch.sigmoid(output.float())
preds = np.max(preds.cpu().numpy().astype(np.float32), axis=0)
preds = np.clip(preds, 0.01, 0.99)
all_preds.append(preds)
all_preds = np.array(all_preds)
all_targets = np.array(all_targets)
if distributed:
all_preds = all_gather(all_preds)
all_preds = np.concatenate(all_preds, axis=0)
all_targets = all_gather(all_targets)
all_targets = np.concatenate(all_targets, axis=0)
result = 10
if local_rank == 0:
y = torch.from_numpy(all_targets).float()
loss = BCELoss(reduction='none')(torch.from_numpy(all_preds).float(), y)
weights = self.competition_weights['positive'] * y + \
self.competition_weights['negative'] * (1 - y)
loss = (loss * weights).sum(axis=1)
loss = loss / weights.sum(axis=1)
result = (loss.sum() / y.shape[0]).item()
print(classification_report(all_targets > 0.5, all_preds > 0.5))
if distributed:
dist.barrier()
torch.cuda.empty_cache()
return {"loss": result}
def get_improved_metrics(self, prev_metrics: Dict, current_metrics: Dict) -> Dict:
improved = {}
best_loss = prev_metrics["loss"]
if current_metrics["loss"] < prev_metrics["loss"]:
print("Loss improved from {:.4f} to {:.4f}".format(prev_metrics["loss"], current_metrics["loss"]))
improved["loss"] = current_metrics["loss"]
best_loss = current_metrics["loss"]
print("Best Loss {:.4f} current {:.4f}".format(best_loss, current_metrics["loss"]))
return improved
def parse_args():
parser = argparse.ArgumentParser("Pipeline")
arg = parser.add_argument
arg('--config', metavar='CONFIG_FILE', help='path to configuration file', default="configs/vgg512.json")
arg('--workers', type=int, default=6, help='number of cpu threads to use')
arg('--gpu', type=str, default='1', help='List of GPUs for parallel training, e.g. 0,1,2,3')
arg('--output-dir', type=str, default='weights/')
arg('--resume', type=str, default='')
arg('--fold', type=int, default=0)
arg('--prefix', type=str, default='')
arg('--data-dir', type=str, default="/home/selim/datasets/rsna/")
arg('--folds-csv', type=str, default="folds.csv")
arg('--logdir', type=str, default='logs')
arg('--zero-score', action='store_true', default=False)
arg('--from-zero', action='store_true', default=False)
arg('--fp16', action='store_true', default=False)
arg('--distributed', action='store_true', default=False)
arg("--local_rank", default=0, type=int)
arg("--world-size", default=1, type=int)
arg("--test_every", type=int, default=1)
arg('--freeze-epochs', type=int, default=0)
arg('--pred-dir', type=str, default="../oof")
arg("--val", action='store_true', default=False)
arg("--freeze-bn", action='store_true', default=False)
args = parser.parse_args()
return args
def create_data_datasets(args):
conf = load_config(args.config)
slice_size = conf.get("slice_size", 32)
val_slice_size = conf.get("val_slice_size", slice_size)
crop_size = conf.get("crop_size", 160)
augmentations = augmentation_sets.__dict__[conf["augmentations"]]() # type: FractureAugmentations
dataset_type = dataset_cls.__dict__[conf["dataset"]["type"]]
params = conf["dataset"].get("params", {})
print(f"Using augmentations: {augmentations.__class__.__name__} with Dataset: {dataset_type.__name__}")
train_dataset = dataset_type(mode="train",
dataset_dir=args.data_dir,
fold=args.fold,
crop_size=crop_size,
folds_csv=args.folds_csv,
transforms=augmentations.get_train_augmentations(conf),
slice_size=slice_size,
multiplier=conf.get("multiplier", 1), **params)
val_dataset = dataset_type(mode="val", dataset_dir=args.data_dir, fold=args.fold,
folds_csv=args.folds_csv,
crop_size=crop_size,
slice_size=val_slice_size,
transforms=augmentations.get_val_augmentations(conf), **params)
return train_dataset, val_dataset
def main():
args = parse_args()
trainer_config = TrainConfiguration(
config_path=args.config,
gpu=args.gpu,
resume_checkpoint=args.resume,
prefix=args.prefix,
world_size=args.world_size,
test_every=args.test_every,
local_rank=args.local_rank,
distributed=args.distributed,
freeze_epochs=args.freeze_epochs,
log_dir=args.logdir,
output_dir=args.output_dir,
workers=args.workers,
from_zero=args.from_zero,
zero_score=args.zero_score,
fp16=args.fp16,
freeze_bn=args.freeze_bn
)
data_train, data_val = create_data_datasets(args)
seg_evaluator = SegFractureEvaluator(args)
trainer = PytorchTrainer(train_config=trainer_config, evaluator=seg_evaluator, fold=args.fold,
train_data=data_train, val_data=data_val)
if args.val:
trainer.validate()
return
trainer.fit()
if __name__ == '__main__':
main()