-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
309 lines (272 loc) · 12 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
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
from __future__ import annotations
import argparse
import os
import warnings
from typing import Optional
import numpy as np
import pandas as pd
import torch
import torch.nn.functional as F
import torchvision
import transformers
import wandb
import yaml
from pytorch_lightning import LightningDataModule, LightningModule, Trainer
from pytorch_lightning import loggers as pl_loggers
from pytorch_lightning.callbacks import LearningRateMonitor, ModelCheckpoint
from sklearn.model_selection import KFold
from torch.utils.data import DataLoader
from config.config import Config, load_config
from src.dataset import ContrailDataset
from src.nn import Segmentor2d, Segmentor25d, SegmentorMid25d, SegmentorMid25dDouble
def parse() -> argparse.Namespace:
parser = argparse.ArgumentParser(description="Training for Kaggle Contrail")
parser.add_argument("--out_base_dir", default="result")
parser.add_argument("--in_base_dir", default="input")
parser.add_argument("--exp_name", default="tmp")
parser.add_argument("--project_name", default="kaggle-contrail")
parser.add_argument("--load_snapshot", action="store_true")
parser.add_argument("--save_checkpoint", action="store_true")
parser.add_argument("--save_model", action="store_true")
parser.add_argument("--wandb_logger", action="store_true")
parser.add_argument("--config_path", default="config/debug.yaml")
return parser.parse_args()
class ContrailDataModule(LightningDataModule):
def __init__(
self,
cfg: Config,
fold: int,
use_val_for_train: bool,
data_dir: str,
):
super().__init__()
self.cfg = cfg
self.fold = fold
self.n_gpus = torch.cuda.device_count()
train_df = pd.DataFrame(
{"image_dir": [f"{data_dir}/train/{dir}" for dir in sorted(os.listdir(f"{data_dir}/train"))]}
)
self.val_df = pd.DataFrame(
{"image_dir": [f"{data_dir}/valid/{dir}" for dir in sorted(os.listdir(f"{data_dir}/valid"))]}
)
if fold < 0:
self.train_df = train_df
self.oof_df = train_df[:0]
else:
splitter = KFold(cfg.n_splits, shuffle=True, random_state=0)
train_idx, oof_idx = list(splitter.split(train_df))[fold]
self.train_df = train_df.iloc[train_idx].copy()
self.oof_df = train_df.iloc[oof_idx].copy()
if use_val_for_train:
self.train_df = pd.concat([self.train_df, self.val_df])
self.val_df = self.val_df[:0]
if cfg.n_data != -1:
self.train_df = self.train_df[: cfg.n_data]
print(self.train_df)
print(f"train: {len(self.train_df)}, oof: {len(self.oof_df)}, val: {len(self.val_df)}")
def train_dataloader(self):
return DataLoader(
ContrailDataset(self.train_df, self.cfg, True),
batch_size=self.cfg.batch_size // self.n_gpus,
num_workers=4,
shuffle=True,
persistent_workers=True,
drop_last=True,
)
def val_dataloader(self):
val_loader = DataLoader(
ContrailDataset(self.val_df, self.cfg, False),
batch_size=self.cfg.batch_size // self.n_gpus,
num_workers=4,
persistent_workers=True,
)
oof_loader = DataLoader(
ContrailDataset(self.oof_df, self.cfg, False),
batch_size=self.cfg.batch_size // self.n_gpus,
num_workers=4,
persistent_workers=True,
)
return val_loader, oof_loader
def calc_dice_coef(pred: torch.Tensor, label: torch.Tensor) -> torch.Tensor:
# (N, C, H, W) -> (1,)
pred = pred.flatten()
label = label.flatten()
intersection = (label * pred).sum()
return 2.0 * intersection / (label.sum() + pred.sum())
def custom_dice_loss(pred: torch.Tensor, label: torch.Tensor) -> torch.Tensor:
# (N, C, H, W) -> (1,)
pred = pred.flatten()
label = label.flatten()
intersection = (label * pred).sum()
return -(2.0 * intersection + 700000) / (label.sum() + pred.sum() + 1000000)
def percentile(x: torch.Tensor, percentile: float) -> float:
x = x.flatten()
idx = x.argsort(descending=True)
return x[idx[round(len(idx) * percentile)]].item()
class ContrailModel(LightningModule):
def __init__(self, cfg: dict):
super().__init__()
if not isinstance(cfg, Config):
cfg = Config(cfg)
self.cfg = cfg
self.save_hyperparameters(cfg)
self.segmentor: Segmentor2d | Segmentor25d | SegmentorMid25d | SegmentorMid25dDouble
if cfg.frame_first == cfg.frame_last:
self.segmentor = Segmentor2d(cfg)
else:
if cfg.segmentor_type == "25d":
self.segmentor = Segmentor25d(cfg)
elif cfg.segmentor_type == "mid-25d":
self.segmentor = SegmentorMid25d(cfg)
elif cfg.segmentor_type == "mid-25d-double":
self.segmentor = SegmentorMid25dDouble(cfg)
else:
raise ValueError()
def forward(self, x: torch.Tensor, apply_sigmoid: bool = True) -> torch.Tensor:
out = self.segmentor(x)
if apply_sigmoid:
out = torch.sigmoid(out)
return out
def _loss_func(self, logit: torch.Tensor, label: torch.Tensor) -> torch.Tensor:
if logit.shape[0] == 0:
return torch.zeros(1, device=logit.device)
dice = custom_dice_loss(torch.sigmoid(logit), label)
bce = F.binary_cross_entropy_with_logits(logit, label)
return (dice + bce) / 2
def training_step(self, batch, batch_idx):
all_logit = self(batch["image"], False)
has_soft_mask = batch["has_soft_mask"]
has_soft_ratio = has_soft_mask.to(torch.float16).mean()
loss = (
custom_dice_loss(torch.sigmoid(all_logit[:, 0]), batch["mask"])
+ self._loss_func(all_logit[has_soft_mask, 1], batch["soft_mask"][has_soft_mask]) * has_soft_ratio
) / 2
if self.cfg.pseudo_frame_aux:
before_1 = batch["mask_before_1"][:, 0, 0] >= 0
loss_1 = -custom_dice_loss(torch.sigmoid(all_logit[before_1, 2]), batch["mask_before_1"][before_1])
before_2 = batch["mask_before_2"][:, 0, 0] >= 0
loss_2 = -custom_dice_loss(torch.sigmoid(all_logit[before_2, 3]), batch["mask_before_2"][before_2])
aux_loss = (loss_1 * before_1.to(torch.float16).mean() + loss_2 * before_2.to(torch.float16).mean()) / 2
loss = (loss + aux_loss) / 2
return {
"loss": loss,
}
def validation_step(self, batch, batch_idx, dataloader_idx=0):
all_logit = self(batch["image"], False)
loss = self._loss_func(all_logit[:, 0], batch["mask"].to(torch.float32))
preds = torchvision.transforms.functional.resize(torch.sigmoid(all_logit[:, :2]), size=(256, 256))
return {
"label": batch["orig_mask"].detach().cpu(),
"preds": preds.detach().cpu(),
"loss": loss.detach().cpu(),
}
def _gather_devices_and_steps(self, outputs: list[dict[str, torch.Tensor]]) -> Optional[dict[str, torch.Tensor]]:
outputs = self.all_gather(outputs)
assert self.trainer is not None
if self.trainer.global_rank != 0 or len(outputs) == 0:
return None
epoch_results: dict[str, torch.Tensor] = {}
for key in outputs[0].keys():
if self.trainer.num_devices > 1:
result = torch.cat(
[(x[key].unsqueeze(1) if x[key].dim() == 1 else x[key]) for x in outputs], dim=1
).flatten(end_dim=1)
else:
result = torch.cat([(x[key].unsqueeze(0) if x[key].dim() == 0 else x[key]) for x in outputs], dim=0)
epoch_results[key] = result.detach().cpu()
return epoch_results
def _epoch_end(self, step_outputs: list[dict[str, torch.Tensor]], phase: str):
epoch_results = self._gather_devices_and_steps(step_outputs)
if epoch_results is None:
return
d = {
f"{phase}/loss": epoch_results["loss"].mean().cpu(),
}
if phase != "train":
preds = epoch_results["preds"].cuda()
label = epoch_results["label"].cuda()
for i, pred in enumerate([preds[:, 0], preds[:, 1], preds.mean(1)]):
# Threshold search.
thresholds = np.arange(0.1, 1.0, 0.01)
dice_coefs = [calc_dice_coef(pred > threshold, label).item() for threshold in thresholds]
max_idx = np.argmax(dice_coefs)
d[f"{phase}/best_gobal_dice_coef_{i}"] = dice_coefs[max_idx]
# d[f"{phase}/best_threshold"] = thresholds[max_idx]
d[f"{phase}/gobal_dice_coef_018_{i}"] = calc_dice_coef(pred > percentile(pred, 0.0018), label).item()
print(d)
self.log_dict(d, on_epoch=True)
def training_epoch_end(self, training_step_outputs) -> None:
self._epoch_end(training_step_outputs, "train")
def validation_epoch_end(self, validation_step_outputs):
self._epoch_end(validation_step_outputs[0], "val")
self._epoch_end(validation_step_outputs[1], "oof")
def _get_total_steps(self) -> int:
if not hasattr(self, "_total_steps"):
train_loader = self.trainer._data_connector._train_dataloader_source.dataloader()
accum = max(1, self.trainer.num_devices) * self.trainer.accumulate_grad_batches
self._total_steps = len(train_loader) // accum * self.trainer.max_epochs
return self._total_steps
def configure_optimizers(self):
optimizer = torch.optim.AdamW(self.parameters(), self.cfg.lr)
total_steps = self._get_total_steps()
warmup_steps = round(total_steps * self.hparams.warmup_steps_ratio)
print(f"lr warmup step: {warmup_steps} / {total_steps}")
scheduler = transformers.get_cosine_schedule_with_warmup(optimizer, warmup_steps, total_steps)
return [optimizer], [{"scheduler": scheduler, "interval": "step"}]
def train(
args: argparse.Namespace,
cfg: Config,
fold: int,
use_val_for_train: bool,
) -> float:
out_dir = f"{args.out_base_dir}/{args.exp_name}/{fold}"
if use_val_for_train:
out_dir += "-all"
model = ContrailModel(cfg)
if cfg.pretrained_model_path is not None:
state_dict = torch.load(cfg.pretrained_model_path, map_location="cpu")
model.load_state_dict(state_dict, strict=False)
model.segmentor.set_grad_checkpointing(cfg.grad_checkpointing)
data_module = ContrailDataModule(cfg, fold, use_val_for_train, args.in_base_dir)
loggers: list[pl_loggers.LightningLoggerBase] = [pl_loggers.CSVLogger(out_dir)]
if args.wandb_logger:
loggers.append(
pl_loggers.WandbLogger(
project=args.project_name,
group=args.exp_name,
name=f"{args.exp_name}/{fold}",
save_dir=out_dir,
)
)
callbacks = [LearningRateMonitor("epoch")]
if args.save_checkpoint:
callbacks.append(ModelCheckpoint(out_dir, save_last=True, save_top_k=0))
n_gpus = torch.cuda.device_count()
trainer = Trainer(
gpus=n_gpus,
max_epochs=cfg.max_epochs,
logger=loggers,
callbacks=callbacks,
enable_checkpointing=args.save_checkpoint,
precision=cfg.precision,
gradient_clip_val=0.7,
strategy="ddp_find_unused_parameters_false" if n_gpus > 1 else None,
)
ckpt_path: Optional[str] = f"{out_dir}/last.ckpt"
if not os.path.exists(ckpt_path) or not args.load_snapshot:
ckpt_path = None
trainer.fit(model, ckpt_path=ckpt_path, datamodule=data_module)
if args.save_model:
torch.save(model.state_dict(), f"{out_dir}/model.pt")
with open(f"{out_dir}/config.yaml", "w") as f:
yaml.dump(dict(cfg), f)
if args.wandb_logger:
wandb.finish()
def main():
args = parse()
warnings.filterwarnings("ignore", ".*does not have many workers.*")
cfg = load_config(args.config_path, "config/default.yaml")
print(cfg)
train(args, cfg, -1, False)
if __name__ == "__main__":
main()