-
Notifications
You must be signed in to change notification settings - Fork 0
/
utils.py
76 lines (63 loc) · 2.24 KB
/
utils.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
from pathlib import Path
from datetime import datetime, timedelta
import numpy as np
from pytorch_lightning.utilities import rank_zero_only
from pytorch_lightning import Trainer
from pytorch_lightning.loggers import TensorBoardLogger
from pytorch_lightning.callbacks import (
ModelCheckpoint,
EarlyStopping,
LearningRateMonitor,
DeviceStatsMonitor,
BatchSizeFinder,
LearningRateFinder,
)
################# TRAIN #################
@rank_zero_only
def maybe_save_checkpoint(trainer: Trainer):
print("Saving checkpoint...")
ckpt_path = (
Path(trainer.log_dir)
/ "checkpoints"
/ f"e{trainer.current_epoch}_last_at_{get_curr_time_w_random_shift()}.ckpt"
)
if trainer.global_rank == 0:
trainer.save_checkpoint(ckpt_path)
def init_log_directory(timestamp: str, log_dir: str = "./logs") -> tuple:
log_dir = Path(log_dir) / timestamp
ckpt_dir = log_dir / "checkpoints"
for d in [log_dir, ckpt_dir]:
if not d.exists():
d.mkdir(parents=True)
return log_dir, ckpt_dir
def get_logger(log_dir: Path, name: str = "", version: str = "") -> TensorBoardLogger:
logger = TensorBoardLogger(
save_dir=log_dir.as_posix(),
name=name,
version=version,
log_graph=False,
default_hp_metric=False,
)
return logger
def get_callbacks(ckpt_dir: Path, save_top_k: int = 3) -> list:
callbacks = [
LearningRateMonitor(logging_interval="step"),
ModelCheckpoint(
dirpath=ckpt_dir.as_posix(),
filename="{epoch}-{step}-{val_loss:.3f}",
monitor="val_loss_epoch",
mode="min",
save_top_k=save_top_k,
save_on_train_epoch_end=True,
),
EarlyStopping(monitor="val_loss_epoch", patience=3, mode="min", verbose=True),
# DeviceStatsMonitor(cpu_stats=True),
# BatchSizeFinder(),
# LearningRateFinder(num_training_steps=200),
]
return callbacks
################# MISC #################
def get_curr_time_w_random_shift() -> str:
# shifting for a random number of seconds so that exp folder names coincide less often
now = datetime.now() - timedelta(seconds=np.random.randint(60))
return now.strftime("%y-%m-%dT%H-%M-%S")