Adds support for hydra
Browse files
train.py
CHANGED
@@ -1,61 +1,83 @@
|
|
|
|
|
|
|
|
1 |
import lightning as L
|
2 |
-
import torch
|
3 |
from lightning.pytorch.callbacks import (
|
4 |
-
ModelCheckpoint,
|
5 |
-
LearningRateMonitor,
|
6 |
EarlyStopping,
|
|
|
|
|
7 |
)
|
8 |
from lightning.pytorch.loggers import TensorBoardLogger
|
9 |
-
|
10 |
-
from src.
|
11 |
from src.model import DRModel
|
|
|
12 |
|
13 |
-
# seed everything for reproducibility
|
14 |
-
SEED = 42
|
15 |
-
L.seed_everything(SEED, workers=True)
|
16 |
-
torch.set_float32_matmul_precision("high")
|
17 |
|
|
|
|
|
|
|
|
|
18 |
|
19 |
-
#
|
20 |
-
|
21 |
-
|
22 |
|
23 |
-
#
|
24 |
-
|
25 |
-
|
26 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
27 |
|
28 |
-
# Init
|
29 |
-
|
30 |
-
|
31 |
-
|
32 |
-
|
33 |
-
|
34 |
-
|
35 |
-
|
36 |
-
filename="{epoch}-{step}-{val_loss:.2f}-{val_acc:.2f}-{val_kappa:.2f}",
|
37 |
-
)
|
38 |
|
39 |
-
# Init
|
40 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
41 |
|
42 |
-
#
|
43 |
-
|
44 |
-
|
45 |
-
|
46 |
-
|
47 |
-
|
48 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
49 |
|
50 |
-
# Init trainer
|
51 |
-
trainer = L.Trainer(
|
52 |
-
max_epochs=50,
|
53 |
-
accelerator="auto",
|
54 |
-
devices="auto",
|
55 |
-
logger=logger,
|
56 |
-
callbacks=[checkpoint_callback, lr_monitor, early_stopping],
|
57 |
-
# check_val_every_n_epoch=4,
|
58 |
-
)
|
59 |
|
60 |
-
|
61 |
-
|
|
|
1 |
+
from os.path import join
|
2 |
+
|
3 |
+
import hydra
|
4 |
import lightning as L
|
|
|
5 |
from lightning.pytorch.callbacks import (
|
|
|
|
|
6 |
EarlyStopping,
|
7 |
+
LearningRateMonitor,
|
8 |
+
ModelCheckpoint,
|
9 |
)
|
10 |
from lightning.pytorch.loggers import TensorBoardLogger
|
11 |
+
from omegaconf import DictConfig
|
12 |
+
from src.data_module import DRDataModule
|
13 |
from src.model import DRModel
|
14 |
+
from src.utils import generate_run_id
|
15 |
|
|
|
|
|
|
|
|
|
16 |
|
17 |
+
@hydra.main(version_base=None, config_path="conf", config_name="config")
|
18 |
+
def train(cfg: DictConfig) -> None:
|
19 |
+
# generate unique run id based on current date & time
|
20 |
+
run_id = generate_run_id()
|
21 |
|
22 |
+
# Seed everything for reproducibility
|
23 |
+
L.seed_everything(cfg.seed, workers=True)
|
24 |
+
# torch.set_float32_matmul_precision("high")
|
25 |
|
26 |
+
# Initialize DataModule
|
27 |
+
dm = DRDataModule(
|
28 |
+
train_csv_path=cfg.train_csv_path,
|
29 |
+
val_csv_path=cfg.val_csv_path,
|
30 |
+
image_size=cfg.image_size,
|
31 |
+
batch_size=cfg.batch_size,
|
32 |
+
num_workers=cfg.num_workers,
|
33 |
+
use_class_weighting=cfg.use_class_weighting,
|
34 |
+
use_weighted_sampler=cfg.use_weighted_sampler,
|
35 |
+
)
|
36 |
+
dm.setup()
|
37 |
|
38 |
+
# Init model from datamodule's attributes
|
39 |
+
model = DRModel(
|
40 |
+
num_classes=dm.num_classes,
|
41 |
+
model_name=cfg.model_name,
|
42 |
+
learning_rate=cfg.learning_rate,
|
43 |
+
class_weights=dm.class_weights,
|
44 |
+
use_scheduler=cfg.use_scheduler,
|
45 |
+
)
|
|
|
|
|
46 |
|
47 |
+
# Init logger
|
48 |
+
logger = TensorBoardLogger(save_dir=cfg.logs_dir, name="", version=run_id)
|
49 |
+
# Init callbacks
|
50 |
+
checkpoint_callback = ModelCheckpoint(
|
51 |
+
monitor="val_loss",
|
52 |
+
mode="min",
|
53 |
+
save_top_k=2,
|
54 |
+
dirpath=join(cfg.checkpoint_dirpath, run_id),
|
55 |
+
filename="{epoch}-{step}-{val_loss:.2f}-{val_acc:.2f}-{val_kappa:.2f}",
|
56 |
+
)
|
57 |
|
58 |
+
# Init LearningRateMonitor
|
59 |
+
lr_monitor = LearningRateMonitor(logging_interval="step")
|
60 |
+
|
61 |
+
# early stopping
|
62 |
+
early_stopping = EarlyStopping(
|
63 |
+
monitor="val_loss",
|
64 |
+
patience=10,
|
65 |
+
verbose=True,
|
66 |
+
mode="min",
|
67 |
+
)
|
68 |
+
|
69 |
+
# Initialize Trainer
|
70 |
+
trainer = L.Trainer(
|
71 |
+
max_epochs=cfg.max_epochs,
|
72 |
+
accelerator="auto",
|
73 |
+
devices="auto",
|
74 |
+
logger=logger,
|
75 |
+
callbacks=[checkpoint_callback, lr_monitor, early_stopping],
|
76 |
+
)
|
77 |
+
|
78 |
+
# Train the model
|
79 |
+
trainer.fit(model, dm)
|
80 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
81 |
|
82 |
+
if __name__ == "__main__":
|
83 |
+
train()
|