idd-solafune/train.py at main · simsam8/idd-solafune · GitHub

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
import os
from argparse import ArgumentParser
from pathlib import Path

import albumentations as albu
import lightning as pl
import torch
from lightning.pytorch import seed_everything
from lightning.pytorch.callbacks import LearningRateMonitor, ModelCheckpoint
from lightning.pytorch.loggers import TensorBoardLogger

import configs
from datasets import IDDDataModule
from Models import Model

seed_everything(367)


def get_augmentations(model_type):
    shift_scale_rotate = [
        # shift, scale, and rotate
        albu.ShiftScaleRotate(
            p=0.5,
            shift_limit=0.0625,
            scale_limit=0.1,
            rotate_limit=15,
            border_mode=0,  # constant border
            fill=0,
            fill_mask=0,
            interpolation=2,  # bicubic
        )
    ]
    crop = [
        # random crop
        albu.RandomCrop(
            p=1,
            width=512,
            height=512,
        )
    ]
    flip_transpose_rotate = [
        # flip, transpose, and rotate90
        albu.HorizontalFlip(p=0.5),
        albu.VerticalFlip(p=0.5),
        albu.Transpose(p=0.5),
        albu.RandomRotate90(p=0.5),
    ]
    # Use full image resolution for Vision Transformer
    if model_type == "vit_seg":
        augs = shift_scale_rotate + flip_transpose_rotate
    else:
        augs = shift_scale_rotate + crop + flip_transpose_rotate

    return albu.Compose(augs)


def pl_trainer(
    config, data_dir, epochs=10, frozen_start=False, frozen_epochs=0, channels="full"
):
    config["model_params"]["in_channels"] = 3 if channels == "rgb" else 12
    config = config | getattr(
        configs, "hyper_params"
    )  # Combine model specific parametrs with common parameters

    model = Model(config, epochs)
    train_ouput_dir = (
        data_dir / "training_result" / f"{config['model_name']}_{channels}"
    )
    trainer = pl.Trainer(
        max_epochs=epochs,
        callbacks=[
            ModelCheckpoint(
                dirpath=train_ouput_dir,
                filename="best_f1",
                save_top_k=1,
                monitor="val/f1",
                mode="max",
                save_last=False,
            ),
            LearningRateMonitor(logging_interval="step"),
        ],
        logger=TensorBoardLogger(train_ouput_dir, name=None),
        precision="16-mixed",
        deterministic=True,
        benchmark=False,
        sync_batchnorm=False,
        check_val_every_n_epoch=5,
        default_root_dir=os.getcwd(),
        accelerator="auto",
        devices="auto",
        strategy="auto",
        log_every_n_steps=5,
        enable_progress_bar=True,
        accumulate_grad_batches=config[
            "batch_accumulation"
        ],  # Update on every n batches
    )
    data_module = IDDDataModule(
        data_dir,
        get_augmentations(config["model_type"]),
        config["train_batch_size"],
        config["val_batch_size"],
        config["test_batch_size"],
        config["num_workers"],
        channels,
    )
    if frozen_start:
        trainer.fit_loop.max_epochs = frozen_epochs
        model.model.unlock_encoder(False)
        trainer.fit(model, datamodule=data_module)
        model.model.unlock_encoder(True)
        trainer.fit_loop.max_epochs = epochs
        trainer.fit(model, datamodule=data_module)
    else:
        trainer.fit(model, datamodule=data_module)


def main(args):
    torch.cuda.empty_cache()
    data_path = Path("data/").absolute()
    pl_trainer(
        getattr(configs, args.config),
        data_path,
        epochs=int(args.epochs),
        frozen_start=args.frozen_start,
        frozen_epochs=int(args.f_epochs),
        channels=args.channels,
    )


if __name__ == "__main__":
    parser = ArgumentParser()
    parser.add_argument("--epochs", required=True)
    parser.add_argument("--config", required=True)
    parser.add_argument("--frozen_start", action="store_true", default=False)
    parser.add_argument("--f_epochs", required=False, default=0)
    parser.add_argument("--channels", required=False, default="full")
    main(parser.parse_args())