-
Notifications
You must be signed in to change notification settings - Fork 0
/
main-ml.py
130 lines (105 loc) · 3.89 KB
/
main-ml.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
import logging
import hydra
import matplotlib.pyplot as plt
from omegaconf import DictConfig, OmegaConf
import pytorch_lightning as pl
from pytorch_lightning.loggers import WandbLogger
from cadml.datamodules import ROIClassificationDataModule
from cadml.models import SimpleCNN, ResNet
from cadml.tasks import ROIClassificationTask
from cadml.callbacks import MetricsLoggingCallback
import torch
logger = logging.getLogger(__name__)
def train(
cfg: DictConfig, datamodule: pl.LightningDataModule, task: pl.LightningModule
):
datamodule.setup()
# TODO add metrics writing to the checkpoint??
wandb_logger = WandbLogger(project="cad-ml")
checkpoint_callback = pl.callbacks.ModelCheckpoint(
save_top_k=3,
monitor="f1",
mode="max",
dirpath="output/",
filename=f"cad-{cfg.model.name}-{wandb_logger.experiment.name}-{{epoch:03d}}-{{f1:.4f}}",
)
trainer = pl.Trainer(
accelerator=cfg.task.accelerator,
logger=wandb_logger,
devices=cfg.task.devices,
max_epochs=cfg.task.max_epochs,
# default_root_dir='output',
callbacks=[MetricsLoggingCallback(), checkpoint_callback],
# fast_dev_run=True,
# overfit_batches=10,
# profiler='simple',
)
# if no checkpoint_path is passed, then it is None, thus the model will start from the very beginning
trainer.fit(task, datamodule=datamodule, ckpt_path=cfg.model.checkpoint_path)
trainer.test(task, datamodule=datamodule, ckpt_path="best")
def evaluate(
cfg: DictConfig, datamodule: pl.LightningDataModule, task: pl.LightningModule
):
if cfg.model.checkpoint_path is None:
raise ValueError("no checkpoint path has been passed")
datamodule.setup("test")
trainer = pl.Trainer(
accelerator=cfg.task.accelerator,
devices=cfg.task.devices,
# precision=32
)
trainer.test(task, datamodule=datamodule, ckpt_path=cfg.model.checkpoint_path)
def inference(
cfg: DictConfig, datamodule: pl.LightningDataModule, task: pl.LightningModule
):
datamodule.setup("predict")
raise NotImplementedError()
def explore(
cfg: DictConfig, datamodule: pl.LightningDataModule, task: pl.LightningModule
):
datamodule.setup()
for idx in range(len(datamodule.train_dataset)):
image, label = datamodule.train_dataset[idx]
print(image)
plt.title(f'Label: {label}')
plt.imshow(image.squeeze(0), cmap="gray")
plt.show()
# from coronaryx import plot_scan, plot_branch
# for item in datamodule.train_dataset.original_dataset:
# plot_branch(item)
@hydra.main(config_path="cadml/conf", config_name="config", version_base=None)
def main(cfg: DictConfig):
print(OmegaConf.to_yaml(cfg))
pl.seed_everything(cfg.seed)
# FIXME refactor this?
if cfg.datamodule.name == "roi-classification":
datamodule = ROIClassificationDataModule(cfg)
else:
raise ValueError(
"unknown datamodule, can be either `roi-classification` or ..."
)
if cfg.model.name == "simple-cnn":
model = SimpleCNN(cfg)
elif cfg.model.name == "resnet":
model = ResNet(cfg)
else:
raise ValueError("unknown model, can be either `simple-cnn` or ...")
if cfg.task.name == "roi-classification":
task = ROIClassificationTask(cfg, model)
else:
raise ValueError("unknown task, can be either `roi-classification` or ...")
if cfg.stage == "train":
train(cfg, datamodule, task)
elif cfg.stage == "evaluate":
evaluate(cfg, datamodule, task)
elif cfg.stage == "inference":
inference(cfg, datamodule, task)
elif cfg.stage == 'explore':
explore(cfg, datamodule, task)
else:
raise ValueError(
"unknown stage, can be either `train`, `evaluate` or `inference`"
)
if __name__ == "__main__":
torch.set_default_dtype(torch.float32)
main()