-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain.py
executable file
·173 lines (153 loc) · 8.14 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
import pytorch_lightning as pl
from pytorch_lightning.loggers import TensorBoardLogger
from pytorch_lightning.callbacks import ModelCheckpoint
import subprocess
import argparse
from src.data.BraTS import brats
from src.data.OASIS import oasis
from src.models import PULPo
from evaluate import Evaluate
#################################################################################################################################################################################
########## HYPERPARAMETERS ###################################################################################################################################
#################################################################################################################################################################################
# TODO: just put these into the parser?
accelerator = "cpu"
dataset = "brats"
segs = False
lms = False
mask = False
total_levels = 5
latent_levels = 4
# TODO: put values from paper as default values
beta = 1
batch_size = 12
learning_rate = 1e-4
recon_loss = ["ncc"]
dice_factor=50
ncc_factor=20
similarity_pyramid = False
lamb = 0
regularizer = "jdet"
image_logging_frequency = 1000
feedback = ["samples"]
df_resolution = "level_res" # "full_res" or "level_res"
ndims = 3
# To save the checkpoint with current git hash
def get_git_revision_short_hash() -> str:
return (
subprocess.check_output(["git", "rev-parse", "--short", "HEAD"])
.decode("ascii")
.strip()
)
def main(hparams):
pl.seed_everything(seed=hparams.random_seed)
git_hash = get_git_revision_short_hash()
human_readable_extra = ""
experiment_name = "-".join(
[git_hash, f"seed={hparams.random_seed}", human_readable_extra]
)
if hparams.dataset == "brats":
(
train_loader,
validation_loader,
test_loader,
) = brats.create_data_loaders(batch_size=hparams.batch_size,
segs=hparams.segs,
lms=hparams.lms,
mask=hparams.mask,
ndims=hparams.ndims,
interpatient=hparams.interpatient)
elif hparams.dataset == "oasis":
(
train_loader,
validation_loader,
test_loader_seg,
test_loader_lm
) = oasis.create_data_loaders(batch_size=hparams.batch_size,
segs=hparams.segs,
lms=False,
mask=False,
ndims=hparams.ndims)
else:
raise ValueError("Dataset not recognized.")
input_size = next(iter(train_loader))[0].shape[2:]
model = PULPo(segs=hparams.segs, lms=hparams.lms, mask=hparams.mask, nondiagonal=hparams.nondiagonal, cp_depth=hparams.cp_depth,
total_levels=hparams.total_levels, latent_levels=hparams.latent_levels, input_size=input_size,
beta=hparams.beta, lr=hparams.learning_rate, recon_loss=hparams.recon_loss, dice_factor=hparams.dice_factor,
ncc_factor=hparams.ncc_factor, similarity_pyramid= hparams.similarity_pyramid, lamb=hparams.lamb, regularizer=hparams.regularizer,
image_logging_frequency=hparams.image_logging_frequency, feedback=hparams.feedback,
df_resolution=hparams.df_resolution, n0=hparams.n0)
logger = TensorBoardLogger(
save_dir="./runs", name=experiment_name, default_hp_metric=False
)
checkpoint_callbacks = [
ModelCheckpoint(
monitor="val/total_loss",
filename="best-total-loss-{epoch}-{step}",
),
ModelCheckpoint(
monitor="val/reconstruction_loss",
filename="best-reconstruction-loss-{epoch}-{step}",
)
]
print(f"RUNNING FOR {hparams.max_epochs} EPOCHS.")
trainer = pl.Trainer(
logger=logger,
val_check_interval=0.1,
log_every_n_steps=5,
accelerator=hparams.accelerator,
devices=hparams.devices,
callbacks=checkpoint_callbacks,
max_epochs= hparams.max_epochs,
)
trainer.fit(
model=model, train_dataloaders=train_loader, val_dataloaders=validation_loader
)
print("TRAINING FINISHED, STARTING EVALUATION.")
eval = Evaluate()
eval.run_one_model(model_dir="runs",
git_hash=experiment_name,
version="version_"+str(logger.version),
segs=hparams.segs,
lms=hparams.lms,
mask=hparams.mask,
N=10,
task=hparams.dataset)
if __name__ == "__main__":
parser = argparse.ArgumentParser(description="Main trainer file for all models.")
parser.add_argument(
"--random_seed",
dest="random_seed",
action="store",
default=0,
type=int,
help="Random seed for pl.seed_everything function.",
)
parser.add_argument("--devices", type=int, default=None)
parser.add_argument("--max_epochs", type=int, default=1000)
parser.add_argument("--accelerator", type=str, default=accelerator)
parser.add_argument("--dataset", type=str, default=dataset, help="Dataset to use. Default is brats. Alternative: oasis.")
parser.add_argument("--segs", action='store_true', default=segs, help="Do we load segmentations from the dataset.")
parser.add_argument("--lms", action='store_true', default=lms, help="Do we load landmarks from the dataset.")
parser.add_argument("--mask", action='store_true', default=mask, help="Do we load masks from the dataset.")
parser.add_argument("--total_levels", type=int, default=total_levels)
parser.add_argument("--latent_levels", type=int, default=latent_levels)
parser.add_argument("--beta", type=float, default=beta)
parser.add_argument("--batch_size", type=int, default=batch_size)
parser.add_argument("--learning_rate", type=float, default=learning_rate)
parser.add_argument("--recon_loss", nargs='+', default=recon_loss, help="Losses used in training. Default is mse. Options: mse, ncc, dice.")
parser.add_argument("--dice_factor", type=int, default=dice_factor, help="Factor to scale dice up to MSE magnitude.")
parser.add_argument("--ncc_factor", type=int, default=ncc_factor, help="Factor to scale ncc up to MSE magnitude.")
parser.add_argument("--similarity_pyramid", action='store_true', default=similarity_pyramid, help="Whether to use a similarity pyramid or not.")
parser.add_argument("--lambda", type=float, default=lamb, dest="lamb", help="Lambda of regularization. Setting to 0 equals no regularization.")
parser.add_argument("--regularizer", type=str, default=regularizer, help="Regularizer to use. Default is jdet. Alternatives: L2.")
parser.add_argument("--image_logging_frequency", type=int, default=image_logging_frequency)
parser.add_argument("--feedback", nargs='+', default=feedback, help="Feedback connection between sampling layers. Default is combined_df. Options: samples, control_points, individual_dfs, combined_dfs, final_dfs, transformed.")
parser.add_argument("--df_resolution", type=str, default=df_resolution, help="Whether the dfs and thus transformed images are created at the resolution of 2x the sampling or at full resolution. Options: full_res, level_res.")
parser.add_argument("--n0", type=int, default=batch_size)
parser.add_argument("--ndims", type=int, default=ndims, help="Choose here if you want to work with volumes (3) or slices (2). Default is 2.")
parser.add_argument("--interpatient", action='store_true', default=False, help="Whether to use the interpatient dataset or not. Only relevant for the BraTS dataset.")
parser.add_argument("--nondiagonal", action='store_true', default=False, help="Whether to use the nondiagonal prior and respective KL loss or not.")
parser.add_argument("--cp_depth", type=int, default=3, help="Depth of the control point layer. Default is 3.")
args = parser.parse_args()
main(args)