def build_args():
parser = ArgumentParser()
# basic args
path_config = pathlib.Path("../../fastmri_dirs.yaml")
backend = "ddp"
num_gpus = 2 if backend == "ddp" else 1
batch_size = 1
# set defaults based on optional directory config
data_path = fetch_dir("knee_path", path_config)
default_root_dir = fetch_dir("log_path",
path_config) / "varnet" / "varnet_demo"
# client arguments
parser.add_argument(
"--mode",
default="train",
choices=("train", "test"),
type=str,
help="Operation mode",
)
# data transform params
parser.add_argument(
"--mask_type",
choices=("random", "equispaced"),
default="equispaced",
type=str,
help="Type of k-space mask",
)
parser.add_argument(
"--center_fractions",
nargs="+",
default=[0.08],
type=float,
help="Number of center lines to use in mask",
)
parser.add_argument(
"--accelerations",
nargs="+",
default=[4],
type=int,
help="Acceleration rates to use for masks",
)
# data config
parser = FastMriDataModule.add_data_specific_args(parser)
parser.set_defaults(
data_path=data_path, # path to fastMRI data
mask_type="equispaced", # VarNet uses equispaced mask
challenge="multicoil", # only multicoil implemented for VarNet
batch_size=batch_size, # number of samples per batch
test_path=None, # path for test split, overwrites data_path
)
# module config
parser = VarNetModule.add_model_specific_args(parser)
parser.set_defaults(
num_cascades=8, # number of unrolled iterations
pools=4, # number of pooling layers for U-Net
chans=18, # number of top-level channels for U-Net
sens_pools=4, # number of pooling layers for sense est. U-Net
sens_chans=8, # number of top-level channels for sense est. U-Net
lr=0.001, # Adam learning rate
lr_step_size=40, # epoch at which to decrease learning rate
lr_gamma=0.1, # extent to which to decrease learning rate
weight_decay=0.0, # weight regularization strength
)
# trainer config
parser = pl.Trainer.add_argparse_args(parser)
parser.set_defaults(
gpus=num_gpus, # number of gpus to use
replace_sampler_ddp=
False, # this is necessary for volume dispatch during val
accelerator=backend, # what distributed version to use
seed=42, # random seed
deterministic=True, # makes things slower, but deterministic
default_root_dir=default_root_dir, # directory for logs and checkpoints
max_epochs=50, # max number of epochs
)
args = parser.parse_args()
# configure checkpointing in checkpoint_dir
checkpoint_dir = args.default_root_dir / "checkpoints"
if not checkpoint_dir.exists():
checkpoint_dir.mkdir(parents=True)
args.checkpoint_callback = pl.callbacks.ModelCheckpoint(
filepath=args.default_root_dir / "checkpoints",
save_top_k=True,
verbose=True,
monitor="validation_loss",
mode="min",
prefix="",
)
# set default checkpoint if one exists in our checkpoint directory
if args.resume_from_checkpoint is None:
ckpt_list = sorted(checkpoint_dir.glob("*.ckpt"), key=os.path.getmtime)
if ckpt_list:
args.resume_from_checkpoint = str(ckpt_list[-1])
return args
def build_varnet_args(data_path, logdir, backend):
parser = ArgumentParser()
num_gpus = 0
batch_size = 1
# data transform params
parser.add_argument(
"--mask_type",
choices=("random", "equispaced"),
default="equispaced",
type=str,
help="Type of k-space mask",
)
parser.add_argument(
"--center_fractions",
nargs="+",
default=[0.08],
type=float,
help="Number of center lines to use in mask",
)
parser.add_argument(
"--accelerations",
nargs="+",
default=[4],
type=int,
help="Acceleration rates to use for masks",
)
# data config
parser = FastMriDataModule.add_data_specific_args(parser)
parser.set_defaults(
data_path=data_path,
mask_type="equispaced",
challenge="multicoil",
batch_size=batch_size,
)
# module config
parser = VarNetModule.add_model_specific_args(parser)
parser.set_defaults(
num_cascades=4,
pools=2,
chans=8,
sens_pools=2,
sens_chans=4,
lr=0.001,
lr_step_size=40,
lr_gamma=0.1,
weight_decay=0.0,
)
# trainer config
parser = Trainer.add_argparse_args(parser)
parser.set_defaults(
gpus=num_gpus,
default_root_dir=logdir,
replace_sampler_ddp=(backend != "ddp"),
accelerator=backend,
)
parser.add_argument("--mode", default="train", type=str)
args = parser.parse_args([])
return args
def build_args():
parser = ArgumentParser()
# basic args
path_config = pathlib.Path("../../fastmri_dirs.yaml")
batch_size = 1 if backend == "ddp" else num_gpus
# set defaults based on optional directory config
data_path = fetch_dir("knee_path", path_config)
default_root_dir = fetch_dir("log_path", path_config) / "nnret" / "nnret_demo"
# client arguments
parser.add_argument(
"--mode",
default="train",
choices=("train", "test"),
type=str,
help="Operation mode",
)
# data transform params
parser.add_argument(
"--mask_type",
choices=("random", "equispaced"),
default="random",
type=str,
help="Type of k-space mask",
)
parser.add_argument(
"--center_fractions",
nargs="+",
default=[0.08],
type=float,
help="Number of center lines to use in mask",
)
parser.add_argument(
"--accelerations",
nargs="+",
default=[4],
type=int,
help="Acceleration rates to use for masks",
)
# data config with path to fastMRI data and batch size
parser = FastMriDataModule.add_data_specific_args(parser)
parser.set_defaults(data_path=data_path, batch_size=batch_size, test_path=None)
# module config
parser = NnRetModule.add_model_specific_args(parser)
parser.set_defaults(
in_chans=1, # number of input channels to NNRET
out_chans=1, # number of output chanenls to NNRET
chans=32, # number of top-level NNRET channels
num_pool_layers=4, # number of NNRET pooling layers
drop_prob=0.0, # dropout probability
lr=0.001, # RMSProp learning rate
lr_step_size=40, # epoch at which to decrease learning rate
lr_gamma=0.1, # extent to which to decrease learning rate
weight_decay=0.0, # weight decay regularization strength
)
# trainer config
parser = pl.Trainer.add_argparse_args(parser)
parser.set_defaults(
gpus=num_gpus, # number of gpus to use
replace_sampler_ddp=False, # this is necessary for volume dispatch during val
accelerator=backend, # what distributed version to use
seed=42, # random seed
deterministic=True, # makes things slower, but deterministic
default_root_dir=default_root_dir, # directory for logs and checkpoints
max_epochs=1, # max number of epochs
)
args = parser.parse_args()
return args
def build_args():
parser = ArgumentParser()
# basic args
path_config = pathlib.Path("../../fastmri_dirs.yaml")
num_gpus = 0
backend = "ddp_cpu"
batch_size = 1 if backend == "ddp_cpu" else num_gpus
# set defaults based on optional directory config
data_path = fetch_dir("knee_path", path_config)
default_root_dir = fetch_dir("log_path",
path_config) / "unet" / "unet_demo"
# client arguments
parser.add_argument(
"--mode",
default="train",
choices=("train", "test"),
type=str,
help="Operation mode",
)
# data transform params
parser.add_argument(
"--mask_type",
choices=("random", "equispaced"),
default="random",
type=str,
help="Type of k-space mask",
)
parser.add_argument(
"--center_fractions",
nargs="+",
default=[0.08],
type=float,
help="Number of center lines to use in mask",
)
parser.add_argument(
"--proportion",
default=0.1,
type=float,
help="Proportion of label data",
)
parser.add_argument(
"--accelerations",
nargs="+",
default=[4],
type=int,
help="Acceleration rates to use for masks",
)
# data config with path to fastMRI data and batch size
parser = FastMriDataModule.add_data_specific_args(parser)
parser.set_defaults(data_path=data_path,
batch_size=batch_size,
test_path=None)
# module config
parser = UnetModule.add_model_specific_args(parser)
parser.set_defaults(
in_chans=1, # number of input channels to U-Net
out_chans=1, # number of output chanenls to U-Net
chans=32, # number of top-level U-Net channels
num_pool_layers=4, # number of U-Net pooling layers
drop_prob=0.0, # dropout probability
lr=0.001, # RMSProp learning rate
lr_step_size=40, # epoch at which to decrease learning rate
lr_gamma=0.1, # extent to which to decrease learning rate
weight_decay=0.0, # weight decay regularization strength
)
# trainer config
parser = pl.Trainer.add_argparse_args(parser)
parser.set_defaults(
gpus=num_gpus, # number of gpus to use
replace_sampler_ddp=
False, # this is necessary for volume dispatch during val
accelerator=backend, # what distributed version to use
seed=42, # random seed
deterministic=True, # makes things slower, but deterministic
default_root_dir=default_root_dir, # directory for logs and checkpoints
max_epochs=50, # max number of epochs
)
args = parser.parse_args()
# configure checkpointing in checkpoint_dir
checkpoint_dir = args.default_root_dir / "checkpoints"
if not checkpoint_dir.exists():
checkpoint_dir.mkdir(parents=True)
args.checkpoint_callback = pl.callbacks.ModelCheckpoint(
dirpath=args.default_root_dir / "checkpoints",
save_top_k=True,
verbose=True,
monitor="validation_loss",
mode="min",
prefix="",
)
# set default checkpoint if one exists in our checkpoint directory
if args.resume_from_checkpoint is None:
ckpt_list = sorted(checkpoint_dir.glob("*.ckpt"), key=os.path.getmtime)
if ckpt_list:
args.resume_from_checkpoint = str(ckpt_list[-1])
return args
def build_args():
parser = ArgumentParser()
batch_size = 1
# client arguments
parser.add_argument(
"--mode",
default="train",
choices=("train", "test"),
type=str,
help="Operation mode",
)
# unet module arguments
parser.add_argument(
"--unet_module",
default="unet",
choices=("unet", "nestedunet"),
type=str,
help="Unet module to run with",
)
# data transform params
parser.add_argument(
"--mask_type",
choices=("random", "equispaced"),
default="random",
type=str,
help="Type of k-space mask",
)
parser.add_argument(
"--center_fractions",
nargs="+",
default=[0.08],
type=float,
help="Number of center lines to use in mask",
)
parser.add_argument(
"--accelerations",
nargs="+",
default=[4],
type=int,
help="Acceleration rates to use for masks",
)
parser.add_argument(
"--device",
default="cuda",
type=str,
help="Model to run",
)
parser.add_argument(
"--state_dict_file",
default=None,
type=Path,
help="Path to saved state_dict (will download if not provided)",
)
parser.add_argument(
"--output_path",
type=Path, # directory for logs and checkpoints
default=Path("./fine_tuning"),
help="Path for saving reconstructions",
)
# unet specific
parser.add_argument(
"--in_chans",
default=1,
type=int,
help="number of input channels to U-Net",
)
parser.add_argument(
"--out_chans",
default=1,
type=int,
help="number of output chanenls to U-Net",
)
parser.add_argument(
"--chans",
default=32,
type=int,
help="number of top-level U-Net channels",
)
# RMSProp parameters
parser.add_argument(
"--opt_drop_prob",
default=0.0,
type=float,
help="dropout probability",
)
parser.add_argument(
"--opt_lr",
default=0.001,
type=float,
help="RMSProp learning rate",
)
parser.add_argument(
"--opt_lr_step_size",
default=10,
type=int,
help="epoch at which to decrease learning rate",
)
parser.add_argument(
"--opt_lr_gamma",
default=0.1,
type=float,
help="extent to which to decrease learning rate",
)
parser.add_argument(
"--opt_weight_decay",
default=0.0,
type=float,
help="weight decay regularization strength",
)
parser.add_argument(
"--opt_optimizer",
choices=("RMSprop", "Adam"),
default="RMSprop",
type=str,
help="optimizer (RMSprop, Adam)",
)
# data config with path to fastMRI data and batch size
parser = FastMriDataModule.add_data_specific_args(parser)
parser.set_defaults(data_path="/home/ec2-user/mri",
batch_size=batch_size,
test_path=None)
# trainer config
parser = pl.Trainer.add_argparse_args(parser)
parser.set_defaults(
gpus=0, # number of gpus to use
replace_sampler_ddp=
False, # this is necessary for volume dispatch during val
seed=42, # random seed
deterministic=True, # makes things slower, but deterministic
max_epochs=50, # max number of epochs
unet_module="unet", # "unet" or "nestedunet"
)
args = parser.parse_args()
# module config
if args.unet_module == "unet":
parser = UnetModule.add_model_specific_args(parser)
parser.set_defaults(
num_pool_layers=4, # number of U-Net pooling layers
drop_prob=args.opt_drop_prob, # dropout probability
lr=args.opt_lr, # RMSProp learning rate
lr_step_size=args.
opt_lr_step_size, # epoch at which to decrease learning rate
lr_gamma=args.
opt_lr_gamma, # extent to which to decrease learning rate
weight_decay=args.
opt_weight_decay, # weight decay regularization strength
optmizer=args.opt_optimizer, # optimizer (RMSprop, Adam)
accelerator="ddp_cpu" if args.device == "cpu" else "ddp",
)
elif args.unet_module == "nestedunet":
parser = NestedUnetModule.add_model_specific_args(parser)
parser.set_defaults(
num_pool_layers=4, # number of U-Net pooling layers
drop_prob=args.opt_drop_prob, # dropout probability
lr=args.opt_lr, # RMSProp learning rate
lr_step_size=args.
opt_lr_step_size, # epoch at which to decrease learning rate
lr_gamma=args.
opt_lr_gamma, # extent to which to decrease learning rate
weight_decay=args.
opt_weight_decay, # weight decay regularization strength
optmizer=args.opt_optimizer, # optimizer (RMSprop, Adam)
accelerator="ddp_cpu" if args.device == "cpu" else "ddp",
)
args = parser.parse_args()
# configure checkpointing in checkpoint_dir
checkpoint_dir = args.output_path / "checkpoints"
if not checkpoint_dir.exists():
checkpoint_dir.mkdir(parents=True)
args.checkpoint_callback = pl.callbacks.ModelCheckpoint(
dirpath=args.output_path / "checkpoints",
save_top_k=True,
verbose=True,
monitor="validation_loss",
mode="min",
prefix="",
)
# set default checkpoint if one exists in our checkpoint directory
if args.resume_from_checkpoint is None:
ckpt_list = sorted(checkpoint_dir.glob("*.ckpt"), key=os.path.getmtime)
if ckpt_list:
args.resume_from_checkpoint = str(ckpt_list[-1])
return args