"eps": 1e-5,
"weight_decay": 1e-5
},
],
"scheduler":
torch.optim.lr_scheduler.StepLR,
"scheduler_params": {
"step_size": 1,
"gamma": 1.0
},
"acc_steps": [1, 200],
"train_transform":
torchvision.transforms.Compose(
[ToComplex(),
SimulateMeasurements(OpA),
Jitter(1e1, 0.0, 1.0)]),
"val_transform":
torchvision.transforms.Compose(
[ToComplex(), SimulateMeasurements(OpA)], ),
"train_loader_params": {
"shuffle": True,
"num_workers": 8
},
"val_loader_params": {
"shuffle": False,
"num_workers": 8
},
}
# ----- data configuration -----
},
{
"lr": 5e-5,
"eps": 1e-5,
"weight_decay": 5e-3
},
],
"scheduler":
torch.optim.lr_scheduler.StepLR,
"scheduler_params": {
"step_size": 1,
"gamma": 1.0
},
"acc_steps": [1, 200],
"train_transform":
Jitter(2e0, 0.0, 1.0),
}
# -----data prep -----
X_train, C_train, Y_train = [
tmp.unsqueeze(-2).to(device)
for tmp in load_dataset(config.set_params["path"], subset="train")
]
X_val, C_val, Y_val = [
tmp.unsqueeze(-2).to(device)
for tmp in load_dataset(config.set_params["path"], subset="val")
]
# ------ save hyperparameters -------
os.makedirs(train_params["save_path"][-1], exist_ok=True)
"eps": 1e-4,
"weight_decay": 1e-6
}],
"scheduler":
torch.optim.lr_scheduler.StepLR,
"scheduler_params": {
"step_size": 1,
"gamma": 1.0
},
"acc_steps": [1],
"train_transform":
torchvision.transforms.Compose([
CropOrPadAndResimulate((320, 320)),
Flatten(0, -3),
Normalize(reduction="mean", use_target=True),
Jitter(1.5e2, 0.0, 1.0),
]),
"val_transform":
torchvision.transforms.Compose([
CropOrPadAndResimulate((320, 320)),
Flatten(0, -3),
Normalize(reduction="mean", use_target=True),
], ),
"train_loader_params": {
"shuffle": True,
"num_workers": 8
},
"val_loader_params": {
"shuffle": False,
"num_workers": 8
},
},
{
"lr": 5e-5,
"eps": 1e-5,
"weight_decay": 1e-3
},
],
"scheduler":
torch.optim.lr_scheduler.StepLR,
"scheduler_params": {
"step_size": 1,
"gamma": 1.0
},
"acc_steps": [1, 200],
"train_transform":
Jitter(4e0, 0.0, 1.0),
}
# -----data prep -----
X_train, C_train, Y_train = [
tmp.unsqueeze(-2).to(device)
for tmp in load_dataset(config.set_params["path"], subset="train")
]
X_val, C_val, Y_val = [
tmp.unsqueeze(-2).to(device)
for tmp in load_dataset(config.set_params["path"], subset="val")
]
# ------ save hyperparameters -------
os.makedirs(train_params["save_path"][-1], exist_ok=True)
"Fourier_UNet_it_jitter_v4_"
"train_phase_{}".format((i + 1) % (train_phases + 1)),
)
for i in range(train_phases + 1)
],
"save_epochs": 1,
"optimizer": torch.optim.Adam,
"optimizer_params": [
{"lr": 5e-5, "eps": 2e-4, "weight_decay": 1e-4},
{"lr": 5e-5, "eps": 2e-4, "weight_decay": 1e-4},
],
"scheduler": torch.optim.lr_scheduler.StepLR,
"scheduler_params": {"step_size": 1, "gamma": 1.0},
"acc_steps": [1, 200],
"train_transform": torchvision.transforms.Compose(
[ToComplex(), SimulateMeasurements(OpA), Jitter(1e1, 0.0, 1.0)]
),
"val_transform": torchvision.transforms.Compose(
[ToComplex(), SimulateMeasurements(OpA)],
),
"train_loader_params": {"shuffle": True, "num_workers": 8},
"val_loader_params": {"shuffle": False, "num_workers": 8},
}
# ----- data configuration -----
train_data_params = {
"path": config.DATA_PATH,
}
train_data = IPDataset
"optimizer_params": [{
"lr": 8e-5,
"eps": 2e-4,
"weight_decay": 5e-4
}],
"scheduler":
torch.optim.lr_scheduler.StepLR,
"scheduler_params": {
"step_size": 1,
"gamma": 1.0
},
"acc_steps": [1],
"train_transform":
torchvision.transforms.Compose(
[SimulateMeasurements(OpA),
Jitter(5e2, 0.0, 1.0)]),
"val_transform":
torchvision.transforms.Compose([SimulateMeasurements(OpA)], ),
"train_loader_params": {
"shuffle": True,
"num_workers": 0
},
"val_loader_params": {
"shuffle": False,
"num_workers": 0
},
}
# ----- data configuration -----
train_data_params = {
"Radon_UNet_jitter_v3_"
"train_phase_{}".format((i + 1) % (train_phases + 1)),
)
for i in range(train_phases + 1)
],
"save_epochs": 1,
"optimizer": torch.optim.Adam,
"optimizer_params": [
{"lr": 2e-4, "eps": 1e-4, "weight_decay": 5e-4},
{"lr": 5e-5, "eps": 1e-4, "weight_decay": 1e-4},
],
"scheduler": torch.optim.lr_scheduler.StepLR,
"scheduler_params": {"step_size": 1, "gamma": 1.0},
"acc_steps": [1, 200],
"train_transform": torchvision.transforms.Compose(
[SimulateMeasurements(OpA), Jitter(5e2, 0.0, 1.0)]
),
"val_transform": torchvision.transforms.Compose(
[SimulateMeasurements(OpA)],
),
"train_loader_params": {"shuffle": True, "num_workers": 0},
"val_loader_params": {"shuffle": False, "num_workers": 0},
}
# ----- data configuration -----
train_data_params = {
"path": config.DATA_PATH,
"device": device,
}
train_data = IPDataset
config.RESULTS_PATH,
"unet_jitter_"
"train_phase_{}".format((i + 1) % (train_phases + 1)),
)
for i in range(train_phases + 1)
],
"save_epochs": 1,
"optimizer": torch.optim.Adam,
"optimizer_params": [
{"lr": 8e-5, "eps": 1e-5, "weight_decay": 1e-3},
{"lr": 5e-5, "eps": 1e-5, "weight_decay": 1e-3},
],
"scheduler": torch.optim.lr_scheduler.StepLR,
"scheduler_params": {"step_size": 1, "gamma": 1.0},
"acc_steps": [1, 200],
"train_transform": Jitter(4e0, 0.0, 1.0),
}
# -----data prep -----
X_train, C_train, Y_train = [
tmp.unsqueeze(-2).to(device)
for tmp in load_dataset(config.set_params["path"], subset="train")
]
X_val, C_val, Y_val = [
tmp.unsqueeze(-2).to(device)
for tmp in load_dataset(config.set_params["path"], subset="val")
]