def load_data(config_file_path, val_proportion=0.10, top_k=-1):
'''
Loads all the data for the ESA Kelvin competition (train, val, test, baseline)
Args:
config_file_path: str, paths of configuration file
val_proportion: float, validation/train fraction
top_k: int, number of low-resolution images to read. Default (top_k=-1) reads all low-res images, sorted by clearance.
Returns:
train_dataset: torch.Dataset
val_dataset: torch.Dataset
test_dataset: torch.Dataset
baseline_cpsnrs: dict, shift cPSNR scores of the ESA baseline
'''
with open(config_file_path, "r") as read_file:
config = json.load(read_file)
data_directory = config["paths"]["prefix"]
baseline_cpsnrs = readBaselineCPSNR(
os.path.join(data_directory, "norm.csv"))
train_set_directories = getImageSetDirectories(
os.path.join(data_directory, "train"))
test_set_directories = getImageSetDirectories(
os.path.join(data_directory, "test"))
# val_proportion = 0.10
train_list, val_list = train_test_split(train_set_directories,
test_size=val_proportion,
random_state=1,
shuffle=True)
config["training"]["create_patches"] = False
train_dataset = ImagesetDataset(imset_dir=train_list,
config=config["training"],
top_k=top_k)
val_dataset = ImagesetDataset(imset_dir=val_list,
config=config["training"],
top_k=top_k)
test_dataset = ImagesetDataset(imset_dir=test_set_directories,
config=config["training"],
top_k=top_k)
return train_dataset, val_dataset, test_dataset, baseline_cpsnrs
def main():
'''
Calls save_clearance on train and test set.
'''
parser = argparse.ArgumentParser()
parser.add_argument("--prefix",
help="root dir of the dataset",
default='data/')
args = parser.parse_args()
prefix = args.prefix
assert os.path.isdir(prefix)
if os.path.exists(os.path.join(prefix, "train")):
train_set_directories = getImageSetDirectories(
os.path.join(prefix, "train"))
save_clearance_scores(train_set_directories) # train data
if os.path.exists(os.path.join(prefix, "test")):
test_set_directories = getImageSetDirectories(
os.path.join(prefix, "test"))
save_clearance_scores(test_set_directories) # test data
def main(config):
"""
Given a configuration, trains HRNet and ShiftNet for Multi-Frame Super Resolution (MFSR), and saves best model.
Args:
config: dict, configuration file
"""
# Reproducibility options
np.random.seed(0) # RNG seeds
torch.manual_seed(0)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
# Initialize the network based on the network configuration
fusion_model = HRNet(config["network"])
regis_model = ShiftNet()
optimizer = optim.Adam(list(fusion_model.parameters()) +
list(regis_model.parameters()),
lr=config["training"]["lr"]) # optim
# ESA dataset
data_directory = config["paths"]["prefix"]
baseline_cpsnrs = None
if os.path.exists(os.path.join(data_directory, "norm.csv")):
baseline_cpsnrs = readBaselineCPSNR(
os.path.join(data_directory, "norm.csv"))
train_set_directories = getImageSetDirectories(
os.path.join(data_directory, "train"))
val_proportion = config['training']['val_proportion']
train_list, val_list = train_test_split(train_set_directories,
test_size=val_proportion,
random_state=1,
shuffle=True)
# Dataloaders
batch_size = config["training"]["batch_size"]
n_workers = config["training"]["n_workers"]
n_views = config["training"]["n_views"]
min_L = config["training"]["min_L"] # minimum number of views
beta = config["training"]["beta"]
train_dataset = ImagesetDataset(imset_dir=train_list,
config=config["training"],
top_k=n_views,
beta=beta)
train_dataloader = DataLoader(train_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=n_workers,
collate_fn=collateFunction(min_L=min_L),
pin_memory=True)
config["training"]["create_patches"] = False
val_dataset = ImagesetDataset(imset_dir=val_list,
config=config["training"],
top_k=n_views,
beta=beta)
val_dataloader = DataLoader(val_dataset,
batch_size=1,
shuffle=False,
num_workers=n_workers,
collate_fn=collateFunction(min_L=min_L),
pin_memory=True)
dataloaders = {'train': train_dataloader, 'val': val_dataloader}
# Train model
torch.cuda.empty_cache()
#fusion_model.load_state_dict(torch.load("/home/ubadmin/Documents/Scripts/highres_net/HighRes-net-master/models/weights/training_8b_full_ESA/HRNet.pth"))
#regis_model.load_state_dict(torch.load("/home/ubadmin/Documents/Scripts/highres_net/HighRes-net-master/models/weights/training_8b_full_ESA/ShiftNet.pth"))
trainAndGetBestModel(fusion_model, regis_model, optimizer, dataloaders,
baseline_cpsnrs, config)
def load_data(config_file_path, val_proportion=0.10, top_k=-1):
'''
Loads all the data for the ESA Kelvin competition (train, val, test, baseline)
Args:
config_file_path: str, paths of configuration file
val_proportion: float, validation/train fraction
top_k: int, number of low-resolution images to read. Default (top_k=-1) reads all low-res images, sorted by clearance.
Returns:
train_dataset: torch.Dataset
val_dataset: torch.Dataset
test_dataset: torch.Dataset
baseline_cpsnrs: dict, shift cPSNR scores of the ESA baseline
'''
with open(config_file_path, "r") as read_file:
config = json.load(read_file)
data_directory = config["paths"]["prefix"]
baseline_cpsnrs = readBaselineCPSNR(
os.path.join(data_directory, "norm.csv"))
train_set_directories = getImageSetDirectories(
os.path.join(data_directory, "train"))
test_set_directories = getImageSetDirectories(
os.path.join(data_directory, "test"))
# val_proportion = 0.10
train_list, val_list = train_test_split(train_set_directories,
test_size=val_proportion,
random_state=1,
shuffle=True)
# val_list = ["imgset0000", "imgset0061", "imgset0203", "imgset0280", "imgset0374", "imgset0476", "imgset0585",
# "imgset0692", "imgset0769", "imgset0845", "imgset0960", "imgset1039", "imgset1128",
#"imgset0011", "imgset0072", "imgset0204", "imgset0285", "imgset0382", "imgset0498", "imgset0588", "imgset0711",
# "imgset0771", "imgset0878", "imgset0962", "imgset1052", "imgset1133",
#"imgset0023", "imgset0085", "imgset0205", "imgset0289", "imgset0414", "imgset0499", "imgset0602", "imgset0728",
# "imgset0776", "imgset0884", "imgset0980", "imgset1054", "imgset1134",
#"imgset0035", "imgset0087", "imgset0208", "imgset0313", "imgset0448", "imgset0503", "imgset0604", "imgset0730",
# "imgset0791", "imgset0896", "imgset0998", "imgset1063", "imgset1158",
#"imgset0039", "imgset0114", "imgset0221", "imgset0324", "imgset0450", "imgset0505", "imgset0617", "imgset0734",
# "imgset0793", "imgset0921", "imgset1013", "imgset1068",
#"imgset0047", "imgset0130", "imgset0235", "imgset0328", "imgset0458", "imgset0530", "imgset0618", "imgset0748",
# "imgset0796", "imgset0923", "imgset1015", "imgset1089",
#"imgset0051", "imgset0138", "imgset0255", "imgset0337", "imgset0460", "imgset0534", "imgset0652", "imgset0751",
# "imgset0811", "imgset0933", "imgset1021", "imgset1112",
#"imgset0056", "imgset0164", "imgset0262", "imgset0340", "imgset0465", "imgset0549", "imgset0674", "imgset0758",
# "imgset0814", "imgset0948", "imgset1023", "imgset1121",
#"imgset0057", "imgset0192", "imgset0270", "imgset0361", "imgset0470", "imgset0558", "imgset0687", "imgset0762",
# "imgset0817", "imgset0951", "imgset1034", "imgset1126"]
config["training"]["create_patches"] = False
train_dataset = ImagesetDataset(imset_dir=train_list,
config=config["training"],
top_k=top_k)
val_dataset = ImagesetDataset(imset_dir=val_list,
config=config["training"],
top_k=top_k)
test_dataset = ImagesetDataset(imset_dir=test_set_directories,
config=config["training"],
top_k=top_k)
return train_dataset, val_dataset, test_dataset, baseline_cpsnrs