def __init__(self, cfg_filepath: str):
set_deterministic_seed()
with open(from_root(cfg_filepath), "r") as file:
self.cfg = DictObject(json.load(file))
for sub_dirname in ("logs", "checkpoints", "debug"):
os.makedirs(self.from_out(sub_dirname), exist_ok=True)
self.model = create_resnet(self.cfg)
self.model = self.model.to(self.cfg.model.device)
self.train_loader, self.infer_train_loader, self.infer_val_loader, \
train_mean, train_std = create_dataset(self.cfg)
self.model.normalize.set_parameters(train_mean, train_std)
self.logger = SummaryWriter(log_dir=self.from_out("logs"))
self.optimizer = create_optimizer(self.model, self.logger, self.cfg)
self.scheduler = optim.lr_scheduler.MultiStepLR(
self.optimizer,
self.cfg.scheduler.milestones,
gamma=self.cfg.scheduler.gamma)
if isinstance(self.model, ResNet_Softmax):
self.criterion = lambda logits, targets: F.nll_loss(
logits, targets, reduction="mean")
else:
assert isinstance(self.model, ResNet_Gaussian)
self.criterion = lambda sqr_distances, targets: mmc_loss(
sqr_distances, targets, reduction="mean")
Tracker.reset(self.cfg.optimizer.n_epochs)
if self.cfg.load_checkpoint is not None:
self.load_checkpoint(self.cfg.load_checkpoint)
def main() -> None:
set_deterministic_seed()
dataset = torchvision.datasets.MNIST(DATA_DIRPATH, train=True, download=True)
size = len(dataset)
indices = list(range(size))
random.shuffle(indices)
train_indices = indices[:TRAIN_SIZE]
write_lines(os.path.join(SPLIT_DIRPATH, "train.txt"), train_indices)
val_indices = indices[TRAIN_SIZE:]
write_lines(os.path.join(SPLIT_DIRPATH, "val.txt"), val_indices)
def main() -> None:
set_deterministic_seed()
dataset = torchvision.datasets.CIFAR10(DATA_DIRPATH, train=True, download=True)
size = len(dataset)
indices = list(range(size))
random.shuffle(indices)
train_size = round(TRAIN_PCT * size)
train_indices = indices[:train_size]
write_lines(os.path.join(SPLIT_DIRPATH, "train.txt"), train_indices)
val_indices = indices[train_size:]
write_lines(os.path.join(SPLIT_DIRPATH, "val.txt"), val_indices)
def __init__(self, cfg_filepath: str):
set_deterministic_seed()
with open(from_root(cfg_filepath), "r") as file:
self.cfg = DictObject(json.load(file))
for sub_dirname in ("logs", "checkpoints", "debug"):
os.makedirs(self.from_out(sub_dirname), exist_ok=True)
self.model = RefineNet_4Cascaded()
self.model = self.model.cuda()
state_dict = torch.load("TODO", map_location=model_device(self.model))
self.model.backbone.load_state_dict(state_dict)
self.train_loader = load_pascal_voc_train(16)
self.infer_train_loader = load_pascal_voc_infer("train", 16)
self.infer_val_loader = load_pascal_voc_infer("val", 16)
def main() -> None:
set_deterministic_seed()
dataset = ConcatDataset([
torchvision.datasets.SVHN(DATA_DIRPATH, split="train", download=True),
torchvision.datasets.SVHN(DATA_DIRPATH, split="extra", download=True)
])
size = len(dataset)
indices = list(range(size))
random.shuffle(indices)
train_size = round(TRAIN_PCT * size)
train_indices = indices[:train_size]
write_lines(os.path.join(SPLIT_DIRPATH, "train.txt"), train_indices)
val_indices = indices[train_size:]
write_lines(os.path.join(SPLIT_DIRPATH, "val.txt"), val_indices)
def __init__(self, cfg_filepath: str, data_loader: DataLoader,
adversary: Optional[Adversary], out_name: str,
visualize_adversary: int):
set_deterministic_seed()
with open(from_root(cfg_filepath), "r") as file:
self.cfg = DictObject(json.load(file))
os.makedirs(self.from_out("inference"), exist_ok=True)
os.makedirs(self.from_out(f"inference_debug/{out_name}"),
exist_ok=True)
self.model = create_resnet(self.cfg)
self.model = self.model.to(self.cfg.model.device)
self.load_best_epoch()
self.loader = data_loader
self.adversary = adversary
self.out_name = out_name
self.visualize_adversary = visualize_adversary
Tracker.reset()
def setUp(self) -> None:
set_deterministic_seed()