CityscapesDataModule#

class torch_uncertainty.datamodules.segmentation.CityscapesDataModule(root, batch_size, eval_batch_size=None, crop_size=1024, eval_size=(1024, 2048), train_transform=None, test_transform=None, basic_augment=True, val_split=None, num_workers=1, pin_memory=True, persistent_workers=True)[source]#

DataModule for the Cityscapes dataset.

Parameters:

  • root (str | Path) – Root directory of the datasets.

  • batch_size (int) – Number of samples per batch during training.

  • eval_batch_size (int | None) – Number of samples per batch during evaluation (val and test). Set to batch_size if None. Defaults to None.

  • crop_size (Union[int, tuple[int, int]]) – Desired input image and segmentation mask sizes during training. If crop_size is an int instead of sequence like \((H, W)\), a square crop \((\text{size},\text{size})\) is made. If provided a sequence of length \(1\), it will be interpreted as \((\text{size[0]},\text{size[1]})\). Has to be provided if train_transform is not provided. Otherwise has no effect. Defaults to 1024.

  • eval_size (Union[int, tuple[int, int]]) – Desired input image and segmentation mask sizes during evaluation. If size is an int, smaller edge of the images will be matched to this number, i.e., \(\text{height}>\text{width}\), then image will be rescaled to \((\text{size}\times\text{height}/\text{width},\text{size})\). Has to be provided if test_transform is not provided. Otherwise has no effect. Defaults to (1024,2048).

  • train_transform (Module | None) – Custom training transform. Defaults to None. If not provided, a default transform is used.

  • test_transform (Module | None) – Custom test transform. Defaults to None. If not provided, a default transform is used.

  • basic_augment (bool) – Whether to apply base augmentations. Defaults to True. Only used if train_transform is not provided.

  • val_split (float | None) – Share of training samples to use for validation. Defaults to None.

  • num_workers (int) – Number of dataloaders to use. Defaults to 1.

  • pin_memory (bool) – Whether to pin memory. Defaults to True.

  • persistent_workers (bool) – Whether to use persistent workers. Defaults to True.

Note

By default this datamodule injects the following transforms into the training and validation/test datasets:

Training transforms:

from torchvision.transforms import v2

v2.Compose(
    [
        v2.ToImage(),
        RandomRescale(min_scale=0.5, max_scale=2.0, antialias=True),
        v2.RandomCrop(size=crop_size, pad_if_needed=True),
        v2.ColorJitter(brightness=0.5, contrast=0.5, saturation=0.5),
        v2.RandomHorizontalFlip(),
        v2.ToDtype(
            {
                tv_tensors.Image: torch.float32,
                tv_tensors.Mask: torch.int64,
                "others": None,
            },
            scale=True,
        ),
        v2.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
    ]
)

Validation/Test transforms:

from torchvision.transforms import v2

v2.Compose(
    [
        v2.ToImage(),
        v2.Resize(size=eval_size, antialias=True),
        v2.ToDtype(
            {
                tv_tensors.Image: torch.float32,
                tv_tensors.Mask: torch.int64,
                "others": None,
            },
            scale=True,
        ),
        v2.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
    ]
)

This behavior can be modified by setting up train_transform and test_transform at initialization.