Function normalize_transform

monai/networks/utils.py:243–286

Compute an affine matrix according to the input shape. The transform normalizes the homogeneous image coordinates to the range of `[-1, 1]`. Currently the following source coordinates are supported: - `align_corners=False`, `zero_centered=False`, normalizing from ``[-0.5, d-0.

(
    shape,
    device: torch.device | str | None = None,
    dtype: torch.dtype | None = None,
    align_corners: bool = False,
    zero_centered: bool = False,
)

def normalize_transform(
    shape,
    device: torch.device | str | None = None,
    dtype: torch.dtype | None = None,
    align_corners: bool = False,
    zero_centered: bool = False,
) -> torch.Tensor:
    """
    Compute an affine matrix according to the input shape.
    The transform normalizes the homogeneous image coordinates to the
    range of `[-1, 1]`.  Currently the following source coordinates are supported:

        - `align_corners=False`, `zero_centered=False`, normalizing from ``[-0.5, d-0.5]``.
        - `align_corners=True`, `zero_centered=False`, normalizing from ``[0, d-1]``.
        - `align_corners=False`, `zero_centered=True`, normalizing from ``[-(d-1)/2, (d-1)/2]``.
        - `align_corners=True`, `zero_centered=True`, normalizing from ``[-d/2, d/2]``.

    Args:
        shape: input spatial shape, a sequence of integers.
        device: device on which the returned affine will be allocated.
        dtype: data type of the returned affine
        align_corners: if True, consider -1 and 1 to refer to the centers of the
            corner pixels rather than the image corners.
            See also: https://pytorch.org/docs/stable/nn.functional.html#torch.nn.functional.grid_sample
        zero_centered: whether the coordinates are normalized from a zero-centered range, default to `False`.
            Setting this flag and `align_corners` will jointly specify the normalization source range.
    """
    shape = convert_to_tensor(shape, torch.float64, device=device, wrap_sequence=True, track_meta=False)
    norm = shape.clone().detach().to(dtype=torch.float64, device=device)  # no in-place change
    if align_corners:
        norm[norm <= 1.0] = 2.0
        norm = 2.0 / (norm if zero_centered else norm - 1.0)
        norm = torch.diag(torch.cat((norm, torch.ones((1,), dtype=torch.float64, device=device))))
        if not zero_centered:  # else shift is 0
            norm[:-1, -1] = -1.0
    else:
        norm[norm <= 0.0] = 2.0
        norm = 2.0 / (norm - 1.0 if zero_centered else norm)
        norm = torch.diag(torch.cat((norm, torch.ones((1,), dtype=torch.float64, device=device))))
        if not zero_centered:
            norm[:-1, -1] = 1.0 / shape - 1.0
    norm = norm.unsqueeze(0).to(dtype=dtype)
    norm.requires_grad = False
    return norm  # type: ignore


def to_norm_affine(