learning.as_real

bartorch.learning.as_real()

Real and imaginary parts of input stacked along a new leading axis.

The pair is placed in front rather than in the trailing axis used by torch.view_as_real(), matching the channel-first convention of convolutional networks and of torchio.ScalarImage.

Parameters:

input (torch.Tensor) – Complex tensor of any shape. A real tensor is given a zero imaginary part, so that magnitude images and complex k-space can be carried in the same layout.

Returns:

Real tensor of shape (2, *input.shape).

Return type:

torch.Tensor

Notes

torchio.ScalarImage requires four axes, (channels, width, height, depth), so a two-dimensional image is unsqueezed to (2, height, width, 1) before it is passed on.

Examples

>>> volume = learning.as_real(image)                      # (2, x, y, z)
>>> subject = torchio.Subject(image=torchio.ScalarImage(tensor=volume))