bartorch.linop.Hankel

Contents

bartorch.linop.Hankel#

bartorch.linop.Hankel(shape, axis, window)#

A sliding window along axis, BART’s linop_hankelization.

torch.Tensor.unfold(axis, window, 1) as an operator, and the same thing as the trajectory matrix that singular spectrum analysis is built on: an axis of n becomes n - window + 1 positions, each carrying the window samples that start there. The codomain is the domain with that axis shortened and the window added as a last axis, which is where unfold puts it.

BART makes the windows by striding rather than by copying, so the overlap costs nothing to build; the adjoint adds each sample back into every window it appeared in, which is what makes this an operator rather than a view.

Parameters:

  • shape (tuple of int) – The domain, C order.

  • axis (int) – Which axis to slide along.

  • window (int) – How many samples each position carries. At most the length of the axis; equal to it gives one position.

Examples

>>> H = Hankel((64, 8), axis=0, window=16)
>>> H.oshape
(49, 8, 16)
Contents