Fftshift for FourierOp

[ckolbPTB]

No description provided.

[ckolbPTB]

@JoHa0811 and @schuenke: it would be great if you could test this with your data to make sure everything is now nicely centered.

[github-actions]

Coverage Report

File	Stmts	Miss	Cover	Missing
src/mrpro/algorithms
_remove_readout_os.py	30	2	93%	46, 48
src/mrpro/data
_AccelerationFactor.py	13	13	0%	15–36
_AcqInfo.py	64	1	98%	93
_DcfData.py	78	3	96%	71, 97, 178
_IData.py	57	7	88%	43, 48, 109, 114, 122, 128, 132
_IHeader.py	58	2	97%	86, 113
_KData.py	86	6	93%	83–84, 126, 153, 158–159
_KHeader.py	125	9	93%	34, 81, 119, 170, 181, 188–189, 192, 199
_KNoise.py	21	9	57%	48–61
_SpatialDimension.py	39	5	87%	74, 77, 84–86
_TrajectoryDescription.py	14	1	93%	35
src/mrpro/data/traj_calculators
_KTrajectoryCalculator.py	26	3	88%	37, 63, 65
_KTrajectoryIsmrmrd.py	15	2	87%	55, 62
_KTrajectoryPulseq.py	33	1	97%	72
src/mrpro/operators
_FourierOp.py	112	6	95%	43–50, 205, 255
_LinearOperator.py	21	3	86%	28, 47, 52
_ZeroPadOp.py	15	1	93%	45
src/mrpro/utils
_rgetattr.py	3	1	67%	20
_zero_pad_or_crop.py	32	6	81%	37, 41, 64, 67, 70, 73
TOTAL	1451	81	94%

Tests	Skipped	Failures	Errors	Time
104	0 💤	0 ❌	0 🔥	32.429s ⏱️

[schuenke]

@JoHa0811 and @schuenke: it would be great if you could test this with your data to make sure everything is now nicely centered.

I tested the following cases that worked fine:

1. 2D dynamic cartesian pulseq data using KTrajectoryCartesian calculator
2. 2D dynamic radial pulseq data using KTrajectoryRadial2D calculator

I tried to test KTrajectoryPulseq as well, but it is not working at all currently even after fixing #141 and #144

[fzimmermann89]

couple of possible issues:

1. name sounds like a reshape operation.
2. consider data instead of data
3. consider dat_shape_new -> new_shape (or I might call you yoda)
4. what if data has a different number of dimensions than the new shape?

Suggested change

	def change_data_shape(dat: torch.Tensor, dat_shape_new: tuple[int, ...]) -> torch.Tensor:
	def pad_or_crop(data: torch.Tensor, new_shape: tuple[int, ...]|torch.Size) -> torch.Tensor:
	"""Change shape of data by cropping or zero-padding.
	Parameters
	----------
	data
	data
	new_shape
	target shape of data
	Returns
	-------
	data padded or cropped to shape
	"""
	if len(new_shape) != data.ndim:
	raise ValueError("length of shape should match dimensions of data")
	npad=[]
	for old, new in zip(data.shape, new_shape):
	diff = new-old
	# ensures padding and cropping lead to the matching asymmetry for odd shape differences
	before = math.trunc(diff/2)
	after = diff - before
	npad.append(before)
	npad.append(after)
	# npad has to be reversed because pad expects it in reversed order
	if any(npad):
	data = F.pad(data, npad[::-1])
	return data

Or, including the along_dimension functionality:

Suggested change

	def change_data_shape(dat: torch.Tensor, dat_shape_new: tuple[int, ...]) -> torch.Tensor:
	import torch
	import torch.nn.functional as F
	import numpy as np
	import math
	def normalize_index(ndim: int, index:int):
	"""Normlize possibly negative indices
	Parameters
	----------
	ndim
	number of dimensions
	index
	index to normalize. negative indices count from the end.
	Raises
	------
	IndexError
	if index is outside [-ndim,ndim)
	"""
	if 0<index<ndim:
	return index
	elif -ndim<=index<0:
	return ndim + index
	else:
	raise IndexError(f"Invalid index {index} for {ndim} data dimensions")
	def pad_or_crop(data: torch.Tensor, new_shape: tuple[int, ...]|torch.Size, dim:None|tuple[int,...]=None) -> torch.Tensor:
	"""Change shape of data by cropping or zero-padding.
	Parameters
	----------
	data
	data
	new_shape
	desired shape of data
	dim:
	dimensions the new_shape corropsoends to. None (default) is interpreted as last len(new_shape) dimensions.
	Returns
	-------
	data padded or cropped to shape
	"""
	if len(new_shape) > data.ndim:
	raise ValueError("length of new shape should not exceed dimensions of data")
	if dim is None: # Use last dimensions
	new_shape = data.shape[:-len(new_shape)] + new_shape
	else:
	if len(new_shape) != len(dim):
	raise ValueError("length of shape should match length of dim")
	dim = tuple(normalize_index(idx) for idx in dim) # raises if any not in [-data.ndim,data.ndim)
	if len(dim)!=len(set(dim)): # this is why we normalize
	raise ValueError("repeated values are not allowed in dims")
	(new_shape:=torch.tensor(data.shape))[dim]=new_shape
	npad=[]
	for old, new in zip(data.shape, new_shape):
	diff = new-old
	after = math.trunc(diff/2)
	before = diff - after
	npad.append(before)
	npad.append(after)
	if any(npad):
	# F.pad expects paddings in reversed order
	data = F.pad(data, npad[::-1])
	return data

[ckolbPTB]

I incorporated your second suggestion with the dimension option to be able to use it direclty in the PadOp

[fzimmermann89]

why is this a tensor if you don't use tuple indexing in the next line?

[fzimmermann89]

is there a reason for this loop instead of just indexing with dims?

[fzimmermann89]

why are the shapes tensors anyway?

[ckolbPTB]

otherwise the difference in the line below does not work

[fzimmermann89]

the padding is broken for dims=None.