Could you give an example of implementing a two-dimensional Fourier transform by use torchkbnufft?

[lalalashini]

Hello,torchkbnufft is great but the example is not easy.So,I want to know if I want to compute the Fourier transform of a two-dimensional array by use torchkbnufft,what should I do?
example:a = torch.rand((4,4))
fft: b = torch.fft.fft2(a)
torchkbnufft:???

[mmuckley]

Converting this to a Discussion as it does not document a bug.

At a high level the equivalent is

import torchkbnufft as tkbn

a = torch.rand(4, 4)
nufft_ob = tkbn.KbNufft(im_size=(4, 4))
b = nufft_ob(a, ktraj)

You need to define a ktraj or k-space trajectory that defines the Fourier points you want to sample. You can think of a regular FFT as estimating the points on a grid between -pi and pi, so you can define your off-grid points in the ktraj with that intuition.

For a full detailed example for 2D arrays, please see the Basic Example notebook.

[lalalashini]

import torch
import torchkbnufft as tkbn

a = torch.rand(4, 4)+1j
print(torch.fft.fft2(a))

nufft_ob = tkbn.KbNufft(im_size=(4, 4))

ktraj = torch.ones(4,4)*torch.pi
print(ktraj)
b = nufft_ob(a, ktraj)

print(b)

D:\anaconda\envs\pytorchcuda\lib\site-packages\torchkbnufft_nufft\interp.py:385: UserWarning: rfloordiv is deprecated, and its behavior will change in a future version of pytorch. It currently rounds toward 0 (like the 'trunc' function NOT 'floor'). This results in incorrect rounding for negative values. To keep the current behavior, use torch.div(a, b, rounding_mode='trunc'), or for actual floor division, use torch.div(a, b, rounding_mode='floor').
if omega.shape[1] / (num_threads // factor) >= min_kspace_per_fork:
Traceback (most recent call last):
File "D:\pycharm\paper\nufftpy-master1\nufftpy\5.py", line 12, in
b = nufft_ob(a, ktraj)
File "D:\anaconda\envs\pytorchcuda\lib\site-packages\torch\nn\modules\module.py", line 1130, in _call_impl
return forward_call(*input, **kwargs)
File "D:\anaconda\envs\pytorchcuda\lib\site-packages\torchkbnufft\modules\kbnufft.py", line 211, in forward
output = tkbnF.kb_table_nufft(
File "D:\anaconda\envs\pytorchcuda\lib\site-packages\torchkbnufft\functional\nufft.py", line 172, in kb_table_nufft
data = kb_table_interp(
File "D:\anaconda\envs\pytorchcuda\lib\site-packages\torchkbnufft\functional\interp.py", line 117, in kb_table_interp
data = KbTableInterpForward.apply(
File "D:\anaconda\envs\pytorchcuda\lib\site-packages\torchkbnufft_autograd\interp.py", line 87, in forward
output = table_interp(
File "D:\anaconda\envs\pytorchcuda\lib\site-packages\torchkbnufft_nufft\interp.py", line 392, in table_interp
kdat = table_interp_fork_over_kspace(
RuntimeError: The following operation failed in the TorchScript interpreter.
Traceback of TorchScript (most recent call last):
File "D:\anaconda\envs\pytorchcuda\lib\site-packages\torchkbnufft_nufft\interp.py", line 310, in table_interp_fork_over_kspace
# collect the results
for ind, future in enumerate(futures):
kdat[:, :, ind:klength:num_forks] = torch.jit.wait(future)
~~~~~~~~~~~~~~ <--- HERE

return kdat

RuntimeError: The following operation failed in the TorchScript interpreter.
Traceback of TorchScript (most recent call last):
File "D:\anaconda\envs\pytorchcuda\lib\site-packages\torchkbnufft_nufft\interp.py", line 288, in
for omega_chunk in omega_chunks:
futures.append(
torch.jit.fork(
~~~~~~~~~~~~~~~
table_interp_one_batch,
~~~~~~~~~~~~~~~~~~~~~~~
image,
~~~~~~
omega_chunk,
~~~~~~~~~~~~
tables,
~~~~~~~
n_shift,
~~~~~~~~
numpoints,
~~~~~~~~~~
table_oversamp,
~~~~~~~~~~~~~~~
offsets,
~~~~~~~ <--- HERE
)
)
File "D:\anaconda\envs\pytorchcuda\lib\site-packages\torchkbnufft_nufft\interp.py", line 171, in table_interp_one_batch

# convert to normalized freq locs
tm = omega / (2 * np.pi / grid_size.to(omega).unsqueeze(-1))
     ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ <--- HERE

# compute interpolation centers

RuntimeError: The size of tensor a (4) must match the size of tensor b (0) at non-singleton dimension 0

WHY?

[mmuckley]

Sorry, I didn't test my example that I wrote. We need batch and channel dimensions (a PyTorch convention). The notebook should be the primary source of truth. Please consult the notebooks for the future.

import torch
import torchkbnufft as tkbn

nufft_ob = tkbn.KbNufft(im_size=(4, 4))
a = torch.rand(1, 1, 4, 4) + 1j*torch.rand(1, 1, 4, 4)
ktraj = torch.randn(2, 128)
b = nufft_ob(a, ktraj)