Forced Fourier class to output contiguous tensors.

[bwittmann]

Forced Fourier class to output contiguous tensors, which potentially fixes a performance bottleneck.

Description

Some transforms, such as RandKSpaceSpikeNoise, rely on the Fourier class.
In its current state, the Fourier class returns non-contiguous tensors, which potentially limits performance.
For example, when followed by RandHistogramShift, the following warning occurs:

<path_to_monai>/monai/transforms/intensity/array.py:1852: UserWarning: torch.searchsorted(): input value tensor is non-contiguous, this will lower the performance due to extra data copy when converting non-contiguous tensor to contiguous, please use contiguous input value tensor if possible. This message will only appear once per program. (Triggered internally at /opt/conda/conda-bld/pytorch_1716905975447/work/aten/src/ATen/native/BucketizationUtils.h:32.)
  indices = ns.searchsorted(xp.reshape(-1), x.reshape(-1)) - 1

A straightforward fix is to force the Fourier class to output contiguous tensors (see commit).
To reproduce, please run:

from monai.transforms import RandKSpaceSpikeNoise
from monai.transforms.utils import Fourier
import numpy as np

### TEST WITH TRANSFORMS ###   
t = RandKSpaceSpikeNoise(prob=1)

# for torch tensors
a_torch = torch.rand(1, 128, 128, 128)
print(a_torch.is_contiguous())

a_torch_mod = t(a_torch)
print(a_torch_mod.is_contiguous())

# for np arrays
a_np = np.random.rand(1, 128, 128, 128)
print(a_np.flags['C_CONTIGUOUS'])

a_np_mod = t(a_np)  # automatically transformed to torch.tensor
print(a_np_mod.is_contiguous())

### TEST DIRECTLY WITH FOURIER ###
f = Fourier()

# inv_shift_fourier
# for torch tensors
real_torch = torch.randn(1, 128, 128, 128)
im_torch = torch.randn(1, 128, 128, 128)
k_torch = torch.complex(real_torch, im_torch)
print(k_torch.is_contiguous())

out_torch = f.inv_shift_fourier(k_torch, spatial_dims=3)
print(out_torch.is_contiguous())

# for np arrays
real_np = np.random.randn(1, 100, 100, 100)
im_np = np.random.randn(1, 100, 100, 100)
k_np = real_np + 1j * im_np
print(k_np.flags['C_CONTIGUOUS'])

out_np = f.inv_shift_fourier(k_np, spatial_dims=3)
print(out_np.flags['C_CONTIGUOUS'])

# shift_fourier
# for torch tensors
a_torch = torch.rand(1, 128, 128, 128)
print(a_torch.is_contiguous())

out_torch = f.shift_fourier(a_torch, spatial_dims=3)
print(out_torch.is_contiguous())

# for np arrays
a_np = np.random.rand(1, 128, 128, 128)
print(a_np.flags['C_CONTIGUOUS'])

out_np = f.shift_fourier(a_np, spatial_dims=3)
print(out_np.flags['C_CONTIGUOUS'])

Types of changes

• Non-breaking change (fix or new feature that would not break existing functionality).
• Breaking change (fix or new feature that would cause existing functionality to change).
• New tests added to cover the changes.
• Integration tests passed locally by running ./runtests.sh -f -u --net --coverage.
• Quick tests passed locally by running ./runtests.sh --quick --unittests --disttests.
• In-line docstrings updated.
• Documentation updated, tested make html command in the docs/ folder.

[KumoLiu]

Hi @bwittmann, do you plan to finish this PR? If we want to include the change in version 1.4, we only have about one or two weeks left to add new things.

[bwittmann]

Dear @KumoLiu,

following your recommendation, I added as_contiguous as an argument to propose an experimental option.

Best,
Bastian

[KumoLiu]

Thanks for the update, LGTM.

[KumoLiu]

Hi @bwittmann, please help fix the DCO and format issue. Thanks.

[KumoLiu]

You can refer to the instruction here to fix the DCO issue.
https://github.com/Project-MONAI/MONAI/pull/7969/checks?check_run_id=29602735375

[KumoLiu]

/build