fix: centre the Gaussian kernel and correct the CPU Scharr x-kernel

[Hendrik-code]

Stacked on #57 — review that one first; this PR's diff is only its own commit.

Two convolution kernels that were not what their names say. Both silent: the pipeline runs, nothing raises, the output is simply not the operation the config asked for.

The Gaussian was not centred

get_gaussian_kernel1d sampled at arange(kernel_size) — 0, 1, 2 — putting the peak at index 0 rather than the centre tap. get_gaussian_kernel3d is the outer product of three of those, so its maximum sat at corner [0, 0, 0].

RandomGaussianBlurGPU and RandomUnsharpMaskGPU therefore blurred and translated the image by about a voxel. The segmentation mask is never convolved, so it did not move with it — the two came out misaligned. The kernel is now sampled on a centred linspace(-half, half, k).

The 2D CPU Scharr x-kernel was not a gradient operator

kernel_x = torch.tensor([[-3, 0, 3], [-10, 0, -10], [-3, 0, 3]])   # sums to -20
kernel_y = torch.tensor([[-3, -10, -3], [0, 0, 0], [3, 10, 3]])    # sums to 0

The middle row should be [-10, 0, 10]. Summing to −20 means it responds to constant regions. kernel_y on the very next line has always been correct, which is what marks this as a typo rather than an intentional variant. The 3D tables are unaffected.

Testing

unit_tests/test_kernel_correctness.py, 13 tests / 23 subtests. 18 checks fail against the parent commit, verified by stashing the source change and re-running.

The one that captures why the blur bug mattered:

def test_blurring_an_impulse_leaves_its_centre_of_mass_in_place(self):
    # the translation is the part that actually hurt: the mask does not move with it

test_the_old_uncentred_formula_really_was_off_centre pins the old formula explicitly so nobody reintroduces it as a simplification. The Scharr side asserts the defining properties — sums to zero, antisymmetric about its differencing axis, zero response on a constant image — plus a sanity check that it still detects the edge it is for, and a guard that the untouched Laplace branch of the same method still sums to zero.

Full suite: 38 passed, 165 subtests. ruff and mypy clean.

Compatibility

Models trained before this change saw the old behaviour and will not reproduce against it. No config key, parameter or default changed.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[NathanMolinier]

Looks good