v2.0.1 (Jul 24, 2026)

Bug fix

• GaussianFilterOperator now works with PyTorch (and array-api-strict) CPU
  arrays under modern scipy. The CPU path previously passed the input array
  straight to scipy.ndimage.gaussian_filter; with array-API-aware scipy
  (SCIPY_ARRAY_API / scipy ≥ 1.16) this made scipy compute natively in the
  input's namespace, where gaussian_filter1d reverses the kernel with a
  negative-step slice (weights[::-1]) that PyTorch does not support. The
  operator now filters on a NumPy view (zero-copy on CPU) and converts back, so
  it is robust regardless of the scipy version or SCIPY_ARRAY_API. This
  fixes downstream use via e.g. DeepInv.