fix: use accurate lower bound (#132)

Author: flying-sheep

pyproject.toml

@@ -37,7 +37,7 @@ optional-dependencies.doc = [
   "sphinx-autofixture>=0.4.1",
 ]
 optional-dependencies.full = [ "fast-array-utils[accel,dask,sparse]", "h5py", "zarr" ]
-optional-dependencies.sparse = [ "scipy>=1.11" ]
+optional-dependencies.sparse = [ "scipy>=1.13" ]
 optional-dependencies.test = [
   "anndata",
   "fast-array-utils[accel,test-min]",
@@ -102,6 +102,7 @@ overrides.matrix.resolution.features = [
 overrides.matrix.resolution.dependencies = [
   # TODO: move to min dep once this is fixed: https://github.com/tlambert03/hatch-min-requirements/issues/11
   { if = [ "lowest" ], value = "dask==2023.6.1" },
+  { if = [ "lowest" ], value = "scipy==1.13.0" },
 ]
 
 [[tool.hatch.envs.hatch-test.matrix]]
@@ -174,6 +175,7 @@ filterwarnings = [
   "error",
   # codspeed seems to break this dtype added by h5py
   "ignore:.*numpy[.]longdouble:UserWarning",
+  "ignore:FNV hashing is not implemented in Numba:UserWarning",
 ]
 markers = [
   "benchmark: marks tests as benchmark (to run with `--codspeed`)",

src/testing/fast_array_utils/_array_type.py

@@ -7,7 +7,7 @@
 from dataclasses import KW_ONLY, dataclass, field
 from functools import cached_property, partial
 from importlib.metadata import version
-from typing import TYPE_CHECKING, Generic, TypeVar, cast
+from typing import TYPE_CHECKING, Generic, TypedDict, TypeVar, cast
 
 import numpy as np
 from packaging.version import Version
@@ -378,12 +378,20 @@ def random_mat(
     m, n = shape
     return cast(
         "types.CSBase",
-        random_spmat(m, n, density=density, format=format, dtype=dtype, rng=rng)
+        random_spmat(m, n, density=density, format=format, dtype=dtype, **_rng_kw(rng))
         if container == "matrix"
-        else random_sparr(shape, density=density, format=format, dtype=dtype, rng=rng),
+        else random_sparr(shape, density=density, format=format, dtype=dtype, **_rng_kw(rng)),
     )
 
 
+class RngKw(TypedDict):
+    rng: np.random.Generator | None
+
+
+def _rng_kw(rng: np.random.Generator | None) -> RngKw:
+    return RngKw(rng=rng) if Version(version("scipy")) >= Version("1.15") else cast("RngKw", dict(random_state=rng))
+
+
 def _half_chunk_size(a: tuple[int, ...]) -> tuple[int, ...]:
     def half_rounded_up(x: int) -> int:
         div, mod = divmod(x, 2)

tests/test_stats.py

@@ -1,6 +1,7 @@
 # SPDX-License-Identifier: MPL-2.0
 from __future__ import annotations
 
+from importlib.metadata import version
 from importlib.util import find_spec
 from pathlib import Path
 from typing import TYPE_CHECKING, cast
@@ -9,6 +10,7 @@
 import pytest
 import scipy.sparse as sps
 from numpy.exceptions import AxisError
+from packaging.version import Version
 
 from fast_array_utils import stats, types
 from testing.fast_array_utils import SUPPORTED_TYPES, Flags
@@ -39,6 +41,11 @@
 ATS_CUPY_SPARSE = {at for at in SUPPORTED_TYPES if "cupyx.scipy" in str(at)}
 
 
+def _xfail_if_old_scipy(array_type: ArrayType[Any], ndim: Literal[1, 2]) -> pytest.MarkDecorator:
+    cond = ndim == 1 and bool(array_type.flags & Flags.Sparse) and Version(version("scipy")) < Version("1.14")
+    return pytest.mark.xfail(cond, reason="Sparse matrices don’t support 1d arrays")
+
+
 @pytest.fixture(
     scope="session",
     params=[
@@ -134,7 +141,10 @@ def pbmc64k_reduced_raw() -> sps.csr_array[np.float32]:
 @pytest.mark.array_type(skip={*ATS_SPARSE_DS, Flags.Matrix})
 @pytest.mark.parametrize("func", STAT_FUNCS)
 @pytest.mark.parametrize(("ndim", "axis"), [(1, 0), (2, 3), (2, -1)], ids=["1d-ax0", "2d-ax3", "2d-axneg"])
-def test_ndim_error(array_type: ArrayType[Array], func: StatFunNoDtype, ndim: Literal[1, 2], axis: Literal[0, 1] | None) -> None:
+def test_ndim_error(
+    request: pytest.FixtureRequest, array_type: ArrayType[Array], func: StatFunNoDtype, ndim: Literal[1, 2], axis: Literal[0, 1] | None
+) -> None:
+    request.applymarker(_xfail_if_old_scipy(array_type, ndim))
     check_ndim(array_type, ndim)
     # not using the fixture because we don’t need to test multiple dtypes
     np_arr = np.array([[1, 2, 3], [4, 5, 6]], dtype=np.float32)
@@ -148,12 +158,15 @@ def test_ndim_error(array_type: ArrayType[Array], func: StatFunNoDtype, ndim: Li
 
 @pytest.mark.array_type(skip=ATS_SPARSE_DS)
 def test_sum(
+    request: pytest.FixtureRequest,
     array_type: ArrayType[CpuArray | GpuArray | DiskArray | types.DaskArray],
     dtype_in: type[DTypeIn],
     dtype_arg: type[DTypeOut] | None,
     axis: Literal[0, 1] | None,
     np_arr: NDArray[DTypeIn],
+    ndim: Literal[1, 2],
 ) -> None:
+    request.applymarker(_xfail_if_old_scipy(array_type, ndim))
     if np.dtype(dtype_arg).kind in "iu" and (array_type.flags & Flags.Gpu) and (array_type.flags & Flags.Sparse):
         pytest.skip("GPU sparse matrices don’t support int dtypes")
     arr = array_type(np_arr.copy())
@@ -209,7 +222,8 @@ def test_sum_dask_shapes(array_type: ArrayType[types.DaskArray], axis: Literal[0
 
 
 @pytest.mark.array_type(skip=ATS_SPARSE_DS)
-def test_mean(array_type: ArrayType[Array], axis: Literal[0, 1] | None, np_arr: NDArray[DTypeIn]) -> None:
+def test_mean(request: pytest.FixtureRequest, array_type: ArrayType[Array], axis: Literal[0, 1] | None, np_arr: NDArray[DTypeIn], ndim: Literal[1, 2]) -> None:
+    request.applymarker(_xfail_if_old_scipy(array_type, ndim))
     arr = array_type(np_arr)
 
     result = stats.mean(arr, axis=axis)  # type: ignore[arg-type]  # https://github.com/python/mypy/issues/16777
@@ -224,10 +238,13 @@ def test_mean(array_type: ArrayType[Array], axis: Literal[0, 1] | None, np_arr:
 
 @pytest.mark.array_type(skip=Flags.Disk)
 def test_mean_var(
+    request: pytest.FixtureRequest,
     array_type: ArrayType[CpuArray | GpuArray | types.DaskArray],
     axis: Literal[0, 1] | None,
     np_arr: NDArray[DTypeIn],
+    ndim: Literal[1, 2],
 ) -> None:
+    request.applymarker(_xfail_if_old_scipy(array_type, ndim))
     arr = array_type(np_arr)
 
     mean, var = stats.mean_var(arr, axis=axis, correction=1)