ENH: apply_where (migrate _lazywhere from scipy)

Author: crusaderky

docs/api-reference.md

@@ -6,6 +6,7 @@
     :nosignatures:
     :toctree: generated
 
+    apply_where
     at
     atleast_nd
     broadcast_shapes

pixi.lock

@@ -64,6 +64,7 @@ numpydoc = ">=1.8.0,<2"
 array-api-strict = "*"
 numpy = "*"
 pytest = "*"
+hypothesis = "*"
 dask-core = "*"  # No distributed, tornado, etc.
 # NOTE: don't add jax, pytorch, sparse, cupy here
 # as they slow down mypy and are not portable across target OSs
@@ -79,6 +80,7 @@ lint = { depends-on = ["pre-commit", "pylint", "mypy", "pyright"] }
 [tool.pixi.feature.tests.dependencies]
 pytest = ">=6"
 pytest-cov = ">=3"
+hypothesis = "*"
 array-api-strict = "*"
 numpy = "*"
 
@@ -232,6 +234,10 @@ reportMissingTypeStubs = false
 reportUnreachable = false
 # ruff handles this
 reportUnusedParameter = false
+# cyclic imports inside function bodies
+reportImportCycles = false
+# PyRight can't trace types in lambdas
+reportUnknownLambdaType = false
 
 executionEnvironments = [
     { root = "tests", reportPrivateUsage = false },

pyproject.toml

@@ -3,6 +3,7 @@
 from ._delegation import isclose, pad
 from ._lib._at import at
 from ._lib._funcs import (
+    apply_where,
     atleast_nd,
     broadcast_shapes,
     cov,
@@ -19,6 +20,7 @@
 # pylint: disable=duplicate-code
 __all__ = [
     "__version__",
+    "apply_where",
     "at",
     "atleast_nd",
     "broadcast_shapes",

src/array_api_extra/__init__.py

@@ -15,6 +15,7 @@
     is_jax_array,
     is_writeable_array,
 )
+from ._utils._helpers import meta_namespace
 from ._utils._typing import Array, SetIndex
 
 
@@ -263,6 +264,8 @@ def _op(
         Array
             Updated `x`.
         """
+        from ._funcs import apply_where  # pylint: disable=cyclic-import
+
         x, idx = self._x, self._idx
         xp = array_namespace(x, y) if xp is None else xp
 
@@ -295,8 +298,10 @@ def _op(
             y_xp = xp.asarray(y, dtype=x.dtype)
             if y_xp.ndim == 0:
                 if out_of_place_op:  # add(), subtract(), ...
-                    # FIXME: suppress inf warnings on dask with lazywhere
-                    out = xp.where(idx, out_of_place_op(x, y_xp), x)
+                    # suppress inf warnings on Dask
+                    out = apply_where(
+                        idx, (x, y_xp), out_of_place_op, fill_value=x, xp=xp
+                    )
                     # Undo int->float promotion on JAX after _AtOp.DIVIDE
                     out = xp.astype(out, x.dtype, copy=False)
                 else:  # set()
@@ -420,9 +425,16 @@ def min(
         xp: ModuleType | None = None,
     ) -> Array:  # numpydoc ignore=PR01,RT01
         """Apply ``x[idx] = minimum(x[idx], y)`` and return the updated array."""
+        # On Dask, this function runs on the chunks, so we need to determine the
+        # namespace that Dask is wrapping.
+        # Note that da.minimum _incidentally_ works on numpy, cupy, and sparse
+        # thanks to all these meta-namespaces implementing the __array_ufunc__
+        # interface, but there's no guarantee that it will work for other
+        # wrapped libraries in the future.
         xp = array_namespace(self._x) if xp is None else xp
+        mxp = meta_namespace(self._x, xp=xp)
         y = xp.asarray(y)
-        return self._op(_AtOp.MIN, xp.minimum, xp.minimum, y, copy=copy, xp=xp)
+        return self._op(_AtOp.MIN, mxp.minimum, mxp.minimum, y, copy=copy, xp=xp)
 
     def max(
         self,
@@ -432,6 +444,8 @@ def max(
         xp: ModuleType | None = None,
     ) -> Array:  # numpydoc ignore=PR01,RT01
         """Apply ``x[idx] = maximum(x[idx], y)`` and return the updated array."""
+        # See note on min()
         xp = array_namespace(self._x) if xp is None else xp
+        mxp = meta_namespace(self._x, xp=xp)
         y = xp.asarray(y)
-        return self._op(_AtOp.MAX, xp.maximum, xp.maximum, y, copy=copy, xp=xp)
+        return self._op(_AtOp.MAX, mxp.maximum, mxp.maximum, y, copy=copy, xp=xp)

src/array_api_extra/_lib/_at.py

@@ -5,17 +5,23 @@
 
 import math
 import warnings
-from collections.abc import Sequence
+from collections.abc import Callable, Sequence
 from types import ModuleType
-from typing import cast
+from typing import cast, overload
 
 from ._at import at
 from ._utils import _compat, _helpers
-from ._utils._compat import array_namespace, is_jax_array
-from ._utils._helpers import asarrays, eager_shape, ndindex
+from ._utils._compat import (
+    array_namespace,
+    is_dask_namespace,
+    is_jax_array,
+    is_jax_namespace,
+)
+from ._utils._helpers import asarrays, eager_shape, meta_namespace, ndindex
 from ._utils._typing import Array
 
 __all__ = [
+    "apply_where",
     "atleast_nd",
     "broadcast_shapes",
     "cov",
@@ -29,6 +35,146 @@
 ]
 
 
+@overload
+def apply_where(  # type: ignore[no-any-explicit,no-any-decorated] # numpydoc ignore=GL08
+    cond: Array,
+    args: Array | tuple[Array, ...],
+    f1: Callable[..., Array],
+    f2: Callable[..., Array],
+    /,
+    *,
+    xp: ModuleType | None = None,
+) -> Array: ...
+
+
+@overload
+def apply_where(  # type: ignore[no-any-explicit,no-any-decorated] # numpydoc ignore=GL08
+    cond: Array,
+    args: Array | tuple[Array, ...],
+    f1: Callable[..., Array],
+    /,
+    *,
+    fill_value: Array | int | float | complex | bool,
+    xp: ModuleType | None = None,
+) -> Array: ...
+
+
+def apply_where(  # type: ignore[no-any-explicit] # numpydoc ignore=PR01,PR02
+    cond: Array,
+    args: Array | tuple[Array, ...],
+    f1: Callable[..., Array],
+    f2: Callable[..., Array] | None = None,
+    /,
+    *,
+    fill_value: Array | int | float | complex | bool | None = None,
+    xp: ModuleType | None = None,
+) -> Array:
+    """
+    Run one of two elementwise functions depending on a condition.
+
+    Equivalent to ``f1(*args) if cond else fill_value`` performed elementwise
+    when `fill_value` is defined, otherwise to ``f1(*args) if cond else f2(*args)``.
+
+    Parameters
+    ----------
+    cond : array
+        The condition, expressed as a boolean array.
+    args : Array or tuple of Arrays
+        Argument(s) to `f1` (and `f2`). Must be broadcastable with `cond`.
+    f1 : callable
+        Elementwise function of `args`, returning a single array.
+        Where `cond` is True, output will be ``f1(arg0[cond], arg1[cond], ...)``.
+    f2 : callable, optional
+        Elementwise function of `args`, returning a single array.
+        Where `cond` is False, output will be ``f2(arg0[cond], arg1[cond], ...)``.
+        Mutually exclusive with `fill_value`.
+    fill_value : Array or scalar, optional
+        If provided, value with which to fill output array where `cond` is False.
+        It does not need to be scalar; it needs however to be broadcastable with
+        `cond` and `args`.
+        Mutually exclusive with `f2`. You must provide one or the other.
+    xp : array_namespace, optional
+        The standard-compatible namespace for `cond` and `args`. Default: infer.
+
+    Returns
+    -------
+    Array
+        An array with elements from the output of `f1` where `cond` is True and either
+        the output of `f2` or `fill_value` where `cond` is False. The returned array has
+        data type determined by type promotion rules between the output of `f1` and
+        either `fill_value` or the output of `f2`.
+
+    Notes
+    -----
+    ``xp.where(cond, f1(*args), f2(*args))`` requires explicitly evaluating `f1` even
+    when `cond` is False, and `f2` when cond is True. This function evaluates each
+    function only for their matching condition, if the backend allows for it.
+
+    On Dask, `f1` and `f2` are applied to the individual chunks and should use functions
+    from the namespace of the chunks.
+
+    Examples
+    --------
+    >>> a = xp.asarray([5, 4, 3])
+    >>> b = xp.asarray([0, 2, 2])
+    >>> def f(a, b):
+    ...     return a // b
+    >>> apply_where(b != 0, (a, b), f, fill_value=xp.nan)
+    array([ nan,  2., 1.])
+    """
+    # Parse and normalize arguments
+    if (f2 is None) == (fill_value is None):
+        msg = "Exactly one of `fill_value` or `f2` must be given."
+        raise TypeError(msg)
+    args_ = list(args) if isinstance(args, tuple) else [args]
+    del args
+
+    xp = array_namespace(cond, *args_) if xp is None else xp
+
+    if getattr(fill_value, "ndim", 0):
+        cond, fill_value, *args_ = xp.broadcast_arrays(cond, fill_value, *args_)
+    else:
+        cond, *args_ = xp.broadcast_arrays(cond, *args_)
+
+    if is_dask_namespace(xp):
+        meta_xp = meta_namespace(cond, fill_value, *args_, xp=xp)
+        # map_blocks doesn't descend into tuples of Arrays
+        return xp.map_blocks(_apply_where, cond, f1, f2, fill_value, *args_, xp=meta_xp)
+    return _apply_where(cond, f1, f2, fill_value, *args_, xp=xp)
+
+
+def _apply_where(  # type: ignore[no-any-explicit]  # numpydoc ignore=PR01,RT01
+    cond: Array,
+    f1: Callable[..., Array],
+    f2: Callable[..., Array] | None,
+    fill_value: Array | int | float | complex | bool | None,
+    *args: Array,
+    xp: ModuleType,
+) -> Array:
+    """Helper of `apply_where`. On Dask, this runs on a single chunk."""
+
+    if is_jax_namespace(xp):
+        # jax.jit does not support assignment by boolean mask
+        return xp.where(cond, f1(*args), f2(*args) if f2 is not None else fill_value)
+
+    temp1 = f1(*(arr[cond] for arr in args))
+
+    if f2 is None:
+        dtype = xp.result_type(temp1, fill_value)
+        if getattr(fill_value, "ndim", 0):
+            out = xp.astype(fill_value, dtype, copy=True)
+        else:
+            out = xp.full_like(cond, dtype=dtype, fill_value=fill_value)
+    else:
+        ncond = ~cond
+        temp2 = f2(*(arr[ncond] for arr in args))
+        dtype = xp.result_type(temp1, temp2)
+        out = xp.empty_like(cond, dtype=dtype)
+        out = at(out, ncond).set(temp2)
+
+    return at(out, cond).set(temp1)
+
+
 def atleast_nd(x: Array, /, *, ndim: int, xp: ModuleType | None = None) -> Array:
     """
     Recursively expand the dimension of an array to at least `ndim`.
@@ -393,12 +539,15 @@ def isclose(
     a_inexact = xp.isdtype(a.dtype, ("real floating", "complex floating"))
     b_inexact = xp.isdtype(b.dtype, ("real floating", "complex floating"))
     if a_inexact or b_inexact:
-        # FIXME: use scipy's lazywhere to suppress warnings on inf
-        out = xp.where(
+        # prevent warnings on numpy and dask on inf - inf
+        mxp = meta_namespace(a, b, xp=xp)
+        out = apply_where(
             xp.isinf(a) | xp.isinf(b),
-            xp.isinf(a) & xp.isinf(b) & (xp.sign(a) == xp.sign(b)),
+            (a, b),
+            lambda a, b: mxp.isinf(a) & mxp.isinf(b) & (mxp.sign(a) == mxp.sign(b)),  # pyright: ignore[reportUnknownArgumentType]
             # Note: inf <= inf is True!
-            xp.abs(a - b) <= (atol + rtol * xp.abs(b)),
+            lambda a, b: mxp.abs(a - b) <= (atol + rtol * mxp.abs(b)),  # pyright: ignore[reportUnknownArgumentType]
+            xp=xp,
         )
         if equal_nan:
             out = xp.where(xp.isnan(a) & xp.isnan(b), xp.asarray(True), out)

src/array_api_extra/_lib/_funcs.py

@@ -9,7 +9,12 @@
 from typing import TYPE_CHECKING, cast
 
 from . import _compat
-from ._compat import array_namespace, is_array_api_obj, is_numpy_array
+from ._compat import (
+    array_namespace,
+    is_array_api_obj,
+    is_dask_namespace,
+    is_numpy_array,
+)
 from ._typing import Array
 
 if TYPE_CHECKING:  # pragma: no cover
@@ -18,6 +23,7 @@
 
 
 __all__ = ["asarrays", "eager_shape", "in1d", "is_python_scalar", "mean"]
+__all__ = ["asarrays", "in1d", "is_python_scalar", "mean", "meta_namespace"]
 
 
 def in1d(
@@ -230,3 +236,33 @@ def eager_shape(x: Array, /) -> tuple[int, ...]:
         msg = "Unsupported lazy shape"
         raise TypeError(msg)
     return cast(tuple[int, ...], shape)
+
+
+def meta_namespace(
+    *arrays: Array | int | float | complex | bool | None,
+    xp: ModuleType | None = None,
+) -> ModuleType:
+    """
+    Get the namespace of Dask chunks.
+
+    On all other backends, just return the namespace of the arrays.
+
+    Parameters
+    ----------
+    *arrays : Array | int | float | complex | bool | None
+        Input arrays.
+    xp : array_namespace, optional
+        The standard-compatible namespace for the input arrays. Default: infer.
+
+    Returns
+    -------
+    array_namespace
+        If xp is Dask, the namespace of the Dask chunks;
+        otherwise, the namespace of the arrays.
+    """
+    xp = array_namespace(*arrays) if xp is None else xp
+    if not is_dask_namespace(xp):
+        return xp
+    # Quietly skip scalars and None's
+    metas = [cast(Array | None, getattr(a, "_meta", None)) for a in arrays]
+    return array_namespace(*metas)

src/array_api_extra/_lib/_utils/_helpers.py

@@ -277,18 +277,7 @@ def test_bool_mask_nd(xp: ModuleType):
     xp_assert_equal(z, xp.asarray([[0, 2, 3], [4, 0, 0]]))
 
 
-@pytest.mark.parametrize(
-    "bool_mask",
-    [
-        False,
-        pytest.param(
-            True,
-            marks=pytest.mark.xfail_xp_backend(
-                Backend.DASK, reason="FIXME need scipy's lazywhere"
-            ),
-        ),
-    ],
-)
+@pytest.mark.parametrize("bool_mask", [False, True])
 def test_no_inf_warnings(xp: ModuleType, bool_mask: bool):
     x = xp.asarray([math.inf, 1.0, 2.0])
     idx = ~xp.isinf(x) if bool_mask else slice(1, None)