Merge pull request #465 from PyAutoLabs/claude/autoarray-pixel-scales-int-tuple-wfxlnj

fix: widen int / numpy scalar pixel_scales to a tuple

Author: Jammy2211

autoarray/geometry/geometry_util.py

@@ -32,16 +32,27 @@ def convert_shape_native_1d(shape_native: Union[int, Tuple[int]]) -> Tuple[int]:
 
 def convert_pixel_scales_1d(pixel_scales: ty.PixelScales) -> Tuple[float]:
     """
-    Convert an input pixel scale of type `float` to a tuple `(float,)`. If the input is already a
-    `(float,)` tuple it is returned unchanged.
+    Convert an input pixel scale given as a single real scalar to a tuple `(float,)`. If the
+    input is already a `(float,)` tuple it is returned unchanged.
 
-    This enables users to input the pixel scale as a single float and have the type automatically
-    normalised to `(float,)` which is used internally by 1D data structures.
+    This enables users to input the pixel scale as a single number and have the type
+    automatically normalised to `(float,)` which is used internally by 1D data structures.
+
+    Any concrete real scalar is widened — `int`, `float`, `np.integer` and `np.floating` — not
+    just an exact `float`. An `int` is a natural thing to type by hand, and an `np.floating` is
+    what indexing a numpy array or reading a FITS header returns, so both reach this function on
+    paths a user would consider ordinary. The widened value is cast to a Python `float`, so the
+    tuple this returns is always `(float,)` regardless of what went in.
+
+    A `bool` is deliberately *not* treated as a scalar here (see
+    :func:`autoarray.validate.is_concrete_scalar`), and neither is a JAX tracer — a traced value
+    passes through untouched so the function stays safe inside a `jax.jit`.
 
     Parameters
     ----------
     pixel_scales
-        The pixel scale to convert, either as a plain `float` or a 1-element tuple `(float,)`.
+        The pixel scale to convert, either as a plain real scalar or a 1-element tuple
+        `(float,)`.
 
     Returns
     -------
@@ -56,8 +67,8 @@ def convert_pixel_scales_1d(pixel_scales: ty.PixelScales) -> Tuple[float]:
 
     validate.validate_pixel_scales(pixel_scales=pixel_scales)
 
-    if type(pixel_scales) is float:
-        pixel_scales = (pixel_scales,)
+    if validate.is_concrete_scalar(pixel_scales):
+        pixel_scales = (float(pixel_scales),)
 
     return pixel_scales
 
@@ -197,17 +208,28 @@ def scaled_coordinates_1d_from(
 
 def convert_pixel_scales_2d(pixel_scales: ty.PixelScales) -> Tuple[float, float]:
     """
-    Convert an input pixel scale of type `float` to a tuple `(float, float)`. If the input is
-    already type `(float, float)` it is returned unchanged.
+    Convert an input pixel scale given as a single real scalar to a tuple `(float, float)`. If
+    the input is already type `(float, float)` it is returned unchanged.
+
+    This enables users to input the pixel scale as a single number and have the type
+    automatically normalised to `(float, float)` which is used internally for rectangular 2D
+    grids (where both axes share the same pixel scale).
+
+    Any concrete real scalar is widened — `int`, `float`, `np.integer` and `np.floating` — not
+    just an exact `float`. An `int` is a natural thing to type by hand, and an `np.floating` is
+    what indexing a numpy array or reading a FITS header returns, so both reach this function on
+    paths a user would consider ordinary. The widened value is cast to a Python `float`, so the
+    tuple this returns is always `(float, float)` regardless of what went in.
 
-    This enables users to input the pixel scale as a single float and have the type automatically
-    normalised to `(float, float)` which is used internally for rectangular 2D grids (where
-    both axes share the same pixel scale).
+    A `bool` is deliberately *not* treated as a scalar here (see
+    :func:`autoarray.validate.is_concrete_scalar`), and neither is a JAX tracer — a traced value
+    passes through untouched so the function stays safe inside a `jax.jit`.
 
     Parameters
     ----------
     pixel_scales
-        The pixel scale to convert, either as a plain `float` or a 2-element tuple `(float, float)`.
+        The pixel scale to convert, either as a plain real scalar or a 2-element tuple
+        `(float, float)`.
 
     Returns
     -------
@@ -228,8 +250,8 @@ def convert_pixel_scales_2d(pixel_scales: ty.PixelScales) -> Tuple[float, float]
 
     validate.validate_pixel_scales(pixel_scales=pixel_scales)
 
-    if type(pixel_scales) is float:
-        pixel_scales = (pixel_scales, pixel_scales)
+    if validate.is_concrete_scalar(pixel_scales):
+        pixel_scales = (float(pixel_scales), float(pixel_scales))
 
     return pixel_scales
 

autoarray/type.py

@@ -1,7 +1,12 @@
 import numpy as np
 from typing import List, Tuple, Union
 
-PixelScales = Union[Tuple[float], Tuple[float, float], float]
+# A pixel scale may be given per-axis as a tuple, or as a single real scalar which
+# `geometry_util.convert_pixel_scales_{1d,2d}` widens to that tuple. The scalar forms are
+# listed out because the widening accepts any real scalar, not just an exact `float`.
+PixelScales = Union[
+    Tuple[float], Tuple[float, float], float, int, np.floating, np.integer
+]
 
 
 from autoarray.mask.mask_1d import Mask1D

test_autoarray/geometry/test_geometry_util.py

@@ -3,6 +3,89 @@
 import pytest
 
 
+class _NotAConcreteScalar:
+    """
+    Stand-in for a JAX tracer: not a concrete Python/NumPy scalar, and raising if anything
+    tries to resolve it to a bool, exactly as a tracer does inside `jax.jit`. Mirrors the
+    stand-in in `test_autoarray/test_validate.py` — unit tests here are NumPy-only.
+    """
+
+    def __bool__(self):
+        raise AssertionError("a tracer must never be resolved to a bool")
+
+
+@pytest.mark.parametrize(
+    "pixel_scales", [1, 1.0, np.float64(1.0), np.int32(1), np.float32(1.0)]
+)
+def test__convert_pixel_scales_1d__widens_any_real_scalar(pixel_scales):
+    """
+    Any concrete real scalar widens, not just an exact `float`. `int` is what a user types by
+    hand; `np.floating` is what indexing an array or reading a FITS header returns.
+    """
+    assert aa.util.geometry.convert_pixel_scales_1d(pixel_scales=pixel_scales) == (1.0,)
+
+
+@pytest.mark.parametrize(
+    "pixel_scales", [1, 1.0, np.float64(1.0), np.int32(1), np.float32(1.0)]
+)
+def test__convert_pixel_scales_2d__widens_any_real_scalar(pixel_scales):
+    assert aa.util.geometry.convert_pixel_scales_2d(pixel_scales=pixel_scales) == (
+        1.0,
+        1.0,
+    )
+
+
+@pytest.mark.parametrize("pixel_scales", [1, np.float64(1.0), np.int32(1)])
+def test__convert_pixel_scales__widened_entries_are_python_floats(pixel_scales):
+    """
+    The widened value is cast, so an `int` or a NumPy scalar never reaches the geometry
+    stored on a mask. `1 == 1.0` in Python, so the cast has to be asserted on the type.
+    """
+    (entry_1d,) = aa.util.geometry.convert_pixel_scales_1d(pixel_scales=pixel_scales)
+    assert type(entry_1d) is float
+
+    for entry in aa.util.geometry.convert_pixel_scales_2d(pixel_scales=pixel_scales):
+        assert type(entry) is float
+
+
+def test__convert_pixel_scales__tuple_input_is_returned_unchanged():
+    pixel_scales_1d = (1.0,)
+    assert (
+        aa.util.geometry.convert_pixel_scales_1d(pixel_scales=pixel_scales_1d)
+        is pixel_scales_1d
+    )
+
+    pixel_scales_2d = (1.0, 2.0)
+    assert (
+        aa.util.geometry.convert_pixel_scales_2d(pixel_scales=pixel_scales_2d)
+        is pixel_scales_2d
+    )
+
+
+def test__convert_pixel_scales__a_tracer_passes_through_untouched():
+    """Inside a `jax.jit` the value is traced; widening it would resolve it to a bool."""
+    tracer_like = _NotAConcreteScalar()
+
+    assert (
+        aa.util.geometry.convert_pixel_scales_1d(pixel_scales=tracer_like)
+        is tracer_like
+    )
+    assert (
+        aa.util.geometry.convert_pixel_scales_2d(pixel_scales=tracer_like)
+        is tracer_like
+    )
+
+
+def test__convert_pixel_scales__a_bool_is_not_treated_as_a_scalar():
+    """
+    `bool` is a subclass of `int`, but `True` reaching a pixel scale is a different mistake
+    than the ones this widening serves — `validate.is_concrete_scalar` excludes it, so it is
+    not silently accepted as a pixel scale of 1.0.
+    """
+    assert aa.util.geometry.convert_pixel_scales_1d(pixel_scales=True) is True
+    assert aa.util.geometry.convert_pixel_scales_2d(pixel_scales=True) is True
+
+
 def test__central_pixel_coordinates_1d_from():
     central_pixel_coordinates = aa.util.geometry.central_pixel_coordinates_1d_from(
         shape_slim=(3,)

test_autoarray/test_validate.py

@@ -124,6 +124,57 @@ def test__b6__control__valid_pixel_scales_still_build():
     assert array.pixel_scales == (0.1, 0.2)
 
 
+@pytest.mark.parametrize("pixel_scales", [0, -1, np.int32(0), np.int32(-1)])
+def test__b6__the_guards_still_fire_on_integer_scalars(pixel_scales):
+    """
+    The guards run before the widening, so broadening what counts as a scalar must not open a
+    hole for the integer forms of the same bad input.
+    """
+    with pytest.raises(ValueError, match="pixel_scales"):
+        aa.Array2D.no_mask(values=np.ones((5, 5)), pixel_scales=pixel_scales)
+
+
+@pytest.mark.parametrize(
+    "pixel_scales", [np.float64(0.0), np.float64(-0.1), np.float64("nan")]
+)
+def test__b6__the_guards_still_fire_on_numpy_scalars(pixel_scales):
+    with pytest.raises(ValueError, match="pixel_scales"):
+        aa.Array2D.no_mask(values=np.ones((5, 5)), pixel_scales=pixel_scales)
+
+
+@pytest.mark.parametrize("pixel_scales", [(0, 1), (1, -1)])
+def test__b6__the_guards_still_fire_on_integer_entries_of_a_tuple(pixel_scales):
+    with pytest.raises(ValueError, match="pixel_scales"):
+        aa.Array2D.no_mask(values=np.ones((5, 5)), pixel_scales=pixel_scales)
+
+
+@pytest.mark.parametrize("pixel_scales", [1, np.float64(1.0), np.int32(1)])
+def test__b6__control__an_integer_or_numpy_pixel_scale_builds_and_is_widened(
+    pixel_scales,
+):
+    """
+    The defect this closes: only an exact `float` used to be widened, so an `int` or a NumPy
+    scalar was stored on the mask unconverted and every later use of it raised
+    `TypeError: 'int' object is not subscriptable`.
+    """
+    array = aa.Array2D.no_mask(values=np.ones((5, 5)), pixel_scales=pixel_scales)
+    assert array.pixel_scales == (1.0, 1.0)
+    assert array.pixel_scale == 1.0
+
+    mask = aa.Mask2D.circular(
+        shape_native=(5, 5), radius=2.0, pixel_scales=pixel_scales
+    )
+    assert mask.pixel_scales == (1.0, 1.0)
+
+    grid = aa.Grid2D.uniform(shape_native=(5, 5), pixel_scales=pixel_scales)
+    assert grid.pixel_scales == (1.0, 1.0)
+
+
+def test__b6__control__an_integer_pixel_scale_builds_a_1d_structure():
+    array = aa.Array1D.no_mask(values=np.ones((5,)), pixel_scales=1)
+    assert array.pixel_scales == (1.0,)
+
+
 # ======================================================================================
 # B7 — annulus radii must be ordered
 # ======================================================================================