diff --git a/CHANGELOG.md b/CHANGELOG.md index d8edbfe38d09..8fb9bc8029df 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -16,6 +16,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 ### Changed * Allowed input array of `uint64` dtype in `dpnp.bincount` [#2361](https://github.com/IntelPython/dpnp/pull/2361) +* The vector norms `ord={None, 1, 2, inf}` and the matrix norms `ord={None, 1, 2, inf, "fro", "nuc"}` now consistently return zero for empty arrays, which are arrays with at least one axis of size zero. This change affects `dpnp.linalg.norm`, `dpnp.linalg.vector_norm`, and `dpnp.linalg.matrix_norm`. Previously, dpnp would either raise errors or return zero depending on the parameters provided [#2371](https://github.com/IntelPython/dpnp/pull/2371) ### Fixed diff --git a/dpnp/linalg/dpnp_utils_linalg.py b/dpnp/linalg/dpnp_utils_linalg.py index c969131ddcc9..6ea4d42a1852 100644 --- a/dpnp/linalg/dpnp_utils_linalg.py +++ b/dpnp/linalg/dpnp_utils_linalg.py @@ -1185,6 +1185,9 @@ def _norm_int_axis(x, ord, axis, keepdims): """ if ord == dpnp.inf: + if x.shape[axis] == 0: + x = dpnp.moveaxis(x, axis, -1) + return dpnp.zeros_like(x, shape=x.shape[:-1]) return dpnp.abs(x).max(axis=axis, keepdims=keepdims) if ord == -dpnp.inf: return dpnp.abs(x).min(axis=axis, keepdims=keepdims) @@ -1220,6 +1223,10 @@ def _norm_tuple_axis(x, ord, row_axis, col_axis, keepdims): """ axis = (row_axis, col_axis) + flag = x.shape[row_axis] == 0 or x.shape[col_axis] == 0 + if flag and ord in [1, 2, dpnp.inf]: + x = dpnp.moveaxis(x, axis, (-2, -1)) + return dpnp.zeros_like(x, shape=x.shape[:-2]) if row_axis == col_axis: raise ValueError("Duplicate axes given.") if ord == 2: diff --git a/dpnp/tests/test_linalg.py b/dpnp/tests/test_linalg.py index 46fbad6dcb9b..07533f159d76 100644 --- a/dpnp/tests/test_linalg.py +++ b/dpnp/tests/test_linalg.py @@ -8,6 +8,7 @@ assert_allclose, assert_almost_equal, assert_array_equal, + assert_equal, assert_raises, assert_raises_regex, suppress_warnings, @@ -24,6 +25,7 @@ has_support_aspect64, is_cpu_device, is_cuda_device, + numpy_version, ) from .third_party.cupy import testing @@ -2074,9 +2076,6 @@ def test_matrix_transpose(): class TestNorm: - def setup_method(self): - numpy.random.seed(42) - @pytest.mark.usefixtures("suppress_divide_numpy_warnings") @pytest.mark.parametrize( "shape", [(0,), (5, 0), (2, 0, 3)], ids=["(0,)", "(5, 0)", "(2, 0, 3)"] @@ -2087,29 +2086,36 @@ def setup_method(self): def test_empty(self, shape, ord, axis, keepdims): a = numpy.empty(shape) ia = dpnp.array(a) + kwarg = {"ord": ord, "axis": axis, "keepdims": keepdims} + if axis is None and a.ndim > 1 and ord in [0, 3]: # Invalid norm order for matrices (a.ndim == 2) or # Improper number of dimensions to norm (a.ndim>2) - with pytest.raises(ValueError): - dpnp.linalg.norm(ia, ord=ord, axis=axis, keepdims=keepdims) + assert_raises(ValueError, dpnp.linalg.norm, ia, **kwarg) + assert_raises(ValueError, numpy.linalg.norm, a, **kwarg) elif axis is None and a.ndim > 2 and ord is not None: # Improper number of dimensions to norm - with pytest.raises(ValueError): - dpnp.linalg.norm(ia, ord=ord, axis=axis, keepdims=keepdims) + assert_raises(ValueError, dpnp.linalg.norm, ia, **kwarg) + assert_raises(ValueError, numpy.linalg.norm, a, **kwarg) elif ( axis is None and ord is not None and a.ndim != 1 and a.shape[-1] == 0 ): - # reduction cannot be performed over zero-size axes - with pytest.raises(ValueError): - dpnp.linalg.norm(ia, ord=ord, axis=axis, keepdims=keepdims) + if ord in [-2, -1, 0, 3]: + # reduction cannot be performed over zero-size axes + assert_raises(ValueError, dpnp.linalg.norm, ia, **kwarg) + assert_raises(ValueError, numpy.linalg.norm, a, **kwarg) + else: + # TODO: when similar changes in numpy are available, instead + # of assert_equal with zero, we should compare with numpy + # ord in [None, 1, 2] + assert_equal(dpnp.linalg.norm(ia, **kwarg), 0) + assert_raises(ValueError, numpy.linalg.norm, a, **kwarg) else: - result = dpnp.linalg.norm(ia, ord=ord, axis=axis, keepdims=keepdims) - expected = numpy.linalg.norm( - a, ord=ord, axis=axis, keepdims=keepdims - ) + result = dpnp.linalg.norm(ia, **kwarg) + expected = numpy.linalg.norm(a, **kwarg) assert_dtype_allclose(result, expected) @pytest.mark.parametrize( @@ -2121,11 +2127,11 @@ def test_0D(self, ord, axis): ia = dpnp.array(a) if axis is None and ord is not None: # Improper number of dimensions to norm - with pytest.raises(ValueError): - dpnp.linalg.norm(ia, ord=ord, axis=axis) + assert_raises(ValueError, dpnp.linalg.norm, ia, ord=ord, axis=axis) + assert_raises(ValueError, numpy.linalg.norm, a, ord=ord, axis=axis) elif axis is not None: - with pytest.raises(AxisError): - dpnp.linalg.norm(ia, ord=ord, axis=axis) + assert_raises(IndexError, dpnp.linalg.norm, ia, ord=ord, axis=axis) + assert_raises(AxisError, numpy.linalg.norm, a, ord=ord, axis=axis) else: result = dpnp.linalg.norm(ia, ord=ord, axis=axis) expected = numpy.linalg.norm(a, ord=ord, axis=axis) @@ -2158,24 +2164,21 @@ def test_1D(self, dtype, ord, axis, keepdims): def test_2D(self, dtype, ord, axis, keepdims): a = generate_random_numpy_array((3, 5), dtype) ia = dpnp.array(a) + kwarg = {"ord": ord, "axis": axis, "keepdims": keepdims} + if (axis in [-1, 0, 1] and ord in ["nuc", "fro"]) or ( (isinstance(axis, tuple) or axis is None) and ord == 3 ): # Invalid norm order for vectors - with pytest.raises(ValueError): - dpnp.linalg.norm(ia, ord=ord, axis=axis, keepdims=keepdims) + assert_raises(ValueError, dpnp.linalg.norm, ia, **kwarg) + assert_raises(ValueError, numpy.linalg.norm, a, **kwarg) else: - result = dpnp.linalg.norm(ia, ord=ord, axis=axis, keepdims=keepdims) - expected = numpy.linalg.norm( - a, ord=ord, axis=axis, keepdims=keepdims - ) + result = dpnp.linalg.norm(ia, **kwarg) + expected = numpy.linalg.norm(a, **kwarg) assert_dtype_allclose(result, expected) @pytest.mark.usefixtures("suppress_divide_numpy_warnings") - @pytest.mark.parametrize( - "dtype", - get_all_dtypes(no_none=True), - ) + @pytest.mark.parametrize("dtype", get_all_dtypes(no_none=True)) @pytest.mark.parametrize( "ord", [None, -dpnp.inf, -2, -1, 1, 2, 3, dpnp.inf, "fro", "nuc"] ) @@ -2188,21 +2191,21 @@ def test_2D(self, dtype, ord, axis, keepdims): def test_ND(self, dtype, ord, axis, keepdims): a = generate_random_numpy_array((2, 3, 4, 5), dtype) ia = dpnp.array(a) + kwarg = {"ord": ord, "axis": axis, "keepdims": keepdims} + if (axis in [-1, 0, 1] and ord in ["nuc", "fro"]) or ( isinstance(axis, tuple) and ord == 3 ): # Invalid norm order for vectors - with pytest.raises(ValueError): - dpnp.linalg.norm(ia, ord=ord, axis=axis, keepdims=keepdims) + assert_raises(ValueError, dpnp.linalg.norm, ia, **kwarg) + assert_raises(ValueError, numpy.linalg.norm, a, **kwarg) elif axis is None and ord is not None: # Improper number of dimensions to norm - with pytest.raises(ValueError): - dpnp.linalg.norm(ia, ord=ord, axis=axis, keepdims=keepdims) + assert_raises(ValueError, dpnp.linalg.norm, ia, **kwarg) + assert_raises(ValueError, numpy.linalg.norm, a, **kwarg) else: - result = dpnp.linalg.norm(ia, ord=ord, axis=axis, keepdims=keepdims) - expected = numpy.linalg.norm( - a, ord=ord, axis=axis, keepdims=keepdims - ) + result = dpnp.linalg.norm(ia, **kwarg) + expected = numpy.linalg.norm(a, **kwarg) assert_dtype_allclose(result, expected) @pytest.mark.usefixtures("suppress_divide_numpy_warnings") @@ -2219,21 +2222,21 @@ def test_ND(self, dtype, ord, axis, keepdims): def test_usm_ndarray(self, dtype, ord, axis, keepdims): a = generate_random_numpy_array((2, 3, 4, 5), dtype) ia = dpt.asarray(a) + kwarg = {"ord": ord, "axis": axis, "keepdims": keepdims} + if (axis in [-1, 0, 1] and ord in ["nuc", "fro"]) or ( isinstance(axis, tuple) and ord == 3 ): # Invalid norm order for vectors - with pytest.raises(ValueError): - dpnp.linalg.norm(ia, ord=ord, axis=axis, keepdims=keepdims) + assert_raises(ValueError, dpnp.linalg.norm, ia, **kwarg) + assert_raises(ValueError, numpy.linalg.norm, a, **kwarg) elif axis is None and ord is not None: # Improper number of dimensions to norm - with pytest.raises(ValueError): - dpnp.linalg.norm(ia, ord=ord, axis=axis, keepdims=keepdims) + assert_raises(ValueError, dpnp.linalg.norm, ia, **kwarg) + assert_raises(ValueError, numpy.linalg.norm, a, **kwarg) else: - result = dpnp.linalg.norm(ia, ord=ord, axis=axis, keepdims=keepdims) - expected = numpy.linalg.norm( - a, ord=ord, axis=axis, keepdims=keepdims - ) + result = dpnp.linalg.norm(ia, **kwarg) + expected = numpy.linalg.norm(a, **kwarg) assert_dtype_allclose(result, expected) @pytest.mark.parametrize("stride", [3, -1, -5]) @@ -2257,8 +2260,7 @@ def test_strided_2D(self, axis, stride): A = numpy.random.rand(20, 30) B = dpnp.asarray(A) slices = tuple(slice(None, None, stride[i]) for i in range(A.ndim)) - a = A[slices] - b = B[slices] + a, b = A[slices], B[slices] result = dpnp.linalg.norm(b, axis=axis) expected = numpy.linalg.norm(a, axis=axis) @@ -2278,8 +2280,7 @@ def test_strided_ND(self, axis, stride): A = numpy.random.rand(12, 16, 20, 24) B = dpnp.asarray(A) slices = tuple(slice(None, None, stride[i]) for i in range(A.ndim)) - a = A[slices] - b = B[slices] + a, b = A[slices], B[slices] result = dpnp.linalg.norm(b, axis=axis) expected = numpy.linalg.norm(a, axis=axis) @@ -2299,6 +2300,49 @@ def test_matrix_norm(self, ord, keepdims): expected = numpy.linalg.matrix_norm(a, ord=ord, keepdims=keepdims) assert_dtype_allclose(result, expected) + @pytest.mark.parametrize( + "xp", + [ + dpnp, + pytest.param( + numpy, + marks=pytest.mark.skipif( + numpy_version() < "2.3.0", + reason="numpy raises an error", + ), + ), + ], + ) + @pytest.mark.parametrize("dtype", [dpnp.float32, dpnp.int32]) + @pytest.mark.parametrize( + "shape_axis", [[(2, 0), None], [(2, 0), (0, 1)], [(0, 2), (0, 1)]] + ) + @pytest.mark.parametrize("ord", [None, "fro", "nuc", 1, 2, dpnp.inf]) + def test_matrix_norm_empty(self, xp, dtype, shape_axis, ord): + shape, axis = shape_axis[0], shape_axis[1] + x = xp.zeros(shape, dtype=dtype) + assert_equal(xp.linalg.norm(x, axis=axis, ord=ord), 0) + + @pytest.mark.parametrize( + "xp", + [ + dpnp, + pytest.param( + numpy, + marks=pytest.mark.skipif( + numpy_version() < "2.3.0", + reason="numpy raises an error", + ), + ), + ], + ) + @pytest.mark.parametrize("dtype", [dpnp.float32, dpnp.int32]) + @pytest.mark.parametrize("axis", [None, 0]) + @pytest.mark.parametrize("ord", [None, 1, 2, dpnp.inf]) + def test_vector_norm_empty(self, xp, dtype, axis, ord): + x = xp.zeros(0, dtype=dtype) + assert_equal(xp.linalg.vector_norm(x, axis=axis, ord=ord), 0) + @testing.with_requires("numpy>=2.0") @pytest.mark.parametrize( "ord", [None, -dpnp.inf, -2, -1, 0, 1, 2, 3.5, dpnp.inf] @@ -2320,13 +2364,10 @@ def test_vector_norm_0D(self, ord): def test_vector_norm_1D(self, ord, axis, keepdims): a = generate_random_numpy_array(10) ia = dpnp.array(a) + kwarg = {"ord": ord, "axis": axis, "keepdims": keepdims} - result = dpnp.linalg.vector_norm( - ia, ord=ord, axis=axis, keepdims=keepdims - ) - expected = numpy.linalg.vector_norm( - a, ord=ord, axis=axis, keepdims=keepdims - ) + result = dpnp.linalg.vector_norm(ia, **kwarg) + expected = numpy.linalg.vector_norm(a, **kwarg) assert_dtype_allclose(result, expected) @testing.with_requires("numpy>=2.0") @@ -2343,29 +2384,26 @@ def test_vector_norm_1D(self, ord, axis, keepdims): def test_vector_norm_ND(self, ord, axis, keepdims): a = numpy.arange(120).reshape(2, 3, 4, 5) ia = dpnp.array(a) + kwarg = {"ord": ord, "axis": axis, "keepdims": keepdims} - result = dpnp.linalg.vector_norm( - ia, ord=ord, axis=axis, keepdims=keepdims - ) - expected = numpy.linalg.vector_norm( - a, ord=ord, axis=axis, keepdims=keepdims - ) + result = dpnp.linalg.vector_norm(ia, **kwarg) + expected = numpy.linalg.vector_norm(a, **kwarg) assert_dtype_allclose(result, expected) def test_error(self): - ia = dpnp.arange(120).reshape(2, 3, 4, 5) + a = numpy.arange(120).reshape(2, 3, 4, 5) + ia = dpnp.array(a) # Duplicate axes given - with pytest.raises(ValueError): - dpnp.linalg.norm(ia, axis=(2, 2)) + assert_raises(ValueError, dpnp.linalg.norm, ia, axis=(2, 2)) + assert_raises(ValueError, numpy.linalg.norm, a, axis=(2, 2)) #'axis' must be None, an integer or a tuple of integers - with pytest.raises(TypeError): - dpnp.linalg.norm(ia, axis=[2]) + assert_raises(TypeError, dpnp.linalg.norm, ia, axis=[2]) + assert_raises(TypeError, numpy.linalg.norm, a, axis=[2]) # Invalid norm order for vectors - with pytest.raises(ValueError): - dpnp.linalg.norm(ia, axis=1, ord=[3]) + assert_raises(ValueError, dpnp.linalg.norm, ia, axis=1, ord=[3]) class TestQr: