Skip to content

Fix test_sum to be more numerically correct #290

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 3 commits into from
Oct 2, 2024
Merged

Fix test_sum to be more numerically correct #290

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
aefe195
Split out assert_scalar_equals and assert_scalar_isclose
asmeurer Aug 8, 2024
8f240f6
Pass arguments through properly
asmeurer Sep 12, 2024
621316b
Use condition numbers to filter out ill-conditioned value testing for…
asmeurer Sep 12, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
40 changes: 37 additions & 3 deletions array_api_tests/pytest_helpers.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -397,7 +397,7 @@ def assert_scalar_equals(
kw: dict = {},
):
"""
Assert a 0d array, convered to a scalar, is as expected, e.g.
Assert a 0d array, converted to a scalar, is as expected, e.g.

>>> x = xp.ones(5, dtype=xp.uint8)
>>> out = xp.sum(x)
Expand All @@ -407,6 +407,8 @@ def assert_scalar_equals(

>>> assert int(out) == 5

NOTE: This function does *exact* comparison, even for floats. For
approximate float comparisons use assert_scalar_isclose
"""
__tracebackhide__ = True
repr_name = repr_name if idx == () else f"{repr_name}[{idx}]"
Expand All @@ -418,8 +420,40 @@ def assert_scalar_equals(
msg = f"{repr_name}={out}, but should be {expected} [{f_func}]"
assert cmath.isnan(out), msg
else:
msg = f"{repr_name}={out}, but should be roughly {expected} [{f_func}]"
assert cmath.isclose(out, expected, rel_tol=0.25, abs_tol=1), msg
msg = f"{repr_name}={out}, but should be {expected} [{f_func}]"
assert out == expected, msg


def assert_scalar_isclose(
func_name: str,
*,
rel_tol: float = 0.25,
abs_tol: float = 1,
type_: ScalarType,
idx: Shape,
out: Scalar,
expected: Scalar,
repr_name: str = "out",
kw: dict = {},
):
"""
Assert a 0d array, converted to a scalar, is close to the expected value, e.g.

>>> x = xp.ones(5., dtype=xp.float64)
>>> out = xp.sum(x)
>>> assert_scalar_isclose('sum', type_int, out=(), out=int(out), expected=5.)

is equivalent to

>>> assert math.isclose(float(out) == 5.)

"""
__tracebackhide__ = True
repr_name = repr_name if idx == () else f"{repr_name}[{idx}]"
f_func = f"{func_name}({fmt_kw(kw)})"
msg = f"{repr_name}={out}, but should be roughly {expected} [{f_func}]"
assert type_ in [float, complex] # Sanity check
assert cmath.isclose(out, expected, rel_tol=rel_tol, abs_tol=abs_tol), msg


def assert_fill(
Expand Down
52 changes: 45 additions & 7 deletions array_api_tests/test_statistical_functions.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -17,6 +17,7 @@


@pytest.mark.min_version("2023.12")
@pytest.mark.unvectorized
@given(
x=hh.arrays(
dtype=hh.numeric_dtypes,
Expand Down Expand Up @@ -80,10 +81,15 @@ def test_cumulative_sum(x, data):
if dh.is_int_dtype(out.dtype):
m, M = dh.dtype_ranges[out.dtype]
assume(m <= expected <= M)
ph.assert_scalar_equals("cumulative_sum", type_=scalar_type,
idx=out_idx.raw, out=out_val,
expected=expected)

ph.assert_scalar_equals("cumulative_sum", type_=scalar_type,
idx=out_idx.raw, out=out_val,
expected=expected)
else:
condition_number = _sum_condition_number(elements)
assume(condition_number < 1e6)
ph.assert_scalar_isclose("cumulative_sum", type_=scalar_type,
idx=out_idx.raw, out=out_val,
expected=expected)

def kwarg_dtypes(dtype: DataType) -> st.SearchStrategy[Optional[DataType]]:
dtypes = [d2 for d1, d2 in dh.promotion_table if d1 == dtype]
Expand Down Expand Up @@ -176,6 +182,16 @@ def test_min(x, data):
ph.assert_scalar_equals("min", type_=scalar_type, idx=out_idx, out=min_, expected=expected)


def _prod_condition_number(elements):
# Relative condition number using the infinity norm
abs_max = max([abs(i) for i in elements])
abs_min = min([abs(i) for i in elements])

if abs_min == 0:
return float('inf')

return abs_max / abs_min

@pytest.mark.unvectorized
@given(
x=hh.arrays(
Expand Down Expand Up @@ -225,7 +241,13 @@ def test_prod(x, data):
if dh.is_int_dtype(out.dtype):
m, M = dh.dtype_ranges[out.dtype]
assume(m <= expected <= M)
ph.assert_scalar_equals("prod", type_=scalar_type, idx=out_idx, out=prod, expected=expected)
ph.assert_scalar_equals("prod", type_=scalar_type, idx=out_idx,
out=prod, expected=expected)
else:
condition_number = _prod_condition_number(elements)
assume(condition_number < 1e15)
ph.assert_scalar_isclose("prod", type_=scalar_type, idx=out_idx,
out=prod, expected=expected)


@pytest.mark.skip(reason="flaky") # TODO: fix!
Expand Down Expand Up @@ -264,8 +286,16 @@ def test_std(x, data):
)
# We can't easily test the result(s) as standard deviation methods vary a lot

def _sum_condition_number(elements):
sum_abs = sum([abs(i) for i in elements])
abs_sum = abs(sum(elements))

@pytest.mark.unvectorized
if abs_sum == 0:
return float('inf')

return sum_abs / abs_sum

# @pytest.mark.unvectorized
@given(
x=hh.arrays(
dtype=hh.numeric_dtypes,
Expand Down Expand Up @@ -314,7 +344,15 @@ def test_sum(x, data):
if dh.is_int_dtype(out.dtype):
m, M = dh.dtype_ranges[out.dtype]
assume(m <= expected <= M)
ph.assert_scalar_equals("sum", type_=scalar_type, idx=out_idx, out=sum_, expected=expected)
ph.assert_scalar_equals("sum", type_=scalar_type, idx=out_idx,
out=sum_, expected=expected)
else:
# Avoid value testing for ill conditioned summations. See
# https://en.wikipedia.org/wiki/Kahan_summation_algorithm#Accuracy and
# https://en.wikipedia.org/wiki/Condition_number.
condition_number = _sum_condition_number(elements)
assume(condition_number < 1e6)
ph.assert_scalar_isclose("sum", type_=scalar_type, idx=out_idx, out=sum_, expected=expected)


@pytest.mark.unvectorized
Expand Down
Loading