Explicit shape comparison for dpnp and numpy outputs
#2295
Conversation
|
View rendered docs @ https://intelpython.github.io/dpnp/pull/2295/index.html |
|
Array API standard conformance tests for dpnp=0.18.0dev1=py312he4f9c94_17 ran successfully. |
vtavana
changed the title
dpnp and numpy outputs
vtavana
requested review from
antonwolfy,
AlexanderKalistratov and
vlad-perevezentsev
as code owners
vtavana
force-pushed
the
compare-shape
branch
from
2badc2f to
135df75
Compare
vtavana
force-pushed
the
compare-shape
branch
7 times, most recently
from
a47ee75 to
94c792e
Compare
dpnp/dpnp_iface_linearalgebra.py
Outdated
| We need to adjust the behavior of the multiply function when it is | ||
| used for special cases of scalar-array dots. |
There was a problem hiding this comment.
It is a bit unclear what kind of adjustment is implementing. To have the same dtype as numpy, i.e. dpnp.dot(a, sc).dtype == dpnp.float64, or what?
There was a problem hiding this comment.
Do we need to state that change in the changelog?
There was a problem hiding this comment.
Updated the explanation and added an entry to change log. Note that in this PR, only the behavior of dpnp.outer is changed because of using this function. For other functions that fall back on _call_multiply nothing is changed in this PR.
| @@ -292,7 +292,7 @@ def test_indexing_array_negative_strides(self): | |||
|
|
|||
| slices = (slice(None), dpnp.array([0, 1, 2, 3])) | |||
| arr[slices] = 10 | |||
| assert_array_equal(arr, 10.0) | |||
| assert_equal(arr, 10.0, strict=False) | |||
There was a problem hiding this comment.
Why don't we need strict here?
There was a problem hiding this comment.
It is comparing an array with a scalar, if strict is True, shape comparison will fail.
There was a problem hiding this comment.
But I thought in that case we will check that dpnp has shape () and it will not raise any exception.
There was a problem hiding this comment.
The logic that is currently used is for comparing shape of the two arrays and the exception is for a case when dpnp return a 0D array and numpy return a scalar, for which, the code validates dpnp output is indeed a 0D array.
We do not have any check-point to avoid shape comparison of a non-0D dpnp array (here the array has a shape of (4, 4)) and a scalar.
In the test suite, the result arrays from
dpnpandnumpywere compared, but their shapes were not explicitly checked for equality. As a result, tests could pass even if the shapes differed, such as one being(1,)and the other().This PR adds an explicit check to ensure that the output shapes of
numpyanddpnpmatch. Additionally, tests have been updated accordingly to pass with this new check.