Implementation of convolve #2205
Conversation
AlexanderKalistratov
requested review from
antonwolfy,
vlad-perevezentsev and
vtavana
as code owners
AlexanderKalistratov
force-pushed
the
convolve
branch
from
84ad9a8 to
6d8689f
Compare
|
View rendered docs @ https://intelpython.github.io/dpnp/pull/2205/index.html |
AlexanderKalistratov
force-pushed
the
convolve
branch
from
6d8689f to
c6dde99
Compare
AlexanderKalistratov
force-pushed
the
correlate_fft
branch
from
5f429be to
6eb94b2
Compare
AlexanderKalistratov
force-pushed
the
convolve
branch
from
c6dde99 to
b39e5c9
Compare
AlexanderKalistratov
force-pushed
the
convolve
branch
from
4168f27 to
3a03fdc
Compare
AlexanderKalistratov
force-pushed
the
convolve
branch
from
3a03fdc to
1c52d67
Compare
|
View rendered docs @ https://intelpython.github.io/dpnp/pull/2205/index.html |
|
Array API standard conformance tests for dpnp=0.18.0dev0=py312he4f9c94_53 ran successfully. |
|
@antonwolfy please review |
| ) | ||
|
|
||
| if a.ndim == 0: | ||
| a = dpnp.reshape(a, (1,)) |
| if a.ndim == 0: | ||
| a = dpnp.reshape(a, (1,)) | ||
| if v.ndim == 0: | ||
| v = dpnp.reshape(v, (1,)) |
dpnp/tests/test_statistics.py
Outdated
| if dtype == dpnp.bool: | ||
| an = numpy.random.rand(a_size) > 0.9 | ||
| vn = numpy.random.rand(v_size) > 0.9 | ||
| else: | ||
| an = (100 * numpy.random.rand(a_size)).astype(dtype) | ||
| vn = (100 * numpy.random.rand(v_size)).astype(dtype) | ||
|
|
||
| if dpnp.issubdtype(dtype, dpnp.complexfloating): | ||
| an = an + 1j * (100 * numpy.random.rand(a_size)).astype(dtype) | ||
| vn = vn + 1j * (100 * numpy.random.rand(v_size)).astype(dtype) |
There was a problem hiding this comment.
Can we use generate_random_numpy_array from helper here?
| dpnp.issubdtype(dtype, dpnp.integer) or dtype == dpnp.bool | ||
| ): | ||
| # For 'direct' and 'auto' methods, we expect exact results for integer types | ||
| assert_array_equal(result, expected) |
There was a problem hiding this comment.
Why can we expect exact result for method="auto" that might fall back on "fft" but cannot expect it for "fft" itself?
dpnp/tests/test_statistics.py
Outdated
| if dpnp.issubdtype(rdtype, dpnp.integer): | ||
| rdtype = dpnp.default_float_type(ad.device) |
There was a problem hiding this comment.
Why do we need to change output dtypes (and also below at lines 202, and 223)? If numpy and dpnp outputs have different dtyps for a known reason, could not we use check_type or check_only_type_kind of assert_dtype_allclose to handle it?
| # For these outliers, the relative error can be significant. | ||
| # We can tolerate a few such outliers. | ||
| max_outliers = 8 if expected.size > 1 else 0 | ||
| if invalid.sum() > max_outliers: |
There was a problem hiding this comment.
I do not get the logic here! First, it seems the code never goes inside the if condition here which mean we do not compare the outputs at all. Second, as I understand it, invalid is True when the difference between result and expected is large (outlier), so this condition indicates if there are more than 8 outliers the outputs should be compared. Shouldn't we avoid comparison when there is an outlier (or manually modify the values of outliers and then compare)?
Co-authored-by: Vahid Tavanashad <[email protected]>
Add implementation of
dpnp.convolve. Implementation is mostly based on already existing functionality developed fordpnp.correlateSimilar to scipy.signal.convolve
methodkeyword is introduced, but unlikescipy.signal.convolvedpnp.convolveworks only with 1-d arrays.