Fabian Höring created ARROW-5651:
------------------------------------
Summary: [Python] Incorrect conversion from strided Numpy array
when other type is specified
Key: ARROW-5651
URL: https://issues.apache.org/jira/browse/ARROW-5651
Project: Apache Arrow
Issue Type: Improvement
Affects Versions: 0.12.0
Reporter: Fabian Höring
In the example below the pyarrow array gives wrong results for strided numpy
arrays:
{code}
>> import pyarrow as pa
>> import numpy as np
>> p_s = pd.Series(np.arange(0, 10, dtype=np.float32)[1:-1:2])
>> pa.array(p_s, type=pa.float64())
<pyarrow.lib.DoubleArray object at 0x7f8453de8138>
[
1,
2,
3,
4
]
{code}
When copying the numpy array to a new location is gives the expected output:
{code}
>> import pyarrow as pa
>> import numpy as np
>> import pandas as pd
>> p_s = pd.Series(np.array(np.arange(0, 10, dtype=np.float32)[1:-1:2]))
>> pa.array(p_s, type=pa.float64())
<pyarrow.lib.DoubleArray object at 0x7f5a0af0a4a8>
[
1,
3,
5,
7
]
{code}
Looking at the
[code|https://github.com/apache/arrow/blob/7a5562174cffb21b16f990f64d114c1a94a30556/cpp/src/arrow/python/numpy_to_arrow.cc#L407]
it seems like to determine the number of elements it uses the target type
instead of the initial numpy type.
In this case the stride is 8 bytes which corresponds to 2 elements in float32
whereas the codes tries to determine the number of elements with the target
type which gives 1 element of float64 and therefore it reads the array one by
one instead of every 2 elements until reaching the total number of elements.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
