Skip to content

Commit

Permalink
Enable implicit casting when concatenating mixed types (#8276)
Browse files Browse the repository at this point in the history 
This enables implicit casting when decimal columns are concatenated with numeric columns by casting the numeric columns to decimal columns.

Closes #8264

Authors:
  - https://github.com/ChrisJar

Approvers:
  - Ashwin Srinath (https://github.com/shwina)
  - https://github.com/brandon-b-miller

URL: #8276
  • Loading branch information
@ChrisJar
ChrisJar authored May 23, 2021
1 parent de579a5 commit 696902d
Show file tree
Hide file tree
Showing 4 changed files with 291 additions and 28 deletions.
19 changes: 6 additions & 13 deletions python/cudf/cudf/core/frame.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -32,6 +32,7 @@
is_numerical_dtype,
is_scalar,
min_scalar_type,
find_common_type,
)

T = TypeVar("T", bound="Frame")
Expand Down Expand Up @@ -4029,8 +4030,11 @@ def _find_common_dtypes_and_categories(non_null_columns, dtypes):
# default to the first non-null dtype
dtypes[idx] = cols[0].dtype
# If all the non-null dtypes are int/float, find a common dtype
if all(is_numerical_dtype(col.dtype) for col in cols):
dtypes[idx] = np.find_common_type([col.dtype for col in cols], [])
if all(
is_numerical_dtype(col.dtype) or is_decimal_dtype(col.dtype)
for col in cols
):
dtypes[idx] = find_common_type([col.dtype for col in cols])
# If all categorical dtypes, combine the categories
elif all(
isinstance(col, cudf.core.column.CategoricalColumn) for col in cols
Expand All @@ -4045,17 +4049,6 @@ def _find_common_dtypes_and_categories(non_null_columns, dtypes):
# Set the column dtype to the codes' dtype. The categories
# will be re-assigned at the end
dtypes[idx] = min_scalar_type(len(categories[idx]))
elif all(
isinstance(col, cudf.core.column.DecimalColumn) for col in cols
):
# Find the largest scale and the largest difference between
# precision and scale of the columns to be concatenated
s = max([col.dtype.scale for col in cols])
lhs = max([col.dtype.precision - col.dtype.scale for col in cols])
# Combine to get the necessary precision and clip at the maximum
# precision
p = min(cudf.Decimal64Dtype.MAX_PRECISION, s + lhs)
dtypes[idx] = cudf.Decimal64Dtype(p, s)
# Otherwise raise an error if columns have different dtypes
elif not all(is_dtype_equal(c.dtype, dtypes[idx]) for c in cols):
raise ValueError("All columns must be the same type")
Expand Down
9 changes: 3 additions & 6 deletions python/cudf/cudf/core/series.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -45,15 +45,14 @@
from cudf.utils import cudautils, docutils, ioutils
from cudf.utils.docutils import copy_docstring
from cudf.utils.dtypes import (
_decimal_normalize_types,
can_convert_to_column,
is_decimal_dtype,
is_list_dtype,
is_list_like,
is_mixed_with_object_dtype,
is_scalar,
min_scalar_type,
numeric_normalize_types,
find_common_type,
)
from cudf.utils.utils import (
get_appropriate_dispatched_func,
Expand Down Expand Up @@ -2402,10 +2401,8 @@ def _concat(cls, objs, axis=0, index=True):
)

if dtype_mismatch:
if isinstance(objs[0]._column, cudf.core.column.DecimalColumn):
objs = _decimal_normalize_types(*objs)
else:
objs = numeric_normalize_types(*objs)
common_dtype = find_common_type([obj.dtype for obj in objs])
objs = [obj.astype(common_dtype) for obj in objs]

col = _concat_columns([o._column for o in objs])

Expand Down
265 changes: 265 additions & 0 deletions python/cudf/cudf/tests/test_concat.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -5,6 +5,7 @@
import numpy as np
import pandas as pd
import pytest
from decimal import Decimal

import cudf as gd
from cudf.tests.utils import assert_eq, assert_exceptions_equal
Expand Down Expand Up @@ -1262,3 +1263,267 @@ def test_concat_decimal_series(ltype, rtype):
expected = pd.concat([ps1, ps2])

assert_eq(expected, got)


@pytest.mark.parametrize(
"df1, df2, df3, expected",
[
(
gd.DataFrame(
{"val": [Decimal("42.5"), Decimal("8.7")]},
dtype=Decimal64Dtype(5, 2),
),
gd.DataFrame(
{"val": [Decimal("9.23"), Decimal("-67.49")]},
dtype=Decimal64Dtype(6, 4),
),
gd.DataFrame({"val": [8, -5]}, dtype="int32"),
gd.DataFrame(
{
"val": [
Decimal("42.5"),
Decimal("8.7"),
Decimal("9.23"),
Decimal("-67.49"),
Decimal("8"),
Decimal("-5"),
]
},
dtype=Decimal64Dtype(7, 4),
index=[0, 1, 0, 1, 0, 1],
),
),
(
gd.DataFrame(
{"val": [Decimal("95.2"), Decimal("23.4")]},
dtype=Decimal64Dtype(5, 2),
),
gd.DataFrame({"val": [54, 509]}, dtype="uint16"),
gd.DataFrame({"val": [24, -48]}, dtype="int32"),
gd.DataFrame(
{
"val": [
Decimal("95.2"),
Decimal("23.4"),
Decimal("54"),
Decimal("509"),
Decimal("24"),
Decimal("-48"),
]
},
dtype=Decimal64Dtype(5, 2),
index=[0, 1, 0, 1, 0, 1],
),
),
(
gd.DataFrame(
{"val": [Decimal("36.56"), Decimal("-59.24")]},
dtype=Decimal64Dtype(9, 4),
),
gd.DataFrame({"val": [403.21, 45.13]}, dtype="float32"),
gd.DataFrame({"val": [52.262, -49.25]}, dtype="float64"),
gd.DataFrame(
{
"val": [
Decimal("36.56"),
Decimal("-59.24"),
Decimal("403.21"),
Decimal("45.13"),
Decimal("52.262"),
Decimal("-49.25"),
]
},
dtype=Decimal64Dtype(9, 4),
index=[0, 1, 0, 1, 0, 1],
),
),
(
gd.DataFrame(
{"val": [Decimal("9563.24"), Decimal("236.633")]},
dtype=Decimal64Dtype(9, 4),
),
gd.DataFrame({"val": [5393, -95832]}, dtype="int64"),
gd.DataFrame({"val": [-29.234, -31.945]}, dtype="float64"),
gd.DataFrame(
{
"val": [
Decimal("9563.24"),
Decimal("236.633"),
Decimal("5393"),
Decimal("-95832"),
Decimal("-29.234"),
Decimal("-31.945"),
]
},
dtype=Decimal64Dtype(9, 4),
index=[0, 1, 0, 1, 0, 1],
),
),
],
)
def test_concat_decimal_numeric_dataframe(df1, df2, df3, expected):
df = gd.concat([df1, df2, df3])
assert_eq(df, expected)
assert_eq(df.val.dtype, expected.val.dtype)


@pytest.mark.parametrize(
"s1, s2, s3, expected",
[
(
gd.Series(
[Decimal("32.8"), Decimal("-87.7")], dtype=Decimal64Dtype(6, 2)
),
gd.Series(
[Decimal("101.243"), Decimal("-92.449")],
dtype=Decimal64Dtype(9, 6),
),
gd.Series([94, -22], dtype="int32"),
gd.Series(
[
Decimal("32.8"),
Decimal("-87.7"),
Decimal("101.243"),
Decimal("-92.449"),
Decimal("94"),
Decimal("-22"),
],
dtype=Decimal64Dtype(10, 6),
index=[0, 1, 0, 1, 0, 1],
),
),
(
gd.Series(
[Decimal("7.2"), Decimal("122.1")], dtype=Decimal64Dtype(5, 2)
),
gd.Series([33, 984], dtype="uint32"),
gd.Series([593, -702], dtype="int32"),
gd.Series(
[
Decimal("7.2"),
Decimal("122.1"),
Decimal("33"),
Decimal("984"),
Decimal("593"),
Decimal("-702"),
],
dtype=Decimal64Dtype(5, 2),
index=[0, 1, 0, 1, 0, 1],
),
),
(
gd.Series(
[Decimal("982.94"), Decimal("-493.626")],
dtype=Decimal64Dtype(9, 4),
),
gd.Series([847.98, 254.442], dtype="float32"),
gd.Series([5299.262, -2049.25], dtype="float64"),
gd.Series(
[
Decimal("982.94"),
Decimal("-493.626"),
Decimal("847.98"),
Decimal("254.442"),
Decimal("5299.262"),
Decimal("-2049.25"),
],
dtype=Decimal64Dtype(9, 4),
index=[0, 1, 0, 1, 0, 1],
),
),
(
gd.Series(
[Decimal("492.204"), Decimal("-72824.455")],
dtype=Decimal64Dtype(9, 4),
),
gd.Series([8438, -27462], dtype="int64"),
gd.Series([-40.292, 49202.953], dtype="float64"),
gd.Series(
[
Decimal("492.204"),
Decimal("-72824.455"),
Decimal("8438"),
Decimal("-27462"),
Decimal("-40.292"),
Decimal("49202.953"),
],
dtype=Decimal64Dtype(9, 4),
index=[0, 1, 0, 1, 0, 1],
),
),
],
)
def test_concat_decimal_numeric_series(s1, s2, s3, expected):
s = gd.concat([s1, s2, s3])
assert_eq(s, expected)


@pytest.mark.parametrize(
"s1, s2, expected",
[
(
gd.Series(
[Decimal("955.22"), Decimal("8.2")], dtype=Decimal64Dtype(5, 2)
),
gd.Series(["2007-06-12", "2006-03-14"], dtype="datetime64"),
gd.Series(
[
"955.22",
"8.20",
"2007-06-12 00:00:00",
"2006-03-14 00:00:00",
],
index=[0, 1, 0, 1],
),
),
(
gd.Series(
[Decimal("-52.44"), Decimal("365.22")],
dtype=Decimal64Dtype(5, 2),
),
gd.Series(
np.arange(
"2005-02-01T12", "2005-02-01T15", dtype="datetime64[h]"
),
dtype="datetime64[s]",
),
gd.Series(
[
"-52.44",
"365.22",
"2005-02-01 12:00:00",
"2005-02-01 13:00:00",
"2005-02-01 14:00:00",
],
index=[0, 1, 0, 1, 2],
),
),
(
gd.Series(
[Decimal("753.0"), Decimal("94.22")],
dtype=Decimal64Dtype(5, 2),
),
gd.Series([np.timedelta64(111, "s"), np.timedelta64(509, "s")]),
gd.Series(
["753.00", "94.22", "0 days 00:01:51", "0 days 00:08:29"],
index=[0, 1, 0, 1],
),
),
(
gd.Series(
[Decimal("753.0"), Decimal("94.22")],
dtype=Decimal64Dtype(5, 2),
),
gd.Series(
[np.timedelta64(940252, "s"), np.timedelta64(758385, "s")]
),
gd.Series(
["753.00", "94.22", "10 days 21:10:52", "8 days 18:39:45"],
index=[0, 1, 0, 1],
),
),
],
)
def test_concat_decimal_non_numeric(s1, s2, expected):
s = gd.concat([s1, s2])
assert_eq(s, expected)
26 changes: 17 additions & 9 deletions python/cudf/cudf/utils/dtypes.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -290,13 +290,15 @@ def is_decimal_dtype(obj):
)


def _decimal_normalize_types(*args):
s = max([a.dtype.scale for a in args])
lhs = max([a.dtype.precision - a.dtype.scale for a in args])
def _find_common_type_decimal(dtypes):
# Find the largest scale and the largest difference between
# precision and scale of the columns to be concatenated
s = max([dtype.scale for dtype in dtypes])
lhs = max([dtype.precision - dtype.scale for dtype in dtypes])
# Combine to get the necessary precision and clip at the maximum
# precision
p = min(cudf.Decimal64Dtype.MAX_PRECISION, s + lhs)
dtype = cudf.Decimal64Dtype(p, s)

return [a.astype(dtype) for a in args]
return cudf.Decimal64Dtype(p, s)


def cudf_dtype_from_pydata_dtype(dtype):
Expand Down Expand Up @@ -690,9 +692,15 @@ def find_common_type(dtypes):
dtypes = set(dtypes)

if any(is_decimal_dtype(dtype) for dtype in dtypes):
raise NotImplementedError(
"DecimalDtype is not yet supported in find_common_type"
)
if all(
is_decimal_dtype(dtype) or is_numerical_dtype(dtype)
for dtype in dtypes
):
return _find_common_type_decimal(
[dtype for dtype in dtypes if is_decimal_dtype(dtype)]
)
else:
return np.dtype("O")

# Corner case 1:
# Resort to np.result_type to handle "M" and "m" types separately
Expand Down

0 comments on commit 696902d

Please sign in to comment.