1

How can I use or manipulate (monkey-patch) pandas in order, to keep always the same major-order on the resulting object for copy and groupby aggregations?

I use pandas.DataFrame as datastructure within a business application (risk model) and need fast aggregation of multidimensional data. Aggregation with pandas depends crucially on the major-ordering scheme in use on the underlying numpy array.

Unfortunatly, pandas (version 0.23.4) changes the major-order of the underlying numpy array when I create a copy or when I perform an aggregation with groupby and sum.

The impact is:

case 1: 17.2 seconds

case 2: 5 min 46 s seconds

on a DataFrame and its copy with 45023 rows and 100000 columns. Aggregation was performed on the index. The index is a pd.MultiIndex with 15 levels. Aggregation keeps three levels and leads to about 239 groups.

I work typically on DataFrames with 45000 rows and 100000 columns. On the row I have a pandas.MultiIndex with about 15 levels. To compute statistics on various hierarchy nodes I need to aggregate (sum) on the index dimension.

Aggregation is fast, if the underlying numpy array is c_contiguous, hence held in column-major-order (C order). It is very slow if it is f_contiguous, hence in row-major-order (F order).

Unfortunatly, pandas changes the the major-order from C to F when

  • creating a copy of a DataFrame and even when,

  • performing an aggregation via a grouby and and taking the sum on the grouper. Hence the resulting DataFrame has a differnt major-order (!)

Sure, I could stick to another 'datamodel', just by keeping the MultiIndex on the columns. Then the current pandas version would always work to my favor. But this is a no go. I think, that one can expect, that for the two operations under consideration (groupby-sum and copy) the major-order should not be changed.

import numpy as np
import pandas as pd

print("pandas version: ", pd.__version__)

array = np.array([[1, 2, 3], [4, 5, 6], [7, 8, 9]])
array.flags
print("Numpy array is C-contiguous: ", data.flags.c_contiguous)

dataframe = pd.DataFrame(array, index = pd.MultiIndex.from_tuples([('A', 'U'), ('A', 'V'), ('B', 'W')], names=['dim_one', 'dim_two']))
print("DataFrame is C-contiguous: ", dataframe.values.flags.c_contiguous)

dataframe_copy = dataframe.copy()
print("Copy of DataFrame is C-contiguous: ", dataframe_copy.values.flags.c_contiguous)

aggregated_dataframe = dataframe.groupby('dim_one').sum()
print("Aggregated DataFrame is C-contiguous: ", aggregated_dataframe.values.flags.c_contiguous)


## Output in Jupyter Notebook
# pandas version:  0.23.4
# Numpy array is C-contiguous:  True
# DataFrame is C-contiguous:  True
# Copy of DataFrame is C-contiguous:  False
# Aggregated DataFrame is C-contiguous:  False

The major order of the data should be preserved. If pandas likes to switch to an implicit preference, then it should allow to overwrite this. Numpy allows to input the order when creating a copy.

A patched version of pandas should result in

## Output in Jupyter Notebook
# pandas version:  0.23.4
# Numpy array is C-contiguous:  True
# DataFrame is C-contiguous:  True
# Copy of DataFrame is C-contiguous:  True
# Aggregated DataFrame is C-contiguous:  True

for the example code snipped above.

4

1 回答 1

0

熊猫猴子补丁(0.23.4 可能还有其他版本)

我创建了一个补丁,我想与你分享。它导致上述问题中提到的性能提高。

它适用于熊猫版本 0.23.4。对于其他版本,您需要尝试它是否仍然有效。

需要以下两个模块,您可以根据放置它​​们的位置调整导入。

memory_layout.py   
memory.py

要修补您的代码,您只需在程序或笔记本的最开始导入以下内容并设置内存布局参数。它会对 pandas 进行修补,并确保 DataFrames 的副本具有所要求的布局。

from memory_layout import memory_layout
# memory_layout.order = 'F'  # assert F-order on copy
# memory_layout.order = 'K'  # Keep given layout on copy 
memory_layout.order = 'C'  # assert C-order on copy

memory_layout.py

使用以下内容创建文件 memory_layout.py。

import numpy as np
from pandas.core.internals import Block
from memory import memory_layout

# memory_layout.order = 'F'  # set memory layout order to 'F' for np.ndarrays in DataFrame copies (fortran/row order)
# memory_layout.order = 'K'  # keep memory layout order for np.ndarrays in DataFrame copies (order out is order in)
memory_layout.order = 'C'  # set memory layout order to 'C' for np.ndarrays in DataFrame copies (C/column order)


def copy(self, deep=True, mgr=None):
    """
    Copy patch on Blocks to set or keep the memory layout
    on copies.

    :param self: `pandas.core.internals.Block`
    :param deep: `bool`
    :param mgr: `BlockManager`
    :return: copy of `pandas.core.internals.Block`
    """
    values = self.values
    if deep:
        if isinstance(values, np.ndarray):
memory_layout))
            values = memory_layout.copy_transposed(values)
memory_layout))
        else:
            values = values.copy()
    return self.make_block_same_class(values)


Block.copy = copy  # Block for pandas 0.23.4: in pandas.core.internals.Block

memory.py

使用以下内容创建文件 memory.py。

"""
Implements MemoryLayout copy factory to change memory layout
of `numpy.ndarrays`.
Depending on the use case, operations on DataFrames can be much
faster if the appropriate memory layout is set and preserved.

The implementation allows for changing the desired layout. Changes apply when
copies or new objects are created, as for example, when slicing or aggregating
via groupby ...

This implementation tries to solve the issue raised on GitHub
https://github.com/pandas-dev/pandas/issues/26502

"""
import numpy as np

_DEFAULT_MEMORY_LAYOUT = 'K'


class MemoryLayout(object):
    """
    Memory layout management for numpy.ndarrays.

    Singleton implementation.

    Example:
    >>> from memory import memory_layout
    >>> memory_layout.order = 'K'  #
    >>> # K ... keep array layout from input
    >>> # C ... set to c-contiguous / column order
    >>> # F ... set to f-contiguous / row order
    >>> array = memory_layout.apply(array)
    >>> array = memory_layout.apply(array, 'C')
    >>> array = memory_layout.copy(array)
    >>> array = memory_layout.apply_on_transpose(array)

    """

    _order = _DEFAULT_MEMORY_LAYOUT
    _instance = None

    @property
    def order(self):
        """
        Return memory layout ordering.

        :return: `str`
        """
        if self.__class__._order is None:
            raise AssertionError("Array layout order not set.")
        return self.__class__._order

    @order.setter
    def order(self, order):
        """
        Set memory layout order.
        Allowed values are 'C', 'F', and 'K'. Raises AssertionError
        when trying to set other values.

        :param order: `str`
        :return: `None`
        """
        assert order in ['C', 'F', 'K'], "Only 'C', 'F' and 'K' supported."
        self.__class__._order = order

    def __new__(cls):
        """
        Create only one instance throughout the lifetime of this process.

        :return: `MemoryLayout` instance as singleton
        """
        if cls._instance is None:
            cls._instance = super(MemoryLayout, cls).__new__(MemoryLayout)
        return cls._instance

    @staticmethod
    def get_from(array):
        """
        Get memory layout from array

        Possible values:
           'C' ... only C-contiguous or column order
           'F' ... only F-contiguous or row order
           'O' ... other: both, C- and F-contiguous or both
           not C- or F-contiguous (as on empty arrays).

        :param array: `numpy.ndarray`
        :return: `str`
        """
        if array.flags.c_contiguous == array.flags.f_contiguous:
            return 'O'
        return {True: 'C', False: 'F'}[array.flags.c_contiguous]

    def apply(self, array, order=None):
        """
        Apply the order set or the order given as input on the array
        given as input.

        Possible values:
           'C' ... apply C-contiguous layout or column order
           'F' ... apply F-contiguous layout or row order
           'K' ... keep the given layout

        :param array: `numpy.ndarray`
        :param order: `str`
        :return: `np.ndarray`
        """
        order = self.__class__._order if order is None else order

        if order == 'K':
            return array

        array_order = MemoryLayout.get_from(array)
        if array_order == order:
            return array

        return np.reshape(np.ravel(array), array.shape, order=order)

    def copy(self, array, order=None):
        """
        Return a copy of the input array with the memory layout set.
        Layout set:
           'C' ... return C-contiguous copy
           'F' ... return F-contiguous copy
           'K' ... return copy with same layout as
           given by the input array.

        :param array: `np.ndarray`
        :return: `np.ndarray`
        """
        order = order if order is not None else self.__class__._order
        return array.copy(order=self.get_from(array)) if order == 'K' \
            else array.copy(order=order)

    def copy_transposed(self, array):
        """
        Return a copy of the input array in order that its transpose
        has the memory layout set.

        Note: numpy simply changes the memory layout from row to column
        order instead of reshuffling the data in memory.

        Layout set:
           'C' ... return F-contiguous copy
           'F' ... return C-contiguous copy
           'K' ... return copy with oposite (C versus F) layout as
           given by the input array.

        :param array: `np.ndarray`
        :return: `np.ndarray`

        :param array:
        :return:
        """
        if self.__class__._order == 'K':
            return array.copy(
                order={'C': 'C', 'F': 'F', 'O': None}[self.get_from(array)])
        else:
            return array.copy(
                order={'C': 'F', 'F': 'C'}[self.__class__._order])

    def __str__(self):
        return str(self.__class__._order)


memory_layout = MemoryLayout()  # Singleton
于 2020-04-29T09:20:21.833 回答