循环遍历 pandas DataFrame 的过滤条件列表并进行一些计算的最快方法是什么？

Question

我经常发现自己有一个需要应用于 pandas 数据框的过滤器列表。我应用每个过滤器并进行一些计算，但这通常会导致代码变慢。我想优化性能。我创建了一个慢速解决方案的示例，它在日期范围列表上过滤数据框并计算与我的日期范围匹配的行的列的总和，然后将此值分配给与日期范围开始匹配的日期：

import numpy as np
import pandas as pd
import datetime


def generateTestDataFrame(N=50, windowSizeInDays=5):
    dd = {"AsOfDate" : [],
            "WindowEndDate" : [],
            "X" : []}

    d = datetime.date.today()

    for i in range(N):

        dd["AsOfDate"].append(d)
        dd["WindowEndDate"].append(d + datetime.timedelta(days=windowSizeInDays))
        dd["X"].append(float(i))

        d = d + datetime.timedelta(days=1)

    newDf = pd.DataFrame(dd)
    return newDf

def run():
    numRows = 50
    windowSizeInDays = 5

    print "NumRows: %s" % (numRows)
    print "WindowSizeInDays: %s" % (windowSizeInDays)

    df = generateTestDataFrame(numRows, windowSizeInDays)

    newAggColumnName = "SumOverNdays"
    df[newAggColumnName] = np.nan  # Initialize the column to nan

    for i in range(df.shape[0]):
        row_i = df.iloc[i]
        startDate = row_i["AsOfDate"]
        endDate = row_i["WindowEndDate"]
        sumAggOverNdays = df.loc[ (df["AsOfDate"] >= startDate) & (df["AsOfDate"] < endDate) ]["X"].sum()
        df.loc[df["AsOfDate"] == startDate, newAggColumnName] = sumAggOverNdays  

    print df.head(10)

if __name__ == "__main__":
    run()

这会产生以下输出：

NumRows: 50
WindowSizeInDays: 5
     AsOfDate WindowEndDate    X  SumOverNdays
0  2019-01-15    2019-01-20  0.0          10.0
1  2019-01-16    2019-01-21  1.0          15.0
2  2019-01-17    2019-01-22  2.0          20.0
3  2019-01-18    2019-01-23  3.0          25.0
4  2019-01-19    2019-01-24  4.0          30.0
5  2019-01-20    2019-01-25  5.0          35.0
6  2019-01-21    2019-01-26  6.0          40.0
7  2019-01-22    2019-01-27  7.0          45.0
8  2019-01-23    2019-01-28  8.0          50.0
9  2019-01-24    2019-01-29  9.0          55.0

Answer 1

尝试使用 pandas.DataFrame.apply() 进行计算。

文档：https://pandas.pydata.org/pandas-docs/stable/generated/pandas.DataFrame.apply.html

使用您的代码：

%%timeit
run()
205 ms ± 33.7 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

适应：

%%timeit
windowSizeInDays = 5
rows = 50
df_ = pd.DataFrame(index=range(rows),columns=['AsOfDate','WindowEndDate','X','SumOverNdays'])
asofdate = [datetime.date.today() + datetime.timedelta(days=i) for i in range(rows)]
windowenddate = [i + datetime.timedelta(days=windowSizeInDays) for i in asofdate]

df_['AsOfDate'] = asofdate
df_['WindowEndDate'] = windowenddate
df_['X'] = np.arange(float(df_.shape[0]))
df_['SumOverNdays'] = df_.apply(lambda x: df_.loc[ (df_["AsOfDate"] >= x['AsOfDate']) & (df_["AsOfDate"] < x['WindowEndDate']) ]["X"].sum(), axis=1)
df_
112 ms ± 3.69 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

没有大差异，但在这个特定的例子中，我们不能做得比这更好......