Question

给出以下两个pandas数据帧

Dataframe 1

       open     high      low    close
0  340.649  340.829  340.374  340.511
1  340.454  340.843  340.442  340.843
2  340.521  340.751  340.241  340.474
3  340.197  340.698  340.145  340.420
4  340.332  340.609  340.123  340.128
5  340.092  340.462  339.993  340.207
6  340.179  340.437  339.810  339.983
7  340.296  340.498  339.977  340.036
8  340.461  340.641  340.189  340.367
9  340.404  340.820  340.338  340.589

Dataframe 2

       ohlc
0  0.374309
1  0.712707
2  0.791436
3  0.761050
4  0.779006
5  0.765193
6  0.578729
7  0.469613
8  0.385359
9  0.511050

以及以下函数，它接受两个数据帧并进行一些规范化和比较

def normalizeAndCompare(df1, df2):
    highest = df1["high"].max()
    lowest = df1["low"].min()

    df1["high"] = ((df1["high"] - lowest) / (highest - lowest))
    df1["low"] = ((df1["low"] - lowest) / (highest - lowest))
    df1["open"] = ((df1["open"] - lowest) / (highest - lowest))
    df1["close"] = ((df1["close"] - lowest) / (highest - lowest))
    df1["ohlc"] = (df1["open"] + df1["high"] + df1["low"] +df1["close"] ) / 4


    df1["highstd"] = df1["high"] + df1["ohlc"].rolling(window=10).std()
    df1["lowstd"] = df1["low"] - df1["ohlc"].rolling(window=10).std()

    df1["highpercent"] = df1["high"] + (df1["high"] * 0.05)
    df1["lowpercent"] = df1["low"] - (df1["low"] * 0.05)

    df1["highstd"] = df1['highstd'].fillna(value=df1['highpercent'])
    df1["lowstd"] = df1['lowstd'].fillna(value=df1['lowpercent'])

    result = (np.where(((df2["ohlc"] <= df1['highstd']) & (df2["ohlc"] >= df1['lowstd'])), 1, 0)).sum()

    return result

如何更改此功能，以便更有效地运行并更快地返回相同的结果？

鉴于我是python的新手，我非常感谢一些帮助。这是我的设置。也许还有提高效率的方法。我在dataframe1上运行循环：

pd_result = pd.DataFrame(columns=('rowNr', 'result'))
batchSize = 10
for rowNr in  range(len(dataframe1)):
    df1_temp = dataframe1[rowNr: rowNr + batchSize]
    df1_temp = df1_temp.reset_index(drop=True)
    result= normalizeAndCompare(df1_temp, dataframe2)
    pd_result.loc[rowNr] = [rowNr , result]

我的最终结果应该是pd_result。还有一点需要注意，dataframe1很大，有几百万行。

Answer 1

这是一个相当快速的从大多数熊猫函数转换为一个大多数numpy函数（rolling仍然在熊猫，但其余的是numpy）。对于10,000行，这大约快10倍。

def norm_comp(df1, df2):

    open  = df1['open'].values
    high  = df1['high'].values
    low   = df1['low'].values
    close = df1['close'].values

    highest = high.max()
    lowest  = low.min()

    high  = ((high  - lowest) / (highest - lowest))
    low   = ((low   - lowest) / (highest - lowest))
    open  = ((open  - lowest) / (highest - lowest))
    close = ((close - lowest) / (highest - lowest))

    ohlc = (open + high + low + close) / 4

    roll_std = pd.Series(ohlc).rolling(10).std().values

    highstd = np.where( np.isnan(roll_std), high * 1.05, high + roll_std )
    lowstd  = np.where( np.isnan(roll_std), low  *  .95, low  - roll_std )

    return np.where(((df2.ohlc.values <= highstd) & 
                     (df2.ohlc.values >= lowstd)), 1, 0).sum()

我将您的示例数据扩展为10,024行，如下所示：

for i in range(10): 
    df1 = df1.append(df1).reset_index(drop=True)
    df2 = df2.append(df2).reset_index(drop=True)

以下是时间安排：

%timeit normalizeAndCompare(df1,df2)
100 loops, best of 3: 9.93 ms per loop

%timeit norm_comp(df1,df2)
1000 loops, best of 3: 957 µs per loop

将pandas数据帧函数转换为更高效的函数

1 个答案: