在重新索引/上采样时间序列时，有没有办法防止 dtype 从 Int64 更改为 float64？答案

【问题标题】：Is there a way to prevent dtype from changing from Int64 to float64 when reindexing/upsampling a time-series?在重新索引/上采样时间序列时，有没有办法防止 dtype 从 Int64 更改为 float64？
【发布时间】：2017-01-06 05:43:45
【问题描述】：

我正在使用 pandas 0.17.0 并且有一个类似于这个的 df：

df.head()
Out[339]: 
                       A     B  C
DATE_TIME                        
2016-10-08 13:57:00  in   5.61  1
2016-10-08 14:02:00  in   8.05  1
2016-10-08 14:07:00  in   7.92  0
2016-10-08 14:12:00  in   7.98  0
2016-10-08 14:17:00  out  8.18  0

df.tail()
Out[340]: 
                       A     B  C
DATE_TIME                        
2016-11-08 13:42:00  in   8.00  0
2016-11-08 13:47:00  in   7.99  0
2016-11-08 13:52:00  out  7.97  0
2016-11-08 13:57:00  in   8.14  1
2016-11-08 14:02:00  in   8.16  1

关注dtypes:

print (df.dtypes)
A     object
B    float64
C      int64
dtype: object

当我将df 重新索引到分钟间隔时，所有列int64 都更改为float64。

index = pd.date_range(df.index[0], df.index[-1], freq="min") 
df2 = df.reindex(index)

print (df2.dtypes)
A     object
B    float64
C    float64
dtype: object

另外，如果我尝试重新采样

df3 = df.resample('Min')

int64 将变成float64，由于某种原因我失去了我的object 列。

print (df3.dtypes)

print (df3.dtypes)
B    float64
C    float64
dtype: object

由于我想在后续步骤中根据这种区别对列进行不同的插值（在将df 与另一个df 连接后），我需要它们保持原来的dtype。我真正的df 每种类型的列要多得多，因此我正在寻找一种不依赖于通过标签单独调用列的解决方案。

有没有办法在整个重新索引过程中保持他们的dtype？或者有没有办法我可以在之后为它们分配dtype（它们是除了 NAN 之外唯一只包含整数的列）？有人可以帮帮我吗？

【问题讨论】：

标签： python pandas types resampling reindex

【解决方案1】：

它是impossible，因为如果在某列中至少有一个NaN 值，int 将转换为float。

index = pd.date_range(df.index[0], df.index[-1], freq="min") 
df2 = df.reindex(index)

print (df2)
                       A     B    C
2016-10-08 13:57:00   in  5.61  1.0
2016-10-08 13:58:00  NaN   NaN  NaN
2016-10-08 13:59:00  NaN   NaN  NaN
2016-10-08 14:00:00  NaN   NaN  NaN
2016-10-08 14:01:00  NaN   NaN  NaN
2016-10-08 14:02:00   in  8.05  1.0
2016-10-08 14:03:00  NaN   NaN  NaN
2016-10-08 14:04:00  NaN   NaN  NaN
2016-10-08 14:05:00  NaN   NaN  NaN
2016-10-08 14:06:00  NaN   NaN  NaN
2016-10-08 14:07:00   in  7.92  0.0
2016-10-08 14:08:00  NaN   NaN  NaN
2016-10-08 14:09:00  NaN   NaN  NaN
2016-10-08 14:10:00  NaN   NaN  NaN
2016-10-08 14:11:00  NaN   NaN  NaN
2016-10-08 14:12:00   in  7.98  0.0
2016-10-08 14:13:00  NaN   NaN  NaN
2016-10-08 14:14:00  NaN   NaN  NaN
2016-10-08 14:15:00  NaN   NaN  NaN
2016-10-08 14:16:00  NaN   NaN  NaN
2016-10-08 14:17:00  out  8.18  0.0

print (df2.dtypes)
A     object
B    float64
C    float64
dtype: object

但如果在reindex中使用参数fill_value，dtypes不会改变：

index = pd.date_range(df.index[0], df.index[-1], freq="min") 
df2 = df.reindex(index, fill_value=0)

print (df2)
                       A     B  C
2016-10-08 13:57:00   in  5.61  1
2016-10-08 13:58:00    0  0.00  0
2016-10-08 13:59:00    0  0.00  0
2016-10-08 14:00:00    0  0.00  0
2016-10-08 14:01:00    0  0.00  0
2016-10-08 14:02:00   in  8.05  1
2016-10-08 14:03:00    0  0.00  0
2016-10-08 14:04:00    0  0.00  0
2016-10-08 14:05:00    0  0.00  0
2016-10-08 14:06:00    0  0.00  0
2016-10-08 14:07:00   in  7.92  0
2016-10-08 14:08:00    0  0.00  0
2016-10-08 14:09:00    0  0.00  0
2016-10-08 14:10:00    0  0.00  0
2016-10-08 14:11:00    0  0.00  0
2016-10-08 14:12:00   in  7.98  0
2016-10-08 14:13:00    0  0.00  0
2016-10-08 14:14:00    0  0.00  0
2016-10-08 14:15:00    0  0.00  0
2016-10-08 14:16:00    0  0.00  0
2016-10-08 14:17:00  out  8.18  0

print (df2.dtypes)
A     object
B    float64
C      int64
dtype: object

在reindex 中使用method='ffill 更好：

index = pd.date_range(df.index[0], df.index[-1], freq="min") 
df2 = df.reindex(index, method='ffill')

print (df2)
                       A     B  C
2016-10-08 13:57:00   in  5.61  1
2016-10-08 13:58:00   in  5.61  1
2016-10-08 13:59:00   in  5.61  1
2016-10-08 14:00:00   in  5.61  1
2016-10-08 14:01:00   in  5.61  1
2016-10-08 14:02:00   in  8.05  1
2016-10-08 14:03:00   in  8.05  1
2016-10-08 14:04:00   in  8.05  1
2016-10-08 14:05:00   in  8.05  1
2016-10-08 14:06:00   in  8.05  1
2016-10-08 14:07:00   in  7.92  0
2016-10-08 14:08:00   in  7.92  0
2016-10-08 14:09:00   in  7.92  0
2016-10-08 14:10:00   in  7.92  0
2016-10-08 14:11:00   in  7.92  0
2016-10-08 14:12:00   in  7.98  0
2016-10-08 14:13:00   in  7.98  0
2016-10-08 14:14:00   in  7.98  0
2016-10-08 14:15:00   in  7.98  0
2016-10-08 14:16:00   in  7.98  0
2016-10-08 14:17:00  out  8.18  0

print (df2.dtypes)
A     object
B    float64
C      int64
dtype: object

如果使用resample，你可以通过unstack和stack返回A列，但不幸的是float仍然存在问题：

df3 = df.set_index('A', append=True)
        .unstack()
        .resample('Min', fill_method='ffill')
        .stack()
        .reset_index(level=1)
print (df3)
                       A     B    C
DATE_TIME                          
2016-10-08 13:57:00   in  5.61  1.0
2016-10-08 13:58:00   in  5.61  1.0
2016-10-08 13:59:00   in  5.61  1.0
2016-10-08 14:00:00   in  5.61  1.0
2016-10-08 14:01:00   in  5.61  1.0
2016-10-08 14:02:00   in  8.05  1.0
2016-10-08 14:03:00   in  8.05  1.0
2016-10-08 14:04:00   in  8.05  1.0
2016-10-08 14:05:00   in  8.05  1.0
2016-10-08 14:06:00   in  8.05  1.0
2016-10-08 14:07:00   in  7.92  0.0
2016-10-08 14:08:00   in  7.92  0.0
2016-10-08 14:09:00   in  7.92  0.0
2016-10-08 14:10:00   in  7.92  0.0
2016-10-08 14:11:00   in  7.92  0.0
2016-10-08 14:12:00   in  7.98  0.0
2016-10-08 14:13:00   in  7.98  0.0
2016-10-08 14:14:00   in  7.98  0.0
2016-10-08 14:15:00   in  7.98  0.0
2016-10-08 14:16:00   in  7.98  0.0
2016-10-08 14:17:00  out  8.18  0.0

print (df3.dtypes)
A     object
B    float64
C    float64
dtype: object

我尝试修改之前的 answer 以转换为 `int:

int_cols = df.select_dtypes(['int64']).columns
print (int_cols)
Index(['C'], dtype='object')

index = pd.date_range(df.index[0], df.index[-1], freq="s")
df2 = df.reindex(index)

for col in df2:
    if col == int_cols: 
        df2[col].ffill(inplace=True)
        df2[col] = df2[col].astype(int)
    elif df2[col].dtype == float:
        df2[col].interpolate(inplace=True)
    else:
        df2[col].ffill(inplace=True)

#print (df2)

print (df2.dtypes)
A     object
B    float64
C      int32
dtype: object

【讨论】：

非常感谢！我想我现在的问题是，我无法将float64 作为下一步线性插值，因为所有新创建的时间步都已填满。 link。这是我原来问题的链接，我今天才意识到接受的解决方案中的 dtype 存在问题，这就是我发布这个新问题的原因。你觉得我真正想要的有可能吗？
我添加了解决方案，请查看上次更新。感谢您的接受！
酷！该代码在我的示例df 上的工作方式与此类似，但要使其与我的真实df 一起使用，每种类型的多个列我必须修改以下行if col == int_cols.all():，因为我确实收到了错误：ValueError: The truth value of an array with more than one element is ambiguous. Use a.any() or a.all()。现在它似乎工作得很好！再次感谢！