Python - Iterate over a list of attributes

问题:

I have a feature in my data set that is a pandas timestamp object. It has (among many others) the following attributes: year, hour, dayofweek, month.

I can create new features based on these attributes using some brute force methods:

df["year"] = df["timeStamp"].apply(lambda x : x.year)

df["hour"] = df["timeStamp"].apply(lambda x : x.hour)

. . .

However, I want to iterate over a list:

nomtimes = ["year", "hour", "month", "dayofweek"]


for i in nomtimes:

  df[i] = df["timeStamp"].apply(lambda x : x.i)

I get the following AttributeError: 'Timestamp' object has no attribute 'i', and I get it and understand why I am having this error.

How can I get the quoted string to unquote so that I can pass it as an attribute?

回答1:

You just need getattr():

df[i] = df["timeStamp"].apply(lambda x : getattr(x, i))

回答2:

Don't use .apply here, pandas has various built-in utilities for handling datetime objects, use the dt attribute on the series objects:

In [11]: start = datetime(2011, 1, 1)
    ...: end = datetime(2012, 1, 1)
    ...:

In [12]: df = pd.DataFrame({'data':pd.date_range(start, end)})

In [13]: df.dtypes
Out[13]:
data    datetime64[ns]
dtype: object

In [14]: df['year'] = df.data.dt.year

In [15]: df['hour'] = df.data.dt.hour

In [16]: df['month'] = df.data.dt.month

In [17]: df['dayofweek'] = df.data.dt.dayofweek

In [18]: df.head()
Out[18]:
        data  year  hour  month  dayofweek
0 2011-01-01  2011     0      1          5
1 2011-01-02  2011     0      1          6
2 2011-01-03  2011     0      1          0
3 2011-01-04  2011     0      1          1
4 2011-01-05  2011     0      1          2

Or, dynamically as you wanted using getattr:

In [24]: df = pd.DataFrame({'data':pd.date_range(start, end)})

In [25]: nomtimes = ["year", "hour", "month", "dayofweek"]
    ...:

In [26]: df.head()
Out[26]:
        data
0 2011-01-01
1 2011-01-02
2 2011-01-03
3 2011-01-04
4 2011-01-05

In [27]: for t in nomtimes:
    ...:     df[t] = getattr(df.data.dt, t)
    ...:

In [28]: df.head()
Out[28]:
        data  year  hour  month  dayofweek
0 2011-01-01  2011     0      1          5
1 2011-01-02  2011     0      1          6
2 2011-01-03  2011     0      1          0
3 2011-01-04  2011     0      1          1
4 2011-01-05  2011     0      1          2

And if you must use a one-liner, go with:

In [30]: df = pd.DataFrame({'data':pd.date_range(start, end)})

In [31]: df.head()
Out[31]:
        data
0 2011-01-01
1 2011-01-02
2 2011-01-03
3 2011-01-04
4 2011-01-05

In [32]: df = df.assign(**{t:getattr(df.data.dt,t) for t in nomtimes})

In [33]: df.head()
Out[33]:
        data  dayofweek  hour  month  year
0 2011-01-01          5     0      1  2011
1 2011-01-02          6     0      1  2011
2 2011-01-03          0     0      1  2011
3 2011-01-04          1     0      1  2011
4 2011-01-05          2     0      1  2011

回答3:

operator.attrgetter

You can extract attributes in a loop:

from operator import attrgetter

for i in nomtimes:
    df[i] = df['timeStamp'].apply(attrgetter(i))

Here's a complete example:

df = pd.DataFrame({'timeStamp': ['2018-05-05 15:00', '2015-01-30 11:00']})
df['timeStamp'] = pd.to_datetime(df['timeStamp'])

nomtimes = ['year', 'hour', 'month', 'dayofweek']

for i in nomtimes:
    df[i] = df['timeStamp'].apply(attrgetter(i))

print(df)

            timeStamp  year  hour  month  dayofweek
0 2018-05-05 15:00:00  2018    15      5          5
1 2015-01-30 11:00:00  2015    11      1          4

Your code will not work because you are attempting to pass a string rather than extracting an attribute by name. Yet this isn't what's happening: the syntax does not feed the string but tries to access i directly, as demonstrated in your first example.

Getting rid of the for loop

You might ask if there's any way to extract all attributes from a datetime object in one go rather than sequentially. The benefit of attrgetter is you can specify multiple attributes directly to avoid a for loop altogether:

attributes = df['timeStamp'].apply(attrgetter(*nomtimes))
df[nomtimes] = pd.DataFrame(attributes.values.tolist())

Using dt accessor instead of apply

But pd.Series.apply is just a thinly veiled loop. Often, it's not necessary. Borrowing @juanpa.arrivillaga's idea, you an access attributes directly via the pd.Series.dt accessor:

attributes = pd.concat(attrgetter(*nomtimes)(df['timeStamp'].dt), axis=1, keys=nomtimes)
df = df.join(attributes)