Essential basic functionality#
Here we discuss a lot of the essential functionality common to the pandas data structures. To begin, let’s create some example objects like we did in the 10 minutes to pandas section
In [1]: index = pd.date_range("1/1/2000", periods=8)
In [2]: s = pd.Series(np.random.randn(5), index=["a", "b", "c", "d", "e"])
In [3]: df = pd.DataFrame(np.random.randn(8, 3), index=index, columns=["A", "B", "C"])
Head and tail#
To view a small sample of a Series or DataFrame object, use the head() and tail() methods. The default number of elements to display is five, but you may pass a custom number.
In [4]: long_series = pd.Series(np.random.randn(1000))
In [5]: long_series.head()
Out[5]:
0 -1.157892
1 -1.344312
2 0.844885
3 1.075770
4 -0.109050
dtype: float64
In [6]: long_series.tail(3)
Out[6]:
997 -0.289388
998 -1.020544
999 0.589993
dtype: float64
Attributes and underlying data#
pandas objects have a number of attributes enabling you to access the metadata.
shape: gives the axis dimensions of the object, consistent with ndarray
- Axis labels
Series: index (only axis)
DataFrame: index (rows) and columns
Note, these attributes can be safely assigned to!
In [7]: df[:2]
Out[7]:
A B C
2000-01-01 -0.173215 0.119209 -1.044236
2000-01-02 -0.861849 -2.104569 -0.494929
In [8]: df.columns = [x.lower() for x in df.columns]
In [9]: df
Out[9]:
a b c
2000-01-01 -0.173215 0.119209 -1.044236
2000-01-02 -0.861849 -2.104569 -0.494929
2000-01-03 1.071804 0.721555 -0.706771
2000-01-04 -1.039575 0.271860 -0.424972
2000-01-05 0.567020 0.276232 -1.087401
2000-01-06 -0.673690 0.113648 -1.478427
2000-01-07 0.524988 0.404705 0.577046
2000-01-08 -1.715002 -1.039268 -0.370647
pandas objects (Index, Series, DataFrame) can be thought of as containers for arrays, which hold the actual data and do the actual computation. For many types, the underlying array is a numpy.ndarray. However, pandas and 3rd party libraries may extend NumPy’s type system to add support for custom arrays (see dtypes).
To get the actual data inside a Index or Series, use the .array property.
In [10]: s.array
Out[10]:
<NumpyExtensionArray>
[ 0.4691122999071863, -0.2828633443286633, -1.5090585031735124,
-1.1356323710171934, 1.2121120250208506]
Length: 5, dtype: float64
In [11]: s.index.array
Out[11]:
<ArrowStringArray>
['a', 'b', 'c', 'd', 'e']
Length: 5, dtype: str
array will always be an ExtensionArray. The exact details of what an ExtensionArray is and why pandas uses them are a bit beyond the scope of this introduction. See dtypes for more.
If you know you need a NumPy array, use to_numpy() or numpy.asarray().
In [12]: s.to_numpy()
Out[12]: array([ 0.4691, -0.2829, -1.5091, -1.1356, 1.2121])
In [13]: np.asarray(s)
Out[13]: array([ 0.4691, -0.2829, -1.5091, -1.1356, 1.2121])
When the Series or Index is backed by an ExtensionArray, to_numpy() may involve copying data and coercing values. See dtypes for more.
to_numpy() gives some control over the dtype of the resulting numpy.ndarray. For example, consider datetimes with timezones. NumPy doesn’t have a dtype to represent timezone-aware datetimes, so there are two possibly useful representations
An object-dtype
numpy.ndarraywithTimestampobjects, each with the correcttz.A
datetime64[ns]-dtypenumpy.ndarray, where the values have been converted to UTC and the timezone discarded.
Timezones may be preserved with dtype=object
In [14]: ser = pd.Series(pd.date_range("2000", periods=2, tz="CET"))
In [15]: ser.to_numpy(dtype=object)
Out[15]:
array([Timestamp('2000-01-01 00:00:00+0100', tz='CET'),
Timestamp('2000-01-02 00:00:00+0100', tz='CET')], dtype=object)
Or thrown away with dtype='datetime64[ns]'
In [16]: ser.to_numpy(dtype="datetime64[ns]")
Out[16]:
array(['1999-12-31T23:00:00.000000000', '2000-01-01T23:00:00.000000000'],
dtype='datetime64[ns]')
Getting the “raw data” inside a DataFrame is possibly a bit more complex. When your DataFrame only has a single data type for all the columns, DataFrame.to_numpy() will return the underlying data
In [17]: df.to_numpy()
Out[17]:
array([[-0.1732, 0.1192, -1.0442],
[-0.8618, -2.1046, -0.4949],
[ 1.0718, 0.7216, -0.7068],
[-1.0396, 0.2719, -0.425 ],
[ 0.567 , 0.2762, -1.0874],
[-0.6737, 0.1136, -1.4784],
[ 0.525 , 0.4047, 0.577 ],
[-1.715 , -1.0393, -0.3706]])
If a DataFrame contains homogeneously-typed data, the ndarray can actually be modified in-place, and the changes will be reflected in the data structure. For heterogeneous data (e.g. some of the DataFrame’s columns are not all the same dtype), this will not be the case. The values attribute itself, unlike the axis labels, cannot be assigned to.
注意
When working with heterogeneous data, the dtype of the resulting ndarray will be chosen to accommodate all of the data involved. For example, if strings are involved, the result will be of object dtype. If there are only floats and integers, the resulting array will be of float dtype.
In the past, pandas recommended Series.values or DataFrame.values for extracting the data from a Series or DataFrame. You’ll still find references to these in old code bases and online. Going forward, we recommend avoiding .values and using .array or .to_numpy(). .values has the following drawbacks
When your Series contains an extension type, it’s unclear whether
Series.valuesreturns a NumPy array or the extension array.Series.arraywill always return anExtensionArray, and will never copy data.Series.to_numpy()will always return a NumPy array, potentially at the cost of copying / coercing values.When your DataFrame contains a mixture of data types,
DataFrame.valuesmay involve copying data and coercing values to a common dtype, a relatively expensive operation.DataFrame.to_numpy(), being a method, makes it clearer that the returned NumPy array may not be a view on the same data in the DataFrame.
Accelerated operations#
pandas has support for accelerating certain types of binary numerical and boolean operations using the numexpr library and the bottleneck libraries.
These libraries are especially useful when dealing with large data sets, and provide large speedups. numexpr uses smart chunking, caching, and multiple cores. bottleneck is a set of specialized cython routines that are especially fast when dealing with arrays that have nans.
You are highly encouraged to install both libraries. See the section Recommended Dependencies for more installation info.
These are both enabled to be used by default, you can control this by setting the options
pd.set_option("compute.use_bottleneck", False)
pd.set_option("compute.use_numexpr", False)
Flexible binary operations#
With binary operations between pandas data structures, there are two key points of interest
Broadcasting behavior between higher- (e.g. DataFrame) and lower-dimensional (e.g. Series) objects.
Missing data in computations.
We will demonstrate how to manage these issues independently, though they can be handled simultaneously.
Matching / broadcasting behavior#
DataFrame has the methods add(), sub(), mul(), div() and related functions radd(), rsub(), … for carrying out binary operations. For broadcasting behavior, Series input is of primary interest. Using these functions, you can use to either match on the index or columns via the axis keyword
In [18]: df = pd.DataFrame(
....: {
....: "one": pd.Series(np.random.randn(3), index=["a", "b", "c"]),
....: "two": pd.Series(np.random.randn(4), index=["a", "b", "c", "d"]),
....: "three": pd.Series(np.random.randn(3), index=["b", "c", "d"]),
....: }
....: )
....:
In [19]: df
Out[19]:
one two three
a 1.394981 1.772517 NaN
b 0.343054 1.912123 -0.050390
c 0.695246 1.478369 1.227435
d NaN 0.279344 -0.613172
In [20]: row = df.iloc[1]
In [21]: column = df["two"]
In [22]: df.sub(row, axis="columns")
Out[22]:
one two three
a 1.051928 -0.139606 NaN
b 0.000000 0.000000 0.000000
c 0.352192 -0.433754 1.277825
d NaN -1.632779 -0.562782
In [23]: df.sub(row, axis=1)
Out[23]:
one two three
a 1.051928 -0.139606 NaN
b 0.000000 0.000000 0.000000
c 0.352192 -0.433754 1.277825
d NaN -1.632779 -0.562782
In [24]: df.sub(column, axis="index")
Out[24]:
one two three
a -0.377535 0.0 NaN
b -1.569069 0.0 -1.962513
c -0.783123 0.0 -0.250933
d NaN 0.0 -0.892516
In [25]: df.sub(column, axis=0)
Out[25]:
one two three
a -0.377535 0.0 NaN
b -1.569069 0.0 -1.962513
c -0.783123 0.0 -0.250933
d NaN 0.0 -0.892516
Furthermore you can align a level of a MultiIndexed DataFrame with a Series.
In [26]: dfmi = df.copy()
In [27]: dfmi.index = pd.MultiIndex.from_tuples(
....: [(1, "a"), (1, "b"), (1, "c"), (2, "a")], names=["first", "second"]
....: )
....:
In [28]: dfmi.sub(column, axis=0, level="second")
Out[28]:
one two three
first second
1 a -0.377535 0.000000 NaN
b -1.569069 0.000000 -1.962513
c -0.783123 0.000000 -0.250933
2 a NaN -1.493173 -2.385688
Series and Index also support the divmod() builtin. This function takes the floor division and modulo operation at the same time returning a two-tuple of the same type as the left hand side. For example
In [29]: s = pd.Series(np.arange(10))
In [30]: s
Out[30]:
0 0
1 1
2 2
3 3
4 4
5 5
6 6
7 7
8 8
9 9
dtype: int64
In [31]: div, rem = divmod(s, 3)
In [32]: div
Out[32]:
0 0
1 0
2 0
3 1
4 1
5 1
6 2
7 2
8 2
9 3
dtype: int64
In [33]: rem
Out[33]:
0 0
1 1
2 2
3 0
4 1
5 2
6 0
7 1
8 2
9 0
dtype: int64
In [34]: idx = pd.Index(np.arange(10))
In [35]: idx
Out[35]: Index([0, 1, 2, 3, 4, 5, 6, 7, 8, 9], dtype='int64')
In [36]: div, rem = divmod(idx, 3)
In [37]: div
Out[37]: Index([0, 0, 0, 1, 1, 1, 2, 2, 2, 3], dtype='int64')
In [38]: rem
Out[38]: Index([0, 1, 2, 0, 1, 2, 0, 1, 2, 0], dtype='int64')
We can also do elementwise divmod()
In [39]: div, rem = divmod(s, [2, 2, 3, 3, 4, 4, 5, 5, 6, 6])
In [40]: div
Out[40]:
0 0
1 0
2 0
3 1
4 1
5 1
6 1
7 1
8 1
9 1
dtype: int64
In [41]: rem
Out[41]:
0 0
1 1
2 2
3 0
4 0
5 1
6 1
7 2
8 2
9 3
dtype: int64
Missing data / operations with fill values#
In Series and DataFrame, the arithmetic functions have the option of inputting a fill_value, namely a value to substitute when at most one of the values at a location are missing. For example, when adding two DataFrame objects, you may wish to treat NaN as 0 unless both DataFrames are missing that value, in which case the result will be NaN (you can later replace NaN with some other value using fillna if you wish).
In [42]: df2 = df.copy()
In [43]: df2.loc["a", "three"] = 1.0
In [44]: df
Out[44]:
one two three
a 1.394981 1.772517 NaN
b 0.343054 1.912123 -0.050390
c 0.695246 1.478369 1.227435
d NaN 0.279344 -0.613172
In [45]: df2
Out[45]:
one two three
a 1.394981 1.772517 1.000000
b 0.343054 1.912123 -0.050390
c 0.695246 1.478369 1.227435
d NaN 0.279344 -0.613172
In [46]: df + df2
Out[46]:
one two three
a 2.789963 3.545034 NaN
b 0.686107 3.824246 -0.100780
c 1.390491 2.956737 2.454870
d NaN 0.558688 -1.226343
In [47]: df.add(df2, fill_value=0)
Out[47]:
one two three
a 2.789963 3.545034 1.000000
b 0.686107 3.824246 -0.100780
c 1.390491 2.956737 2.454870
d NaN 0.558688 -1.226343
Flexible comparisons#
Series and DataFrame have the binary comparison methods eq, ne, lt, gt, le, and ge whose behavior is analogous to the binary arithmetic operations described above
In [48]: df.gt(df2)
Out[48]:
one two three
a False False False
b False False False
c False False False
d False False False
In [49]: df2.ne(df)
Out[49]:
one two three
a False False True
b False False False
c False False False
d True False False
These operations produce a pandas object of the same type as the left-hand-side input that is of dtype bool. These boolean objects can be used in indexing operations, see the section on Boolean indexing.
Boolean reductions#
You can apply the reductions: empty, any(), all().
In [50]: (df > 0).all()
Out[50]:
one False
two True
three False
dtype: bool
In [51]: (df > 0).any()
Out[51]:
one True
two True
three True
dtype: bool
You can reduce to a final boolean value.
In [52]: (df > 0).any().any()
Out[52]: np.True_
You can test if a pandas object is empty, via the empty property.
In [53]: df.empty
Out[53]: False
In [54]: pd.DataFrame(columns=list("ABC")).empty
Out[54]: True
警告
Asserting the truthiness of a pandas object will raise an error, as the testing of the emptiness or values is ambiguous.
In [55]: if df:
....: print(True)
....:
---------------------------------------------------------------------------
ValueError Traceback (most recent call last)
<ipython-input-55-318d08b2571a> in ?()
----> 1 if df:
2 print(True)
~/work/pandas/pandas/pandas/core/generic.py in ?(self)
1511 @final
1512 def __bool__(self) -> NoReturn:
-> 1513 raise ValueError(
1514 f"The truth value of a {type(self).__name__} is ambiguous. "
1515 "Use a.empty, a.bool(), a.item(), a.any() or a.all()."
1516 )
ValueError: The truth value of a DataFrame is ambiguous. Use a.empty, a.bool(), a.item(), a.any() or a.all().
In [56]: df and df2
---------------------------------------------------------------------------
ValueError Traceback (most recent call last)
<ipython-input-56-b241b64bb471> in ?()
----> 1 df and df2
~/work/pandas/pandas/pandas/core/generic.py in ?(self)
1511 @final
1512 def __bool__(self) -> NoReturn:
-> 1513 raise ValueError(
1514 f"The truth value of a {type(self).__name__} is ambiguous. "
1515 "Use a.empty, a.bool(), a.item(), a.any() or a.all()."
1516 )
ValueError: The truth value of a DataFrame is ambiguous. Use a.empty, a.bool(), a.item(), a.any() or a.all().
See gotchas for a more detailed discussion.
Comparing if objects are equivalent#
Often you may find that there is more than one way to compute the same result. As a simple example, consider df + df and df * 2. To test that these two computations produce the same result, given the tools shown above, you might imagine using (df + df == df * 2).all(). But in fact, this expression is False
In [57]: df + df == df * 2
Out[57]:
one two three
a True True False
b True True True
c True True True
d False True True
In [58]: (df + df == df * 2).all()
Out[58]:
one False
two True
three False
dtype: bool
Notice that the boolean DataFrame df + df == df * 2 contains some False values! This is because NaNs do not compare as equals
In [59]: np.nan == np.nan
Out[59]: False
So, NDFrames (such as Series and DataFrames) have an equals() method for testing equality, with NaNs in corresponding locations treated as equal.
In [60]: (df + df).equals(df * 2)
Out[60]: True
Note that the Series or DataFrame index needs to be in the same order for equality to be True
In [61]: df1 = pd.DataFrame({"col": ["foo", 0, np.nan]})
In [62]: df2 = pd.DataFrame({"col": [np.nan, 0, "foo"]}, index=[2, 1, 0])
In [63]: df1.equals(df2)
Out[63]: False
In [64]: df1.equals(df2.sort_index())
Out[64]: True
Comparing array-like objects#
You can conveniently perform element-wise comparisons when comparing a pandas data structure with a scalar value
In [65]: pd.Series(["foo", "bar", "baz"]) == "foo"
Out[65]:
0 True
1 False
2 False
dtype: bool
In [66]: pd.Index(["foo", "bar", "baz"]) == "foo"
Out[66]: array([ True, False, False])
pandas also handles element-wise comparisons between different array-like objects of the same length
In [67]: pd.Series(["foo", "bar", "baz"]) == pd.Index(["foo", "bar", "qux"])
Out[67]:
0 True
1 True
2 False
dtype: bool
In [68]: pd.Series(["foo", "bar", "baz"]) == np.array(["foo", "bar", "qux"])
Out[68]:
0 True
1 True
2 False
dtype: bool
Trying to compare Index or Series objects of different lengths will raise a ValueError
In [69]: pd.Series(['foo', 'bar', 'baz']) == pd.Series(['foo', 'bar'])
---------------------------------------------------------------------------
ValueError Traceback (most recent call last)
Cell In[69], line 1
----> 1 pd.Series(['foo', 'bar', 'baz']) == pd.Series(['foo', 'bar'])
File ~/work/pandas/pandas/pandas/core/ops/common.py:85, in _unpack_zerodim_and_defer.<locals>.new_method(self, other)
82 other = sanitize_array(other, None)
83 other = ensure_wrapped_if_datetimelike(other)
---> 85 return method(self, other)
File ~/work/pandas/pandas/pandas/core/arraylike.py:42, in OpsMixin.__eq__(self, other)
40 @unpack_zerodim_and_defer("__eq__")
41 def __eq__(self, other):
---> 42 return self._cmp_method(other, operator.eq)
File ~/work/pandas/pandas/pandas/core/series.py:6730, in Series._cmp_method(self, other, op)
6727 res_name = ops.get_op_result_name(self, other)
6729 if isinstance(other, Series) and not self._indexed_same(other):
-> 6730 raise ValueError("Can only compare identically-labeled Series objects")
6732 lvalues = self._values
6733 rvalues = extract_array(other, extract_numpy=True, extract_range=True)
ValueError: Can only compare identically-labeled Series objects
In [70]: pd.Series(['foo', 'bar', 'baz']) == pd.Series(['foo'])
---------------------------------------------------------------------------
ValueError Traceback (most recent call last)
Cell In[70], line 1
----> 1 pd.Series(['foo', 'bar', 'baz']) == pd.Series(['foo'])
File ~/work/pandas/pandas/pandas/core/ops/common.py:85, in _unpack_zerodim_and_defer.<locals>.new_method(self, other)
82 other = sanitize_array(other, None)
83 other = ensure_wrapped_if_datetimelike(other)
---> 85 return method(self, other)
File ~/work/pandas/pandas/pandas/core/arraylike.py:42, in OpsMixin.__eq__(self, other)
40 @unpack_zerodim_and_defer("__eq__")
41 def __eq__(self, other):
---> 42 return self._cmp_method(other, operator.eq)
File ~/work/pandas/pandas/pandas/core/series.py:6730, in Series._cmp_method(self, other, op)
6727 res_name = ops.get_op_result_name(self, other)
6729 if isinstance(other, Series) and not self._indexed_same(other):
-> 6730 raise ValueError("Can only compare identically-labeled Series objects")
6732 lvalues = self._values
6733 rvalues = extract_array(other, extract_numpy=True, extract_range=True)
ValueError: Can only compare identically-labeled Series objects
Combining overlapping data sets#
A problem occasionally arising is the combination of two similar data sets where values in one are preferred over the other. An example would be two data series representing a particular economic indicator where one is considered to be of “higher quality”. However, the lower quality series might extend further back in history or have more complete data coverage. As such, we would like to combine two DataFrame objects where missing values in one DataFrame are conditionally filled with like-labeled values from the other DataFrame. The function implementing this operation is combine_first(), which we illustrate
In [71]: df1 = pd.DataFrame(
....: {"A": [1.0, np.nan, 3.0, 5.0, np.nan], "B": [np.nan, 2.0, 3.0, np.nan, 6.0]}
....: )
....:
In [72]: df2 = pd.DataFrame(
....: {
....: "A": [5.0, 2.0, 4.0, np.nan, 3.0, 7.0],
....: "B": [np.nan, np.nan, 3.0, 4.0, 6.0, 8.0],
....: }
....: )
....:
In [73]: df1
Out[73]:
A B
0 1.0 NaN
1 NaN 2.0
2 3.0 3.0
3 5.0 NaN
4 NaN 6.0
In [74]: df2
Out[74]:
A B
0 5.0 NaN
1 2.0 NaN
2 4.0 3.0
3 NaN 4.0
4 3.0 6.0
5 7.0 8.0
In [75]: df1.combine_first(df2)
Out[75]:
A B
0 1.0 NaN
1 2.0 2.0
2 3.0 3.0
3 5.0 4.0
4 3.0 6.0
5 7.0 8.0
General DataFrame combine#
The combine_first() method above calls the more general DataFrame.combine(). This method takes another DataFrame and a combiner function, aligns the input DataFrame and then passes the combiner function pairs of Series (i.e., columns whose names are the same).
So, for instance, to reproduce combine_first() as above
In [76]: def combiner(x, y):
....: return np.where(pd.isna(x), y, x)
....:
In [77]: df1.combine(df2, combiner)
Out[77]:
A B
0 1.0 NaN
1 2.0 2.0
2 3.0 3.0
3 5.0 4.0
4 3.0 6.0
5 7.0 8.0
Descriptive statistics#
There exists a large number of methods for computing descriptive statistics and other related operations on Series, DataFrame. Most of these are aggregations (hence producing a lower-dimensional result) like sum(), mean(), and quantile(), but some of them, like cumsum() and cumprod(), produce an object of the same size. Generally speaking, these methods take an axis argument, just like ndarray.{sum, std, …}, but the axis can be specified by name or integer
Series: no axis argument needed
DataFrame: “index” (axis=0, default), “columns” (axis=1)
例如：
In [78]: df
Out[78]:
one two three
a 1.394981 1.772517 NaN
b 0.343054 1.912123 -0.050390
c 0.695246 1.478369 1.227435
d NaN 0.279344 -0.613172
In [79]: df.mean(axis=0)
Out[79]:
one 0.811094
two 1.360588
three 0.187958
dtype: float64
In [80]: df.mean(axis=1)
Out[80]:
a 1.583749
b 0.734929
c 1.133683
d -0.166914
dtype: float64
All such methods have a skipna option signaling whether to exclude missing data (True by default)
In [81]: df.sum(axis=0, skipna=False)
Out[81]:
one NaN
two 5.442353
three NaN
dtype: float64
In [82]: df.sum(axis=1, skipna=True)
Out[82]:
a 3.167498
b 2.204786
c 3.401050
d -0.333828
dtype: float64
Combined with the broadcasting / arithmetic behavior, one can describe various statistical procedures, like standardization (rendering data zero mean and standard deviation of 1), very concisely
In [83]: ts_stand = (df - df.mean()) / df.std()
In [84]: ts_stand.std()
Out[84]:
one 1.0
two 1.0
three 1.0
dtype: float64
In [85]: xs_stand = df.sub(df.mean(axis=1), axis=0).div(df.std(axis=1), axis=0)
In [86]: xs_stand.std(axis=1)
Out[86]:
a 1.0
b 1.0
c 1.0
d 1.0
dtype: float64
Note that methods like cumsum() and cumprod() preserve the location of NaN values. This is somewhat different from expanding() and rolling() since NaN behavior is furthermore dictated by a min_periods parameter.
In [87]: df.cumsum()
Out[87]:
one two three
a 1.394981 1.772517 NaN
b 1.738035 3.684640 -0.050390
c 2.433281 5.163008 1.177045
d NaN 5.442353 0.563873
Here is a quick reference summary table of common functions. Each also takes an optional level parameter which applies only if the object has a hierarchical index.
Function |
描述 |
|---|---|
|
Number of non-NA observations |
|
Sum of values |
|
Mean of values |
|
Arithmetic median of values |
|
Minimum |
|
Maximum |
|
Mode |
|
Absolute Value |
|
Product of values |
|
Bessel-corrected sample standard deviation |
|
Unbiased variance |
|
Standard error of the mean |
|
Sample skewness (3rd moment) |
|
Sample kurtosis (4th moment) |
|
Sample quantile (value at %) |
|
Cumulative sum |
|
Cumulative product |
|
Cumulative maximum |
|
Cumulative minimum |
Note that by chance some NumPy methods, like mean, std, and sum, will exclude NAs on Series input by default
In [88]: np.mean(df["one"])
Out[88]: np.float64(0.8110935116651192)
In [89]: np.mean(df["one"].to_numpy())
Out[89]: np.float64(nan)
Series.nunique() will return the number of unique non-NA values in a Series
In [90]: series = pd.Series(np.random.randn(500))
In [91]: series[20:500] = np.nan
In [92]: series[10:20] = 5
In [93]: series.nunique()
Out[93]: 11
Summarizing data: describe#
There is a convenient describe() function which computes a variety of summary statistics about a Series or the columns of a DataFrame (excluding NAs of course)
In [94]: series = pd.Series(np.random.randn(1000))
In [95]: series[::2] = np.nan
In [96]: series.describe()
Out[96]:
count 500.000000
mean -0.021292
std 1.015906
min -2.683763
25% -0.699070
50% -0.069718
75% 0.714483
max 3.160915
dtype: float64
In [97]: frame = pd.DataFrame(np.random.randn(1000, 5), columns=["a", "b", "c", "d", "e"])
In [98]: frame.iloc[::2] = np.nan
In [99]: frame.describe()
Out[99]:
a b c d e
count 500.000000 500.000000 500.000000 500.000000 500.000000
mean 0.033387 0.030045 -0.043719 -0.051686 0.005979
std 1.017152 0.978743 1.025270 1.015988 1.006695
min -3.000951 -2.637901 -3.303099 -3.159200 -3.188821
25% -0.647623 -0.576449 -0.712369 -0.691338 -0.691115
50% 0.047578 -0.021499 -0.023888 -0.032652 -0.025363
75% 0.729907 0.775880 0.618896 0.670047 0.649748
max 2.740139 2.752332 3.004229 2.728702 3.240991
You can select specific percentiles to include in the output
In [100]: series.describe(percentiles=[0.05, 0.25, 0.75, 0.95])
Out[100]:
count 500.000000
mean -0.021292
std 1.015906
min -2.683763
5% -1.645423
25% -0.699070
75% 0.714483
95% 1.711409
max 3.160915
dtype: float64
By default, the median is always included.
For a non-numerical Series object, describe() will give a simple summary of the number of unique values and most frequently occurring values
In [101]: s = pd.Series(["a", "a", "b", "b", "a", "a", np.nan, "c", "d", "a"])
In [102]: s.describe()
Out[102]:
count 9
unique 4
top a
freq 5
dtype: object
Note that on a mixed-type DataFrame object, describe() will restrict the summary to include only numerical columns or, if none are, only categorical columns
In [103]: frame = pd.DataFrame({"a": ["Yes", "Yes", "No", "No"], "b": range(4)})
In [104]: frame.describe()
Out[104]:
b
count 4.000000
mean 1.500000
std 1.290994
min 0.000000
25% 0.750000
50% 1.500000
75% 2.250000
max 3.000000
This behavior can be controlled by providing a list of types as include/exclude arguments. The special value all can also be used
In [105]: frame.describe(include=["str"])
Out[105]:
a
count 4
unique 2
top Yes
freq 2
In [106]: frame.describe(include=["number"])
Out[106]:
b
count 4.000000
mean 1.500000
std 1.290994
min 0.000000
25% 0.750000
50% 1.500000
75% 2.250000
max 3.000000
In [107]: frame.describe(include="all")
Out[107]:
a b
count 4 4.000000
unique 2 NaN
top Yes NaN
freq 2 NaN
mean NaN 1.500000
std NaN 1.290994
min NaN 0.000000
25% NaN 0.750000
50% NaN 1.500000
75% NaN 2.250000
max NaN 3.000000
That feature relies on select_dtypes. Refer to there for details about accepted inputs.
Index of min/max values#
The idxmin() and idxmax() functions on Series and DataFrame compute the index labels with the minimum and maximum corresponding values
In [108]: s1 = pd.Series(np.random.randn(5))
In [109]: s1
Out[109]:
0 1.118076
1 -0.352051
2 -1.242883
3 -1.277155
4 -0.641184
dtype: float64
In [110]: s1.idxmin(), s1.idxmax()
Out[110]: (3, 0)
In [111]: df1 = pd.DataFrame(np.random.randn(5, 3), columns=["A", "B", "C"])
In [112]: df1
Out[112]:
A B C
0 -0.327863 -0.946180 -0.137570
1 -0.186235 -0.257213 -0.486567
2 -0.507027 -0.871259 -0.111110
3 2.000339 -2.430505 0.089759
4 -0.321434 -0.033695 0.096271
In [113]: df1.idxmin(axis=0)
Out[113]:
A 2
B 3
C 1
dtype: int64
In [114]: df1.idxmax(axis=1)
Out[114]:
0 C
1 A
2 C
3 A
4 C
dtype: str
When there are multiple rows (or columns) matching the minimum or maximum value, idxmin() and idxmax() return the first matching index
In [115]: df3 = pd.DataFrame([2, 1, 1, 3, np.nan], columns=["A"], index=list("edcba"))
In [116]: df3
Out[116]:
A
e 2.0
d 1.0
c 1.0
b 3.0
a NaN
In [117]: df3["A"].idxmin()
Out[117]: 'd'
注意
idxmin and idxmax are called argmin and argmax in NumPy.
Value counts (histogramming) / mode#
The value_counts() Series method computes a histogram of a 1D array of values. It can also be used as a function on regular arrays
In [118]: data = np.random.randint(0, 7, size=50)
In [119]: data
Out[119]:
array([6, 6, 2, 3, 5, 3, 2, 5, 4, 5, 4, 3, 4, 5, 0, 2, 0, 4, 2, 0, 3, 2,
2, 5, 6, 5, 3, 4, 6, 4, 3, 5, 6, 4, 3, 6, 2, 6, 6, 2, 3, 4, 2, 1,
6, 2, 6, 1, 5, 4])
In [120]: s = pd.Series(data)
In [121]: s.value_counts()
Out[121]:
6 10
2 10
4 9
3 8
5 8
0 3
1 2
Name: count, dtype: int64
The value_counts() method can be used to count combinations across multiple columns. By default all columns are used but a subset can be selected using the subset argument.
In [122]: data = {"a": [1, 2, 3, 4], "b": ["x", "x", "y", "y"]}
In [123]: frame = pd.DataFrame(data)
In [124]: frame.value_counts()
Out[124]:
a b
1 x 1
2 x 1
3 y 1
4 y 1
Name: count, dtype: int64
Similarly, you can get the most frequently occurring value(s), i.e. the mode, of the values in a Series or DataFrame
In [125]: s5 = pd.Series([1, 1, 3, 3, 3, 5, 5, 7, 7, 7])
In [126]: s5.mode()
Out[126]:
0 3
1 7
dtype: int64
In [127]: df5 = pd.DataFrame(
.....: {
.....: "A": np.random.randint(0, 7, size=50),
.....: "B": np.random.randint(-10, 15, size=50),
.....: }
.....: )
.....:
In [128]: df5.mode()
Out[128]:
A B
0 1.0 -9
1 NaN 10
2 NaN 13
Discretization and quantiling#
Continuous values can be discretized using the cut() (bins based on values) and qcut() (bins based on sample quantiles) functions
In [129]: arr = np.random.randn(20)
In [130]: factor = pd.cut(arr, 4)
In [131]: factor
Out[131]:
[(-0.251, 0.464], (-0.968, -0.251], (0.464, 1.179], (-0.251, 0.464], (-0.968, -0.251], ..., (-0.251, 0.464], (-0.968, -0.251], (-0.968, -0.251], (-0.968, -0.251], (-0.968, -0.251]]
Length: 20
Categories (4, interval[float64, right]): [(-0.968, -0.251] < (-0.251, 0.464] < (0.464, 1.179] <
(1.179, 1.893]]
In [132]: factor = pd.cut(arr, [-5, -1, 0, 1, 5])
In [133]: factor
Out[133]:
[(0, 1], (-1, 0], (0, 1], (0, 1], (-1, 0], ..., (-1, 0], (-1, 0], (-1, 0], (-1, 0], (-1, 0]]
Length: 20
Categories (4, interval[int64, right]): [(-5, -1] < (-1, 0] < (0, 1] < (1, 5]]
qcut() computes sample quantiles. For example, we could slice up some normally distributed data into equal-size quartiles like so
In [134]: arr = np.random.randn(30)
In [135]: factor = pd.qcut(arr, [0, 0.25, 0.5, 0.75, 1])
In [136]: factor
Out[136]:
[(0.569, 1.184], (-2.278, -0.301], (-2.278, -0.301], (0.569, 1.184], (0.569, 1.184], ..., (-0.301, 0.569], (1.184, 2.346], (1.184, 2.346], (-0.301, 0.569], (-2.278, -0.301]]
Length: 30
Categories (4, interval[float64, right]): [(-2.278, -0.301] < (-0.301, 0.569] < (0.569, 1.184] <
(1.184, 2.346]]
We can also pass infinite values to define the bins
In [137]: arr = np.random.randn(20)
In [138]: factor = pd.cut(arr, [-np.inf, 0, np.inf])
In [139]: factor
Out[139]:
[(-inf, 0.0], (0.0, inf], (0.0, inf], (-inf, 0.0], (-inf, 0.0], ..., (-inf, 0.0], (-inf, 0.0], (-inf, 0.0], (0.0, inf], (0.0, inf]]
Length: 20
Categories (2, interval[float64, right]): [(-inf, 0.0] < (0.0, inf]]
函数应用#
To apply your own or another library’s functions to pandas objects, you should be aware of the three methods below. The appropriate method to use depends on whether your function expects to operate on an entire DataFrame or Series, row- or column-wise, or elementwise.
Tablewise function application#
DataFrames and Series can be passed into functions. However, if the function needs to be called in a chain, consider using the pipe() method.
First some setup
In [140]: def extract_city_name(df):
.....: """
.....: Chicago, IL -> Chicago for city_name column
.....: """
.....: df["city_name"] = df["city_and_code"].str.split(",").str.get(0)
.....: return df
.....:
In [141]: def add_country_name(df, country_name=None):
.....: """
.....: Chicago -> Chicago-US for city_name column
.....: """
.....: col = "city_name"
.....: df["city_and_country"] = df[col] + country_name
.....: return df
.....:
In [142]: df_p = pd.DataFrame({"city_and_code": ["Chicago, IL"]})
extract_city_name and add_country_name are functions taking and returning DataFrames.
Now compare the following
In [143]: add_country_name(extract_city_name(df_p), country_name="US")
Out[143]:
city_and_code city_name city_and_country
0 Chicago, IL Chicago ChicagoUS
Is equivalent to
In [144]: df_p.pipe(extract_city_name).pipe(add_country_name, country_name="US")
Out[144]:
city_and_code city_name city_and_country
0 Chicago, IL Chicago ChicagoUS
pandas encourages the second style, which is known as method chaining. pipe makes it easy to use your own or another library’s functions in method chains, alongside pandas’ methods.
In the example above, the functions extract_city_name and add_country_name each expected a DataFrame as the first positional argument. What if the function you wish to apply takes its data as, say, the second argument? In this case, provide pipe with a tuple of (callable, data_keyword). .pipe will route the DataFrame to the argument specified in the tuple.
例如，我们可以使用statsmodels拟合一个回归。它们的API首先需要一个公式，然后是第二个参数，即一个DataFrame，名为data。我们将函数、关键字对(sm.ols, 'data')传递给pipe。
In [147]: import statsmodels.formula.api as sm
In [148]: bb = pd.read_csv("data/baseball.csv", index_col="id")
In [149]: (
.....: bb.query("h > 0")
.....: .assign(ln_h=lambda df: np.log(df.h))
.....: .pipe((sm.ols, "data"), "hr ~ ln_h + year + g + C(lg)")
.....: .fit()
.....: .summary()
.....: )
.....:
Out[149]:
<class 'statsmodels.iolib.summary.Summary'>
"""
OLS Regression Results
==============================================================================
Dep. Variable: hr R-squared: 0.685
Model: OLS Adj. R-squared: 0.665
Method: Least Squares F-statistic: 34.28
Date: Tue, 22 Nov 2022 Prob (F-statistic): 3.48e-15
Time: 05:34:17 Log-Likelihood: -205.92
No. Observations: 68 AIC: 421.8
Df Residuals: 63 BIC: 432.9
Df Model: 4
Covariance Type: nonrobust
===============================================================================
coef std err t P>|t| [0.025 0.975]
-------------------------------------------------------------------------------
Intercept -8484.7720 4664.146 -1.819 0.074 -1.78e+04 835.780
C(lg)[T.NL] -2.2736 1.325 -1.716 0.091 -4.922 0.375
ln_h -1.3542 0.875 -1.547 0.127 -3.103 0.395
year 4.2277 2.324 1.819 0.074 -0.417 8.872
g 0.1841 0.029 6.258 0.000 0.125 0.243
==============================================================================
Omnibus: 10.875 Durbin-Watson: 1.999
Prob(Omnibus): 0.004 Jarque-Bera (JB): 17.298
Skew: 0.537 Prob(JB): 0.000175
Kurtosis: 5.225 Cond. No. 1.49e+07
==============================================================================
Notes:
[1] Standard Errors assume that the covariance matrix of the errors is correctly specified.
[2] The condition number is large, 1.49e+07. This might indicate that there are
strong multicollinearity or other numerical problems.
"""
pipe方法受到了Unix管道以及最近的dplyr和magrittr的启发，它们为R引入了流行的(%>%)（读取管道）运算符。这里的pipe实现非常简洁，在Python中感觉很自然。我们鼓励您查看pipe()的源代码。
按行或按列应用函数#
可以使用apply()方法沿DataFrame的轴应用任意函数，该方法与描述性统计方法一样，接受一个可选的axis参数。
In [145]: df.apply(lambda x: np.mean(x))
Out[145]:
one 0.811094
two 1.360588
three 0.187958
dtype: float64
In [146]: df.apply(lambda x: np.mean(x), axis=1)
Out[146]:
a 1.583749
b 0.734929
c 1.133683
d -0.166914
dtype: float64
In [147]: df.apply(lambda x: x.max() - x.min())
Out[147]:
one 1.051928
two 1.632779
three 1.840607
dtype: float64
In [148]: df.apply(np.cumsum)
Out[148]:
one two three
a 1.394981 1.772517 NaN
b 1.738035 3.684640 -0.050390
c 2.433281 5.163008 1.177045
d NaN 5.442353 0.563873
In [149]: df.apply(np.exp)
Out[149]:
one two three
a 4.034899 5.885648 NaN
b 1.409244 6.767440 0.950858
c 2.004201 4.385785 3.412466
d NaN 1.322262 0.541630
apply()方法还将根据字符串方法名称进行分派。
In [150]: df.apply("mean")
Out[150]:
one 0.811094
two 1.360588
three 0.187958
dtype: float64
In [151]: df.apply("mean", axis=1)
Out[151]:
a 1.583749
b 0.734929
c 1.133683
d -0.166914
dtype: float64
传递给apply()的函数的返回类型会影响DataFrame.apply的默认行为的最终输出类型。
如果应用函数返回一个
Series，则最终输出是一个DataFrame。列与应用函数返回的Series的索引匹配。如果应用函数返回任何其他类型，则最终输出是一个
Series。
可以通过result_type来覆盖此默认行为，它接受三个选项：reduce、broadcast和expand。这些选项将决定列表类返回值如何（或不）展开为DataFrame。
apply()结合一些巧妙的技巧可以用来回答关于数据集的许多问题。例如，假设我们想提取每列最大值出现的日期。
In [152]: tsdf = pd.DataFrame(
.....: np.random.randn(1000, 3),
.....: columns=["A", "B", "C"],
.....: index=pd.date_range("1/1/2000", periods=1000),
.....: )
.....:
In [153]: tsdf.apply(lambda x: x.idxmax())
Out[153]:
A 2000-08-06
B 2001-01-18
C 2001-07-18
dtype: datetime64[us]
您还可以将其他参数和关键字参数传递给apply()方法。
In [154]: def subtract_and_divide(x, sub, divide=1):
.....: return (x - sub) / divide
.....:
In [155]: df_udf = pd.DataFrame(np.ones((2, 2)))
In [156]: df_udf.apply(subtract_and_divide, args=(5,), divide=3)
Out[156]:
0 1
0 -1.333333 -1.333333
1 -1.333333 -1.333333
另一个有用的功能是能够传递Series方法来对每列或每行执行一些Series操作。
In [157]: tsdf = pd.DataFrame(
.....: np.random.randn(10, 3),
.....: columns=["A", "B", "C"],
.....: index=pd.date_range("1/1/2000", periods=10),
.....: )
.....:
In [158]: tsdf.iloc[3:7] = np.nan
In [159]: tsdf
Out[159]:
A B C
2000-01-01 -0.158131 -0.232466 0.321604
2000-01-02 -1.810340 -3.105758 0.433834
2000-01-03 -1.209847 -1.156793 -0.136794
2000-01-04 NaN NaN NaN
2000-01-05 NaN NaN NaN
2000-01-06 NaN NaN NaN
2000-01-07 NaN NaN NaN
2000-01-08 -0.653602 0.178875 1.008298
2000-01-09 1.007996 0.462824 0.254472
2000-01-10 0.307473 0.600337 1.643950
In [160]: tsdf.apply(pd.Series.interpolate)
Out[160]:
A B C
2000-01-01 -0.158131 -0.232466 0.321604
2000-01-02 -1.810340 -3.105758 0.433834
2000-01-03 -1.209847 -1.156793 -0.136794
2000-01-04 -1.098598 -0.889659 0.092225
2000-01-05 -0.987349 -0.622526 0.321243
2000-01-06 -0.876100 -0.355392 0.550262
2000-01-07 -0.764851 -0.088259 0.779280
2000-01-08 -0.653602 0.178875 1.008298
2000-01-09 1.007996 0.462824 0.254472
2000-01-10 0.307473 0.600337 1.643950
最后，apply()接受一个raw参数，该参数默认为False，它会在应用函数之前将每行或每列转换为Series。当设置为True时，传入的函数将接收一个ndarray对象，如果您不需要索引功能，这会带来性能上的好处。
聚合API#
聚合API允许以一种简洁的方式表达可能多个聚合操作。此API在pandas对象之间是相似的，请参见groupby API、window API和resample API。聚合的入口点是DataFrame.aggregate()，或其别名DataFrame.agg()。
我们将使用上面类似的起始帧。
In [161]: tsdf = pd.DataFrame(
.....: np.random.randn(10, 3),
.....: columns=["A", "B", "C"],
.....: index=pd.date_range("1/1/2000", periods=10),
.....: )
.....:
In [162]: tsdf.iloc[3:7] = np.nan
In [163]: tsdf
Out[163]:
A B C
2000-01-01 1.257606 1.004194 0.167574
2000-01-02 -0.749892 0.288112 -0.757304
2000-01-03 -0.207550 -0.298599 0.116018
2000-01-04 NaN NaN NaN
2000-01-05 NaN NaN NaN
2000-01-06 NaN NaN NaN
2000-01-07 NaN NaN NaN
2000-01-08 0.814347 -0.257623 0.869226
2000-01-09 -0.250663 -1.206601 0.896839
2000-01-10 2.169758 -1.333363 0.283157
使用单个函数等同于apply()。您也可以将命名方法作为字符串传递。这些将返回一个聚合输出的Series。
In [164]: tsdf.agg(lambda x: np.sum(x))
Out[164]:
A 3.033606
B -1.803879
C 1.575510
dtype: float64
In [165]: tsdf.agg("sum")
Out[165]:
A 3.033606
B -1.803879
C 1.575510
dtype: float64
# these are equivalent to a ``.sum()`` because we are aggregating
# on a single function
In [166]: tsdf.sum()
Out[166]:
A 3.033606
B -1.803879
C 1.575510
dtype: float64
在Series上进行单次聚合将返回一个标量值。
In [167]: tsdf["A"].agg("sum")
Out[167]: np.float64(3.033606102414146)
使用多个函数进行聚合#
您可以将多个聚合参数作为列表传递。传递的每个函数的计算结果将作为结果DataFrame中的一行。这些自然地从聚合函数命名。
In [168]: tsdf.agg(["sum"])
Out[168]:
A B C
sum 3.033606 -1.803879 1.57551
多个函数产生多行。
In [169]: tsdf.agg(["sum", "mean"])
Out[169]:
A B C
sum 3.033606 -1.803879 1.575510
mean 0.505601 -0.300647 0.262585
在Series上，多个函数将返回一个Series，其索引是函数名称。
In [170]: tsdf["A"].agg(["sum", "mean"])
Out[170]:
sum 3.033606
mean 0.505601
Name: A, dtype: float64
传递一个lambda函数将产生一个名为<lambda>的行。
In [171]: tsdf["A"].agg(["sum", lambda x: x.mean()])
Out[171]:
sum 3.033606
<lambda> 0.505601
Name: A, dtype: float64
传递一个命名函数将产生该名称作为行。
In [172]: def mymean(x):
.....: return x.mean()
.....:
In [173]: tsdf["A"].agg(["sum", mymean])
Out[173]:
sum 3.033606
mymean 0.505601
Name: A, dtype: float64
使用字典进行聚合#
将列名字典传递给标量或标量列表，以DataFrame.agg，可以定制哪些函数应用于哪些列。请注意，结果的顺序不固定，您可以使用OrderedDict来保证顺序。
In [174]: tsdf.agg({"A": "mean", "B": "sum"})
Out[174]:
A 0.505601
B -1.803879
dtype: float64
传递列表类将生成一个DataFrame输出。您将获得所有聚合器的矩阵式输出。输出将包含所有唯一的函数。对于特定列未指定的函数将为NaN。
In [175]: tsdf.agg({"A": ["mean", "min"], "B": "sum"})
Out[175]:
A B
mean 0.505601 NaN
min -0.749892 NaN
sum NaN -1.803879
自定义describe#
使用.agg()可以轻松创建一个自定义的describe函数，类似于内置的describe函数。
In [176]: from functools import partial
In [177]: q_25 = partial(pd.Series.quantile, q=0.25)
In [178]: q_25.__name__ = "25%"
In [179]: q_75 = partial(pd.Series.quantile, q=0.75)
In [180]: q_75.__name__ = "75%"
In [181]: tsdf.agg(["count", "mean", "std", "min", q_25, "median", q_75, "max"])
Out[181]:
A B C
count 6.000000 6.000000 6.000000
mean 0.505601 -0.300647 0.262585
std 1.103362 0.887508 0.606860
min -0.749892 -1.333363 -0.757304
25% -0.239885 -0.979600 0.128907
median 0.303398 -0.278111 0.225365
75% 1.146791 0.151678 0.722709
max 2.169758 1.004194 0.896839
Transform API#
该transform()方法返回一个与原始对象索引相同的对象（大小相同）。此API允许您一次提供*多个*操作，而不是一次一个。其API与.agg API非常相似。
我们创建一个与上面各节中使用的类似的框架。
In [182]: tsdf = pd.DataFrame(
.....: np.random.randn(10, 3),
.....: columns=["A", "B", "C"],
.....: index=pd.date_range("1/1/2000", periods=10),
.....: )
.....:
In [183]: tsdf.iloc[3:7] = np.nan
In [184]: tsdf
Out[184]:
A B C
2000-01-01 -0.428759 -0.864890 -0.675341
2000-01-02 -0.168731 1.338144 -1.279321
2000-01-03 -1.621034 0.438107 0.903794
2000-01-04 NaN NaN NaN
2000-01-05 NaN NaN NaN
2000-01-06 NaN NaN NaN
2000-01-07 NaN NaN NaN
2000-01-08 0.254374 -1.240447 -0.201052
2000-01-09 -0.157795 0.791197 -1.144209
2000-01-10 -0.030876 0.371900 0.061932
转换整个框架。.transform()允许函数输入为：NumPy函数、字符串函数名称或用户定义的函数。
In [185]: tsdf.transform(np.abs)
Out[185]:
A B C
2000-01-01 0.428759 0.864890 0.675341
2000-01-02 0.168731 1.338144 1.279321
2000-01-03 1.621034 0.438107 0.903794
2000-01-04 NaN NaN NaN
2000-01-05 NaN NaN NaN
2000-01-06 NaN NaN NaN
2000-01-07 NaN NaN NaN
2000-01-08 0.254374 1.240447 0.201052
2000-01-09 0.157795 0.791197 1.144209
2000-01-10 0.030876 0.371900 0.061932
In [186]: tsdf.transform("abs")
Out[186]:
A B C
2000-01-01 0.428759 0.864890 0.675341
2000-01-02 0.168731 1.338144 1.279321
2000-01-03 1.621034 0.438107 0.903794
2000-01-04 NaN NaN NaN
2000-01-05 NaN NaN NaN
2000-01-06 NaN NaN NaN
2000-01-07 NaN NaN NaN
2000-01-08 0.254374 1.240447 0.201052
2000-01-09 0.157795 0.791197 1.144209
2000-01-10 0.030876 0.371900 0.061932
In [187]: tsdf.transform(lambda x: x.abs())
Out[187]:
A B C
2000-01-01 0.428759 0.864890 0.675341
2000-01-02 0.168731 1.338144 1.279321
2000-01-03 1.621034 0.438107 0.903794
2000-01-04 NaN NaN NaN
2000-01-05 NaN NaN NaN
2000-01-06 NaN NaN NaN
2000-01-07 NaN NaN NaN
2000-01-08 0.254374 1.240447 0.201052
2000-01-09 0.157795 0.791197 1.144209
2000-01-10 0.030876 0.371900 0.061932
这里transform()接收了一个函数；这相当于一个ufunc应用。
In [188]: np.abs(tsdf)
Out[188]:
A B C
2000-01-01 0.428759 0.864890 0.675341
2000-01-02 0.168731 1.338144 1.279321
2000-01-03 1.621034 0.438107 0.903794
2000-01-04 NaN NaN NaN
2000-01-05 NaN NaN NaN
2000-01-06 NaN NaN NaN
2000-01-07 NaN NaN NaN
2000-01-08 0.254374 1.240447 0.201052
2000-01-09 0.157795 0.791197 1.144209
2000-01-10 0.030876 0.371900 0.061932
将单个函数传递给.transform()，并使用Series，将返回一个单独的Series。
In [189]: tsdf["A"].transform(np.abs)
Out[189]:
2000-01-01 0.428759
2000-01-02 0.168731
2000-01-03 1.621034
2000-01-04 NaN
2000-01-05 NaN
2000-01-06 NaN
2000-01-07 NaN
2000-01-08 0.254374
2000-01-09 0.157795
2000-01-10 0.030876
Freq: D, Name: A, dtype: float64
使用多个函数进行转换#
传递多个函数将生成一个列MultiIndexed DataFrame。第一级将是原始框架的列名；第二级将是转换函数名称。
In [190]: tsdf.transform([np.abs, lambda x: x + 1])
Out[190]:
A B C
absolute <lambda> absolute <lambda> absolute <lambda>
2000-01-01 0.428759 0.571241 0.864890 0.135110 0.675341 0.324659
2000-01-02 0.168731 0.831269 1.338144 2.338144 1.279321 -0.279321
2000-01-03 1.621034 -0.621034 0.438107 1.438107 0.903794 1.903794
2000-01-04 NaN NaN NaN NaN NaN NaN
2000-01-05 NaN NaN NaN NaN NaN NaN
2000-01-06 NaN NaN NaN NaN NaN NaN
2000-01-07 NaN NaN NaN NaN NaN NaN
2000-01-08 0.254374 1.254374 1.240447 -0.240447 0.201052 0.798948
2000-01-09 0.157795 0.842205 0.791197 1.791197 1.144209 -0.144209
2000-01-10 0.030876 0.969124 0.371900 1.371900 0.061932 1.061932
将多个函数传递给Series将生成一个DataFrame。结果的列名将是转换函数。
In [191]: tsdf["A"].transform([np.abs, lambda x: x + 1])
Out[191]:
absolute <lambda>
2000-01-01 0.428759 0.571241
2000-01-02 0.168731 0.831269
2000-01-03 1.621034 -0.621034
2000-01-04 NaN NaN
2000-01-05 NaN NaN
2000-01-06 NaN NaN
2000-01-07 NaN NaN
2000-01-08 0.254374 1.254374
2000-01-09 0.157795 0.842205
2000-01-10 0.030876 0.969124
使用字典进行转换#
传递函数字典将允许按列进行选择性转换。
In [192]: tsdf.transform({"A": np.abs, "B": lambda x: x + 1})
Out[192]:
A B
2000-01-01 0.428759 0.135110
2000-01-02 0.168731 2.338144
2000-01-03 1.621034 1.438107
2000-01-04 NaN NaN
2000-01-05 NaN NaN
2000-01-06 NaN NaN
2000-01-07 NaN NaN
2000-01-08 0.254374 -0.240447
2000-01-09 0.157795 1.791197
2000-01-10 0.030876 1.371900
传递列表字典将生成一个具有这些选择性转换的多重索引DataFrame。
In [193]: tsdf.transform({"A": np.abs, "B": [lambda x: x + 1, "sqrt"]})
Out[193]:
A B
absolute <lambda> sqrt
2000-01-01 0.428759 0.135110 NaN
2000-01-02 0.168731 2.338144 1.156782
2000-01-03 1.621034 1.438107 0.661897
2000-01-04 NaN NaN NaN
2000-01-05 NaN NaN NaN
2000-01-06 NaN NaN NaN
2000-01-07 NaN NaN NaN
2000-01-08 0.254374 -0.240447 NaN
2000-01-09 0.157795 1.791197 0.889493
2000-01-10 0.030876 1.371900 0.609836
应用逐元素函数#
由于并非所有函数都可以向量化（接受NumPy数组并返回另一个数组或值），因此DataFrame上的map()方法和Series上类似的map()方法接受任何接受单个值并返回单个值的Python函数。例如：
In [194]: df4 = df.copy()
In [195]: df4
Out[195]:
one two three
a 1.394981 1.772517 NaN
b 0.343054 1.912123 -0.050390
c 0.695246 1.478369 1.227435
d NaN 0.279344 -0.613172
In [196]: def f(x):
.....: return len(str(x))
.....:
In [197]: df4["one"].map(f)
Out[197]:
a 18
b 19
c 18
d 3
Name: one, dtype: int64
In [198]: df4.map(f)
Out[198]:
one two three
a 18 17 3
b 19 18 20
c 18 18 16
d 3 19 19
Series.map()有一个额外功能；它可以用来轻松地“链接”或“映射”由第二个Series定义的值。这与合并/连接功能密切相关。
In [199]: s = pd.Series(
.....: ["six", "seven", "six", "seven", "six"], index=["a", "b", "c", "d", "e"]
.....: )
.....:
In [200]: t = pd.Series({"six": 6.0, "seven": 7.0})
In [201]: s
Out[201]:
a six
b seven
c six
d seven
e six
dtype: str
In [202]: s.map(t)
Out[202]:
a 6.0
b 7.0
c 6.0
d 7.0
e 6.0
dtype: float64
重新索引和更改标签#
reindex()是pandas中基本的数据对齐方法。它用于实现几乎所有依赖于标签对齐功能的其他特性。*重新索引*意味着使数据与特定轴上的一组给定标签匹配。这实现了几个目的：
重新排序现有数据以匹配新的标签集。
在不存在该标签数据的标签位置插入缺失值（NA）标记。
如果指定，则使用逻辑（与处理时间序列数据高度相关）**填充**缺失标签的数据。
下面是一个简单的例子。
In [203]: s = pd.Series(np.random.randn(5), index=["a", "b", "c", "d", "e"])
In [204]: s
Out[204]:
a 1.695148
b 1.328614
c 1.234686
d -0.385845
e -1.326508
dtype: float64
In [205]: s.reindex(["e", "b", "f", "d"])
Out[205]:
e -1.326508
b 1.328614
f NaN
d -0.385845
dtype: float64
在这里，f标签未包含在Series中，因此在结果中显示为NaN。
使用DataFrame时，您可以同时重新索引索引和列。
In [206]: df
Out[206]:
one two three
a 1.394981 1.772517 NaN
b 0.343054 1.912123 -0.050390
c 0.695246 1.478369 1.227435
d NaN 0.279344 -0.613172
In [207]: df.reindex(index=["c", "f", "b"], columns=["three", "two", "one"])
Out[207]:
three two one
c 1.227435 1.478369 0.695246
f NaN NaN NaN
b -0.050390 1.912123 0.343054
请注意，包含实际轴标签的Index对象可以在对象之间**共享**。因此，如果我们有一个Series和一个DataFrame，可以这样做：
In [208]: rs = s.reindex(df.index)
In [209]: rs
Out[209]:
a 1.695148
b 1.328614
c 1.234686
d -0.385845
dtype: float64
In [210]: rs.index is df.index
Out[210]: True
这意味着重新索引的Series的索引与DataFrame的索引是同一个Python对象。
DataFrame.reindex()还支持“轴样式”调用约定，在这种约定中，您指定一个labels参数以及它适用的axis。
In [211]: df.reindex(["c", "f", "b"], axis="index")
Out[211]:
one two three
c 0.695246 1.478369 1.227435
f NaN NaN NaN
b 0.343054 1.912123 -0.050390
In [212]: df.reindex(["three", "two", "one"], axis="columns")
Out[212]:
three two one
a NaN 1.772517 1.394981
b -0.050390 1.912123 0.343054
c 1.227435 1.478369 0.695246
d -0.613172 0.279344 NaN
另请参阅
MultiIndex / Advanced Indexing是进行重新索引的更简洁的方式。
注意
在编写性能敏感的代码时，花一些时间成为重新索引的忍者是有好处的：**在预对齐的数据上执行许多操作更快**。添加两个未对齐的DataFrame在内部会触发一个重新索引步骤。对于探索性分析，您几乎注意不到差异（因为reindex已经经过了高度优化），但当CPU周期很重要时，在这里和那里添加一些显式的reindex调用可能会产生影响。
重新索引以与另一个对象对齐#
您可能希望获取一个对象并将其轴重新索引为与另一个对象相同的标签。虽然这方面的语法直接但冗长，但这是一个足够常见的操作，因此提供了reindex_like()方法来简化此操作。
In [213]: df2 = df.reindex(["a", "b", "c"], columns=["one", "two"])
In [214]: df3 = df2 - df2.mean()
In [215]: df2
Out[215]:
one two
a 1.394981 1.772517
b 0.343054 1.912123
c 0.695246 1.478369
In [216]: df3
Out[216]:
one two
a 0.583888 0.051514
b -0.468040 0.191120
c -0.115848 -0.242634
In [217]: df.reindex_like(df2)
Out[217]:
one two
a 1.394981 1.772517
b 0.343054 1.912123
c 0.695246 1.478369
使用align将对象彼此对齐#
该align()方法是同时对齐两个对象的最快方法。它支持一个join参数（与连接和合并相关）。
join='outer'：取索引的并集（默认）。
join='left'：使用调用对象的索引。
join='right'：使用传递对象的索引。
join='inner'：取索引的交集。
它返回包含两个重新索引的Series的元组。
In [218]: s = pd.Series(np.random.randn(5), index=["a", "b", "c", "d", "e"])
In [219]: s1 = s[:4]
In [220]: s2 = s[1:]
In [221]: s1.align(s2)
Out[221]:
(a -0.186646
b -1.692424
c -0.303893
d -1.425662
e NaN
dtype: float64,
a NaN
b -1.692424
c -0.303893
d -1.425662
e 1.114285
dtype: float64)
In [222]: s1.align(s2, join="inner")
Out[222]:
(b -1.692424
c -0.303893
d -1.425662
dtype: float64,
b -1.692424
c -0.303893
d -1.425662
dtype: float64)
In [223]: s1.align(s2, join="left")
Out[223]:
(a -0.186646
b -1.692424
c -0.303893
d -1.425662
dtype: float64,
a NaN
b -1.692424
c -0.303893
d -1.425662
dtype: float64)
对于DataFrames，join方法默认将应用于索引和列。
In [224]: df.align(df2, join="inner")
Out[224]:
( one two
a 1.394981 1.772517
b 0.343054 1.912123
c 0.695246 1.478369,
one two
a 1.394981 1.772517
b 0.343054 1.912123
c 0.695246 1.478369)
您还可以传递一个axis选项以仅在指定轴上对齐。
In [225]: df.align(df2, join="inner", axis=0)
Out[225]:
( one two three
a 1.394981 1.772517 NaN
b 0.343054 1.912123 -0.050390
c 0.695246 1.478369 1.227435,
one two
a 1.394981 1.772517
b 0.343054 1.912123
c 0.695246 1.478369)
如果您将Series传递给DataFrame.align()，您可以使用axis参数选择在DataFrame的索引或列上对齐两个对象。
In [226]: df.align(df2.iloc[0], axis=1)
Out[226]:
( one three two
a 1.394981 NaN 1.772517
b 0.343054 -0.050390 1.912123
c 0.695246 1.227435 1.478369
d NaN -0.613172 0.279344,
one 1.394981
three NaN
two 1.772517
Name: a, dtype: float64)
重新索引时填充#
reindex()接受一个可选参数method，这是一个填充方法，从下表中选择：
方法 |
Action |
|---|---|
ffill |
向前填充值 |
bfill |
向后填充值 |
nearest |
从最近的索引值填充。 |
我们在一个简单的Series上说明这些填充方法。
In [227]: rng = pd.date_range("1/3/2000", periods=8)
In [228]: ts = pd.Series(np.random.randn(8), index=rng)
In [229]: ts2 = ts.iloc[[0, 3, 6]]
In [230]: ts
Out[230]:
2000-01-03 0.183051
2000-01-04 0.400528
2000-01-05 -0.015083
2000-01-06 2.395489
2000-01-07 1.414806
2000-01-08 0.118428
2000-01-09 0.733639
2000-01-10 -0.936077
Freq: D, dtype: float64
In [231]: ts2
Out[231]:
2000-01-03 0.183051
2000-01-06 2.395489
2000-01-09 0.733639
Freq: 3D, dtype: float64
In [232]: ts2.reindex(ts.index)
Out[232]:
2000-01-03 0.183051
2000-01-04 NaN
2000-01-05 NaN
2000-01-06 2.395489
2000-01-07 NaN
2000-01-08 NaN
2000-01-09 0.733639
2000-01-10 NaN
Freq: D, dtype: float64
In [233]: ts2.reindex(ts.index, method="ffill")
Out[233]:
2000-01-03 0.183051
2000-01-04 0.183051
2000-01-05 0.183051
2000-01-06 2.395489
2000-01-07 2.395489
2000-01-08 2.395489
2000-01-09 0.733639
2000-01-10 0.733639
Freq: D, dtype: float64
In [234]: ts2.reindex(ts.index, method="bfill")
Out[234]:
2000-01-03 0.183051
2000-01-04 2.395489
2000-01-05 2.395489
2000-01-06 2.395489
2000-01-07 0.733639
2000-01-08 0.733639
2000-01-09 0.733639
2000-01-10 NaN
Freq: D, dtype: float64
In [235]: ts2.reindex(ts.index, method="nearest")
Out[235]:
2000-01-03 0.183051
2000-01-04 0.183051
2000-01-05 2.395489
2000-01-06 2.395489
2000-01-07 2.395489
2000-01-08 0.733639
2000-01-09 0.733639
2000-01-10 0.733639
Freq: D, dtype: float64
这些方法要求索引是**有序**的递增或递减。
请注意，使用ffill（method='nearest'除外）或interpolate也可以达到相同的效果。
In [236]: ts2.reindex(ts.index).ffill()
Out[236]:
2000-01-03 0.183051
2000-01-04 0.183051
2000-01-05 0.183051
2000-01-06 2.395489
2000-01-07 2.395489
2000-01-08 2.395489
2000-01-09 0.733639
2000-01-10 0.733639
Freq: D, dtype: float64
reindex()如果索引不是单调递增或递减的，将引发ValueError。fillna()和interpolate()不会执行任何索引顺序检查。
重新索引时填充的限制#
limit和tolerance参数在重新索引时填充时提供了额外的控制。Limit指定连续匹配的最大计数。
In [237]: ts2.reindex(ts.index, method="ffill", limit=1)
Out[237]:
2000-01-03 0.183051
2000-01-04 0.183051
2000-01-05 NaN
2000-01-06 2.395489
2000-01-07 2.395489
2000-01-08 NaN
2000-01-09 0.733639
2000-01-10 0.733639
Freq: D, dtype: float64
相比之下，tolerance指定索引和索引值之间的最大距离。
In [238]: ts2.reindex(ts.index, method="ffill", tolerance="1 day")
Out[238]:
2000-01-03 0.183051
2000-01-04 0.183051
2000-01-05 NaN
2000-01-06 2.395489
2000-01-07 2.395489
2000-01-08 NaN
2000-01-09 0.733639
2000-01-10 0.733639
Freq: D, dtype: float64
请注意，当在DatetimeIndex、TimedeltaIndex或PeriodIndex上使用时，如果可能，tolerance将被强制转换为Timedelta。这允许您使用适当的字符串指定容差。
从轴删除标签#
与reindex密切相关的方法是drop()函数。它从轴中删除一组标签。
In [239]: df
Out[239]:
one two three
a 1.394981 1.772517 NaN
b 0.343054 1.912123 -0.050390
c 0.695246 1.478369 1.227435
d NaN 0.279344 -0.613172
In [240]: df.drop(["a", "d"], axis=0)
Out[240]:
one two three
b 0.343054 1.912123 -0.050390
c 0.695246 1.478369 1.227435
In [241]: df.drop(["one"], axis=1)
Out[241]:
two three
a 1.772517 NaN
b 1.912123 -0.050390
c 1.478369 1.227435
d 0.279344 -0.613172
请注意，以下操作也有效，但不太明显/清晰：
In [242]: df.reindex(df.index.difference(["a", "d"]))
Out[242]:
one two three
b 0.343054 1.912123 -0.050390
c 0.695246 1.478369 1.227435
重命名/映射标签#
该rename()方法允许您基于某些映射（字典或Series）或任意函数来重命名轴。
In [243]: s
Out[243]:
a -0.186646
b -1.692424
c -0.303893
d -1.425662
e 1.114285
dtype: float64
In [244]: s.rename(str.upper)
Out[244]:
A -0.186646
B -1.692424
C -0.303893
D -1.425662
E 1.114285
dtype: float64
如果您传递一个函数，它在调用任何标签时都必须返回一个值（并且必须产生一组唯一值）。也可以使用字典或Series。
In [245]: df.rename(
.....: columns={"one": "foo", "two": "bar"},
.....: index={"a": "apple", "b": "banana", "d": "durian"},
.....: )
.....:
Out[245]:
foo bar three
apple 1.394981 1.772517 NaN
banana 0.343054 1.912123 -0.050390
c 0.695246 1.478369 1.227435
durian NaN 0.279344 -0.613172
如果映射不包含列/索引标签，则不会重命名。请注意，映射中多余的标签不会引发错误。
DataFrame.rename()还支持“轴样式”调用约定，在这种约定中，您指定一个mapper以及要应用该映射的axis。
In [246]: df.rename({"one": "foo", "two": "bar"}, axis="columns")
Out[246]:
foo bar three
a 1.394981 1.772517 NaN
b 0.343054 1.912123 -0.050390
c 0.695246 1.478369 1.227435
d NaN 0.279344 -0.613172
In [247]: df.rename({"a": "apple", "b": "banana", "d": "durian"}, axis="index")
Out[247]:
one two three
apple 1.394981 1.772517 NaN
banana 0.343054 1.912123 -0.050390
c 0.695246 1.478369 1.227435
durian NaN 0.279344 -0.613172
最后，rename()还接受标量或列表类来修改Series.name属性。
In [248]: s.rename("scalar-name")
Out[248]:
a -0.186646
b -1.692424
c -0.303893
d -1.425662
e 1.114285
Name: scalar-name, dtype: float64
方法DataFrame.rename_axis()和Series.rename_axis()允许更改MultiIndex的特定名称（而不是标签）。
In [249]: df = pd.DataFrame(
.....: {"x": [1, 2, 3, 4, 5, 6], "y": [10, 20, 30, 40, 50, 60]},
.....: index=pd.MultiIndex.from_product(
.....: [["a", "b", "c"], [1, 2]], names=["let", "num"]
.....: ),
.....: )
.....:
In [250]: df
Out[250]:
x y
let num
a 1 1 10
2 2 20
b 1 3 30
2 4 40
c 1 5 50
2 6 60
In [251]: df.rename_axis(index={"let": "abc"})
Out[251]:
x y
abc num
a 1 1 10
2 2 20
b 1 3 30
2 4 40
c 1 5 50
2 6 60
In [252]: df.rename_axis(index=str.upper)
Out[252]:
x y
LET NUM
a 1 1 10
2 2 20
b 1 3 30
2 4 40
c 1 5 50
2 6 60
迭代#
pandas对象的基本迭代行为取决于其类型。迭代Series时，它被视为类数组，基本迭代生成值。DataFrames遵循字典式的约定，即迭代对象的“键”。
简而言之，基本迭代（for i in object）产生：
Series：值
DataFrame：列标签
因此，例如，迭代DataFrame会得到列名。
In [253]: df = pd.DataFrame(
.....: {"col1": np.random.randn(3), "col2": np.random.randn(3)}, index=["a", "b", "c"]
.....: )
.....:
In [254]: for col in df:
.....: print(col)
.....:
col1
col2
pandas对象还具有字典式的items()方法，用于迭代（键，值）对。
要迭代DataFrame的行，可以使用以下方法：
iterrows()：迭代DataFrame的行，作为（索引，Series）对。这会将行转换为Series对象，这可能会更改dtype，并具有一些性能影响。itertuples()：迭代DataFrame的行，作为值的命名元组。这比iterrows()快得多，并且在大多数情况下比迭代DataFrame的值更可取。
警告
迭代pandas对象通常**速度很慢**。在许多情况下，手动迭代行是不必要的，可以通过以下方法之一避免：
警告
您**切勿修改**您正在迭代的内容。这不能保证在所有情况下都有效。根据数据类型，迭代器返回的是副本而不是视图，对其进行写入将无效！
例如，在以下情况下，设置值无效：
In [255]: df = pd.DataFrame({"a": [1, 2, 3], "b": ["a", "b", "c"]})
In [256]: for index, row in df.iterrows():
.....: row["a"] = 10
.....:
In [257]: df
Out[257]:
a b
0 1 a
1 2 b
2 3 c
items#
与字典式接口一致，items()迭代键值对。
Series：（索引，标量值）对。
DataFrame：（列，Series）对。
例如：
In [258]: for label, ser in df.items():
.....: print(label)
.....: print(ser)
.....:
a
0 1
1 2
2 3
Name: a, dtype: int64
b
0 a
1 b
2 c
Name: b, dtype: str
iterrows#
iterrows()允许您将DataFrame的行作为Series对象进行迭代。它返回一个迭代器，产生每个索引值以及包含每行数据的Series。
In [259]: for row_index, row in df.iterrows():
.....: print(row_index, row, sep="\n")
.....:
0
a 1
b a
Name: 0, dtype: object
1
a 2
b b
Name: 1, dtype: object
2
a 3
b c
Name: 2, dtype: object
注意
因为iterrows()为每行返回一个Series，所以它**不**能跨行保留dtype（DataFrame的dtype是跨列保留的）。例如：
In [260]: df_orig = pd.DataFrame([[1, 1.5]], columns=["int", "float"])
In [261]: df_orig.dtypes
Out[261]:
int int64
float float64
dtype: object
In [262]: row = next(df_orig.iterrows())[1]
In [263]: row
Out[263]:
int 1.0
float 1.5
Name: 0, dtype: float64
在row中返回的所有值，现在都被提升为浮点数，包括列x中的原始整数值。
In [264]: row["int"].dtype
Out[264]: dtype('float64')
In [265]: df_orig["int"].dtype
Out[265]: dtype('int64')
为了在迭代行时保留dtype，最好使用itertuples()，它返回值的命名元组，并且通常比iterrows()快得多。
例如，一种人为地转置DataFrame的方法是：
In [266]: df2 = pd.DataFrame({"x": [1, 2, 3], "y": [4, 5, 6]})
In [267]: print(df2)
x y
0 1 4
1 2 5
2 3 6
In [268]: print(df2.T)
0 1 2
x 1 2 3
y 4 5 6
In [269]: df2_t = pd.DataFrame({idx: values for idx, values in df2.iterrows()})
In [270]: print(df2_t)
0 1 2
x 1 2 3
y 4 5 6
itertuples#
该itertuples()方法将返回一个迭代器，为DataFrame中的每一行生成一个命名元组。元组的第一个元素是行的相应索引值，其余值是行值。
例如：
In [271]: for row in df.itertuples():
.....: print(row)
.....:
Pandas(Index=0, a=1, b='a')
Pandas(Index=1, a=2, b='b')
Pandas(Index=2, a=3, b='c')
此方法不会将行转换为Series对象；它只是将值返回到一个命名元组中。因此，itertuples()保留了值的dtype，并且通常比iterrows()快。
注意
列名将在无效Python标识符、重复或以下划线开头时重命名为位置名称。当列数很多（>255）时，将返回常规元组。
.dt 访问器#
如果Series是datetime/period类型，Series有一个访问器可以简洁地返回datetime类属性（针对Series的*值*）。这将返回一个Series，其索引与现有Series相同。
# datetime
In [272]: s = pd.Series(pd.date_range("20130101 09:10:12", periods=4))
In [273]: s
Out[273]:
0 2013-01-01 09:10:12
1 2013-01-02 09:10:12
2 2013-01-03 09:10:12
3 2013-01-04 09:10:12
dtype: datetime64[us]
In [274]: s.dt.hour
Out[274]:
0 9
1 9
2 9
3 9
dtype: int32
In [275]: s.dt.second
Out[275]:
0 12
1 12
2 12
3 12
dtype: int32
In [276]: s.dt.day
Out[276]:
0 1
1 2
2 3
3 4
dtype: int32
这使得表达式可以这样写：
In [277]: s[s.dt.day == 2]
Out[277]:
1 2013-01-02 09:10:12
dtype: datetime64[us]
您可以轻松地进行时区感知的转换。
In [278]: stz = s.dt.tz_localize("US/Eastern")
In [279]: stz
Out[279]:
0 2013-01-01 09:10:12-05:00
1 2013-01-02 09:10:12-05:00
2 2013-01-03 09:10:12-05:00
3 2013-01-04 09:10:12-05:00
dtype: datetime64[us, US/Eastern]
In [280]: stz.dt.tz
Out[280]: zoneinfo.ZoneInfo(key='US/Eastern')
您还可以链接这些类型的操作。
In [281]: s.dt.tz_localize("UTC").dt.tz_convert("US/Eastern")
Out[281]:
0 2013-01-01 04:10:12-05:00
1 2013-01-02 04:10:12-05:00
2 2013-01-03 04:10:12-05:00
3 2013-01-04 04:10:12-05:00
dtype: datetime64[us, US/Eastern]
您还可以使用Series.dt.strftime()将datetime值格式化为字符串，该方法支持与标准strftime()相同的格式。
# DatetimeIndex
In [282]: s = pd.Series(pd.date_range("20130101", periods=4))
In [283]: s
Out[283]:
0 2013-01-01
1 2013-01-02
2 2013-01-03
3 2013-01-04
dtype: datetime64[us]
In [284]: s.dt.strftime("%Y/%m/%d")
Out[284]:
0 2013/01/01
1 2013/01/02
2 2013/01/03
3 2013/01/04
dtype: str
# PeriodIndex
In [285]: s = pd.Series(pd.period_range("20130101", periods=4))
In [286]: s
Out[286]:
0 2013-01-01
1 2013-01-02
2 2013-01-03
3 2013-01-04
dtype: period[D]
In [287]: s.dt.strftime("%Y/%m/%d")
Out[287]:
0 2013/01/01
1 2013/01/02
2 2013/01/03
3 2013/01/04
dtype: str
该.dt访问器适用于period和timedelta dtypes。
# period
In [288]: s = pd.Series(pd.period_range("20130101", periods=4, freq="D"))
In [289]: s
Out[289]:
0 2013-01-01
1 2013-01-02
2 2013-01-03
3 2013-01-04
dtype: period[D]
In [290]: s.dt.year
Out[290]:
0 2013
1 2013
2 2013
3 2013
dtype: int64
In [291]: s.dt.day
Out[291]:
0 1
1 2
2 3
3 4
dtype: int64
# timedelta
In [292]: s = pd.Series(pd.timedelta_range("1 day 00:00:05", periods=4, freq="s"))
In [293]: s
Out[293]:
0 1 days 00:00:05
1 1 days 00:00:06
2 1 days 00:00:07
3 1 days 00:00:08
dtype: timedelta64[us]
In [294]: s.dt.days
Out[294]:
0 1
1 1
2 1
3 1
dtype: int64
In [295]: s.dt.seconds
Out[295]:
0 5
1 6
2 7
3 8
dtype: int32
In [296]: s.dt.components
Out[296]:
days hours minutes seconds milliseconds microseconds nanoseconds
0 1 0 0 5 0 0 0
1 1 0 0 6 0 0 0
2 1 0 0 7 0 0 0
3 1 0 0 8 0 0 0
注意
如果使用非datetime类值访问Series.dt，它将引发TypeError。
向量化字符串方法#
Series配备了一系列字符串处理方法，可以轻松地对数组的每个元素进行操作。最重要的是，这些方法会自动排除缺失/NA值。它们通过Series的str属性访问，并且通常具有与等效的（标量）内置字符串方法相匹配的名称。例如：
In [297]: s = pd.Series( .....: ["A", "B", "C", "Aaba", "Baca", np.nan, "CABA", "dog", "cat"], dtype="string" .....: ) .....: In [298]: s.str.lower() Out[298]: 0 a 1 b 2 c 3 aaba 4 baca 5 <NA> 6 caba 7 dog 8 cat dtype: string
还提供了强大的模式匹配方法，但请注意，模式匹配默认使用正则表达式（并且在某些情况下始终使用它们）。
注意
在pandas 1.0之前，字符串方法仅在object dtype的Series上可用。pandas 1.0添加了专用于字符串的StringDtype。更多信息请参阅文本数据类型。
请参阅Vectorized String Methods以获取完整描述。
排序#
pandas支持三种排序方式：按索引标签排序、按列值排序以及按两者的组合排序。
按索引#
Series.sort_index()和DataFrame.sort_index()方法用于按其索引级别对pandas对象进行排序。
In [299]: df = pd.DataFrame(
.....: {
.....: "one": pd.Series(np.random.randn(3), index=["a", "b", "c"]),
.....: "two": pd.Series(np.random.randn(4), index=["a", "b", "c", "d"]),
.....: "three": pd.Series(np.random.randn(3), index=["b", "c", "d"]),
.....: }
.....: )
.....:
In [300]: unsorted_df = df.reindex(
.....: index=["a", "d", "c", "b"], columns=["three", "two", "one"]
.....: )
.....:
In [301]: unsorted_df
Out[301]:
three two one
a NaN -1.152244 0.562973
d -0.252916 -0.109597 NaN
c 1.273388 -0.167123 0.640382
b -0.098217 0.009797 -1.299504
# DataFrame
In [302]: unsorted_df.sort_index()
Out[302]:
three two one
a NaN -1.152244 0.562973
b -0.098217 0.009797 -1.299504
c 1.273388 -0.167123 0.640382
d -0.252916 -0.109597 NaN
In [303]: unsorted_df.sort_index(ascending=False)
Out[303]:
three two one
d -0.252916 -0.109597 NaN
c 1.273388 -0.167123 0.640382
b -0.098217 0.009797 -1.299504
a NaN -1.152244 0.562973
In [304]: unsorted_df.sort_index(axis=1)
Out[304]:
one three two
a 0.562973 NaN -1.152244
d NaN -0.252916 -0.109597
c 0.640382 1.273388 -0.167123
b -1.299504 -0.098217 0.009797
# Series
In [305]: unsorted_df["three"].sort_index()
Out[305]:
a NaN
b -0.098217
c 1.273388
d -0.252916
Name: three, dtype: float64
按索引排序还支持一个key参数，该参数接受一个可调用函数，应用于正在排序的索引。对于MultiIndex对象，key按级别应用于由level指定的级别。
In [306]: s1 = pd.DataFrame({"a": ["B", "a", "C"], "b": [1, 2, 3], "c": [2, 3, 4]}).set_index(
.....: list("ab")
.....: )
.....:
In [307]: s1
Out[307]:
c
a b
B 1 2
a 2 3
C 3 4
In [308]: s1.sort_index(level="a")
Out[308]:
c
a b
B 1 2
C 3 4
a 2 3
In [309]: s1.sort_index(level="a", key=lambda idx: idx.str.lower())
Out[309]:
c
a b
a 2 3
B 1 2
C 3 4
有关按值进行键排序的信息，请参阅值排序。
按值#
Series.sort_values()方法用于按其值对Series进行排序。DataFrame.sort_values()方法用于按其列或行值对DataFrame进行排序。DataFrame.sort_values()的可选by参数可用于指定一个或多个用于确定排序顺序的列。
In [310]: df1 = pd.DataFrame(
.....: {"one": [2, 1, 1, 1], "two": [1, 3, 2, 4], "three": [5, 4, 3, 2]}
.....: )
.....:
In [311]: df1.sort_values(by="two")
Out[311]:
one two three
0 2 1 5
2 1 2 3
1 1 3 4
3 1 4 2
by参数可以接受一个列名列表，例如：
In [312]: df1[["one", "two", "three"]].sort_values(by=["one", "two"])
Out[312]:
one two three
2 1 2 3
1 1 3 4
3 1 4 2
0 2 1 5
这些方法通过na_position参数对NA值进行特殊处理。
In [313]: s[2] = np.nan
In [314]: s.sort_values()
Out[314]:
0 A
3 Aaba
1 B
4 Baca
6 CABA
8 cat
7 dog
2 <NA>
5 <NA>
dtype: string
In [315]: s.sort_values(na_position="first")
Out[315]:
2 <NA>
5 <NA>
0 A
3 Aaba
1 B
4 Baca
6 CABA
8 cat
7 dog
dtype: string
排序还支持一个key参数，该参数接受一个可调用函数，应用于正在排序的值。
In [316]: s1 = pd.Series(["B", "a", "C"])
In [317]: s1.sort_values()
Out[317]:
0 B
2 C
1 a
dtype: str
In [318]: s1.sort_values(key=lambda x: x.str.lower())
Out[318]:
1 a
0 B
2 C
dtype: str
key将接收Series值，并应返回一个形状相同的Series或数组，其中包含转换后的值。对于DataFrame对象，key按列应用，因此key仍应接收Series并返回Series，例如：
In [319]: df = pd.DataFrame({"a": ["B", "a", "C"], "b": [1, 2, 3]})
In [320]: df.sort_values(by="a")
Out[320]:
a b
0 B 1
2 C 3
1 a 2
In [321]: df.sort_values(by="a", key=lambda col: col.str.lower())
Out[321]:
a b
1 a 2
0 B 1
2 C 3
每列的名称或类型可用于将不同的函数应用于不同的列。
按索引和值#
传递给DataFrame.sort_values()的by参数中的字符串可以指代列名或索引级别名称。
# Build MultiIndex
In [322]: idx = pd.MultiIndex.from_tuples(
.....: [("a", 1), ("a", 2), ("a", 2), ("b", 2), ("b", 1), ("b", 1)]
.....: )
.....:
In [323]: idx.names = ["first", "second"]
# Build DataFrame
In [324]: df_multi = pd.DataFrame({"A": np.arange(6, 0, -1)}, index=idx)
In [325]: df_multi
Out[325]:
A
first second
a 1 6
2 5
2 4
b 2 3
1 2
1 1
按“second”（索引）和“A”（列）排序。
In [326]: df_multi.sort_values(by=["second", "A"])
Out[326]:
A
first second
b 1 1
1 2
a 1 6
b 2 3
a 2 4
2 5
注意
如果一个字符串同时匹配列名和索引级别名称，则会发出警告，并且列将优先。这将在未来版本中导致歧义错误。
searchsorted#
Series 具有 searchsorted() 方法，其功能类似于 numpy.ndarray.searchsorted()。
In [327]: ser = pd.Series([1, 2, 3])
In [328]: ser.searchsorted([0, 3])
Out[328]: array([0, 2])
In [329]: ser.searchsorted([0, 4])
Out[329]: array([0, 3])
In [330]: ser.searchsorted([1, 3], side="right")
Out[330]: array([1, 3])
In [331]: ser.searchsorted([1, 3], side="left")
Out[331]: array([0, 2])
In [332]: ser = pd.Series([3, 1, 2])
In [333]: ser.searchsorted([0, 3], sorter=np.argsort(ser))
Out[333]: array([0, 2])
最小/最大值#
Series 具有 nsmallest() 和 nlargest() 方法，它们返回最小或最大的 \(n\) 个值。对于大型 Series，这可能比对整个 Series 进行排序然后调用 head(n) 快得多。
In [334]: s = pd.Series(np.random.permutation(10))
In [335]: s
Out[335]:
0 2
1 0
2 3
3 7
4 1
5 5
6 9
7 6
8 8
9 4
dtype: int64
In [336]: s.sort_values()
Out[336]:
1 0
4 1
0 2
2 3
9 4
5 5
7 6
3 7
8 8
6 9
dtype: int64
In [337]: s.nsmallest(3)
Out[337]:
1 0
4 1
0 2
dtype: int64
In [338]: s.nlargest(3)
Out[338]:
6 9
8 8
3 7
dtype: int64
DataFrame 也具有 nlargest 和 nsmallest 方法。
In [339]: df = pd.DataFrame(
.....: {
.....: "a": [-2, -1, 1, 10, 8, 11, -1],
.....: "b": list("abdceff"),
.....: "c": [1.0, 2.0, 4.0, 3.2, np.nan, 3.0, 4.0],
.....: }
.....: )
.....:
In [340]: df.nlargest(3, "a")
Out[340]:
a b c
5 11 f 3.0
3 10 c 3.2
4 8 e NaN
In [341]: df.nlargest(5, ["a", "c"])
Out[341]:
a b c
5 11 f 3.0
3 10 c 3.2
4 8 e NaN
2 1 d 4.0
6 -1 f 4.0
In [342]: df.nsmallest(3, "a")
Out[342]:
a b c
0 -2 a 1.0
1 -1 b 2.0
6 -1 f 4.0
In [343]: df.nsmallest(5, ["a", "c"])
Out[343]:
a b c
0 -2 a 1.0
1 -1 b 2.0
6 -1 f 4.0
2 1 d 4.0
4 8 e NaN
按 MultiIndex 列排序#
当列是 MultiIndex 时，您必须明确指定排序，并在 `by` 中完整指定所有级别。
In [344]: df1.columns = pd.MultiIndex.from_tuples(
.....: [("a", "one"), ("a", "two"), ("b", "three")]
.....: )
.....:
In [345]: df1.sort_values(by=("a", "two"))
Out[345]:
a b
one two three
0 2 1 5
2 1 2 3
1 1 3 4
3 1 4 2
复制#
pandas 对象上的 copy() 方法会复制底层数据（但不复制轴索引，因为它们是不可变的）并返回一个新对象。请注意，**很少有必要复制对象**。例如，只有少数几种方法可以*就地*修改 DataFrame
插入、删除或修改列。
赋值给
index或columns属性。对于同质数据，可以通过 `values` 属性或高级索引直接修改值。
明确地说，没有 pandas 方法会产生修改数据的副作用；几乎所有方法都会返回一个新对象，而原始对象保持不变。如果数据被修改，那是因为您明确地进行了修改。
数据类型#
在大多数情况下，pandas 会为 Series 或 DataFrame 的单个列使用 NumPy 数组和数据类型。NumPy 支持 `float`、`int`、`bool`、`timedelta64[ns]` 和 `datetime64[ns]`（请注意，NumPy 不支持带时区的日期时间）。
pandas 和第三方库在一些地方*扩展*了 NumPy 的类型系统。本节描述了 pandas 内部进行的扩展。有关如何编写自己的与 pandas 配合使用的扩展，请参阅 扩展类型。有关已实现扩展的第三方库列表，请参阅 生态系统页面。
下表列出了所有 pandas 扩展类型。对于需要 `dtype` 参数的方法，可以按照指示指定字符串。有关每种类型的更多信息，请参阅相应的文档部分。
数据种类 |
数据类型 |
标量 |
数组 |
字符串别名 |
||
|---|---|---|---|---|---|---|
|
||||||
(无) |
|
|||||
|
|
|||||
(无) |
|
|||||
|
||||||
|
(无) |
|
||||
|
(无) |
|
||||
|
||||||
|
||||||
pandas 有两种存储字符串的方法。
object数据类型，它可以容纳任何 Python 对象，包括字符串。StringDtype，它专门用于字符串。
通常，我们建议使用 StringDtype。有关更多信息，请参阅 文本数据类型。
最后，可以使用 `object` 数据类型存储任意对象，但应尽可能避免（为了性能以及与其他库和方法的互操作性。有关详细信息，请参阅 对象转换）。
DataFrame 的一个方便的 dtypes 属性会返回一个 Series，其中包含每列的数据类型。
In [346]: dft = pd.DataFrame(
.....: {
.....: "A": np.random.rand(3),
.....: "B": 1,
.....: "C": "foo",
.....: "D": pd.Timestamp("20010102"),
.....: "E": pd.Series([1.0] * 3).astype("float32"),
.....: "F": False,
.....: "G": pd.Series([1] * 3, dtype="int8"),
.....: }
.....: )
.....:
In [347]: dft
Out[347]:
A B C D E F G
0 0.035962 1 foo 2001-01-02 1.0 False 1
1 0.701379 1 foo 2001-01-02 1.0 False 1
2 0.281885 1 foo 2001-01-02 1.0 False 1
In [348]: dft.dtypes
Out[348]:
A float64
B int64
C str
D datetime64[us]
E float32
F bool
G int8
dtype: object
在 `Series` 对象上，使用 dtype 属性。
In [349]: dft["A"].dtype
Out[349]: dtype('float64')
如果 pandas 对象在*单个列中*包含多种数据类型的数据，则该列的数据类型将选择为能够容纳所有数据类型（`object` 是最通用的）。
# these ints are coerced to floats
In [350]: pd.Series([1, 2, 3, 4, 5, 6.0])
Out[350]:
0 1.0
1 2.0
2 3.0
3 4.0
4 5.0
5 6.0
dtype: float64
# string data forces an ``object`` dtype
In [351]: pd.Series([1, 2, 3, 6.0, "foo"])
Out[351]:
0 1
1 2
2 3
3 6.0
4 foo
dtype: object
可以通过调用 `DataFrame.dtypes.value_counts()` 来查找 DataFrame 中每种类型列的数量。
In [352]: dft.dtypes.value_counts()
Out[352]:
float64 1
int64 1
str 1
datetime64[us] 1
float32 1
bool 1
int8 1
Name: count, dtype: int64
数值数据类型将传播并可以共存于 DataFrame 中。如果传递了数据类型（直接通过 `dtype` 关键字，或传递的 `ndarray`，或传递的 `Series`），则它将在 DataFrame 操作中保留。此外，不同的数值数据类型*不会*合并。以下示例将为您提供一些启发。
In [353]: df1 = pd.DataFrame(np.random.randn(8, 1), columns=["A"], dtype="float64")
In [354]: df1
Out[354]:
A
0 0.224364
1 1.890546
2 0.182879
3 0.787847
4 -0.188449
5 0.667715
6 -0.011736
7 -0.399073
In [355]: df1.dtypes
Out[355]:
A float64
dtype: object
In [356]: df2 = pd.DataFrame(
.....: {
.....: "A": pd.Series(np.random.randn(8), dtype="float32"),
.....: "B": pd.Series(np.random.randn(8)),
.....: "C": pd.Series(np.random.randint(0, 255, size=8), dtype="uint8"), # [0,255] (range of uint8)
.....: }
.....: )
.....:
In [357]: df2
Out[357]:
A B C
0 0.823270 0.256090 26
1 1.607861 1.426469 86
2 -0.333669 -0.416203 46
3 -0.063467 1.139976 212
4 -1.014637 -1.193477 26
5 0.678758 0.096706 7
6 -0.040865 -1.956850 184
7 -0.357346 -0.714337 206
In [358]: df2.dtypes
Out[358]:
A float32
B float64
C uint8
dtype: object
默认值#
默认情况下，整数类型为 `int64`，浮点数类型为 `float64`，*无论*平台如何（32 位或 64 位）。以下所有操作都将产生 `int64` 数据类型。
In [359]: pd.DataFrame([1, 2], columns=["a"]).dtypes
Out[359]:
a int64
dtype: object
In [360]: pd.DataFrame({"a": [1, 2]}).dtypes
Out[360]:
a int64
dtype: object
In [361]: pd.DataFrame({"a": 1}, index=list(range(2))).dtypes
Out[361]:
a int64
dtype: object
请注意，NumPy 在创建数组时会选择*平台相关的*类型。以下操作在 32 位平台上*将*产生 `int32`。
In [362]: frame = pd.DataFrame(np.array([1, 2]))
向上转型#
当与其他类型组合时，类型可能会被*向上转型*，这意味着它们会从当前类型（例如 `int` 变为 `float`）提升。
In [363]: df3 = df1.reindex_like(df2).fillna(value=0.0) + df2
In [364]: df3
Out[364]:
A B C
0 1.047634 0.256090 26.0
1 3.498407 1.426469 86.0
2 -0.150790 -0.416203 46.0
3 0.724380 1.139976 212.0
4 -1.203087 -1.193477 26.0
5 1.346474 0.096706 7.0
6 -0.052601 -1.956850 184.0
7 -0.756419 -0.714337 206.0
In [365]: df3.dtypes
Out[365]:
A float64
B float64
C float64
dtype: object
DataFrame.to_numpy() 将返回数据类型的*低共同分母*，即能够容纳结果同质数据类型 NumPy 数组中*所有*类型的类型。这可能会强制进行一些*向上转型*。
In [366]: df3.to_numpy().dtype
Out[366]: dtype('float64')
astype#
您可以使用 `astype()` 方法将数据类型显式地从一种转换为另一种。默认情况下，这些将返回一个副本，即使数据类型未更改（传递 `copy=False` 可以更改此行为）。此外，如果 `astype` 操作无效，它们将引发异常。
向上转型始终遵循 **NumPy** 规则。如果操作涉及两种不同的数据类型，则在该操作的结果中使用更*通用*的类型。
In [367]: df3
Out[367]:
A B C
0 1.047634 0.256090 26.0
1 3.498407 1.426469 86.0
2 -0.150790 -0.416203 46.0
3 0.724380 1.139976 212.0
4 -1.203087 -1.193477 26.0
5 1.346474 0.096706 7.0
6 -0.052601 -1.956850 184.0
7 -0.756419 -0.714337 206.0
In [368]: df3.dtypes
Out[368]:
A float64
B float64
C float64
dtype: object
# conversion of dtypes
In [369]: df3.astype("float32").dtypes
Out[369]:
A float32
B float32
C float32
dtype: object
使用 astype() 将列的子集转换为指定类型。
In [370]: dft = pd.DataFrame({"a": [1, 2, 3], "b": [4, 5, 6], "c": [7, 8, 9]})
In [371]: dft[["a", "b"]] = dft[["a", "b"]].astype(np.uint8)
In [372]: dft
Out[372]:
a b c
0 1 4 7
1 2 5 8
2 3 6 9
In [373]: dft.dtypes
Out[373]:
a uint8
b uint8
c int64
dtype: object
通过将字典传递给 astype() 将某些列转换为特定数据类型。
In [374]: dft1 = pd.DataFrame({"a": [1, 0, 1], "b": [4, 5, 6], "c": [7, 8, 9]})
In [375]: dft1 = dft1.astype({"a": np.bool_, "c": np.float64})
In [376]: dft1
Out[376]:
a b c
0 True 4 7.0
1 False 5 8.0
2 True 6 9.0
In [377]: dft1.dtypes
Out[377]:
a bool
b int64
c float64
dtype: object
注意
当尝试使用 astype() 和 loc() 将列的子集转换为指定类型时，会发生向上转型。
loc() 尝试将要分配的内容拟合到当前数据类型中，而 `[]` 会覆盖它们，从右侧获取数据类型。因此，以下代码会产生意外的结果。
In [378]: dft = pd.DataFrame({"a": [1, 2, 3], "b": [4, 5, 6], "c": [7, 8, 9]})
In [379]: dft.loc[:, ["a", "b"]].astype(np.uint8).dtypes
Out[379]:
a uint8
b uint8
dtype: object
In [380]: dft.loc[:, ["a", "b"]] = dft.loc[:, ["a", "b"]].astype(np.uint8)
In [381]: dft.dtypes
Out[381]:
a int64
b int64
c int64
dtype: object
对象转换#
pandas 提供了各种函数来尝试将 `object` 数据类型中的类型强制转换为其他类型。在数据已是正确类型但存储在 `object` 数组中的情况下，可以使用 DataFrame.infer_objects() 和 Series.infer_objects() 方法进行软转换到正确类型。
In [382]: import datetime In [383]: df = pd.DataFrame( .....: [ .....: [1, 2], .....: ["a", "b"], .....: [datetime.datetime(2016, 3, 2), datetime.datetime(2016, 3, 2)], .....: ] .....: ) .....: In [384]: df = df.T In [385]: df Out[385]: 0 1 2 0 1 a 2016-03-02 00:00:00 1 2 b 2016-03-02 00:00:00 In [386]: df.dtypes Out[386]: 0 object 1 object 2 object dtype: object
由于数据被转置，原始推断将所有列存储为对象，这将由 `infer_objects` 进行更正。
In [387]: df.infer_objects().dtypes Out[387]: 0 int64 1 str 2 datetime64[us] dtype: object
以下函数可用于一维对象数组或标量，以执行对象的硬转换到指定类型
to_numeric()（转换为数值数据类型）In [388]: m = ["1.1", 2, 3] In [389]: pd.to_numeric(m) Out[389]: array([1.1, 2. , 3. ])
to_datetime()（转换为日期时间对象）In [390]: import datetime In [391]: m = ["2016-07-09", datetime.datetime(2016, 3, 2)] In [392]: pd.to_datetime(m) Out[392]: DatetimeIndex(['2016-07-09', '2016-03-02'], dtype='datetime64[us]', freq=None)
to_timedelta()（转换为时间差对象）In [393]: m = ["5us", pd.Timedelta("1day")] In [394]: pd.to_timedelta(m) Out[394]: TimedeltaIndex(['0 days 00:00:00.000005', '1 days 00:00:00'], dtype='timedelta64[us]', freq=None)
要强制进行转换，我们可以传递一个 `errors` 参数，该参数指定 pandas 如何处理无法转换为所需数据类型或对象的元素。默认情况下，`errors='raise'`，这意味着在转换过程中会引发遇到的任何错误。但是，如果 `errors='coerce'`，这些错误将被忽略，pandas 会将有问题的元素转换为 `pd.NaT`（对于日期时间和时间差）或 `np.nan`（对于数值）。如果您正在读取的数据大部分是所需的数据类型（例如，数字、日期时间），但偶尔会混入不符合要求的元素，并且您希望将其表示为缺失，这可能会很有用。
In [395]: import datetime
In [396]: m = ["apple", datetime.datetime(2016, 3, 2)]
In [397]: pd.to_datetime(m, errors="coerce")
Out[397]: DatetimeIndex(['NaT', '2016-03-02'], dtype='datetime64[us]', freq=None)
In [398]: m = ["apple", 2, 3]
In [399]: pd.to_numeric(m, errors="coerce")
Out[399]: array([nan, 2., 3.])
In [400]: m = ["apple", pd.Timedelta("1day")]
In [401]: pd.to_timedelta(m, errors="coerce")
Out[401]: TimedeltaIndex([NaT, '1 days'], dtype='timedelta64[us]', freq=None)
除了对象转换之外，to_numeric() 还提供了一个名为 `downcast` 的参数，该参数可以选择将新（或已）转换的数字数据向下转换为较小的数据类型，从而节省内存。
In [402]: m = ["1", 2, 3]
In [403]: pd.to_numeric(m, downcast="integer") # smallest signed int dtype
Out[403]: array([1, 2, 3], dtype=int8)
In [404]: pd.to_numeric(m, downcast="signed") # same as 'integer'
Out[404]: array([1, 2, 3], dtype=int8)
In [405]: pd.to_numeric(m, downcast="unsigned") # smallest unsigned int dtype
Out[405]: array([1, 2, 3], dtype=uint8)
In [406]: pd.to_numeric(m, downcast="float") # smallest float dtype
Out[406]: array([1., 2., 3.], dtype=float32)
由于这些方法仅适用于一维数组、列表或标量；它们不能直接用于 DataFrame 等多维对象。但是，通过 apply()，我们可以高效地“对”每个列应用函数。
In [407]: import datetime
In [408]: df = pd.DataFrame([["2016-07-09", datetime.datetime(2016, 3, 2)]] * 2, dtype="O")
In [409]: df
Out[409]:
0 1
0 2016-07-09 2016-03-02 00:00:00
1 2016-07-09 2016-03-02 00:00:00
In [410]: df.apply(pd.to_datetime)
Out[410]:
0 1
0 2016-07-09 2016-03-02
1 2016-07-09 2016-03-02
In [411]: df = pd.DataFrame([["1.1", 2, 3]] * 2, dtype="O")
In [412]: df
Out[412]:
0 1 2
0 1.1 2 3
1 1.1 2 3
In [413]: df.apply(pd.to_numeric)
Out[413]:
0 1 2
0 1.1 2 3
1 1.1 2 3
In [414]: df = pd.DataFrame([["5us", pd.Timedelta("1day")]] * 2, dtype="O")
In [415]: df
Out[415]:
0 1
0 5us 1 days 00:00:00
1 5us 1 days 00:00:00
In [416]: df.apply(pd.to_timedelta)
Out[416]:
0 1
0 0 days 00:00:00.000005 1 days
1 0 days 00:00:00.000005 1 days
注意事项#
对*整数*类型数据执行选择操作很容易将数据向上转型为*浮点数*。在不引入 `nan` 的情况下，输入数据的数据类型将被保留。另请参阅 整数 NA 支持。
In [417]: dfi = df3.astype("int32")
In [418]: dfi["E"] = 1
In [419]: dfi
Out[419]:
A B C E
0 1 0 26 1
1 3 1 86 1
2 0 0 46 1
3 0 1 212 1
4 -1 -1 26 1
5 1 0 7 1
6 0 -1 184 1
7 0 0 206 1
In [420]: dfi.dtypes
Out[420]:
A int32
B int32
C int32
E int64
dtype: object
In [421]: casted = dfi[dfi > 0]
In [422]: casted
Out[422]:
A B C E
0 1.0 NaN 26 1
1 3.0 1.0 86 1
2 NaN NaN 46 1
3 NaN 1.0 212 1
4 NaN NaN 26 1
5 1.0 NaN 7 1
6 NaN NaN 184 1
7 NaN NaN 206 1
In [423]: casted.dtypes
Out[423]:
A float64
B float64
C int32
E int64
dtype: object
浮点数数据类型保持不变。
In [424]: dfa = df3.copy()
In [425]: dfa["A"] = dfa["A"].astype("float32")
In [426]: dfa.dtypes
Out[426]:
A float32
B float64
C float64
dtype: object
In [427]: casted = dfa[df2 > 0]
In [428]: casted
Out[428]:
A B C
0 1.047634 0.256090 26.0
1 3.498407 1.426469 86.0
2 NaN NaN 46.0
3 NaN 1.139976 212.0
4 NaN NaN 26.0
5 1.346474 0.096706 7.0
6 NaN NaN 184.0
7 NaN NaN 206.0
In [429]: casted.dtypes
Out[429]:
A float32
B float64
C float64
dtype: object
按数据类型选择列#
`select_dtypes()` 方法实现了基于其 `dtype` 对列进行子集选择。
首先，让我们创建一个包含各种不同数据类型的 `DataFrame`。
In [430]: df = pd.DataFrame(
.....: {
.....: "string": list("abc"),
.....: "int64": list(range(1, 4)),
.....: "uint8": np.arange(3, 6).astype("u1"),
.....: "float64": np.arange(4.0, 7.0),
.....: "bool1": [True, False, True],
.....: "bool2": [False, True, False],
.....: "dates": pd.date_range("now", periods=3),
.....: "category": pd.Series(list("ABC")).astype("category"),
.....: }
.....: )
.....:
In [431]: df["tdeltas"] = df.dates.diff()
In [432]: df["uint64"] = np.arange(3, 6).astype("u8")
In [433]: df["other_dates"] = pd.date_range("20130101", periods=3)
In [434]: df["tz_aware_dates"] = pd.date_range("20130101", periods=3, tz="US/Eastern")
In [435]: df
Out[435]:
string int64 uint8 ... uint64 other_dates tz_aware_dates
0 a 1 3 ... 3 2013-01-01 2013-01-01 00:00:00-05:00
1 b 2 4 ... 4 2013-01-02 2013-01-02 00:00:00-05:00
2 c 3 5 ... 5 2013-01-03 2013-01-03 00:00:00-05:00
[3 rows x 12 columns]
以及数据类型
In [436]: df.dtypes
Out[436]:
string str
int64 int64
uint8 uint8
float64 float64
bool1 bool
bool2 bool
dates datetime64[us]
category category
tdeltas timedelta64[us]
uint64 uint64
other_dates datetime64[us]
tz_aware_dates datetime64[us, US/Eastern]
dtype: object
select_dtypes() 有两个参数 `include` 和 `exclude`，允许您指定“给我*包含*这些数据类型的列”（`include`）和/或“给我*不包含*这些数据类型的列”（`exclude`）。
例如，选择 `bool` 列
In [437]: df.select_dtypes(include=[bool])
Out[437]:
bool1 bool2
0 True False
1 False True
2 True False
您还可以传递一个数据类型的名称，该名称位于 NumPy 数据类型层次结构中。
In [438]: df.select_dtypes(include=["bool"])
Out[438]:
bool1 bool2
0 True False
1 False True
2 True False
select_dtypes() 也适用于通用数据类型。
例如，选择所有数值和布尔列，同时排除无符号整数
In [439]: df.select_dtypes(include=["number", "bool"], exclude=["unsignedinteger"])
Out[439]:
int64 float64 bool1 bool2 tdeltas
0 1 4.0 True False NaT
1 2 5.0 False True 1 days
2 3 6.0 True False 1 days
要选择字符串列，请包含 `str`
In [440]: df.select_dtypes(include=[str])
Out[440]:
string
0 a
1 b
2 c
注意
这是 pandas 3.0 中的一项更改。以前，字符串存储在 `object` 数据类型列中，因此可以通过 `include=[object]` 来选择。有关如何编写同时适用于两个版本的代码的详细信息，请参阅 迁移指南。
要查看通用 `dtype`（如 `numpy.number`）的所有子数据类型，您可以定义一个返回子数据类型树的函数。
In [441]: def subdtypes(dtype):
.....: subs = dtype.__subclasses__()
.....: if not subs:
.....: return dtype
.....: return [dtype, [subdtypes(dt) for dt in subs]]
.....:
所有 NumPy 数据类型都是 `numpy.generic` 的子类。
In [442]: subdtypes(np.generic)
Out[442]:
[numpy.generic,
[[numpy.number,
[[numpy.integer,
[[numpy.signedinteger,
[numpy.int8,
numpy.int16,
numpy.int32,
numpy.int64,
numpy.longlong,
numpy.timedelta64]],
[numpy.unsignedinteger,
[numpy.uint8,
numpy.uint16,
numpy.uint32,
numpy.uint64,
numpy.ulonglong]]]],
[numpy.inexact,
[[numpy.floating,
[numpy.float16, numpy.float32, numpy.float64, numpy.longdouble]],
[numpy.complexfloating,
[numpy.complex64, numpy.complex128, numpy.clongdouble]]]]]],
[numpy.flexible,
[[numpy.character, [numpy.bytes_, numpy.str_]],
[numpy.void, [numpy.record]]]],
numpy.bool,
numpy.datetime64,
numpy.object_]]
注意
pandas 还定义了 `category` 和 `datetime64[ns, tz]` 等类型，它们未集成到正常的 NumPy 层次结构中，并且不会出现在上述函数中。