pandas Guide#
This is an elegant-design introduction to pandas, geared mainly for new users. You can see more complex recipes at pandas. Note that this introduction refers to the official documentation, so pandas has copyright to part of the content. The author enhances by adding more detailed examples and explanations for easier comprehension.
Customarily, we import as follows:
In [1]: import numpy as np
In [2]: import pandas as pd
In [3]: import seaborn as sns
Object creation#
Creating a Series by passing a list of values, letting pandas create a default integer index:
In [4]: s = pd.Series([1, 3, 5, np.nan, 6, 8]); s
Out[4]:
0 1.0
1 3.0
2 5.0
3 NaN
4 6.0
5 8.0
dtype: float64
Creating a DataFrame by passing a Numpy array, with a datetime index and labeled columns:
In [5]: dates = pd.date_range("20210801", periods=6)
In [6]: dates
Out[6]:
DatetimeIndex(['2021-08-01', '2021-08-02', '2021-08-03', '2021-08-04',
'2021-08-05', '2021-08-06'],
dtype='datetime64[ns]', freq='D')
In [7]: df = pd.DataFrame(np.random.randn(6, 4), index=dates, columns=list("ABCD"))
In [8]: df
Out[8]:
A B C D
2021-08-01 0.217120 0.353121 -0.551206 0.840941
2021-08-02 0.151220 1.155111 0.077040 -0.548415
2021-08-03 -0.239416 0.997688 -1.006559 0.673164
2021-08-04 -0.640489 0.219262 0.606844 -0.652033
2021-08-05 -0.177799 0.023023 1.535067 -0.051603
2021-08-06 -0.070985 0.068657 -0.068792 -1.131507
Creating a DataFrame by passing a dict of objects that can be converted to series-like:
In [9]: df2 = pd.DataFrame(
...: {
...: "A": 4.0,
...: "B": pd.Timestamp("20210801"),
...: "C": pd.Series(1, index=list(range(4)), dtype="float64"),
...: "D": np.array([3] * 4, dtype="int64"),
...: "E": pd.Categorical(["test", "train", "test", "train"]),
...: "F": "cat",
...: }
...: )
...:
In [10]: df2
Out[10]:
A B C D E F
0 4.0 2021-08-01 1.0 3 test cat
1 4.0 2021-08-01 1.0 3 train cat
2 4.0 2021-08-01 1.0 3 test cat
3 4.0 2021-08-01 1.0 3 train cat
The columns of the resulting DataFrame have different dtypes:
In [11]: df2.dtypes
Out[11]:
A float64
B datetime64[ns]
C float64
D int64
E category
F object
dtype: object
or the data type for a specific column:
In [12]: df2.A.dtype
Out[12]: dtype('float64')
If you’re using IPython, tab completion for public attributes is automatically enabled. Here’s a subset of the attributes that will be completed:
In [13]: df2.<TAB> # noqa: E225, E999
df2.A df2.bool
df2.abs df2.boxplot
df2.add df2.C
df2.add_prefix df2.clip
df2.add_suffix df2.columns
df2.align df2.copy
df2.all df2.count
df2.any df2.combine
df2.append df2.D
df2.apply df2.describe
df2.applymap df2.diff
df2.B df2.duplicated
Viewing data#
Here is how to view the top and bottom rows of the frame:
In [14]: df.head()
Out[14]:
A B C D
2021-08-01 0.217120 0.353121 -0.551206 0.840941
2021-08-02 0.151220 1.155111 0.077040 -0.548415
2021-08-03 -0.239416 0.997688 -1.006559 0.673164
2021-08-04 -0.640489 0.219262 0.606844 -0.652033
2021-08-05 -0.177799 0.023023 1.535067 -0.051603
In [15]: df.tail(3)
Out[15]:
A B C D
2021-08-04 -0.640489 0.219262 0.606844 -0.652033
2021-08-05 -0.177799 0.023023 1.535067 -0.051603
2021-08-06 -0.070985 0.068657 -0.068792 -1.131507
Display the index, columns:
In [16]: df.index
Out[16]:
DatetimeIndex(['2021-08-01', '2021-08-02', '2021-08-03', '2021-08-04',
'2021-08-05', '2021-08-06'],
dtype='datetime64[ns]', freq='D')
In [17]: df.columns
Out[17]: Index(['A', 'B', 'C', 'D'], dtype='object')
Display the shape, dimension:
In [18]: df.shape
Out[18]: (6, 4)
In [19]: df.ndim
Out[19]: 2
Note
DataFrame.to_numpy gives a Numpy representation of the underlying data.Note that this can be an expensive operation when your DataFrame has columns with different data types, which comes down to a fundamental difference between pandas and Numpy: Numpy arrays have one dtype for the entire array, while pandas DataFrames have one dtype per column. When you call DataFrame.to_numpy, pandas will find the Numpy dtype that can hold all of the dtypes in the DataFrame. This may end up being object, which requires casting every value to a Python object.
For df, our DataFrame of all floating-point values, DataFrame.to_numpy is fast and doesn’t require copying data.
In [20]: df.to_numpy()
Out[20]:
array([[ 0.2171, 0.3531, -0.5512, 0.8409],
[ 0.1512, 1.1551, 0.077 , -0.5484],
[-0.2394, 0.9977, -1.0066, 0.6732],
[-0.6405, 0.2193, 0.6068, -0.652 ],
[-0.1778, 0.023 , 1.5351, -0.0516],
[-0.071 , 0.0687, -0.0688, -1.1315]])
For df2, the DataFrame with multiple dtypes, DataFrame.to_numpy is relatively expensive.
In [21]: df2.to_numpy()
Out[21]:
array([[4.0, Timestamp('2021-08-01 00:00:00'), 1.0, 3, 'test', 'cat'],
[4.0, Timestamp('2021-08-01 00:00:00'), 1.0, 3, 'train', 'cat'],
[4.0, Timestamp('2021-08-01 00:00:00'), 1.0, 3, 'test', 'cat'],
[4.0, Timestamp('2021-08-01 00:00:00'), 1.0, 3, 'train', 'cat']],
dtype=object)
Note
DataFrame.to_numpy does not include the index or column labels in the output.
DataFrame.describe shows a quick statistic summary of your data:
In [22]: df.describe()
Out[22]:
A B C D
count 6.000000 6.000000 6.000000 6.000000
mean -0.126725 0.469477 0.098732 -0.144909
std 0.309144 0.486873 0.893740 0.780156
min -0.640489 0.023023 -1.006559 -1.131507
25% -0.224012 0.106308 -0.430602 -0.626129
50% -0.124392 0.286192 0.004124 -0.300009
75% 0.095669 0.836547 0.474393 0.491972
max 0.217120 1.155111 1.535067 0.840941
DateFrame.info show information about dataset
In [23]: df.info()
<class 'pandas.core.frame.DataFrame'>
DatetimeIndex: 6 entries, 2021-08-01 to 2021-08-06
Freq: D
Data columns (total 4 columns):
# Column Non-Null Count Dtype
--- ------ -------------- -----
0 A 6 non-null float64
1 B 6 non-null float64
2 C 6 non-null float64
3 D 6 non-null float64
dtypes: float64(4)
memory usage: 240.0 bytes
Locate and aggregate NaN values column-wise:
In [24]: df.isnull().sum().sort_values(ascending=False)
Out[24]:
A 0
B 0
C 0
D 0
dtype: int64
Find the total number of NaN values
In [25]: df.isna().sum().sum() # identical to isnull() here
Out[25]: 0
Transposing your data:
In [26]: df.T
Out[26]:
2021-08-01 2021-08-02 2021-08-03 2021-08-04 2021-08-05 2021-08-06
A 0.217120 0.151220 -0.239416 -0.640489 -0.177799 -0.070985
B 0.353121 1.155111 0.997688 0.219262 0.023023 0.068657
C -0.551206 0.077040 -1.006559 0.606844 1.535067 -0.068792
D 0.840941 -0.548415 0.673164 -0.652033 -0.051603 -1.131507
Sorting by an axis:
In [27]: df.sort_index(axis=1, ascending=False)
Out[27]:
D C B A
2021-08-01 0.840941 -0.551206 0.353121 0.217120
2021-08-02 -0.548415 0.077040 1.155111 0.151220
2021-08-03 0.673164 -1.006559 0.997688 -0.239416
2021-08-04 -0.652033 0.606844 0.219262 -0.640489
2021-08-05 -0.051603 1.535067 0.023023 -0.177799
2021-08-06 -1.131507 -0.068792 0.068657 -0.070985
Sorting by values:
In [28]: df.sort_values(by="B")
Out[28]:
A B C D
2021-08-05 -0.177799 0.023023 1.535067 -0.051603
2021-08-06 -0.070985 0.068657 -0.068792 -1.131507
2021-08-04 -0.640489 0.219262 0.606844 -0.652033
2021-08-01 0.217120 0.353121 -0.551206 0.840941
2021-08-03 -0.239416 0.997688 -1.006559 0.673164
2021-08-02 0.151220 1.155111 0.077040 -0.548415
Selection#
Note
While standard Python / Numpy expressions for selecting and setting are intuitive and come in handy for interactive work, for production code, we recommend the optimized pandas data access methods, .at, .iat, .loc and .iloc.
Getting#
Selecting a single column, which yields a Series, equivalent to df.A:
In [29]: df["A"]
Out[29]:
2021-08-01 0.217120
2021-08-02 0.151220
2021-08-03 -0.239416
2021-08-04 -0.640489
2021-08-05 -0.177799
2021-08-06 -0.070985
Freq: D, Name: A, dtype: float64
Selecting via [], which slices the rows.
In [30]: df[0:3]
Out[30]:
A B C D
2021-08-01 0.217120 0.353121 -0.551206 0.840941
2021-08-02 0.151220 1.155111 0.077040 -0.548415
2021-08-03 -0.239416 0.997688 -1.006559 0.673164
In [31]: df["20210802":"20210804"]
Out[31]:
A B C D
2021-08-02 0.151220 1.155111 0.077040 -0.548415
2021-08-03 -0.239416 0.997688 -1.006559 0.673164
2021-08-04 -0.640489 0.219262 0.606844 -0.652033
Selection by label#
For getting a cross section using a label:
In [32]: df.loc[dates[0]]
Out[32]:
A 0.217120
B 0.353121
C -0.551206
D 0.840941
Name: 2021-08-01 00:00:00, dtype: float64
Selecting on a multi-axis by label:
In [33]: df.loc[:, ["A", "B"]]
Out[33]:
A B
2021-08-01 0.217120 0.353121
2021-08-02 0.151220 1.155111
2021-08-03 -0.239416 0.997688
2021-08-04 -0.640489 0.219262
2021-08-05 -0.177799 0.023023
2021-08-06 -0.070985 0.068657
Showing label slicing, both endpoints are included:
In [34]: df.loc["20210802":"20210804", ["A", "B"]]
Out[34]:
A B
2021-08-02 0.151220 1.155111
2021-08-03 -0.239416 0.997688
2021-08-04 -0.640489 0.219262
Reduction in the dimensions of the returned object:
In [35]: df.loc["20210802", ["A", "B"]]
Out[35]:
A 0.151220
B 1.155111
Name: 2021-08-02 00:00:00, dtype: float64
For getting a scalar value:
In [36]: df.loc[dates[0], "A"]
Out[36]: 0.21712016704323
For getting fast access to a scalar (equivalent to the prior method):
In [37]: df.at[dates[0], "A"]
Out[37]: 0.21712016704323
Selection by position#
Select via the position of the passed integers:
In [38]: df.iloc[3]
Out[38]:
A -0.640489
B 0.219262
C 0.606844
D -0.652033
Name: 2021-08-04 00:00:00, dtype: float64
By integer slices, acting similar to Numpy/Python:
In [39]: df.iloc[3:5, 0:2]
Out[39]:
A B
2021-08-04 -0.640489 0.219262
2021-08-05 -0.177799 0.023023
By lists of integer position locations, similar to the Numpy/Python style:
In [40]: df.iloc[[1, 2, 4], [0, 2]]
Out[40]:
A C
2021-08-02 0.151220 0.077040
2021-08-03 -0.239416 -1.006559
2021-08-05 -0.177799 1.535067
For slicing rows explicitly:
In [41]: df.iloc[1:3, :]
Out[41]:
A B C D
2021-08-02 0.151220 1.155111 0.077040 -0.548415
2021-08-03 -0.239416 0.997688 -1.006559 0.673164
For slicing columns explicitly:
In [42]: df.iloc[:, 1:3]
Out[42]:
B C
2021-08-01 0.353121 -0.551206
2021-08-02 1.155111 0.077040
2021-08-03 0.997688 -1.006559
2021-08-04 0.219262 0.606844
2021-08-05 0.023023 1.535067
2021-08-06 0.068657 -0.068792
For getting a value explicitly:
In [43]: df.iloc[1, 1]
Out[43]: 1.1551105853091492
For getting fast access to a scalar (equivalent to the prior method):
In [44]: df.iat[1, 1]
Out[44]: 1.1551105853091492
Selection by dtype#
In [45]: titanic = sns.load_dataset('titanic')
In [46]: titanic.head()
Out[46]:
survived pclass sex age sibsp parch fare embarked class who adult_male deck embark_town alive alone
0 0 3 male 22.0 1 0 7.2500 S Third man True NaN Southampton no False
1 1 1 female 38.0 1 0 71.2833 C First woman False C Cherbourg yes False
2 1 3 female 26.0 0 0 7.9250 S Third woman False NaN Southampton yes True
3 1 1 female 35.0 1 0 53.1000 S First woman False C Southampton yes False
4 0 3 male 35.0 0 0 8.0500 S Third man True NaN Southampton no True
In [47]: titanic.select_dtypes(include=['datetime', 'number']).head()
Out[47]:
survived pclass age sibsp parch fare
0 0 3 22.0 1 0 7.2500
1 1 1 38.0 1 0 71.2833
2 1 3 26.0 0 0 7.9250
3 1 1 35.0 1 0 53.1000
4 0 3 35.0 0 0 8.0500
In [48]: titanic.select_dtypes(exclude=['object', 'double']).head()
Out[48]:
survived pclass sibsp parch class adult_male deck alone
0 0 3 1 0 Third True NaN False
1 1 1 1 0 First False C False
2 1 3 0 0 Third False NaN True
3 1 1 1 0 First False C False
4 0 3 0 0 Third True NaN True
Boolean indexing#
Using a single column’s values to select data.
In [49]: df[df["A"] > 0]
Out[49]:
A B C D
2021-08-01 0.21712 0.353121 -0.551206 0.840941
2021-08-02 0.15122 1.155111 0.077040 -0.548415
Selecting values from a DataFrame where a boolean condition is met.
In [50]: df[df > 0]
Out[50]:
A B C D
2021-08-01 0.21712 0.353121 NaN 0.840941
2021-08-02 0.15122 1.155111 0.077040 NaN
2021-08-03 NaN 0.997688 NaN 0.673164
2021-08-04 NaN 0.219262 0.606844 NaN
2021-08-05 NaN 0.023023 1.535067 NaN
2021-08-06 NaN 0.068657 NaN NaN
Using the Series.isin method for filtering:
In [51]: df2[df2["E"].isin(["test"])]
Out[51]:
A B C D E F
0 4.0 2021-08-01 1.0 3 test cat
2 4.0 2021-08-01 1.0 3 test cat
Setting#
Setting a new column automatically aligns the data by the indexes.
In [52]: s1 = pd.Series(np.arange(6)+1, index=pd.date_range("20210801", periods=6))
In [53]: s1
Out[53]:
2021-08-01 1
2021-08-02 2
2021-08-03 3
2021-08-04 4
2021-08-05 5
2021-08-06 6
Freq: D, dtype: int64
In [54]: df["F"] = s1
Setting values by label:
In [55]: df.at[dates[0], "A"] = 0
Setting values by position:
In [56]: df.iat[0, 1] = 0
Setting by assigning with a Numpy array:
In [57]: df["D"] = np.array([6] * len(df))
The result of the prior setting operations.
In [58]: df
Out[58]:
A B C D F
2021-08-01 0.000000 0.000000 -0.551206 6 1
2021-08-02 0.151220 1.155111 0.077040 6 2
2021-08-03 -0.239416 0.997688 -1.006559 6 3
2021-08-04 -0.640489 0.219262 0.606844 6 4
2021-08-05 -0.177799 0.023023 1.535067 6 5
2021-08-06 -0.070985 0.068657 -0.068792 6 6
A where operation with setting.
In [59]: df2 = df.copy()
In [60]: df2[df2 > 0] = -df2
In [61]: df2
Out[61]:
A B C D F
2021-08-01 0.000000 0.000000 -0.551206 -6 -1
2021-08-02 -0.151220 -1.155111 -0.077040 -6 -2
2021-08-03 -0.239416 -0.997688 -1.006559 -6 -3
2021-08-04 -0.640489 -0.219262 -0.606844 -6 -4
2021-08-05 -0.177799 -0.023023 -1.535067 -6 -5
2021-08-06 -0.070985 -0.068657 -0.068792 -6 -6
Missing data#
pandas primarily uses the value np.nan to represent missing data. It is by default not included in computations.
Reindexing allows you to change/add/delete the index on a specified axis. This returns a copy of the data.
In [62]: df1 = df.reindex(index=dates[0:4], columns=list(df.columns) + ["E"])
In [63]: df1
Out[63]:
A B C D F E
2021-08-01 0.000000 0.000000 -0.551206 6 1 NaN
2021-08-02 0.151220 1.155111 0.077040 6 2 NaN
2021-08-03 -0.239416 0.997688 -1.006559 6 3 NaN
2021-08-04 -0.640489 0.219262 0.606844 6 4 NaN
In [64]: df1.loc[dates[0] : dates[1], "E"] = 1
In [65]: df1
Out[65]:
A B C D F E
2021-08-01 0.000000 0.000000 -0.551206 6 1 1.0
2021-08-02 0.151220 1.155111 0.077040 6 2 1.0
2021-08-03 -0.239416 0.997688 -1.006559 6 3 NaN
2021-08-04 -0.640489 0.219262 0.606844 6 4 NaN
Drop any rows that have missing data.
In [66]: df1.dropna(how="any")
Out[66]:
A B C D F E
2021-08-01 0.00000 0.000000 -0.551206 6 1 1.0
2021-08-02 0.15122 1.155111 0.077040 6 2 1.0
Drop any columns that exceed certain threshold。
In [67]: df1.dropna(thresh=len(df)*0.9, axis=1)
Out[67]:
Empty DataFrame
Columns: []
Index: [2021-08-01 00:00:00, 2021-08-02 00:00:00, 2021-08-03 00:00:00, 2021-08-04 00:00:00]
Get the boolean mask where values are nan.
In [68]: pd.isna(df1)
Out[68]:
A B C D F E
2021-08-01 False False False False False False
2021-08-02 False False False False False False
2021-08-03 False False False False False True
2021-08-04 False False False False False True
Replace NaN in a column with a value.
In [69]: df1.E.replace(np.nan, 6.0, inplace=True)
In [70]: df1
Out[70]:
A B C D F E
2021-08-01 0.000000 0.000000 -0.551206 6 1 1.0
2021-08-02 0.151220 1.155111 0.077040 6 2 1.0
2021-08-03 -0.239416 0.997688 -1.006559 6 3 6.0
2021-08-04 -0.640489 0.219262 0.606844 6 4 6.0
Filling missing data with closest previous valid value with preceding direction
In [71]: df1.fillna(axis=0, method='ffill', inplace=True) # up
In [72]: df1.fillna(axis=1, method='ffill') # left
Out[72]:
A B C D F E
2021-08-01 0.000000 0.000000 -0.551206 6.0 1.0 1.0
2021-08-02 0.151220 1.155111 0.077040 6.0 2.0 1.0
2021-08-03 -0.239416 0.997688 -1.006559 6.0 3.0 6.0
2021-08-04 -0.640489 0.219262 0.606844 6.0 4.0 6.0
In [73]: df1.fillna(axis=0, method='bfill') # down
Out[73]:
A B C D F E
2021-08-01 0.000000 0.000000 -0.551206 6 1 1.0
2021-08-02 0.151220 1.155111 0.077040 6 2 1.0
2021-08-03 -0.239416 0.997688 -1.006559 6 3 6.0
2021-08-04 -0.640489 0.219262 0.606844 6 4 6.0
In [74]: df1.fillna(axis=1, method='bfill') # right
Out[74]:
A B C D F E
2021-08-01 0.000000 0.000000 -0.551206 6.0 1.0 1.0
2021-08-02 0.151220 1.155111 0.077040 6.0 2.0 1.0
2021-08-03 -0.239416 0.997688 -1.006559 6.0 3.0 6.0
2021-08-04 -0.640489 0.219262 0.606844 6.0 4.0 6.0
Drop duplicate values in a column.
In [75]: df1.E.drop_duplicates(keep="last")
Out[75]:
2021-08-02 1.0
2021-08-04 6.0
Name: E, dtype: float64
Drop rows which contains a value.
In [76]: df1[((df1 != 2) & (df1 != 3)).all(axis=1)]
Out[76]:
A B C D F E
2021-08-01 0.000000 0.000000 -0.551206 6 1 1.0
2021-08-04 -0.640489 0.219262 0.606844 6 4 6.0
Change certain column names.
In [77]: df1.rename(columns={"F":"Rank", "E":"Variance"})
Out[77]:
A B C D Rank Variance
2021-08-01 0.000000 0.000000 -0.551206 6 1 1.0
2021-08-02 0.151220 1.155111 0.077040 6 2 1.0
2021-08-03 -0.239416 0.997688 -1.006559 6 3 6.0
2021-08-04 -0.640489 0.219262 0.606844 6 4 6.0
Operations#
Stats#
Operations in general exclude missing data.
Performing a descriptive statistic:
In [78]: df.mean()
Out[78]:
A -0.162912
B 0.410623
C 0.098732
D 6.000000
F 3.500000
dtype: float64
Same operation on the other axis:
In [79]: df.mean(1)
Out[79]:
2021-08-01 1.289759
2021-08-02 1.876674
2021-08-03 1.750343
2021-08-04 2.037123
2021-08-05 2.476058
2021-08-06 2.385776
Freq: D, dtype: float64
Operating with objects that have different dimensionality and need alignment. In addition, pandas automatically broadcasts along the specified dimension.
In [80]: s = pd.Series([1, 3, 5, np.nan, 6, 8], index=dates).shift(2)
In [81]: s
Out[81]:
2021-08-01 NaN
2021-08-02 NaN
2021-08-03 1.0
2021-08-04 3.0
2021-08-05 5.0
2021-08-06 NaN
Freq: D, dtype: float64
In [82]: df.sub(s, axis="index")
Out[82]:
A B C D F
2021-08-01 NaN NaN NaN NaN NaN
2021-08-02 NaN NaN NaN NaN NaN
2021-08-03 -1.239416 -0.002312 -2.006559 5.0 2.0
2021-08-04 -3.640489 -2.780738 -2.393156 3.0 1.0
2021-08-05 -5.177799 -4.976977 -3.464933 1.0 0.0
2021-08-06 NaN NaN NaN NaN NaN
Apply#
Applying functions to the data:
In [83]: df.apply(np.cumsum)
Out[83]:
A B C D F
2021-08-01 0.000000 0.000000 -0.551206 6 1
2021-08-02 0.151220 1.155111 -0.474166 12 3
2021-08-03 -0.088196 2.152799 -1.480724 18 6
2021-08-04 -0.728685 2.372060 -0.873880 24 10
2021-08-05 -0.906484 2.395084 0.661187 30 15
2021-08-06 -0.977470 2.463741 0.592395 36 21
In [84]: df.apply(lambda x: x.max() - x.min())
Out[84]:
A 0.791708
B 1.155111
C 2.541626
D 0.000000
F 5.000000
dtype: float64
String Methods#
Series is equipped with a set of string processing methods in the str attribute that make it easy to operate on each element of the array, as in the code snippet below. Note that pattern-matching in str generally uses regular expressions by default (and in some cases always uses them).
In [85]: s = pd.Series(["A", "B", "C", np.nan, "CABA", "dog", "cat"])
In [86]: s.str.lower()
Out[86]:
0 a
1 b
2 c
3 NaN
4 caba
5 dog
6 cat
dtype: object
More string methods are provided like str.title, str.capitalize, str.upper.
Data types#
Several methods to convert data types are available:
In [87]: df = pd.DataFrame({
....: 'product': ['A', 'B', 'C', 'D'],
....: 'price': ['10', '20', '30', '40'],
....: 'sales': ['20', '-', '60', '-']
....: })
....:
In [88]: df
Out[88]:
product price sales
0 A 10 20
1 B 20 -
2 C 30 60
3 D 40 -
In [89]: df = df.astype({'price': 'int'})
But when we use astype to convert sales to integer, ValueError will occur because of invalid literal ‘-’. Fortunately, to_numeric solves the issue with parameter errors:
In [90]: df['sales'] = pd.to_numeric(df['sales'], errors='coerce')
In [91]: df.dtypes
Out[91]:
product object
price int64
sales float64
dtype: object
Merge#
Join#
We illustrate a more detailed example in a background of business. Suppose we have three dataframes:
In [92]: df1 = pd.DataFrame({
....: "id": np.arange(1, 4),
....: "price": [99, 105, 50]
....: })
....:
In [93]: df1
Out[93]:
id price
0 1 99
1 2 105
2 3 50
In [94]: df2 = pd.DataFrame({
....: "id": np.arange(3, 5),
....: "count": [12, 15]
....: })
....:
In [95]: df2
Out[95]:
id count
0 3 12
1 4 15
In [96]: df3 = pd.DataFrame({
....: "id": np.arange(1, 7),
....: "price": [99, 105, 50, 60, 30, 40],
....: "count": [10, 20, 12, 15, 100, 50],
....: "date": pd.date_range("20210801", periods=6)
....: })
....:
In [97]: df3
Out[97]:
id price count date
0 1 99 10 2021-08-01
1 2 105 20 2021-08-02
2 3 50 12 2021-08-03
3 4 60 15 2021-08-04
4 5 30 100 2021-08-05
5 6 40 50 2021-08-06
SQL style merges :
In [98]: pd.merge(df1, df2, on="id", how="left")
Out[98]:
id price count
0 1 99 NaN
1 2 105 NaN
2 3 50 12.0
In [99]: pd.merge(df1, df2, on="id", how="right")
Out[99]:
id price count
0 3 50.0 12
1 4 NaN 15
In [100]: pd.merge(df1, df2, on="id", how="outer")
Out[100]:
id price count
0 1 99.0 NaN
1 2 105.0 NaN
2 3 50.0 12.0
3 4 NaN 15.0
In [101]: pd.merge(df1, df2, on="id", how="inner")
Out[101]:
id price count
0 3 50 12
In [102]: df1.merge(df2, on="id", how="inner") # alternative syntax
Out[102]:
id price count
0 3 50 12
If we are to merge two dataframes with identical column names, directly using DataFrame.join will produce an ValueError said columns overlap but no suffix specified. One approach is to specify suffix for both dataframes with respect to columns that cause the conflict:
In [103]: df1.join(df2, lsuffix="_df1", rsuffix="_df2")
Out[103]:
id_df1 price id_df2 count
0 1 99 3.0 12.0
1 2 105 4.0 15.0
2 3 50 NaN NaN
Another approach is to set ahead index for both dataframes:
In [104]: df1.set_index("id").join(df2.set_index("id"))
Out[104]:
price count
id
1 99 NaN
2 105 NaN
3 50 12.0
Reset index while adding the old index as a new column:
In [105]: df1.set_index("id", inplace=True)
In [106]: df1
Out[106]:
price
id
1 99
2 105
3 50
In [107]: df1.reset_index(drop=True)
Out[107]:
price
0 99
1 105
2 50
In [108]: df1.reset_index(drop=False, inplace=True)
Concat#
pandas provides various facilities for easily combining together Series and DataFrame objects with various kinds of set logic for the indexes and relational algebra functionality in the case of join / merge-type operations.
Split into pieces and concatenate back:
In [109]: pieces = [df3[:2], df3[2:4], df3[4:]]
In [110]: pieces
Out[110]:
[ id price count date
0 1 99 10 2021-08-01
1 2 105 20 2021-08-02,
id price count date
2 3 50 12 2021-08-03
3 4 60 15 2021-08-04,
id price count date
4 5 30 100 2021-08-05
5 6 40 50 2021-08-06]
In [111]: pd.concat(pieces)
Out[111]:
id price count date
0 1 99 10 2021-08-01
1 2 105 20 2021-08-02
2 3 50 12 2021-08-03
3 4 60 15 2021-08-04
4 5 30 100 2021-08-05
5 6 40 50 2021-08-06
Note
Adding a column to a DataFrame is relatively fast. However, adding
a row requires a copy, and may be expensive. We recommend passing a
pre-built list of records to the DataFrame constructor instead
of building a DataFrame by iteratively appending records to it.
Extend this example to consolidate former knowledge. Determine whether to supplement products:
In [112]: df3["enough"] = np.where(df3["count"] >= 50, "yes", "no")
Add level of price to expensive or cheap:
df3.loc[(df3[“price”] > 50), “level”] = “high” df3.loc[(df3[“price”] < 51), “level”] = “low”
Use cut to categorize data by bins:
In [113]: df3['ageGroup'] = pd.cut(df3['price'], bins=[0, 40, 70, 100],
.....: labels=['low', 'mid', 'high'])
.....:
Split date into detailed unit and horizontally append to dataframe:
In [114]: df3.date = df3.date.astype(str) # avoid AttributeError
In [115]: subdate = pd.DataFrame((x.split('-') for x in df3.date), index=df3.index,
.....: columns=['year','month','day'])
.....:
In [116]: pd.concat([df3,subdate], axis=1, keys=["df3", "dates"])
Out[116]:
df3 dates
id price count date enough ageGroup year month day
0 1 99 10 2021-08-01 no high 2021 08 01
1 2 105 20 2021-08-02 no NaN 2021 08 02
2 3 50 12 2021-08-03 no mid 2021 08 03
3 4 60 15 2021-08-04 no mid 2021 08 04
4 5 30 100 2021-08-05 yes low 2021 08 05
5 6 40 50 2021-08-06 yes low 2021 08 06
Warning
Convert data to string dtype before conducting string manipulation, or AttributeError will occur. If we do not convert df3.date ahead, an error will raise as: AttributeError: ‘Timestamp’ object has no attribute ‘split’.
Grouping#
By “group by” we are referring to a process involving one or more of the following steps:
Splitting the data into groups based on some criteria
Applying a function to each group independently
Combining the results into a data structure
In [117]: df = pd.DataFrame({
.....: "A": ["foo", "bar", "foo", "bar", "foo", "bar", "foo", "foo"],
.....: "B": ["one", "one", "two", "three", "two", "two", "one", "three"],
.....: "C": np.random.randn(8),
.....: "D": np.random.randn(8),
.....: })
.....:
In [118]: df
Out[118]:
A B C D
0 foo one 0.066567 1.593288
1 bar one -0.858266 -0.002256
2 foo two 1.090977 0.556972
3 bar three -0.914114 -0.365743
4 foo two -0.518657 1.653892
5 bar two -1.252172 0.866307
6 foo one -0.179143 0.288360
7 foo three -0.780076 0.609170
Grouping and then applying the pandas.core.groupby.GroupBy.sum function to the resulting groups.
In [119]: df.groupby("A").sum(numeric_only=True)
Out[119]:
C D
A
bar -3.024551 0.498308
foo -0.320333 4.701682
Grouping by multiple columns forms a hierarchical index, and again we can apply the sum function.
In [120]: df.groupby(["A", "B"]).sum(numeric_only=True)
Out[120]:
C D
A B
bar one -0.858266 -0.002256
three -0.914114 -0.365743
two -1.252172 0.866307
foo one -0.112576 1.881648
three -0.780076 0.609170
two 0.572319 2.210864
Reshaping#
Stack#
In [121]: tuples = list(
.....: zip(
.....: *[
.....: ["bar", "bar", "baz", "baz", "foo", "foo", "qux", "qux"],
.....: ["one", "two", "one", "two", "one", "two", "one", "two"],
.....: ]
.....: )
.....: )
.....:
In [122]: tuples
Out[122]:
[('bar', 'one'),
('bar', 'two'),
('baz', 'one'),
('baz', 'two'),
('foo', 'one'),
('foo', 'two'),
('qux', 'one'),
('qux', 'two')]
In [123]: index = pd.MultiIndex.from_tuples(tuples, names=["first", "second"])
In [124]: index
Out[124]:
MultiIndex([('bar', 'one'),
('bar', 'two'),
('baz', 'one'),
('baz', 'two'),
('foo', 'one'),
('foo', 'two'),
('qux', 'one'),
('qux', 'two')],
names=['first', 'second'])
In [125]: df = pd.DataFrame(np.random.randn(8, 2), index=index, columns=["A", "B"])
In [126]: df
Out[126]:
A B
first second
bar one 0.942738 0.473426
two -0.481780 0.217468
baz one -0.935083 -2.314436
two -0.756820 0.134721
foo one 1.626621 -0.189159
two 0.163252 -2.167390
qux one 0.872637 -0.399940
two 1.590034 -1.410947
In [127]: df2 = df[:4]
In [128]: df2
Out[128]:
A B
first second
bar one 0.942738 0.473426
two -0.481780 0.217468
baz one -0.935083 -2.314436
two -0.756820 0.134721
The DataFrame.stack method compresses a level in the DataFrame’s columns.
In [129]: stacked = df2.stack()
In [130]: stacked
Out[130]:
first second
bar one A 0.942738
B 0.473426
two A -0.481780
B 0.217468
baz one A -0.935083
B -2.314436
two A -0.756820
B 0.134721
dtype: float64
With a stacked DataFrame or Series (having a MultiIndex as the index), the inverse operation of DataFrame.stack is DataFrame.unstack, which by default unstacks the last level:
In [131]: stacked.unstack() # unstack on the last level 2
Out[131]:
A B
first second
bar one 0.942738 0.473426
two -0.481780 0.217468
baz one -0.935083 -2.314436
two -0.756820 0.134721
In [132]: stacked.unstack(1)
Out[132]:
second one two
first
bar A 0.942738 -0.481780
B 0.473426 0.217468
baz A -0.935083 -0.756820
B -2.314436 0.134721
In [133]: stacked.unstack(0)
Out[133]:
first bar baz
second
one A 0.942738 -0.935083
B 0.473426 -2.314436
two A -0.481780 -0.756820
B 0.217468 0.134721
Pivot tables#
In [134]: df = pd.DataFrame(
.....: {
.....: "A": ["one", "one", "two", "three"] * 3,
.....: "B": ["A", "B", "C"] * 4,
.....: "C": ["foo", "foo", "foo", "bar", "bar", "bar"] * 2,
.....: "D": np.random.randn(12),
.....: "E": np.random.randn(12),
.....: }
.....: )
.....:
In [135]: df
Out[135]:
A B C D E
0 one A foo -0.322460 -1.200878
1 one B foo 2.964687 0.539208
2 two C foo 0.039817 -0.656564
3 three A bar -1.170292 -0.248316
4 one B bar 0.268392 -0.265521
.. ... .. ... ... ...
7 three B foo -1.341226 0.041183
8 one C foo 0.073688 -1.352641
9 one A bar 0.102749 -0.313030
10 two B bar -0.321690 0.616352
11 three C bar -0.551272 -0.504566
[12 rows x 5 columns]
We can produce pivot tables from this data very easily:
In [136]: pd.pivot_table(df, values="D", index=["A", "B"], columns=["C"])
Out[136]:
C bar foo
A B
one A 0.102749 -0.322460
B 0.268392 2.964687
C 0.273466 0.073688
three A -1.170292 NaN
B NaN -1.341226
C -0.551272 NaN
two A NaN -0.074330
B -0.321690 NaN
C NaN 0.039817
Time series#
pandas has simple, powerful, and efficient functionality for performing resampling operations during frequency conversion (e.g., converting secondly data into 5-minutely data). This is extremely common in, but not limited to, financial applications.
In [137]: rng = pd.date_range("1/1/2012", periods=100, freq="S")
In [138]: ts = pd.Series(np.random.randint(0, 500, len(rng)), index=rng)
In [139]: ts
Out[139]:
2012-01-01 00:00:00 182
2012-01-01 00:00:01 250
2012-01-01 00:00:02 357
2012-01-01 00:00:03 212
2012-01-01 00:00:04 192
...
2012-01-01 00:01:35 466
2012-01-01 00:01:36 462
2012-01-01 00:01:37 206
2012-01-01 00:01:38 465
2012-01-01 00:01:39 491
Freq: S, Length: 100, dtype: int64
In [140]: ts.resample("5Min").sum()
Out[140]:
2012-01-01 28000
Freq: 5T, dtype: int64
Time zone representation:
In [141]: rng = pd.date_range("3/6/2012 00:00", periods=5, freq="D")
In [142]: ts = pd.Series(np.random.randn(len(rng)), rng)
In [143]: ts
Out[143]:
2012-03-06 0.856107
2012-03-07 -0.822492
2012-03-08 -0.274845
2012-03-09 -1.456409
2012-03-10 0.765602
Freq: D, dtype: float64
In [144]: ts_utc = ts.tz_localize("UTC")
In [145]: ts_utc
Out[145]:
2012-03-06 00:00:00+00:00 0.856107
2012-03-07 00:00:00+00:00 -0.822492
2012-03-08 00:00:00+00:00 -0.274845
2012-03-09 00:00:00+00:00 -1.456409
2012-03-10 00:00:00+00:00 0.765602
Freq: D, dtype: float64
Converting to another time zone:
In [146]: ts_utc.tz_convert("US/Eastern")
Out[146]:
2012-03-05 19:00:00-05:00 0.856107
2012-03-06 19:00:00-05:00 -0.822492
2012-03-07 19:00:00-05:00 -0.274845
2012-03-08 19:00:00-05:00 -1.456409
2012-03-09 19:00:00-05:00 0.765602
Freq: D, dtype: float64
Converting between time span representations:
In [147]: rng = pd.date_range("1/1/2012", periods=5, freq="M")
In [148]: ts = pd.Series(np.random.randn(len(rng)), index=rng)
In [149]: ts
Out[149]:
2012-01-31 -0.454218
2012-02-29 0.945687
2012-03-31 -1.095913
2012-04-30 -0.868136
2012-05-31 -1.558698
Freq: M, dtype: float64
In [150]: ps = ts.to_period()
In [151]: ps
Out[151]:
2012-01 -0.454218
2012-02 0.945687
2012-03 -1.095913
2012-04 -0.868136
2012-05 -1.558698
Freq: M, dtype: float64
In [152]: ps.to_timestamp()
Out[152]:
2012-01-01 -0.454218
2012-02-01 0.945687
2012-03-01 -1.095913
2012-04-01 -0.868136
2012-05-01 -1.558698
Freq: MS, dtype: float64
Converting between period and timestamp enables some convenient arithmetic functions to be used. In the following example, we convert a quarterly frequency with year ending in November to 9am of the end of the month following the quarter end:
In [153]: prng = pd.period_range("1990Q1", "2000Q4", freq="Q-NOV")
In [154]: ts = pd.Series(np.random.randn(len(prng)), prng)
In [155]: ts.index = (prng.asfreq("M", "e") + 1).asfreq("H", "s") + 9
In [156]: ts.head()
Out[156]:
1990-03-01 09:00 0.498611
1990-06-01 09:00 2.248593
1990-09-01 09:00 -0.207991
1990-12-01 09:00 1.639503
1991-03-01 09:00 -0.182373
Freq: H, dtype: float64
Categoricals#
pandas can include categorical data in a DataFrame.
In [157]: df = pd.DataFrame(
.....: {
.....: "id": [1, 2, 3, 4, 5, 6],
.....: "raw_grade": ["a", "b", "b", "a", "a", "e"]
.....: }
.....: )
.....:
Convert the raw grades to a categorical data type.
In [158]: df["grade"] = df["raw_grade"].astype("category")
In [159]: df["grade"]
Out[159]:
0 a
1 b
2 b
3 a
4 a
5 e
Name: grade, dtype: category
Categories (3, object): ['a', 'b', 'e']
Rename the categories to more meaningful names (assigning to Series.cat.rename_categories is in place!).
In [160]: df["grade"].cat.rename_categories = ["very good", "good", "very bad"]
Reorder the categories and simultaneously add the missing categories (methods under Series.cat return a new Series by default).
In [161]: df["grade"] = df["grade"].cat.set_categories(
.....: ["very bad", "bad", "medium", "good", "very good"]
.....: )
.....:
In [162]: df["grade"]
Out[162]:
0 NaN
1 NaN
2 NaN
3 NaN
4 NaN
5 NaN
Name: grade, dtype: category
Categories (5, object): ['very bad', 'bad', 'medium', 'good', 'very good']
Sorting is per order in the categories, not lexical order.
In [163]: df.sort_values(by="grade")
Out[163]:
id raw_grade grade
0 1 a NaN
1 2 b NaN
2 3 b NaN
3 4 a NaN
4 5 a NaN
5 6 e NaN
Grouping by a categorical column also shows empty categories.
In [164]: df.groupby("grade").size()
Out[164]:
grade
very bad 0
bad 0
medium 0
good 0
very good 0
dtype: int64
Plotting#
We use the standard convention for referencing the matplotlib API:
In [165]: import matplotlib.pyplot as plt
In [166]: plt.close("all")
The plt.close method is used to close a figure window.
In [167]: ts = pd.Series(np.random.randn(1000), index=pd.date_range("1/1/2000", periods=1000))
In [168]: ts = ts.cumsum()
In [169]: ts.plot();
On a DataFrame, the DataFrame.plot method is a convenience to plot all of the columns with labels:
In [170]: df = pd.DataFrame(
.....: np.random.randn(1000, 4), index=ts.index, columns=["A", "B", "C", "D"]
.....: )
.....:
In [171]: df = df.cumsum()
In [172]: plt.figure();
In [173]: df.plot();
In [174]: plt.legend(loc='best');
IO Management#
CSV#
Writing to a csv file.
In [175]: df.to_csv("foo.csv")
Reading from a csv file.
In [176]: pd.read_csv("foo.csv")
Out[176]:
Unnamed: 0 A B C D
0 2000-01-01 4.293040 -0.263270 -0.449204 -0.107300
1 2000-01-02 3.683266 0.697799 -0.522885 -0.617935
2 2000-01-03 3.394994 0.889145 -0.891287 -1.544124
3 2000-01-04 2.585954 -0.395924 0.029842 -3.537277
4 2000-01-05 2.372088 -0.382483 -0.185600 -2.740550
.. ... ... ... ... ...
995 2002-09-22 21.785727 -37.691960 17.134546 -28.607807
996 2002-09-23 20.752620 -37.406264 16.697785 -29.420493
997 2002-09-24 21.213086 -38.618105 16.832882 -31.164423
998 2002-09-25 18.640338 -38.324714 16.479689 -33.188935
999 2002-09-26 17.734826 -38.764208 17.033938 -31.751839
[1000 rows x 5 columns]
Excel#
Reading and writing to MS Excel.
Writing to an excel file.
In [177]: df.to_excel("foo.xlsx", sheet_name="Sheet1")
Reading from an excel file.
In [178]: pd.read_excel("foo.xlsx", "Sheet1", index_col=None, na_values=["NA"])
Out[178]:
Unnamed: 0 A B C D
0 2000-01-01 4.293040 -0.263270 -0.449204 -0.107300
1 2000-01-02 3.683266 0.697799 -0.522885 -0.617935
2 2000-01-03 3.394994 0.889145 -0.891287 -1.544124
3 2000-01-04 2.585954 -0.395924 0.029842 -3.537277
4 2000-01-05 2.372088 -0.382483 -0.185600 -2.740550
.. ... ... ... ... ...
995 2002-09-22 21.785727 -37.691960 17.134546 -28.607807
996 2002-09-23 20.752620 -37.406264 16.697785 -29.420493
997 2002-09-24 21.213086 -38.618105 16.832882 -31.164423
998 2002-09-25 18.640338 -38.324714 16.479689 -33.188935
999 2002-09-26 17.734826 -38.764208 17.033938 -31.751839
[1000 rows x 5 columns]
Composition#
To compose one dataframe from multiple files, use glob library to get paths of files with random order and concatenate vertically and horizontally.
from glob import glob
files = sorted(glob("data/prefix_*.csv"))
pd.concat((pd.read_csv(file) for file in files), ignore_index=True)
pd.concat((pd.read_csv(file) for file in files), axis=1)
Miscellaneous#
Trim white spaces for object-dtype element.
for i in df: # traverse each column
if pd.api.types.is_object_dtype(df[i]): # if dtype is object
df[i] = df[i].str.strip() # remove white space
Copy from applications like Excel and paste through pandas as a DataFrame.
pd.read_clipboard()
Data wrangling with pandas can be referenced through an official cheatsheet.