10 minutes to Mars DataFrame#
This is a short introduction to Mars DataFrame which is originated from 10 minutes to pandas.
Customarily, we import as follows:
In [1]: import mars
In [2]: import mars.tensor as mt
In [3]: import mars.dataframe as md
Now create a new default session.
In [4]: mars.new_session()
Out[4]: <mars.deploy.oscar.session.SyncSession at 0x7f30738e4e50>
Object creation#
Creating a Series by passing a list of values, letting it create
a default integer index:
In [5]: s = md.Series([1, 3, 5, mt.nan, 6, 8])
In [6]: s.execute()
Out[6]:
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 Mars tensor, with a datetime index
and labeled columns:
In [7]: dates = md.date_range('20130101', periods=6)
In [8]: dates.execute()
Out[8]:
DatetimeIndex(['2013-01-01', '2013-01-02', '2013-01-03', '2013-01-04',
'2013-01-05', '2013-01-06'],
dtype='datetime64[ns]', freq='D')
In [9]: df = md.DataFrame(mt.random.randn(6, 4), index=dates, columns=list('ABCD'))
In [10]: df.execute()
Out[10]:
A B C D
2013-01-01 0.691614 -1.105920 -0.100480 1.124014
2013-01-02 -0.563252 0.615544 -0.629121 -1.613738
2013-01-03 0.117056 -1.281171 -0.048552 -0.403358
2013-01-04 0.265501 -0.504812 0.317328 1.150110
2013-01-05 -0.660179 0.723665 -1.577349 -0.502156
2013-01-06 0.413119 2.180548 0.522483 -0.214173
Creating a DataFrame by passing a dict of objects that can be converted to series-like.
In [11]: df2 = md.DataFrame({'A': 1.,
....: 'B': md.Timestamp('20130102'),
....: 'C': md.Series(1, index=list(range(4)), dtype='float32'),
....: 'D': mt.array([3] * 4, dtype='int32'),
....: 'E': 'foo'})
....:
In [12]: df2.execute()
Out[12]:
A B C D E
0 1.0 2013-01-02 1.0 3 foo
1 1.0 2013-01-02 1.0 3 foo
2 1.0 2013-01-02 1.0 3 foo
3 1.0 2013-01-02 1.0 3 foo
The columns of the resulting DataFrame have different dtypes.
In [13]: df2.dtypes
Out[13]:
A float64
B datetime64[ns]
C float32
D int32
E object
dtype: object
Viewing data#
Here is how to view the top and bottom rows of the frame:
In [14]: df.head().execute()
Out[14]:
A B C D
2013-01-01 0.691614 -1.105920 -0.100480 1.124014
2013-01-02 -0.563252 0.615544 -0.629121 -1.613738
2013-01-03 0.117056 -1.281171 -0.048552 -0.403358
2013-01-04 0.265501 -0.504812 0.317328 1.150110
2013-01-05 -0.660179 0.723665 -1.577349 -0.502156
In [15]: df.tail(3).execute()
Out[15]:
A B C D
2013-01-04 0.265501 -0.504812 0.317328 1.150110
2013-01-05 -0.660179 0.723665 -1.577349 -0.502156
2013-01-06 0.413119 2.180548 0.522483 -0.214173
Display the index, columns:
In [16]: df.index.execute()
Out[16]:
DatetimeIndex(['2013-01-01', '2013-01-02', '2013-01-03', '2013-01-04',
'2013-01-05', '2013-01-06'],
dtype='datetime64[ns]', freq='D')
In [17]: df.columns.execute()
Out[17]: Index(['A', 'B', 'C', 'D'], dtype='object')
DataFrame.to_tensor() gives a Mars tensor 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 DataFrame and tensor: tensors have one dtype for the entire tensor,
while DataFrames have one dtype per column. When you call
DataFrame.to_tensor(), Mars DataFrame will find the tensor 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_tensor() is fast and doesn’t require copying data.
In [18]: df.to_tensor().execute()
Out[18]:
array([[ 0.69161369, -1.10591963, -0.10047961, 1.12401441],
[-0.56325231, 0.61554437, -0.62912078, -1.61373767],
[ 0.11705591, -1.28117091, -0.04855218, -0.40335838],
[ 0.26550144, -0.50481205, 0.31732827, 1.15010997],
[-0.66017923, 0.72366454, -1.57734917, -0.50215598],
[ 0.41311872, 2.18054793, 0.52248332, -0.21417311]])
For df2, the DataFrame with multiple dtypes,
DataFrame.to_tensor() is relatively expensive.
In [19]: df2.to_tensor().execute()
Out[19]:
array([[1.0, Timestamp('2013-01-02 00:00:00'), 1.0, 3, 'foo'],
[1.0, Timestamp('2013-01-02 00:00:00'), 1.0, 3, 'foo'],
[1.0, Timestamp('2013-01-02 00:00:00'), 1.0, 3, 'foo'],
[1.0, Timestamp('2013-01-02 00:00:00'), 1.0, 3, 'foo']],
dtype=object)
Note
DataFrame.to_tensor() does not include the index or column
labels in the output.
describe() shows a quick statistic summary of your data:
In [20]: df.describe().execute()
Out[20]:
A B C D
count 6.000000 6.000000 6.000000 6.000000
mean 0.043976 0.104642 -0.252615 -0.076550
std 0.543079 1.319922 0.759879 1.059800
min -0.660179 -1.281171 -1.577349 -1.613738
25% -0.393175 -0.955643 -0.496960 -0.477457
50% 0.191279 0.055366 -0.074516 -0.308766
75% 0.376214 0.696634 0.225858 0.789468
max 0.691614 2.180548 0.522483 1.150110
Sorting by an axis:
In [21]: df.sort_index(axis=1, ascending=False).execute()
Out[21]:
D C B A
2013-01-01 1.124014 -0.100480 -1.105920 0.691614
2013-01-02 -1.613738 -0.629121 0.615544 -0.563252
2013-01-03 -0.403358 -0.048552 -1.281171 0.117056
2013-01-04 1.150110 0.317328 -0.504812 0.265501
2013-01-05 -0.502156 -1.577349 0.723665 -0.660179
2013-01-06 -0.214173 0.522483 2.180548 0.413119
Sorting by values:
In [22]: df.sort_values(by='B').execute()
Out[22]:
A B C D
2013-01-03 0.117056 -1.281171 -0.048552 -0.403358
2013-01-01 0.691614 -1.105920 -0.100480 1.124014
2013-01-04 0.265501 -0.504812 0.317328 1.150110
2013-01-02 -0.563252 0.615544 -0.629121 -1.613738
2013-01-05 -0.660179 0.723665 -1.577349 -0.502156
2013-01-06 0.413119 2.180548 0.522483 -0.214173
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 DataFrame data access methods, .at, .iat,
.loc and .iloc.
Getting#
Selecting a single column, which yields a Series,
equivalent to df.A:
In [23]: df['A'].execute()
Out[23]:
2013-01-01 0.691614
2013-01-02 -0.563252
2013-01-03 0.117056
2013-01-04 0.265501
2013-01-05 -0.660179
2013-01-06 0.413119
Freq: D, Name: A, dtype: float64
Selecting via [], which slices the rows.
In [24]: df[0:3].execute()
Out[24]:
A B C D
2013-01-01 0.691614 -1.105920 -0.100480 1.124014
2013-01-02 -0.563252 0.615544 -0.629121 -1.613738
2013-01-03 0.117056 -1.281171 -0.048552 -0.403358
In [25]: df['20130102':'20130104'].execute()
Out[25]:
A B C D
2013-01-02 -0.563252 0.615544 -0.629121 -1.613738
2013-01-03 0.117056 -1.281171 -0.048552 -0.403358
2013-01-04 0.265501 -0.504812 0.317328 1.150110
Selection by label#
For getting a cross section using a label:
In [26]: df.loc['20130101'].execute()
Out[26]:
A 0.691614
B -1.105920
C -0.100480
D 1.124014
Name: 2013-01-01 00:00:00, dtype: float64
Selecting on a multi-axis by label:
In [27]: df.loc[:, ['A', 'B']].execute()
Out[27]:
A B
2013-01-01 0.691614 -1.105920
2013-01-02 -0.563252 0.615544
2013-01-03 0.117056 -1.281171
2013-01-04 0.265501 -0.504812
2013-01-05 -0.660179 0.723665
2013-01-06 0.413119 2.180548
Showing label slicing, both endpoints are included:
In [28]: df.loc['20130102':'20130104', ['A', 'B']].execute()
Out[28]:
A B
2013-01-02 -0.563252 0.615544
2013-01-03 0.117056 -1.281171
2013-01-04 0.265501 -0.504812
Reduction in the dimensions of the returned object:
In [29]: df.loc['20130102', ['A', 'B']].execute()
Out[29]:
A -0.563252
B 0.615544
Name: 2013-01-02 00:00:00, dtype: float64
For getting a scalar value:
In [30]: df.loc['20130101', 'A'].execute()
Out[30]: 0.6916136887164693
For getting fast access to a scalar (equivalent to the prior method):
In [31]: df.at['20130101', 'A'].execute()
Out[31]: 0.6916136887164693
Selection by position#
Select via the position of the passed integers:
In [32]: df.iloc[3].execute()
Out[32]:
A 0.265501
B -0.504812
C 0.317328
D 1.150110
Name: 2013-01-04 00:00:00, dtype: float64
By integer slices, acting similar to numpy/python:
In [33]: df.iloc[3:5, 0:2].execute()
Out[33]:
A B
2013-01-04 0.265501 -0.504812
2013-01-05 -0.660179 0.723665
By lists of integer position locations, similar to the numpy/python style:
In [34]: df.iloc[[1, 2, 4], [0, 2]].execute()
Out[34]:
A C
2013-01-02 -0.563252 -0.629121
2013-01-03 0.117056 -0.048552
2013-01-05 -0.660179 -1.577349
For slicing rows explicitly:
In [35]: df.iloc[1:3, :].execute()
Out[35]:
A B C D
2013-01-02 -0.563252 0.615544 -0.629121 -1.613738
2013-01-03 0.117056 -1.281171 -0.048552 -0.403358
For slicing columns explicitly:
In [36]: df.iloc[:, 1:3].execute()
Out[36]:
B C
2013-01-01 -1.105920 -0.100480
2013-01-02 0.615544 -0.629121
2013-01-03 -1.281171 -0.048552
2013-01-04 -0.504812 0.317328
2013-01-05 0.723665 -1.577349
2013-01-06 2.180548 0.522483
For getting a value explicitly:
In [37]: df.iloc[1, 1].execute()
Out[37]: 0.6155443677589363
For getting fast access to a scalar (equivalent to the prior method):
In [38]: df.iat[1, 1].execute()
Out[38]: 0.6155443677589363
Boolean indexing#
Using a single column’s values to select data.
In [39]: df[df['A'] > 0].execute()
Out[39]:
A B C D
2013-01-01 0.691614 -1.105920 -0.100480 1.124014
2013-01-03 0.117056 -1.281171 -0.048552 -0.403358
2013-01-04 0.265501 -0.504812 0.317328 1.150110
2013-01-06 0.413119 2.180548 0.522483 -0.214173
Selecting values from a DataFrame where a boolean condition is met.
In [40]: df[df > 0].execute()
Out[40]:
A B C D
2013-01-01 0.691614 NaN NaN 1.124014
2013-01-02 NaN 0.615544 NaN NaN
2013-01-03 0.117056 NaN NaN NaN
2013-01-04 0.265501 NaN 0.317328 1.150110
2013-01-05 NaN 0.723665 NaN NaN
2013-01-06 0.413119 2.180548 0.522483 NaN
Operations#
Stats#
Operations in general exclude missing data.
Performing a descriptive statistic:
In [41]: df.mean().execute()
Out[41]:
A 0.043976
B 0.104642
C -0.252615
D -0.076550
dtype: float64
Same operation on the other axis:
In [42]: df.mean(1).execute()
Out[42]:
2013-01-01 0.152307
2013-01-02 -0.547642
2013-01-03 -0.404006
2013-01-04 0.307032
2013-01-05 -0.504005
2013-01-06 0.725494
Freq: D, dtype: float64
Operating with objects that have different dimensionality and need alignment. In addition, Mars DataFrame automatically broadcasts along the specified dimension.
In [43]: s = md.Series([1, 3, 5, mt.nan, 6, 8], index=dates).shift(2)
In [44]: s.execute()
Out[44]:
2013-01-01 NaN
2013-01-02 NaN
2013-01-03 1.0
2013-01-04 3.0
2013-01-05 5.0
2013-01-06 NaN
Freq: D, dtype: float64
In [45]: df.sub(s, axis='index').execute()
Out[45]:
A B C D
2013-01-01 NaN NaN NaN NaN
2013-01-02 NaN NaN NaN NaN
2013-01-03 -0.882944 -2.281171 -1.048552 -1.403358
2013-01-04 -2.734499 -3.504812 -2.682672 -1.849890
2013-01-05 -5.660179 -4.276335 -6.577349 -5.502156
2013-01-06 NaN NaN NaN NaN
Apply#
Applying functions to the data:
In [46]: df.apply(lambda x: x.max() - x.min()).execute()
Out[46]:
A 1.351793
B 3.461719
C 2.099832
D 2.763848
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). See more at Vectorized String Methods.
In [47]: s = md.Series(['A', 'B', 'C', 'Aaba', 'Baca', mt.nan, 'CABA', 'dog', 'cat'])
In [48]: s.str.lower().execute()
Out[48]:
0 a
1 b
2 c
3 aaba
4 baca
5 NaN
6 caba
7 dog
8 cat
dtype: object
Merge#
Concat#
Mars DataFrame 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.
Concatenating DataFrame objects together with concat():
In [49]: df = md.DataFrame(mt.random.randn(10, 4))
In [50]: df.execute()
Out[50]:
0 1 2 3
0 -0.744246 -0.820294 -0.218120 0.362971
1 -1.683835 -0.068356 0.001055 -0.629065
2 -0.851774 0.640352 -0.308788 -0.046406
3 -1.819059 0.263546 1.096967 1.692534
4 -2.515883 -0.037570 0.235536 -0.462660
5 -0.936556 0.693611 1.061610 1.087943
6 1.230238 -0.068336 -0.945819 -1.502882
7 -0.946746 -0.222657 0.410326 1.412804
8 -0.379649 0.089457 -0.378820 -0.848689
9 0.944580 0.227600 1.040578 0.195827
# break it into pieces
In [51]: pieces = [df[:3], df[3:7], df[7:]]
In [52]: md.concat(pieces).execute()
Out[52]:
0 1 2 3
0 -0.744246 -0.820294 -0.218120 0.362971
1 -1.683835 -0.068356 0.001055 -0.629065
2 -0.851774 0.640352 -0.308788 -0.046406
3 -1.819059 0.263546 1.096967 1.692534
4 -2.515883 -0.037570 0.235536 -0.462660
5 -0.936556 0.693611 1.061610 1.087943
6 1.230238 -0.068336 -0.945819 -1.502882
7 -0.946746 -0.222657 0.410326 1.412804
8 -0.379649 0.089457 -0.378820 -0.848689
9 0.944580 0.227600 1.040578 0.195827
Join#
SQL style merges. See the Database style joining section.
In [53]: left = md.DataFrame({'key': ['foo', 'foo'], 'lval': [1, 2]})
In [54]: right = md.DataFrame({'key': ['foo', 'foo'], 'rval': [4, 5]})
In [55]: left.execute()
Out[55]:
key lval
0 foo 1
1 foo 2
In [56]: right.execute()
Out[56]:
key rval
0 foo 4
1 foo 5
In [57]: md.merge(left, right, on='key').execute()
Out[57]:
key lval rval
0 foo 1 4
1 foo 1 5
2 foo 2 4
3 foo 2 5
Another example that can be given is:
In [58]: left = md.DataFrame({'key': ['foo', 'bar'], 'lval': [1, 2]})
In [59]: right = md.DataFrame({'key': ['foo', 'bar'], 'rval': [4, 5]})
In [60]: left.execute()
Out[60]:
key lval
0 foo 1
1 bar 2
In [61]: right.execute()
Out[61]:
key rval
0 foo 4
1 bar 5
In [62]: md.merge(left, right, on='key').execute()
Out[62]:
key lval rval
0 foo 1 4
1 bar 2 5
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 [63]: df = md.DataFrame({'A': ['foo', 'bar', 'foo', 'bar',
....: 'foo', 'bar', 'foo', 'foo'],
....: 'B': ['one', 'one', 'two', 'three',
....: 'two', 'two', 'one', 'three'],
....: 'C': mt.random.randn(8),
....: 'D': mt.random.randn(8)})
....:
In [64]: df.execute()
Out[64]:
A B C D
0 foo one -0.665916 0.938514
1 bar one 0.935748 1.129597
2 foo two -1.309126 0.315047
3 bar three 1.552172 -0.149467
4 foo two -1.291504 0.313293
5 bar two -1.589682 -0.283976
6 foo one 2.244470 -1.790700
7 foo three 0.316835 1.152761
Grouping and then applying the sum() function to the resulting
groups.
In [65]: df.groupby('A').sum().execute()
Out[65]:
C D
A
bar 0.898238 0.696155
foo -0.705240 0.928915
Grouping by multiple columns forms a hierarchical index, and again we can apply the sum function.
In [66]: df.groupby(['A', 'B']).sum().execute()
Out[66]:
C D
A B
bar one 0.935748 1.129597
three 1.552172 -0.149467
two -1.589682 -0.283976
foo one 1.578555 -0.852186
three 0.316835 1.152761
two -2.600630 0.628340
Plotting#
We use the standard convention for referencing the matplotlib API:
In [67]: import matplotlib.pyplot as plt
In [68]: plt.close('all')
In [69]: ts = md.Series(mt.random.randn(1000),
....: index=md.date_range('1/1/2000', periods=1000))
....:
In [70]: ts = ts.cumsum()
In [71]: ts.plot()
Out[71]: <AxesSubplot:>
On a DataFrame, the plot() method is a convenience to plot all
of the columns with labels:
In [72]: df = md.DataFrame(mt.random.randn(1000, 4), index=ts.index,
....: columns=['A', 'B', 'C', 'D'])
....:
In [73]: df = df.cumsum()
In [74]: plt.figure()
Out[74]: <Figure size 640x480 with 0 Axes>
In [75]: df.plot()
Out[75]: <AxesSubplot:>
In [76]: plt.legend(loc='best')
Out[76]: <matplotlib.legend.Legend at 0x7f307842a850>
Getting data in/out#
CSV#
In [77]: df.to_csv('foo.csv').execute()
Out[77]:
Empty DataFrame
Columns: []
Index: []
In [78]: md.read_csv('foo.csv').execute()
Out[78]:
Unnamed: 0 A B C D
0 2000-01-01 -0.148634 0.905379 -1.346986 0.256851
1 2000-01-02 2.242298 0.296500 -0.260887 0.248435
2 2000-01-03 2.828900 -0.649995 -1.834000 -0.136707
3 2000-01-04 0.338252 -0.974826 -0.755381 1.289598
4 2000-01-05 -0.593980 -1.436103 0.351157 1.847797
.. ... ... ... ... ...
995 2002-09-22 17.130050 -66.730427 -41.623043 -8.845876
996 2002-09-23 17.449925 -67.154070 -40.873604 -10.577547
997 2002-09-24 17.491354 -68.654259 -40.193393 -9.598333
998 2002-09-25 16.675418 -67.501809 -39.500102 -7.878921
999 2002-09-26 15.723870 -69.473105 -39.332906 -9.419267
[1000 rows x 5 columns]