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.tensor as mt In [2]: import mars.dataframe as md
Creating a Series by passing a list of values, letting it create a default integer index:
Series
In [3]: s = md.Series([1, 3, 5, mt.nan, 6, 8]) In [4]: s.execute() 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 Mars tensor, with a datetime index and labeled columns:
DataFrame
In [5]: dates = md.date_range('20130101', periods=6) In [6]: dates.execute() Out[6]: 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 [7]: df = md.DataFrame(mt.random.randn(6, 4), index=dates, columns=list('ABCD')) In [8]: df.execute() Out[8]: A B C D 2013-01-01 -0.924740 0.770940 0.848022 0.000714 2013-01-02 -0.115180 -1.186477 1.455773 -1.009835 2013-01-03 0.648884 -1.742358 -2.110255 0.699699 2013-01-04 -0.896207 -0.995048 0.680549 -1.133533 2013-01-05 -0.546100 0.971047 -0.431097 -0.960006 2013-01-06 -0.373037 1.217092 -1.576501 -1.003959
Creating a DataFrame by passing a dict of objects that can be converted to series-like.
In [9]: 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 [10]: df2.execute() Out[10]: 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 [11]: df2.dtypes Out[11]: A float64 B datetime64[ns] C float32 D int32 E object dtype: object
Here is how to view the top and bottom rows of the frame:
In [12]: df.head().execute() Out[12]: A B C D 2013-01-01 -0.924740 0.770940 0.848022 0.000714 2013-01-02 -0.115180 -1.186477 1.455773 -1.009835 2013-01-03 0.648884 -1.742358 -2.110255 0.699699 2013-01-04 -0.896207 -0.995048 0.680549 -1.133533 2013-01-05 -0.546100 0.971047 -0.431097 -0.960006 In [13]: df.tail(3).execute() Out[13]: A B C D 2013-01-04 -0.896207 -0.995048 0.680549 -1.133533 2013-01-05 -0.546100 0.971047 -0.431097 -0.960006 2013-01-06 -0.373037 1.217092 -1.576501 -1.003959
Display the index, columns:
In [14]: df.index.execute() Out[14]: 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 [15]: df.columns.execute() Out[15]: 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.
DataFrame.to_tensor()
object
For df, our DataFrame of all floating-point values, DataFrame.to_tensor() is fast and doesn’t require copying data.
df
In [16]: df.to_tensor().execute() Out[16]: array([[-9.24739857e-01, 7.70939607e-01, 8.48021684e-01, 7.13706225e-04], [-1.15179745e-01, -1.18647693e+00, 1.45577294e+00, -1.00983483e+00], [ 6.48883680e-01, -1.74235772e+00, -2.11025488e+00, 6.99698627e-01], [-8.96206921e-01, -9.95047578e-01, 6.80548829e-01, -1.13353310e+00], [-5.46100357e-01, 9.71047404e-01, -4.31096994e-01, -9.60005508e-01], [-3.73036572e-01, 1.21709221e+00, -1.57650131e+00, -1.00395859e+00]])
For df2, the DataFrame with multiple dtypes, DataFrame.to_tensor() is relatively expensive.
df2
In [17]: df2.to_tensor().execute() Out[17]: 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:
describe()
In [18]: df.describe().execute() Out[18]: A B C D count 6.000000 6.000000 6.000000 6.000000 mean -0.367730 -0.160801 -0.188918 -0.567820 std 0.586290 1.288186 1.429111 0.746896 min -0.924740 -1.742358 -2.110255 -1.133533 25% -0.808680 -1.138620 -1.290150 -1.008366 50% -0.459568 -0.112054 0.124726 -0.981982 75% -0.179644 0.921020 0.806153 -0.239466 max 0.648884 1.217092 1.455773 0.699699
Sorting by an axis:
In [19]: df.sort_index(axis=1, ascending=False).execute() Out[19]: D C B A 2013-01-01 0.000714 0.848022 0.770940 -0.924740 2013-01-02 -1.009835 1.455773 -1.186477 -0.115180 2013-01-03 0.699699 -2.110255 -1.742358 0.648884 2013-01-04 -1.133533 0.680549 -0.995048 -0.896207 2013-01-05 -0.960006 -0.431097 0.971047 -0.546100 2013-01-06 -1.003959 -1.576501 1.217092 -0.373037
Sorting by values:
In [20]: df.sort_values(by='B').execute() Out[20]: A B C D 2013-01-03 0.648884 -1.742358 -2.110255 0.699699 2013-01-02 -0.115180 -1.186477 1.455773 -1.009835 2013-01-04 -0.896207 -0.995048 0.680549 -1.133533 2013-01-01 -0.924740 0.770940 0.848022 0.000714 2013-01-05 -0.546100 0.971047 -0.431097 -0.960006 2013-01-06 -0.373037 1.217092 -1.576501 -1.003959
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.
.at
.iat
.loc
.iloc
Selecting a single column, which yields a Series, equivalent to df.A:
df.A
In [21]: df['A'].execute() Out[21]: 2013-01-01 -0.924740 2013-01-02 -0.115180 2013-01-03 0.648884 2013-01-04 -0.896207 2013-01-05 -0.546100 2013-01-06 -0.373037 Freq: D, Name: A, dtype: float64
Selecting via [], which slices the rows.
[]
In [22]: df[0:3].execute() Out[22]: A B C D 2013-01-01 -0.924740 0.770940 0.848022 0.000714 2013-01-02 -0.115180 -1.186477 1.455773 -1.009835 2013-01-03 0.648884 -1.742358 -2.110255 0.699699 In [23]: df['20130102':'20130104'].execute() Out[23]: A B C D 2013-01-02 -0.115180 -1.186477 1.455773 -1.009835 2013-01-03 0.648884 -1.742358 -2.110255 0.699699 2013-01-04 -0.896207 -0.995048 0.680549 -1.133533
For getting a cross section using a label:
In [24]: df.loc['20130101'].execute() Out[24]: A -0.924740 B 0.770940 C 0.848022 D 0.000714 Name: 2013-01-01 00:00:00, dtype: float64
Selecting on a multi-axis by label:
In [25]: df.loc[:, ['A', 'B']].execute() Out[25]: A B 2013-01-01 -0.924740 0.770940 2013-01-02 -0.115180 -1.186477 2013-01-03 0.648884 -1.742358 2013-01-04 -0.896207 -0.995048 2013-01-05 -0.546100 0.971047 2013-01-06 -0.373037 1.217092
Showing label slicing, both endpoints are included:
In [26]: df.loc['20130102':'20130104', ['A', 'B']].execute() Out[26]: A B 2013-01-02 -0.115180 -1.186477 2013-01-03 0.648884 -1.742358 2013-01-04 -0.896207 -0.995048
Reduction in the dimensions of the returned object:
In [27]: df.loc['20130102', ['A', 'B']].execute() Out[27]: A -0.115180 B -1.186477 Name: 2013-01-02 00:00:00, dtype: float64
For getting a scalar value:
In [28]: df.loc['20130101', 'A'].execute() Out[28]: -0.9247398569374877
For getting fast access to a scalar (equivalent to the prior method):
In [29]: df.at['20130101', 'A'].execute() Out[29]: -0.9247398569374877
Select via the position of the passed integers:
In [30]: df.iloc[3].execute() Out[30]: A -0.896207 B -0.995048 C 0.680549 D -1.133533 Name: 2013-01-04 00:00:00, dtype: float64
By integer slices, acting similar to numpy/python:
In [31]: df.iloc[3:5, 0:2].execute() Out[31]: A B 2013-01-04 -0.896207 -0.995048 2013-01-05 -0.546100 0.971047
By lists of integer position locations, similar to the numpy/python style:
In [32]: df.iloc[[1, 2, 4], [0, 2]].execute() Out[32]: A C 2013-01-02 -0.115180 1.455773 2013-01-03 0.648884 -2.110255 2013-01-05 -0.546100 -0.431097
For slicing rows explicitly:
In [33]: df.iloc[1:3, :].execute() Out[33]: A B C D 2013-01-02 -0.115180 -1.186477 1.455773 -1.009835 2013-01-03 0.648884 -1.742358 -2.110255 0.699699
For slicing columns explicitly:
In [34]: df.iloc[:, 1:3].execute() Out[34]: B C 2013-01-01 0.770940 0.848022 2013-01-02 -1.186477 1.455773 2013-01-03 -1.742358 -2.110255 2013-01-04 -0.995048 0.680549 2013-01-05 0.971047 -0.431097 2013-01-06 1.217092 -1.576501
For getting a value explicitly:
In [35]: df.iloc[1, 1].execute() Out[35]: -1.1864769305715868
In [36]: df.iat[1, 1].execute() Out[36]: -1.1864769305715868
Using a single column’s values to select data.
In [37]: df[df['A'] > 0].execute() Out[37]: A B C D 2013-01-03 0.648884 -1.742358 -2.110255 0.699699
Selecting values from a DataFrame where a boolean condition is met.
In [38]: df[df > 0].execute() Out[38]: A B C D 2013-01-01 NaN 0.770940 0.848022 0.000714 2013-01-02 NaN NaN 1.455773 NaN 2013-01-03 0.648884 NaN NaN 0.699699 2013-01-04 NaN NaN 0.680549 NaN 2013-01-05 NaN 0.971047 NaN NaN 2013-01-06 NaN 1.217092 NaN NaN
Operations in general exclude missing data.
Performing a descriptive statistic:
In [39]: df.mean().execute() Out[39]: A -0.367730 B -0.160801 C -0.188918 D -0.567820 dtype: float64
Same operation on the other axis:
In [40]: df.mean(1).execute() Out[40]: 2013-01-01 0.173734 2013-01-02 -0.213930 2013-01-03 -0.626008 2013-01-04 -0.586060 2013-01-05 -0.241539 2013-01-06 -0.434101 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 [41]: s = md.Series([1, 3, 5, mt.nan, 6, 8], index=dates).shift(2) In [42]: s.execute() Out[42]: 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 [43]: df.sub(s, axis='index').execute() Out[43]: A B C D 2013-01-01 NaN NaN NaN NaN 2013-01-02 NaN NaN NaN NaN 2013-01-03 -0.351116 -2.742358 -3.110255 -0.300301 2013-01-04 -3.896207 -3.995048 -2.319451 -4.133533 2013-01-05 -5.546100 -4.028953 -5.431097 -5.960006 2013-01-06 NaN NaN NaN NaN
Applying functions to the data:
In [44]: df.apply(lambda x: x.max() - x.min()).execute() Out[44]: A 1.573624 B 2.959450 C 3.566028 D 1.833232 dtype: float64
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 [45]: s = md.Series(['A', 'B', 'C', 'Aaba', 'Baca', mt.nan, 'CABA', 'dog', 'cat']) In [46]: s.str.lower().execute() Out[46]: 0 a 1 b 2 c 3 aaba 4 baca 5 NaN 6 caba 7 dog 8 cat dtype: object
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():
concat()
In [47]: df = md.DataFrame(mt.random.randn(10, 4)) In [48]: df.execute() Out[48]: 0 1 2 3 0 -1.192854 -0.162735 0.850603 -0.503124 1 1.850383 0.167894 -0.793017 -0.214691 2 -0.473524 1.010253 0.632031 -0.425816 3 -0.682673 0.125081 1.528010 -2.036973 4 2.118904 -1.092768 0.935487 -1.521369 5 -0.443550 1.159571 1.237090 1.048793 6 -0.674023 -0.225085 0.594600 -0.483192 7 1.289236 2.228827 -2.402175 0.057371 8 -0.314895 2.010411 -0.432206 0.652077 9 1.480547 0.409850 -1.030584 0.328009 # break it into pieces In [49]: pieces = [df[:3], df[3:7], df[7:]] In [50]: md.concat(pieces).execute() Out[50]: 0 1 2 3 0 -1.192854 -0.162735 0.850603 -0.503124 1 1.850383 0.167894 -0.793017 -0.214691 2 -0.473524 1.010253 0.632031 -0.425816 3 -0.682673 0.125081 1.528010 -2.036973 4 2.118904 -1.092768 0.935487 -1.521369 5 -0.443550 1.159571 1.237090 1.048793 6 -0.674023 -0.225085 0.594600 -0.483192 7 1.289236 2.228827 -2.402175 0.057371 8 -0.314895 2.010411 -0.432206 0.652077 9 1.480547 0.409850 -1.030584 0.328009
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.
SQL style merges. See the Database style joining section.
In [51]: left = md.DataFrame({'key': ['foo', 'foo'], 'lval': [1, 2]}) In [52]: right = md.DataFrame({'key': ['foo', 'foo'], 'rval': [4, 5]}) In [53]: left.execute() Out[53]: key lval 0 foo 1 1 foo 2 In [54]: right.execute() Out[54]: key rval 0 foo 4 1 foo 5 In [55]: md.merge(left, right, on='key').execute() Out[55]: 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 [56]: left = md.DataFrame({'key': ['foo', 'bar'], 'lval': [1, 2]}) In [57]: right = md.DataFrame({'key': ['foo', 'bar'], 'rval': [4, 5]}) In [58]: left.execute() Out[58]: key lval 0 foo 1 1 bar 2 In [59]: right.execute() Out[59]: key rval 0 foo 4 1 bar 5 In [60]: md.merge(left, right, on='key').execute() Out[60]: key lval rval 0 foo 1 4 1 bar 2 5
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
Splitting the data into groups based on some criteria
Applying a function to each group independently
Combining the results into a data structure
In [61]: 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 [62]: df.execute() Out[62]: A B C D 0 foo one -0.444180 1.652666 1 bar one -0.053441 0.804783 2 foo two 0.957405 0.073286 3 bar three 0.019604 0.970040 4 foo two 1.706177 -0.796002 5 bar two -0.287955 -0.351346 6 foo one 0.086541 -1.087024 7 foo three 0.545263 -0.327195
Grouping and then applying the sum() function to the resulting groups.
sum()
In [63]: df.groupby('A').sum().execute() Out[63]: C D A bar -0.321792 1.423477 foo 2.851206 -0.484268
Grouping by multiple columns forms a hierarchical index, and again we can apply the sum function.
In [64]: df.groupby(['A', 'B']).sum().execute() Out[64]: C D A B foo one -0.357639 0.565642 two 2.663583 -0.722715 three 0.545263 -0.327195 bar one -0.053441 0.804783 two -0.287955 -0.351346 three 0.019604 0.970040
We use the standard convention for referencing the matplotlib API:
In [65]: import matplotlib.pyplot as plt In [66]: plt.close('all')
In [67]: ts = md.Series(mt.random.randn(1000), ....: index=md.date_range('1/1/2000', periods=1000)) ....: In [68]: ts = ts.cumsum() In [69]: ts.plot() Out[69]: <AxesSubplot:>
On a DataFrame, the plot() method is a convenience to plot all of the columns with labels:
plot()
In [70]: df = md.DataFrame(mt.random.randn(1000, 4), index=ts.index, ....: columns=['A', 'B', 'C', 'D']) ....: In [71]: df = df.cumsum() In [72]: plt.figure() Out[72]: <Figure size 640x480 with 0 Axes> In [73]: df.plot() Out[73]: <AxesSubplot:> In [74]: plt.legend(loc='best') Out[74]: <matplotlib.legend.Legend at 0x7f35267a1b10>
In [75]: df.to_csv('foo.csv').execute() Out[75]: Empty DataFrame Columns: [] Index: []
Reading from a csv file.
In [76]: md.read_csv('foo.csv').execute() Out[76]: Unnamed: 0 A B C D 0 2000-01-01 0.171792 -1.428496 -0.590932 0.212112 1 2000-01-02 1.606502 -1.694335 0.460951 -2.167250 2 2000-01-03 0.108290 -1.976547 0.753892 -1.838001 3 2000-01-04 1.045008 -1.335184 -0.878818 -0.033611 4 2000-01-05 1.395585 -2.295376 -2.711310 -0.185244 .. ... ... ... ... ... 995 2002-09-22 6.698322 92.609805 -30.091714 -26.088067 996 2002-09-23 5.041494 90.951700 -30.017777 -26.234808 997 2002-09-24 5.891865 92.485130 -28.675356 -25.813865 998 2002-09-25 6.164564 93.698065 -28.802962 -26.113199 999 2002-09-26 6.137191 92.348835 -28.728546 -27.055299 [1000 rows x 5 columns]