SciPy 2011 pandas lightning talk
5 likes2,931 views
This document summarizes new and improved features in the pandas library. It describes how pandas provides labeled, structured arrays for heterogeneous data and time series functionality. Major updates include merging DataFrame and DataMatrix into a single optimized DataFrame, better handling of missing data, more robust IO capabilities, and improved pivoting and reshaping of data. Future plans include more powerful group by operations and integration with statsmodels.
![pandas?
In [13]: foo
Out[13]:
methyl1 age edu something indic
0 38.36 30to39 geCollege 1 False
1 37.85 lt30 geCollege 1 False
2 38.57 30to39 geCollege 1 False
3 39.75 30to39 geCollege 1 True
4 43.83 30to39 geCollege 1 True
5 39.08 30to39 ltHS 1 True
Size-mutable “labeled arrays” that
can handle heterogeneous data](https://image.slidesharecdn.com/scipy2011scipylightning-110714172006-phpapp01/85/SciPy-2011-pandas-lightning-talk-2-320.jpg)
![2. Fancier indexing
Mix boolean / integer / label /
slice-based indexing
df.ix[0]
df.ix[date1:date2]
df.ix[:5, ‘A’:’F’]
Setting works too
df.ix[df[‘A’] > 0, [‘B’, ‘C’, ‘D’]] = nan](https://image.slidesharecdn.com/scipy2011scipylightning-110714172006-phpapp01/85/SciPy-2011-pandas-lightning-talk-7-320.jpg)
![3. More robust IO
data_frame = read_csv(‘mydata.csv’)
data_frame2 = read_table(‘mydata.txt’, sep=‘t’,
skiprows=[1,2],
na_values=[‘#N/A NA’])
store = HDFStore(‘pytables.h5’)
store[‘a’] = data_frame
store[‘b’] = data_frame2](https://image.slidesharecdn.com/scipy2011scipylightning-110714172006-phpapp01/85/SciPy-2011-pandas-lightning-talk-8-320.jpg)
![4. Better pivoting / reshaping
In [29]: pivoted = df.pivot('bar', 'foo')
In [30]: pivoted['B']
Out[30]:
one two
a 1.664 0.7857
b 0.8306 -0.8615
c 0.3897 -1.846
d 0.6513 -1.19
e 1.938 -0.3339](https://image.slidesharecdn.com/scipy2011scipylightning-110714172006-phpapp01/85/SciPy-2011-pandas-lightning-talk-10-320.jpg)
![4. Better pivoting / reshaping
In [31]: pivoted.major_xs('a')
Out[31]:
A B C
one -0.0524 1.664 1.171
two 0.2073 0.7857 0.9051
In [32]: pivoted.minor_xs('one')
Out[32]:
A B C
a -0.0524 1.664 1.171
b 0.2514 0.8306 -1.396
c 0.1256 0.3897 0.5227
d -0.9301 0.6513 -0.2313
e 2.037 1.938 -0.3454](https://image.slidesharecdn.com/scipy2011scipylightning-110714172006-phpapp01/85/SciPy-2011-pandas-lightning-talk-11-320.jpg)
![4. Better pivoting / reshaping
In [30]: pivoted['B']
Out[30]:
one two
a 1.664 0.7857
b 0.8306 -0.8615
c 0.3897 -1.846
d 0.6513 -1.19
e 1.938 -0.3339](https://image.slidesharecdn.com/scipy2011scipylightning-110714172006-phpapp01/85/SciPy-2011-pandas-lightning-talk-12-320.jpg)
SciPy 2011 pandas lightning talk
- 1. What’s new and awesome in pandas
- 2. pandas? In [13]: foo Out[13]: methyl1 age edu something indic 0 38.36 30to39 geCollege 1 False 1 37.85 lt30 geCollege 1 False 2 38.57 30to39 geCollege 1 False 3 39.75 30to39 geCollege 1 True 4 43.83 30to39 geCollege 1 True 5 39.08 30to39 ltHS 1 True Size-mutable “labeled arrays” that can handle heterogeneous data
- 3. Kinda like a structured array?? • Automatic data alignment with lots of reshaping and indexing methods • Implicit and explicit handling of missing data • Easy time series functionality – Far less fuss than scikits.timeseries • Lots of in-memory SQL-like operations (group by, join, etc.)
- 4. pandas? • Extremely good for financial data – StackOverflow: “this is a beast of a financial analysis tool” • One of the better relational data munging tools in any language? • But also has maybe 60+% of what R users expect when they come to Python
- 5. 1. Heavily redesigned internals • Merged old DataFrame and DataMatrix into a single DataFrame: retain optimal performance where possible • Internal BlockManager class manages homogeneous ndarrays for optimal performance and reshaping
- 6. 1. Heavily redesigned internals • Better handling of missing data for non-floating point dtypes • Soon: DataFrame variant with N-dim “hyperslabs”
- 7. 2. Fancier indexing Mix boolean / integer / label / slice-based indexing df.ix[0] df.ix[date1:date2] df.ix[:5, ‘A’:’F’] Setting works too df.ix[df[‘A’] > 0, [‘B’, ‘C’, ‘D’]] = nan
- 8. 3. More robust IO data_frame = read_csv(‘mydata.csv’) data_frame2 = read_table(‘mydata.txt’, sep=‘t’, skiprows=[1,2], na_values=[‘#N/A NA’]) store = HDFStore(‘pytables.h5’) store[‘a’] = data_frame store[‘b’] = data_frame2
- 9. 4. Better pivoting / reshaping foo bar A B C 0 one a -0.0524 1.664 1.171 1 one a 0.2514 0.8306 -1.396 2 one b 0.1256 0.3897 0.5227 3 one b -0.9301 0.6513 -0.2313 4 one c 2.037 1.938 -0.3454 5 two a 0.2073 0.7857 0.9051 6 two a -1.032 -0.8615 1.028 7 two b -0.7319 -1.846 0.9294 8 two b 0.1004 -1.19 0.6043 9 two c -1.008 -0.3339 0.09522
- 10. 4. Better pivoting / reshaping In [29]: pivoted = df.pivot('bar', 'foo') In [30]: pivoted['B'] Out[30]: one two a 1.664 0.7857 b 0.8306 -0.8615 c 0.3897 -1.846 d 0.6513 -1.19 e 1.938 -0.3339
- 11. 4. Better pivoting / reshaping In [31]: pivoted.major_xs('a') Out[31]: A B C one -0.0524 1.664 1.171 two 0.2073 0.7857 0.9051 In [32]: pivoted.minor_xs('one') Out[32]: A B C a -0.0524 1.664 1.171 b 0.2514 0.8306 -1.396 c 0.1256 0.3897 0.5227 d -0.9301 0.6513 -0.2313 e 2.037 1.938 -0.3454
- 12. 4. Better pivoting / reshaping In [30]: pivoted['B'] Out[30]: one two a 1.664 0.7857 b 0.8306 -0.8615 c 0.3897 -1.846 d 0.6513 -1.19 e 1.938 -0.3339
- 13. 4. Some other things • “Sparse” (mostly NA) versions of data structures • Time zone support in DateRange • Generic moving window function rolling_apply
- 14. Near future • More powerful Group By • Flexible, fast frequency (time series) conversions • More integration with statsmodels
- 15. Thanks! • Hack: github.com/wesm/pandas • Twitter: @wesmckinn • Blog: blog.wesmckinney.com