This is a minor release from 0.13.0 and includes a small number of API changes, several new features, enhancements, and performance improvements along with a large number of bug fixes. We recommend that all users upgrade to this version.
Highlights include:
Added infer_datetime_format keyword to read_csv/to_datetime to allow speedups for homogeneously formatted datetimes.
infer_datetime_format
read_csv/to_datetime
Will intelligently limit display precision for datetime/timedelta formats.
Enhanced Panel apply() method.
apply()
Suggested tutorials in new Tutorials section.
Our pandas ecosystem is growing, We now feature related projects in a new Pandas Ecosystem section.
Much work has been taking place on improving the docs, and a new Contributing section has been added.
Even though it may only be of interest to devs, we <3 our new CI status page: ScatterCI.
Warning
0.13.1 fixes a bug that was caused by a combination of having numpy < 1.8, and doing chained assignment on a string-like array. Please review the docs, chained indexing can have unexpected results and should generally be avoided.
This would previously segfault:
In [1]: df = pd.DataFrame({"A": np.array(["foo", "bar", "bah", "foo", "bar"])}) In [2]: df["A"].iloc[0] = np.nan In [3]: df Out[3]: A 0 NaN 1 bar 2 bah 3 foo 4 bar
The recommended way to do this type of assignment is:
In [4]: df = pd.DataFrame({"A": np.array(["foo", "bar", "bah", "foo", "bar"])}) In [5]: df.loc[0, "A"] = np.nan In [6]: df Out[6]: A 0 NaN 1 bar 2 bah 3 foo 4 bar
df.info() view now display dtype info per column (GH5682)
df.info() now honors the option max_info_rows, to disable null counts for large frames (GH5974)
max_info_rows
In [7]: max_info_rows = pd.get_option("max_info_rows") In [8]: df = pd.DataFrame( ...: { ...: "A": np.random.randn(10), ...: "B": np.random.randn(10), ...: "C": pd.date_range("20130101", periods=10), ...: } ...: ) ...: In [9]: df.iloc[3:6, [0, 2]] = np.nan
# set to not display the null counts In [10]: pd.set_option("max_info_rows", 0) In [11]: df.info() <class 'pandas.core.frame.DataFrame'> RangeIndex: 10 entries, 0 to 9 Data columns (total 3 columns): # Column Dtype --- ------ ----- 0 A float64 1 B float64 2 C datetime64[ns] dtypes: datetime64[ns](1), float64(2) memory usage: 368.0 bytes
# this is the default (same as in 0.13.0) In [12]: pd.set_option("max_info_rows", max_info_rows) In [13]: df.info() <class 'pandas.core.frame.DataFrame'> RangeIndex: 10 entries, 0 to 9 Data columns (total 3 columns): # Column Non-Null Count Dtype --- ------ -------------- ----- 0 A 7 non-null float64 1 B 10 non-null float64 2 C 7 non-null datetime64[ns] dtypes: datetime64[ns](1), float64(2) memory usage: 368.0 bytes
Add show_dimensions display option for the new DataFrame repr to control whether the dimensions print.
show_dimensions
In [14]: df = pd.DataFrame([[1, 2], [3, 4]]) In [15]: pd.set_option("show_dimensions", False) In [16]: df Out[16]: 0 1 0 1 2 1 3 4 In [17]: pd.set_option("show_dimensions", True) In [18]: df Out[18]: 0 1 0 1 2 1 3 4 [2 rows x 2 columns]
The ArrayFormatter for datetime and timedelta64 now intelligently limit precision based on the values in the array (GH3401)
ArrayFormatter
datetime
timedelta64
Previously output might look like:
age today diff 0 2001-01-01 00:00:00 2013-04-19 00:00:00 4491 days, 00:00:00 1 2004-06-01 00:00:00 2013-04-19 00:00:00 3244 days, 00:00:00
Now the output looks like:
In [19]: df = pd.DataFrame( ....: [pd.Timestamp("20010101"), pd.Timestamp("20040601")], columns=["age"] ....: ) ....: In [20]: df["today"] = pd.Timestamp("20130419") In [21]: df["diff"] = df["today"] - df["age"] In [22]: df Out[22]: age today diff 0 2001-01-01 2013-04-19 4491 days 1 2004-06-01 2013-04-19 3244 days [2 rows x 3 columns]
Add -NaN and -nan to the default set of NA values (GH5952). See NA Values.
-NaN
-nan
Added Series.str.get_dummies vectorized string method (GH6021), to extract dummy/indicator variables for separated string columns:
Series.str.get_dummies
In [23]: s = pd.Series(["a", "a|b", np.nan, "a|c"]) In [24]: s.str.get_dummies(sep="|") Out[24]: a b c 0 1 0 0 1 1 1 0 2 0 0 0 3 1 0 1 [4 rows x 3 columns]
Added the NDFrame.equals() method to compare if two NDFrames are equal have equal axes, dtypes, and values. Added the array_equivalent function to compare if two ndarrays are equal. NaNs in identical locations are treated as equal. (GH5283) See also the docs for a motivating example.
NDFrame.equals()
array_equivalent
df = pd.DataFrame({"col": ["foo", 0, np.nan]}) df2 = pd.DataFrame({"col": [np.nan, 0, "foo"]}, index=[2, 1, 0]) df.equals(df2) df.equals(df2.sort_index())
DataFrame.apply will use the reduce argument to determine whether a Series or a DataFrame should be returned when the DataFrame is empty (GH6007).
DataFrame.apply
reduce
Series
DataFrame
Previously, calling DataFrame.apply an empty DataFrame would return either a DataFrame if there were no columns, or the function being applied would be called with an empty Series to guess whether a Series or DataFrame should be returned:
In [32]: def applied_func(col): ....: print("Apply function being called with: ", col) ....: return col.sum() ....: In [33]: empty = DataFrame(columns=['a', 'b']) In [34]: empty.apply(applied_func) Apply function being called with: Series([], Length: 0, dtype: float64) Out[34]: a NaN b NaN Length: 2, dtype: float64
Now, when apply is called on an empty DataFrame: if the reduce argument is True a Series will returned, if it is False a DataFrame will be returned, and if it is None (the default) the function being applied will be called with an empty series to try and guess the return type.
apply
True
False
None
In [35]: empty.apply(applied_func, reduce=True) Out[35]: a NaN b NaN Length: 2, dtype: float64 In [36]: empty.apply(applied_func, reduce=False) Out[36]: Empty DataFrame Columns: [a, b] Index: [] [0 rows x 2 columns]
There are no announced changes in 0.13 or prior that are taking effect as of 0.13.1
There are no deprecations of prior behavior in 0.13.1
pd.read_csv and pd.to_datetime learned a new infer_datetime_format keyword which greatly improves parsing perf in many cases. Thanks to @lexual for suggesting and @danbirken for rapidly implementing. (GH5490, GH6021)
pd.read_csv
pd.to_datetime
If parse_dates is enabled and this flag is set, pandas will attempt to infer the format of the datetime strings in the columns, and if it can be inferred, switch to a faster method of parsing them. In some cases this can increase the parsing speed by ~5-10x.
parse_dates
# Try to infer the format for the index column df = pd.read_csv( "foo.csv", index_col=0, parse_dates=True, infer_datetime_format=True )
date_format and datetime_format keywords can now be specified when writing to excel files (GH4133)
date_format
datetime_format
excel
MultiIndex.from_product convenience function for creating a MultiIndex from the cartesian product of a set of iterables (GH6055):
MultiIndex.from_product
In [25]: shades = ["light", "dark"] In [26]: colors = ["red", "green", "blue"] In [27]: pd.MultiIndex.from_product([shades, colors], names=["shade", "color"]) Out[27]: MultiIndex([('light', 'red'), ('light', 'green'), ('light', 'blue'), ( 'dark', 'red'), ( 'dark', 'green'), ( 'dark', 'blue')], names=['shade', 'color'])
Panel apply() will work on non-ufuncs. See the docs.
In [28]: import pandas._testing as tm In [29]: panel = tm.makePanel(5) In [30]: panel Out[30]: <class 'pandas.core.panel.Panel'> Dimensions: 3 (items) x 5 (major_axis) x 4 (minor_axis) Items axis: ItemA to ItemC Major_axis axis: 2000-01-03 00:00:00 to 2000-01-07 00:00:00 Minor_axis axis: A to D In [31]: panel['ItemA'] Out[31]: A B C D 2000-01-03 -0.673690 0.577046 -1.344312 -1.469388 2000-01-04 0.113648 -1.715002 0.844885 0.357021 2000-01-05 -1.478427 -1.039268 1.075770 -0.674600 2000-01-06 0.524988 -0.370647 -0.109050 -1.776904 2000-01-07 0.404705 -1.157892 1.643563 -0.968914 [5 rows x 4 columns]
Specifying an apply that operates on a Series (to return a single element)
In [32]: panel.apply(lambda x: x.dtype, axis='items') Out[32]: A B C D 2000-01-03 float64 float64 float64 float64 2000-01-04 float64 float64 float64 float64 2000-01-05 float64 float64 float64 float64 2000-01-06 float64 float64 float64 float64 2000-01-07 float64 float64 float64 float64 [5 rows x 4 columns]
A similar reduction type operation
In [33]: panel.apply(lambda x: x.sum(), axis='major_axis') Out[33]: ItemA ItemB ItemC A -1.108775 -1.090118 -2.984435 B -3.705764 0.409204 1.866240 C 2.110856 2.960500 -0.974967 D -4.532785 0.303202 -3.685193 [4 rows x 3 columns]
This is equivalent to
In [34]: panel.sum('major_axis') Out[34]: ItemA ItemB ItemC A -1.108775 -1.090118 -2.984435 B -3.705764 0.409204 1.866240 C 2.110856 2.960500 -0.974967 D -4.532785 0.303202 -3.685193 [4 rows x 3 columns]
A transformation operation that returns a Panel, but is computing the z-score across the major_axis
In [35]: result = panel.apply(lambda x: (x - x.mean()) / x.std(), ....: axis='major_axis') ....: In [36]: result Out[36]: <class 'pandas.core.panel.Panel'> Dimensions: 3 (items) x 5 (major_axis) x 4 (minor_axis) Items axis: ItemA to ItemC Major_axis axis: 2000-01-03 00:00:00 to 2000-01-07 00:00:00 Minor_axis axis: A to D In [37]: result['ItemA'] # noqa E999 Out[37]: A B C D 2000-01-03 -0.535778 1.500802 -1.506416 -0.681456 2000-01-04 0.397628 -1.108752 0.360481 1.529895 2000-01-05 -1.489811 -0.339412 0.557374 0.280845 2000-01-06 0.885279 0.421830 -0.453013 -1.053785 2000-01-07 0.742682 -0.474468 1.041575 -0.075499 [5 rows x 4 columns]
Panel apply() operating on cross-sectional slabs. (GH1148)
In [38]: def f(x): ....: return ((x.T - x.mean(1)) / x.std(1)).T ....: In [39]: result = panel.apply(f, axis=['items', 'major_axis']) In [40]: result Out[40]: <class 'pandas.core.panel.Panel'> Dimensions: 4 (items) x 5 (major_axis) x 3 (minor_axis) Items axis: A to D Major_axis axis: 2000-01-03 00:00:00 to 2000-01-07 00:00:00 Minor_axis axis: ItemA to ItemC In [41]: result.loc[:, :, 'ItemA'] Out[41]: A B C D 2000-01-03 0.012922 -0.030874 -0.629546 -0.757034 2000-01-04 0.392053 -1.071665 0.163228 0.548188 2000-01-05 -1.093650 -0.640898 0.385734 -1.154310 2000-01-06 1.005446 -1.154593 -0.595615 -0.809185 2000-01-07 0.783051 -0.198053 0.919339 -1.052721 [5 rows x 4 columns]
This is equivalent to the following
In [42]: result = pd.Panel({ax: f(panel.loc[:, :, ax]) for ax in panel.minor_axis}) In [43]: result Out[43]: <class 'pandas.core.panel.Panel'> Dimensions: 4 (items) x 5 (major_axis) x 3 (minor_axis) Items axis: A to D Major_axis axis: 2000-01-03 00:00:00 to 2000-01-07 00:00:00 Minor_axis axis: ItemA to ItemC In [44]: result.loc[:, :, 'ItemA'] Out[44]: A B C D 2000-01-03 0.012922 -0.030874 -0.629546 -0.757034 2000-01-04 0.392053 -1.071665 0.163228 0.548188 2000-01-05 -1.093650 -0.640898 0.385734 -1.154310 2000-01-06 1.005446 -1.154593 -0.595615 -0.809185 2000-01-07 0.783051 -0.198053 0.919339 -1.052721 [5 rows x 4 columns]
Performance improvements for 0.13.1
Series datetime/timedelta binary operations (GH5801)
DataFrame count/dropna for axis=1
count/dropna
axis=1
Series.str.contains now has a regex=False keyword which can be faster for plain (non-regex) string patterns. (GH5879)
regex=False
Series.str.extract (GH5944)
dtypes/ftypes methods (GH5968)
dtypes/ftypes
indexing with object dtypes (GH5968)
DataFrame.apply (GH6013)
Regression in JSON IO (GH5765)
Index construction from Series (GH6150)
There are no experimental changes in 0.13.1
Bug in io.wb.get_countries not including all countries (GH6008)
io.wb.get_countries
Bug in Series replace with timestamp dict (GH5797)
read_csv/read_table now respects the prefix kwarg (GH5732).
prefix
Bug in selection with missing values via .ix from a duplicate indexed DataFrame failing (GH5835)
.ix
Fix issue of boolean comparison on empty DataFrames (GH5808)
Bug in isnull handling NaT in an object array (GH5443)
NaT
Bug in to_datetime when passed a np.nan or integer datelike and a format string (GH5863)
to_datetime
np.nan
Bug in groupby dtype conversion with datetimelike (GH5869)
Regression in handling of empty Series as indexers to Series (GH5877)
Bug in internal caching, related to (GH5727)
Testing bug in reading JSON/msgpack from a non-filepath on windows under py3 (GH5874)
Bug when assigning to .ix[tuple(…)] (GH5896)
Bug in fully reindexing a Panel (GH5905)
Bug in idxmin/max with object dtypes (GH5914)
Bug in BusinessDay when adding n days to a date not on offset when n>5 and n%5==0 (GH5890)
BusinessDay
Bug in assigning to chained series with a series via ix (GH5928)
Bug in creating an empty DataFrame, copying, then assigning (GH5932)
Bug in DataFrame.tail with empty frame (GH5846)
Bug in propagating metadata on resample (GH5862)
resample
Fixed string-representation of NaT to be “NaT” (GH5708)
Fixed string-representation for Timestamp to show nanoseconds if present (GH5912)
pd.match not returning passed sentinel
pd.match
Panel.to_frame() no longer fails when major_axis is a MultiIndex (GH5402).
Panel.to_frame()
major_axis
MultiIndex
Bug in pd.read_msgpack with inferring a DateTimeIndex frequency incorrectly (GH5947)
pd.read_msgpack
DateTimeIndex
Fixed to_datetime for array with both Tz-aware datetimes and NaT’s (GH5961)
Bug in rolling skew/kurtosis when passed a Series with bad data (GH5749)
Bug in scipy interpolate methods with a datetime index (GH5975)
interpolate
Bug in NaT comparison if a mixed datetime/np.datetime64 with NaT were passed (GH5968)
Fixed bug with pd.concat losing dtype information if all inputs are empty (GH5742)
pd.concat
Recent changes in IPython cause warnings to be emitted when using previous versions of pandas in QTConsole, now fixed. If you’re using an older version and need to suppress the warnings, see (GH5922).
Bug in merging timedelta dtypes (GH5695)
timedelta
Bug in plotting.scatter_matrix function. Wrong alignment among diagonal and off-diagonal plots, see (GH5497).
Regression in Series with a MultiIndex via ix (GH6018)
Bug in Series.xs with a MultiIndex (GH6018)
Bug in Series construction of mixed type with datelike and an integer (which should result in object type and not automatic conversion) (GH6028)
Possible segfault when chained indexing with an object array under NumPy 1.7.1 (GH6026, GH6056)
Bug in setting using fancy indexing a single element with a non-scalar (e.g. a list), (GH6043)
to_sql did not respect if_exists (GH4110 GH4304)
to_sql
if_exists
Regression in .get(None) indexing from 0.12 (GH5652)
.get(None)
Subtle iloc indexing bug, surfaced in (GH6059)
iloc
Bug with insert of strings into DatetimeIndex (GH5818)
Fixed unicode bug in to_html/HTML repr (GH6098)
Fixed missing arg validation in get_options_data (GH6105)
Bug in assignment with duplicate columns in a frame where the locations are a slice (e.g. next to each other) (GH6120)
Bug in propagating _ref_locs during construction of a DataFrame with dups index/columns (GH6121)
Bug in DataFrame.apply when using mixed datelike reductions (GH6125)
Bug in DataFrame.append when appending a row with different columns (GH6129)
DataFrame.append
Bug in DataFrame construction with recarray and non-ns datetime dtype (GH6140)
Bug in .loc setitem indexing with a dataframe on rhs, multiple item setting, and a datetimelike (GH6152)
.loc
Fixed a bug in query/eval during lexicographic string comparisons (GH6155).
query
eval
Fixed a bug in query where the index of a single-element Series was being thrown away (GH6148).
Bug in HDFStore on appending a dataframe with MultiIndexed columns to an existing table (GH6167)
HDFStore
Consistency with dtypes in setting an empty DataFrame (GH6171)
Bug in selecting on a MultiIndex HDFStore even in the presence of under specified column spec (GH6169)
Bug in nanops.var with ddof=1 and 1 elements would sometimes return inf rather than nan on some platforms (GH6136)
nanops.var
ddof=1
inf
nan
Bug in Series and DataFrame bar plots ignoring the use_index keyword (GH6209)
use_index
Bug in groupby with mixed str/int under python3 fixed; argsort was failing (GH6212)
argsort
A total of 52 people contributed patches to this release. People with a “+” by their names contributed a patch for the first time.
Alex Rothberg
Alok Singhal +
Andrew Burrows +
Andy Hayden
Bjorn Arneson +
Brad Buran
Caleb Epstein
Chapman Siu
Chase Albert +
Clark Fitzgerald +
DSM
Dan Birken
Daniel Waeber +
David Wolever +
Doran Deluz +
Douglas McNeil +
Douglas Rudd +
Dražen Lučanin
Elliot S +
Felix Lawrence +
George Kuan +
Guillaume Gay +
Jacob Schaer
Jan Wagner +
Jeff Tratner
John McNamara
Joris Van den Bossche
Julia Evans +
Kieran O’Mahony
Michael Schatzow +
Naveen Michaud-Agrawal +
Patrick O’Keeffe +
Phillip Cloud
Roman Pekar
Skipper Seabold
Spencer Lyon
Tom Augspurger +
TomAugspurger
acorbe +
akittredge +
bmu +
bwignall +
chapman siu
danielballan
david +
davidshinn
immerrr +
jreback
lexual
mwaskom +
unutbu
y-p