While pandas provides a rich set of methods, containers, and data types, yourneeds may not be fully satisfied. pandas offers a few options for extendingpandas.
Registering custom accessors#
Libraries can use the decoratorspandas.api.extensions.register_dataframe_accessor(),pandas.api.extensions.register_series_accessor(), andpandas.api.extensions.register_index_accessor(), to add additional“namespaces” to pandas objects. All of these follow a similar convention: youdecorate a class, providing the name of attribute to add. The class’s__init__
method gets the object being decorated. For example:
@pd.api.extensions.register_dataframe_accessor("geo")class GeoAccessor: def __init__(self, pandas_obj): self._validate(pandas_obj) self._obj = pandas_obj @staticmethod def _validate(obj): # verify there is a column latitude and a column longitude if "latitude" not in obj.columns or "longitude" not in obj.columns: raise AttributeError("Must have 'latitude' and 'longitude'.") @property def center(self): # return the geographic center point of this DataFrame lat = self._obj.latitude lon = self._obj.longitude return (float(lon.mean()), float(lat.mean())) def plot(self): # plot this array's data on a map, e.g., using Cartopy pass
Now users can access your methods using the geo
namespace:
>>> ds = pd.DataFrame(... {"longitude": np.linspace(0, 10), "latitude": np.linspace(0, 20)}... )>>> ds.geo.center(5.0, 10.0)>>> ds.geo.plot()# plots data on a map
This can be a convenient way to extend pandas objects without subclassing them.If you write a custom accessor, make a pull request adding it to ourecosystem page.
We highly recommend validating the data in your accessor’s __init__
.In our GeoAccessor
, we validate that the data contains the expected columns,raising an AttributeError
when the validation fails.For a Series
accessor, you should validate the dtype
if the accessorapplies only to certain dtypes.
Extension types#
Note
The pandas.api.extensions.ExtensionDtype and pandas.api.extensions.ExtensionArray APIs wereexperimental prior to pandas 1.5. Starting with version 1.5, future changes will followthe pandas deprecation policy.
pandas defines an interface for implementing data types and arrays that extendNumPy’s type system. pandas itself uses the extension system for some typesthat aren’t built into NumPy (categorical, period, interval, datetime withtimezone).
Libraries can define a custom array and data type. When pandas encounters theseobjects, they will be handled properly (i.e. not converted to an ndarray ofobjects). Many methods like pandas.isna() will dispatch to the extensiontype’s implementation.
If you’re building a library that implements the interface, please publicize iton the ecosystem page.
The interface consists of two classes.
ExtensionDtype#
A pandas.api.extensions.ExtensionDtype is similar to a numpy.dtype
object. It describes thedata type. Implementers are responsible for a few unique items like the name.
One particularly important item is the type
property. This should be theclass that is the scalar type for your data. For example, if you were writing anextension array for IP Address data, this might be ipaddress.IPv4Address
.
See the extension dtype source for interface definition.
pandas.api.extensions.ExtensionDtype can be registered to pandas to allow creation via a string dtype name.This allows one to instantiate Series
and .astype()
with a registered string name, forexample 'category'
is a registered string accessor for the CategoricalDtype
.
See the extension dtype dtypes for more on how to register dtypes.
ExtensionArray#
This class provides all the array-like functionality. ExtensionArrays arelimited to 1 dimension. An ExtensionArray is linked to an ExtensionDtype via thedtype
attribute.
pandas makes no restrictions on how an extension array is created via its__new__
or __init__
, and puts no restrictions on how you store yourdata. We do require that your array be convertible to a NumPy array, even ifthis is relatively expensive (as it is for Categorical
).
They may be backed by none, one, or many NumPy arrays. For example,pandas.Categorical is an extension array backed by two arrays,one for codes and one for categories. An array of IPv6 addresses maybe backed by a NumPy structured array with two fields, one for thelower 64 bits and one for the upper 64 bits. Or they may be backedby some other storage type, like Python lists.
See the extension array source for the interface definition. The docstringsand comments contain guidance for properly implementing the interface.
ExtensionArray operator support#
By default, there are no operators defined for the class ExtensionArray.There are two approaches for providing operator support for your ExtensionArray:
Define each of the operators on your
ExtensionArray
subclass.Use an operator implementation from pandas that depends on operators that are already definedon the underlying elements (scalars) of the ExtensionArray.
Note
Regardless of the approach, you may want to set __array_priority__
if you want your implementation to be called when involved in binary operationswith NumPy arrays.
For the first approach, you define selected operators, e.g., __add__
, __le__
, etc. thatyou want your ExtensionArray
subclass to support.
The second approach assumes that the underlying elements (i.e., scalar type) of the ExtensionArray
have the individual operators already defined. In other words, if your ExtensionArray
named MyExtensionArray
is implemented so that each element is an instanceof the class MyExtensionElement
, then if the operators are definedfor MyExtensionElement
, the second approach will automaticallydefine the operators for MyExtensionArray
.
A mixin class, ExtensionScalarOpsMixin
supports this secondapproach. If developing an ExtensionArray
subclass, for example MyExtensionArray
,can simply include ExtensionScalarOpsMixin
as a parent class of MyExtensionArray
,and then call the methods _add_arithmetic_ops()
and/or_add_comparison_ops()
to hook the operators intoyour MyExtensionArray
class, as follows:
from pandas.api.extensions import ExtensionArray, ExtensionScalarOpsMixinclass MyExtensionArray(ExtensionArray, ExtensionScalarOpsMixin): passMyExtensionArray._add_arithmetic_ops()MyExtensionArray._add_comparison_ops()
Note
Since pandas
automatically calls the underlying operator on eachelement one-by-one, this might not be as performant as implementing your ownversion of the associated operators directly on the ExtensionArray
.
For arithmetic operations, this implementation will try to reconstruct a newExtensionArray
with the result of the element-wise operation. Whetheror not that succeeds depends on whether the operation returns a resultthat’s valid for the ExtensionArray
. If an ExtensionArray
cannotbe reconstructed, an ndarray containing the scalars returned instead.
For ease of implementation and consistency with operations between pandasand NumPy ndarrays, we recommend not handling Series and Indexes in your binary ops.Instead, you should detect these cases and return NotImplemented
.When pandas encounters an operation like op(Series, ExtensionArray)
, pandaswill
unbox the array from the
Series
(Series.array
)call
result = op(values, ExtensionArray)
re-box the result in a
Series
NumPy universal functions#
Series implements __array_ufunc__
. As part of the implementation,pandas unboxes the ExtensionArray
from the Series, applies the ufunc,and re-boxes it if necessary.
If applicable, we highly recommend that you implement __array_ufunc__
in yourextension array to avoid coercion to an ndarray. Seethe NumPy documentationfor an example.
As part of your implementation, we require that you defer to pandas when a pandascontainer (Series, DataFrame, Index) is detected in inputs
.If any of those is present, you should return NotImplemented
. pandas will take care ofunboxing the array from the container and re-calling the ufunc with the unwrapped input.
Testing extension arrays#
We provide a test suite for ensuring that your extension arrays satisfy the expectedbehavior. To use the test suite, you must provide several pytest fixtures and inheritfrom the base test class. The required fixtures are found inpandas-dev/pandas.
To use a test, subclass it:
from pandas.tests.extension import baseclass TestConstructors(base.BaseConstructorsTests): pass
See pandas-dev/pandasfor a list of all the tests available.
Compatibility with Apache Arrow#
An ExtensionArray
can support conversion to / from pyarrow
arrays(and thus support for example serialization to the Parquet file format)by implementing two methods: ExtensionArray.__arrow_array__
andExtensionDtype.__from_arrow__
.
The ExtensionArray.__arrow_array__
ensures that pyarrow
knowns howto convert the specific extension array into a pyarrow.Array
(also whenincluded as a column in a pandas DataFrame):
class MyExtensionArray(ExtensionArray): ... def __arrow_array__(self, type=None): # convert the underlying array values to a pyarrow Array import pyarrow return pyarrow.array(..., type=type)
The ExtensionDtype.__from_arrow__
method then controls the conversionback from pyarrow to a pandas ExtensionArray. This method receives a pyarrowArray
or ChunkedArray
as only argument and is expected to return theappropriate pandas ExtensionArray
for this dtype and the passed values:
class ExtensionDtype: ... def __from_arrow__(self, array: pyarrow.Array/ChunkedArray) -> ExtensionArray: ...
See more in the Arrow documentation.
Those methods have been implemented for the nullable integer and string extensiondtypes included in pandas, and ensure roundtrip to pyarrow and the Parquet file format.
Subclassing pandas data structures#
Warning
There are some easier alternatives before considering subclassing pandas
data structures.
Extensible method chains with pipe
Use composition. See here.
Extending by registering an accessor
Extending by extension type
This section describes how to subclass pandas
data structures to meet more specific needs. There are two points that need attention:
Override constructor properties.
Define original properties
Override constructor properties#
Each data structure has several constructor properties for returning a newdata structure as the result of an operation. By overriding these properties,you can retain subclasses through pandas
data manipulations.
There are 3 possible constructor properties to be defined on a subclass:
DataFrame/Series._constructor
: Used when a manipulation result has the same dimension as the original.DataFrame._constructor_sliced
: Used when aDataFrame
(sub-)class manipulation result should be aSeries
(sub-)class.Series._constructor_expanddim
: Used when aSeries
(sub-)class manipulation result should be aDataFrame
(sub-)class, e.g.Series.to_frame()
.
Below example shows how to define SubclassedSeries
and SubclassedDataFrame
overriding constructor properties.
class SubclassedSeries(pd.Series): @property def _constructor(self): return SubclassedSeries @property def _constructor_expanddim(self): return SubclassedDataFrameclass SubclassedDataFrame(pd.DataFrame): @property def _constructor(self): return SubclassedDataFrame @property def _constructor_sliced(self): return SubclassedSeries
>>> s = SubclassedSeries([1, 2, 3])>>> type(s)<class '__main__.SubclassedSeries'>>>> to_framed = s.to_frame()>>> type(to_framed)<class '__main__.SubclassedDataFrame'>>>> df = SubclassedDataFrame({"A": [1, 2, 3], "B": [4, 5, 6], "C": [7, 8, 9]})>>> df A B C0 1 4 71 2 5 82 3 6 9>>> type(df)<class '__main__.SubclassedDataFrame'>>>> sliced1 = df[["A", "B"]]>>> sliced1 A B0 1 41 2 52 3 6>>> type(sliced1)<class '__main__.SubclassedDataFrame'>>>> sliced2 = df["A"]>>> sliced20 11 22 3Name: A, dtype: int64>>> type(sliced2)<class '__main__.SubclassedSeries'>
Define original properties#
To let original data structures have additional properties, you should let pandas
know what properties are added. pandas
maps unknown properties to data names overriding __getattribute__
. Defining original properties can be done in one of 2 ways:
Define
_internal_names
and_internal_names_set
for temporary properties which WILL NOT be passed to manipulation results.Define
_metadata
for normal properties which will be passed to manipulation results.
Below is an example to define two original properties, “internal_cache” as a temporary property and “added_property” as a normal property
class SubclassedDataFrame2(pd.DataFrame): # temporary properties _internal_names = pd.DataFrame._internal_names + ["internal_cache"] _internal_names_set = set(_internal_names) # normal properties _metadata = ["added_property"] @property def _constructor(self): return SubclassedDataFrame2
>>> df = SubclassedDataFrame2({"A": [1, 2, 3], "B": [4, 5, 6], "C": [7, 8, 9]})>>> df A B C0 1 4 71 2 5 82 3 6 9>>> df.internal_cache = "cached">>> df.added_property = "property">>> df.internal_cachecached>>> df.added_propertyproperty# properties defined in _internal_names is reset after manipulation>>> df[["A", "B"]].internal_cacheAttributeError: 'SubclassedDataFrame2' object has no attribute 'internal_cache'# properties defined in _metadata are retained>>> df[["A", "B"]].added_propertyproperty
Plotting backends#
pandas can be extended with third-party plotting backends. Themain idea is letting users select a plotting backend different than the providedone based on Matplotlib. For example:
>>> pd.set_option("plotting.backend", "backend.module")>>> pd.Series([1, 2, 3]).plot()
This would be more or less equivalent to:
>>> import backend.module>>> backend.module.plot(pd.Series([1, 2, 3]))
The backend module can then use other visualization tools (Bokeh, Altair,…)to generate the plots.
Libraries implementing the plotting backend should use entry pointsto make their backend discoverable to pandas. The key is "pandas_plotting_backends"
. For example, pandasregisters the default “matplotlib” backend as follows.
# in setup.pysetup( # noqa: F821 ..., entry_points={ "pandas_plotting_backends": [ "matplotlib = pandas:plotting._matplotlib", ], },)
More information on how to implement a third-party plotting backend can be found atpandas-dev/pandas.
Arithmetic with 3rd party types#
In order to control how arithmetic works between a custom type and a pandas type,implement __pandas_priority__
. Similar to numpy’s __array_priority__
semantics, arithmetic methods on DataFrame, Series, and Indexobjects will delegate to other
, if it has an attribute __pandas_priority__
with a higher value.
By default, pandas objects try to operate with other objects, even if they are not types known to pandas:
>>> pd.Series([1, 2]) + [10, 20]0 111 22dtype: int64
In the example above, if [10, 20]
was a custom type that can be understood as a list, pandas objects will still operate with it in the same way.
In some cases, it is useful to delegate to the other type the operation. For example, consider I implement acustom list object, and I want the result of adding my custom list with a pandas Series to be an instance of my listand not a Series as seen in the previous example. This is now possible by defining the __pandas_priority__
attributeof my custom list, and setting it to a higher value, than the priority of the pandas objects I want to operate with.
The __pandas_priority__
of DataFrame, Series, and Index are 4000
, 3000
, and 2000
respectively. The base ExtensionArray.__pandas_priority__
is 1000
.
class CustomList(list): __pandas_priority__ = 5000 def __radd__(self, other): # return `self` and not the addition for simplicity return selfcustom = CustomList()series = pd.Series([1, 2, 3])# Series refuses to add custom, since it's an unknown type with higher priorityassert series.__add__(custom) is NotImplemented# This will cause the custom class `__radd__` being used insteadassert series + custom is custom