Skip to content

astropenguin/pandas-dataclasses

Repository files navigation

pandas-dataclasses

Release Python Downloads DOI Tests

pandas data creation made easy by dataclass

Overview

pandas-dataclass makes it easy to create pandas data (DataFrame and Series) by specifying their data types, attributes, and names using the Python's dataclass:

Click to see all imports
from dataclasses import dataclass
from pandas_dataclasses import AsFrame, Data, Index
@dataclass
class Weather(AsFrame):
    """Weather information."""

    year: Index[int]
    month: Index[int]
    temp: Data[float]
    wind: Data[float]


df = Weather.new(
    [2020, 2020, 2021, 2021, 2022],
    [1, 7, 1, 7, 1],
    [7.1, 24.3, 5.4, 25.9, 4.9],
    [2.4, 3.1, 2.3, 2.4, 2.6],
)

where df will become a DataFrame object like:

            temp  wind
year month
2020 1       7.1   2.4
     7      24.3   3.1
2021 1       5.4   2.3
     7      25.9   2.4
2022 1       4.9   2.6

Features

  • Specifying data types and names of each element in pandas data
  • Specifying metadata stored in pandas data attributes (attrs)
  • Support for hierarchical index and columns
  • Support for custom factory for data creation
  • Support for full dataclass features
  • Support for static type check by mypy and Pyright (Pylance)

Installation

pip install pandas-dataclasses

How it works

pandas-dataclasses provides you the following features:

  • Type hints for dataclass fields (Attr, Data, Index) to specify the data type and name of each element in pandas data
  • Mix-in classes for dataclasses (As, AsFrame, AsSeries) to create pandas data by a classmethod (new) that takes the same arguments as dataclass initialization

When you call new, it will first create a dataclass object and then create a Series or DataFrame object from the dataclass object according the type hints and values in it. In the example above, df = Weather.new(...) is thus equivalent to:

Click to see all imports
from pandas_dataclasses import asframe
obj = Weather([2020, ...], [1, ...], [7.1, ...], [2.4, ...])
df = asframe(obj)

where asframe is a conversion function. pandas-dataclasses does not touch the dataclass object creation itself; this allows you to fully customize your dataclass before conversion by the dataclass features (field, __post_init__, ...).

Basic usage

DataFrame creation

As shown in the example above, a dataclass that has the AsFrame (or AsDataFrame as an alias) mix-in will create DataFrame objects:

Click to see all imports
from dataclasses import dataclass
from pandas_dataclasses import AsFrame, Data, Index
@dataclass
class Weather(AsFrame):
    """Weather information."""

    year: Index[int]
    month: Index[int]
    temp: Data[float]
    wind: Data[float]


df = Weather.new(...)

where fields typed by Index are index fields, each value of which will become an index or a part of a hierarchical index of a DataFrame object. Fields typed by Data are data fields, each value of which will become a data column of a DataFrame object. Fields typed by other types are just ignored in the DataFrame creation.

Each data or index will be cast to the data type specified in a type hint like Index[int]. Use Any or None (like Index[Any]) if you do not want type casting. See also data typing rules for more examples.

By default, a field name (i.e. an argument name) is used for the name of corresponding data or index. See also custom naming and naming rules if you want customization.

Series creation

A dataclass that has the AsSeries mix-in will create Series objects:

Click to see all imports
from dataclasses import dataclass
from pandas_dataclasses import AsSeries, Data, Index
@dataclass
class Weather(AsSeries):
    """Weather information."""

    year: Index[int]
    month: Index[int]
    temp: Data[float]


ser = Weather.new(...)

Unlike AsFrame, the second and subsequent data fields are ignored in the Series creation even if they exist. Other rules are the same as for the DataFrame creation.

Advanced usage

Metadata storing

Fields typed by Attr are attribute fields, each value of which will become an item of attributes of a DataFrame or a Series object:

Click to see all imports
from dataclasses import dataclass
from pandas_dataclasses import AsFrame, Attr, Data, Index
@dataclass
class Weather(AsFrame):
    """Weather information."""

    year: Index[int]
    month: Index[int]
    temp: Data[float]
    wind: Data[float]
    loc: Attr[str] = "Tokyo"
    lon: Attr[float] = 139.69167
    lat: Attr[float] = 35.68944


df = Weather.new(...)

where df.attrs will become like:

{"loc": "Tokyo", "lon": 139.69167, "lat": 35.68944}

Custom naming

The name of attribute, data, or index can be explicitly specified by adding a hashable annotation to the corresponding type:

Click to see all imports
from dataclasses import dataclass
from typing import Annotated as Ann
from pandas_dataclasses import AsFrame, Attr, Data, Index
@dataclass
class Weather(AsFrame):
    """Weather information."""

    year: Ann[Index[int], "Year"]
    month: Ann[Index[int], "Month"]
    temp: Ann[Data[float], "Temperature (deg C)"]
    wind: Ann[Data[float], "Wind speed (m/s)"]
    loc: Ann[Attr[str], "Location"] = "Tokyo"
    lon: Ann[Attr[float], "Longitude (deg)"] = 139.69167
    lat: Ann[Attr[float], "Latitude (deg)"] = 35.68944


df = Weather.new(...)

where df and df.attrs will become like:

            Temperature (deg C)  Wind speed (m/s)
Year Month
2020 1                      7.1               2.4
     7                     24.3               3.1
2021 1                      5.4               2.3
     7                     25.9               2.4
2022 1                      4.9               2.6
{"Location": "Tokyo", "Longitude (deg)": 139.69167, "Latitude (deg)": 35.68944}

If an annotation is a format string, it will be formatted by a dataclass object before the data creation:

Click to see all imports
from dataclasses import dataclass
from typing import Annotated as Ann
from pandas_dataclasses import AsFrame, Data, Index
@dataclass
class Weather(AsFrame):
    """Weather information."""

    year: Ann[Index[int], "Year"]
    month: Ann[Index[int], "Month"]
    temp: Ann[Data[float], "Temperature ({.temp_unit})"]
    wind: Ann[Data[float], "Wind speed ({.wind_unit})"]
    temp_unit: str = "deg C"
    wind_unit: str = "m/s"


df = Weather.new(..., temp_unit="deg F", wind_unit="km/h")

where units of the temperature and the wind speed will be dynamically updated (see also naming rules).

Hierarchical columns

Adding tuple annotations to data fields will create DataFrame objects with hierarchical columns:

Click to see all imports
from dataclasses import dataclass
from typing import Annotated as Ann
from pandas_dataclasses import AsFrame, Data, Index
@dataclass
class Weather(AsFrame):
    """Weather information."""

    year: Ann[Index[int], "Year"]
    month: Ann[Index[int], "Month"]
    temp_avg: Ann[Data[float], ("Temperature (deg C)", "Average")]
    temp_max: Ann[Data[float], ("Temperature (deg C)", "Maximum")]
    wind_avg: Ann[Data[float], ("Wind speed (m/s)", "Average")]
    wind_max: Ann[Data[float], ("Wind speed (m/s)", "Maximum")]


df = Weather.new(...)

where df will become like:

           Temperature (deg C)         Wind speed (m/s)
                       Average Maximum          Average Maximum
Year Month
2020 1                     7.1    11.1              2.4     8.8
     7                    24.3    27.7              3.1    10.2
2021 1                     5.4    10.3              2.3    10.7
     7                    25.9    30.3              2.4     9.0
2022 1                     4.9     9.4              2.6     8.8

Column names can be (explicitly) specified by dictionary annotations:

Click to see all imports
from dataclasses import dataclass
from typing import Annotated as Ann
from pandas_dataclasses import AsFrame, Data, Index
def name(meas: str, stat: str) -> dict[str, str]:
    """Create a dictionary annotation for a column name."""
    return {"Measurement": meas, "Statistic": stat}


@dataclass
class Weather(AsFrame):
    """Weather information."""

    year: Ann[Index[int], "Year"]
    month: Ann[Index[int], "Month"]
    temp_avg: Ann[Data[float], name("Temperature (deg C)", "Average")]
    temp_max: Ann[Data[float], name("Temperature (deg C)", "Maximum")]
    wind_avg: Ann[Data[float], name("Wind speed (m/s)", "Average")]
    wind_max: Ann[Data[float], name("Wind speed (m/s)", "Maximum")]


df = Weather.new(...)

where df will become like:

Measurement Temperature (deg C)         Wind speed (m/s)
Statistic               Average Maximum          Average Maximum
Year Month
2020 1                      7.1    11.1              2.4     8.8
     7                     24.3    27.7              3.1    10.2
2021 1                      5.4    10.3              2.3    10.7
     7                     25.9    30.3              2.4     9.0
2022 1                      4.9     9.4              2.6     8.8

If a tuple or dictionary annotation has format strings, they will also be formatted by a dataclass object (see also naming rules).

Multiple-item fields

Multiple (and possibly extra) attributes, data, or indices can be added by fields with corresponding type hints wrapped by Multiple:

Click to see all imports
from dataclasses import dataclass
from pandas_dataclasses import AsFrame, Data, Index, Multiple
@dataclass
class Weather(AsFrame):
    """Weather information."""

    year: Index[int]
    month: Index[int]
    temp: Data[float]
    wind: Data[float]
    extra_index: Multiple[Index[int]]
    extra_data: Multiple[Data[float]]


df = Weather.new(
    [2020, 2020, 2021, 2021, 2022],
    [1, 7, 1, 7, 1],
    [7.1, 24.3, 5.4, 25.9, 4.9],
    [2.4, 3.1, 2.3, 2.4, 2.6],
    extra_index={
        "day": [1, 1, 1, 1, 1],
        "week": [2, 2, 4, 3, 5],
    },
    extra_data={
        "humid": [65, 89, 57, 83, 52],
        "press": [1013.8, 1006.2, 1014.1, 1007.7, 1012.7],
    },
)

where df will become like:

                     temp  wind  humid   press
year month day week
2020 1     1   2      7.1   2.4   65.0  1013.8
     7     1   2     24.3   3.1   89.0  1006.2
2021 1     1   4      5.4   2.3   57.0  1014.1
     7     1   3     25.9   2.4   83.0  1007.7
2022 1     1   5      4.9   2.6   52.0  1012.7

If multiple items of the same name exist, the last-defined one will be finally used. For example, if the extra_index field contains "month": [2, 8, 2, 8, 2], the values given by the month field will be overwritten.

Custom pandas factory

A custom class can be specified as a factory for the Series or DataFrame creation by As, the generic version of AsFrame and AsSeries. Note that the custom class must be a subclass of either pandas.Series or pandas.DataFrame:

Click to see all imports
import pandas as pd
from dataclasses import dataclass
from pandas_dataclasses import As, Data, Index
class CustomSeries(pd.Series):
    """Custom pandas Series."""

    pass


@dataclass
class Temperature(As[CustomSeries]):
    """Temperature information."""

    year: Index[int]
    month: Index[int]
    temp: Data[float]


ser = Temperature.new(...)

where ser is statically regarded as CustomSeries and will become a CustomSeries object.

Generic Series type (Series[T]) is also supported, however, it is only for static the type check in the current pandas versions. In such cases, you can additionally give a factory that must work in runtime as a class argument:

Click to see all imports
import pandas as pd
from dataclasses import dataclass
from pandas_dataclasses import As, Data, Index
@dataclass
class Temperature(As["pd.Series[float]"], factory=pd.Series):
    """Temperature information."""

    year: Index[int]
    month: Index[int]
    temp: Data[float]


ser = Temperature.new(...)

where ser is statically regarded as Series[float] but will become a Series object in runtime.

Appendix

Data typing rules

The data type (dtype) of data or index is determined from the first Data or Index type of the corresponding field, respectively. The following table shows how the data type is inferred:

Click to see all imports
from typing import Any, Annotated as Ann, Literal as L
from pandas_dataclasses import Data
Type hint Inferred data type
Data[Any] None (no type casting)
Data[None] None (no type casting)
Data[int] numpy.int64
Data[int | str] numpy.int64
Data[numpy.int32] numpy.int32
Data[L["datetime64[ns]"]] numpy.dtype("<M8[ns]")
Data[L["category"]] pandas.CategoricalDtype()
Data[int] | str numpy.int64
Data[int] | Data[float] numpy.int64
Ann[Data[int], "spam"] numpy.int64
Data[Ann[int, "spam"]] numpy.int64

Naming rules

The name of attribute, data, or index is determined from the first annotation of the first Attr, Data, or Index type of the corresponding field, respectively. If the annotation is a format string or a tuple that has format strings, it (they) will be formatted by a dataclass object before the data creation. Otherwise, the field name (i.e. argument name) will be used. The following table shows how the name is inferred:

Click to see all imports
from typing import Any, Annotated as Ann
from pandas_dataclasses import Data
Type hint Inferred name
Data[Any] (field name)
Ann[Data[Any], ..., "spam"] (field name)
Ann[Data[Any], "spam"] "spam"
Ann[Data[Any], "spam", "ham"] "spam"
Ann[Data[Any], "spam"] | Ann[str, "ham"] "spam"
Ann[Data[Any], "spam"] | Ann[Data[float], "ham"] "spam"
Ann[Data[Any], "{.name}" "{.name}".format(obj)
Ann[Data[Any], ("spam", "ham")] ("spam", "ham")
Ann[Data[Any], ("{.name}", "ham")] ("{.name}".format(obj), "ham")

where obj is a dataclass object that is expected to have obj.name.

Development roadmap

Release version Features
v0.5 Support for dynamic naming
v0.6 Support for extension array and dtype
v0.7 Support for hierarchical columns
v0.8 Support for mypy and callable pandas factory
v0.9 Support for Ellipsis (...) as an alias of field name
v0.10 Support for union type in type hints
v0.11 Support for Python 3.11 and drop support for Python 3.7
v0.12 Support for multiple items received in a single field
v1.0 Initial major release (freezing public features until v2.0)