Mixin columns

Version 1.0 of astropy introduces a new concept of the “Mixin Column” in tables which allows integration of appropriate non-Column based class objects within a Table object. These mixin column objects are not converted in any way but are used natively.

The available built-in mixin column classes are:

As a first example we can create a table and add a time column:

>>> from astropy.table import Table
>>> from astropy.time import Time
>>> t = Table()
>>> t['index'] = [1, 2]
>>> t['time'] = Time(['2001-01-02T12:34:56', '2001-02-03T00:01:02'])
>>> print(t)
index           time
----- -----------------------
    1 2001-01-02T12:34:56.000
    2 2001-02-03T00:01:02.000

The important point here is that the time column is a bona fide Time object:

>>> t['time']
<Time object: scale='utc' format='isot' value=['2001-01-02T12:34:56.000' '2001-02-03T00:01:02.000']>
>>> t['time'].mjd  
array([51911.52425926, 51943.00071759])

Quantity and QTable

The ability to natively handle Quantity objects within a table makes it easier to manipulate tabular data with units in a natural and robust way. However, this feature introduces an ambiguity because data with a unit (e.g. from a FITS binary table) can be represented as either a Column with a unit attribute or as a Quantity object. In order to retain complete backward compatibility with astropy versions prior to 1.0, a minor variant of the Table class called QTable is available. QTable is exactly the same as Table except that Quantity is the default for any data column with a defined unit.

If you take advantage of the Quantity infrastructure in your analysis then QTable is the preferred way to create tables with units. If instead you use table column units more as a descriptive label then the plain Table class is probably the best class to use.

To illustrate these concepts we first create a standard Table where we supply as input a Time object and a Quantity object with units of m / s. In this case the quantity is converted to a Column (which has a unit attribute but does not have all the features of a Quantity):

>>> import astropy.units as u
>>> t = Table()
>>> t['index'] = [1, 2]
>>> t['time'] = Time(['2001-01-02T12:34:56', '2001-02-03T00:01:02'])
>>> t['velocity'] = [3, 4] * u.m / u.s

>>> print(t)
index           time          velocity
                               m / s
----- ----------------------- --------
    1 2001-01-02T12:34:56.000      3.0
    2 2001-02-03T00:01:02.000      4.0

>>> type(t['velocity'])
<class 'astropy.table.column.Column'>

>>> t['velocity'].unit
Unit("m / s")

>>> (t['velocity'] ** 2).unit  # WRONG because Column is not smart about unit
Unit("m / s")

So instead let’s do the same thing using a quantity table QTable:

>>> from astropy.table import QTable

>>> qt = QTable()
>>> qt['index'] = [1, 2]
>>> qt['time'] = Time(['2001-01-02T12:34:56', '2001-02-03T00:01:02'])
>>> qt['velocity'] = [3, 4] * u.m / u.s

The velocity column is now a Quantity and behaves accordingly:

>>> type(qt['velocity'])
<class 'astropy.units.quantity.Quantity'>

>>> qt['velocity'].unit
Unit("m / s")

>>> (qt['velocity'] ** 2).unit  # GOOD!
Unit("m2 / s2")

You can easily convert Table to QTable and vice-versa:

>>> qt2 = QTable(t)
>>> type(qt2['velocity'])
<class 'astropy.units.quantity.Quantity'>

>>> t2 = Table(qt2)
>>> type(t2['velocity'])
<class 'astropy.table.column.Column'>

Note

To summarize: the only difference between QTable and Table is the behavior when adding a column that has a specified unit. With QTable such a column is always converted to a Quantity object before being added to the table. Likewise if a unit is specified for an existing unit-less Column in a QTable, then the column is converted to Quantity.

The converse is that if one adds a Quantity column to an ordinary Table then it gets converted to an ordinary Column with the corresponding unit attribute.

Mixin Attributes

The usual column attributes name, dtype, unit, format, and description are available in any mixin column via the info property:

>>> qt['velocity'].info.name
'velocity'

This info property is a key bit of glue that allows for a non-Column object to behave much like a column.

The same info property is also available in standard Column objects. These info attributes like t['a'].info.name simply refer to the direct Column attribute (e.g. t['a'].name) and can be used interchangeably. Likewise in a Quantity object, info.dtype attribute refers to the native dtype attribute of the object.

Note

When writing generalized code that handles column objects which might be mixin columns, one must always use the info property to access column attributes.

Details and caveats

Most common table operations behave as expected when mixin columns are part of the table. However, there are limitations in the current implementation.

Adding or inserting a row

Adding or inserting a row works as expected only for mixin classes that are mutable (data can changed internally) and that have an insert() method. Quantity supports insert() but Time and SkyCoord do not. If we try to insert a row into the previously defined table an exception occurs:

>>> qt.add_row((1, '2001-02-03T00:01:02', 5 * u.m / u.s))
Traceback (most recent call last):
  ...
ValueError: Unable to insert row because of exception in column 'time':
'Time' object has no attribute 'insert'

Initializing from a list of rows or a list of dicts

This mode of initializing a table does not work with mixin columns, so both of the following will fail:

>>> qt = QTable([{'a': 1 * u.m, 'b': 2},
...              {'a': 2 * u.m, 'b': 3}])  
Traceback (most recent call last):
 ...
TypeError: only dimensionless scalar quantities can be converted to Python scalars

>>> qt = QTable(rows=[[1 * u.m, 2],
...                   [2 * u.m, 3]])  
Traceback (most recent call last):
 ...
TypeError: only dimensionless scalar quantities can be converted to Python scalars

The problem lies in knowing if and how to assemble the individual elements for each column into an appropriate mixin column. The current code uses numpy to perform this function on numerical or string types, but it obviously does not handle mixin column types like Quantity or SkyCoord.

Masking

Mixin columns do not support masking, but there is limited support for use of mixins within a masked table. In this case a mask attribute is assigned to the mixin column object. This mask is a special object that is a boolean array of False corresponding to the mixin data shape. The mask looks like a normal numpy array but an exception will be raised if True is assigned to any element. The consequences of the limitation are most obvious in the high-level table operations.

High-level table operations

The table below gives a summary of support for high-level operations on tables that contain mixin columns:

Operation Support
Grouped operations Not implemented yet, but no fundamental limitation
Stack vertically Available for Quantity and any other mixin classes that provide an new_like() method in the info descriptor.
Stack horizontally Works if output mixin column supports masking or if no masking is required
Join Works if output mixin column supports masking or if no masking is required; key columns must be subclasses of numpy.ndarray.
Unique rows Not implemented yet, uses grouped operations

ASCII table writing

Mixin columns can be written out to file using the astropy.io.ascii module, but the fast C-based writers are not available. Instead the legacy pure-Python writers will be used.

Mixin protocol

A key idea behind mixin columns is that any class which satisfies a specified protocol can be used. That means many user-defined class objects which handle array-like data can be used natively within a Table. The protocol is relatively simple and requires that a class behave like a minimal numpy array with the following properties:

  • Contains array-like data
  • Implements __getitem__ to support getting data as a single item, slicing, or index array access
  • Has a shape attribute
  • Has a __len__ method for length
  • Has an info class descriptor which is a subclass of the astropy.utils.data_info.MixinInfo class.

The Example: ArrayWrapper section shows a working minimal example of a class which can be used as a mixin column. A pandas.Series object can function as a mixin column as well.

Other interesting possibilities for mixin columns include:

  • Columns which are dynamically computed as a function of other columns (AKA spreadsheet)
  • Columns which are themselves a Table, i.e. nested tables. A proof of concept is available.

new_like() method

In order to support high-level operations like join and vstack, a mixin class must provide a new_like() method in the info class descriptor. A key part of the functionality is to ensure that the input column metadata are merged appropriately and that the columns have consistent properties such as the shape.

A mixin class that provides new_like() must also implement __setitem__ to support setting via a single item, slicing, or index array.

The new_like method has the following signature:

def new_like(self, cols, length, metadata_conflicts='warn', name=None):
    """
    Return a new instance of this class which is consistent with the
    input ``cols`` and has ``length`` rows.

    This is intended for creating an empty column object whose elements can
    be set in-place for table operations like join or vstack.

    Parameters
    ----------
    cols : list
        List of input columns
    length : int
        Length of the output column object
    metadata_conflicts : str ('warn'|'error'|'silent')
        How to handle metadata conflicts
    name : str
        Output column name

    Returns
    -------
    col : object
        New instance of this class consistent with ``cols``
    """

Examples of this are found in the ColumnInfo and QuantityInfo classes.

Example: ArrayWrapper

The code listing below shows a example of a data container class which acts as a mixin column class. This class is a simple wrapper around a numpy array. It is used in the astropy mixin test suite and is fully compliant as a mixin column.

from astropy.utils.data_info import ParentDtypeInfo

class ArrayWrapper(object):
    """
    Minimal mixin using a simple wrapper around a numpy array
    """
    info = ParentDtypeInfo()

    def __init__(self, data):
        self.data = np.array(data)
        if 'info' in getattr(data, '__dict__', ()):
            self.info = data.info

    def __getitem__(self, item):
        if isinstance(item, (int, np.integer)):
            out = self.data[item]
        else:
            out = self.__class__(self.data[item])
            if 'info' in self.__dict__:
                out.info = self.info
        return out

    def __setitem__(self, item, value):
        self.data[item] = value

    def __len__(self):
        return len(self.data)

    @property
    def dtype(self):
        return self.data.dtype

    @property
    def shape(self):
        return self.data.shape

    def __repr__(self):
        return ("<{0} name='{1}' data={2}>"
                .format(self.__class__.__name__, self.info.name, self.data))