ak.broadcast_arrays
-------------------
.. py:module: ak.broadcast_arrays
Defined in `awkward.operations.ak_broadcast_arrays `__ on `line 17 `__.
.. py:function:: ak.broadcast_arrays(*arrays, depth_limit=None, broadcast_parameters_rule='one_to_one', left_broadcast=True, right_broadcast=True, highlevel=True, behavior=None)
:param arrays: Array-like data (anything :py:obj:`ak.to_layout` recognizes).
:param depth_limit: If None, attempt to fully
broadcast the ``arrays`` to all levels. If an int, limit the number
of dimensions that get broadcasted. The minimum value is ``1``,
for no broadcasting.
:type depth_limit: None or int, default is None
:param broadcast_parameters_rule: Rule for broadcasting parameters, one of:
- ``"intersect"``
- ``"all_or_nothing"``
- ``"one_to_one"``
- ``"none"``
:type broadcast_parameters_rule: str
:param left_broadcast: If True, follow rules for implicit
left-broadcasting, as described below.
:type left_broadcast: bool
:param right_broadcast: If True, follow rules for implicit
right-broadcasting, as described below.
:type right_broadcast: bool
:param highlevel: If True, return an :py:obj:`ak.Array`;
otherwise, return a low-level :py:obj:`ak.contents.Content` subclass.
:type highlevel: bool, default is True
:param behavior: Custom :py:obj:`ak.behavior` for the output array, if
high-level.
:type behavior: None or dict
Like NumPy's
`broadcast_arrays `__
function, this function returns the input ``arrays`` with enough elements
duplicated that they can be combined element-by-element.
For NumPy arrays, this means that scalars are replaced with arrays with
the same scalar value repeated at every element of the array, and regular
dimensions are created to increase low-dimensional data into
high-dimensional data.
For example,
.. code-block:: python
>>> ak.broadcast_arrays(5,
... [1, 2, 3, 4, 5])
[,
]
and
.. code-block:: python
>>> ak.broadcast_arrays(np.array([1, 2, 3]),
... np.array([[0.1, 0.2, 0.3], [10, 20, 30]]))
[,
]
Note that in the second example, when the ``3 * int64`` array is expanded
to match the ``2 * 3 * float64`` array, it is the deepest dimension that
is aligned. If we try to match a ``2 * int64`` with the ``2 * 3 * float64``,
.. code-block:: python
>>> ak.broadcast_arrays(np.array([1, 2]),
... np.array([[0.1, 0.2, 0.3], [10, 20, 30]]))
ValueError: while calling
ak.broadcast_arrays(
arrays = (array([1, 2]), array([[ 0.1, 0.2, 0.3],
[10. , 20....
depth_limit = None
broadcast_parameters_rule = 'one_to_one'
left_broadcast = True
right_broadcast = True
highlevel = True
behavior = None
)
Error details: cannot broadcast RegularArray of size 2 with RegularArray of size 3
NumPy has the same behavior: arrays with different numbers of dimensions
are aligned to the right before expansion. One can control this by
explicitly adding a new axis (reshape to add a dimension of length 1)
where the expansion is supposed to take place because a dimension of
length 1 can be expanded like a scalar.
.. code-block:: python
>>> ak.broadcast_arrays(np.array([1, 2])[:, np.newaxis],
... np.array([[0.1, 0.2, 0.3], [10, 20, 30]]))
[,
]
Again, NumPy does the same thing (``np.newaxis`` is equal to None, so this
trick is often shown with None in the slice-tuple). Where the broadcasting
happens can be controlled, but numbers of dimensions that don't match are
implicitly aligned to the right (fitting innermost structure, not
outermost).
While that might be an arbitrary decision for rectilinear arrays, it is
much more natural for implicit broadcasting to align left for tree-like
structures. That is, the root of each data structure should agree and
leaves may be duplicated to match. For example,
.. code-block:: python
>>> ak.broadcast_arrays([ 100, 200, 300],
... [[1.1, 2.2, 3.3], [], [4.4, 5.5]])
[,
]
One typically wants single-item-per-element data to be duplicated to
match multiple-items-per-element data. Operations on the broadcasted
arrays like
.. code-block:: python
one_dimensional + nested_lists
would then have the same effect as the procedural code
.. code-block:: python
for x, outer in zip(one_dimensional, nested_lists):
output = []
for inner in outer:
output.append(x + inner)
yield output
where ``x`` has the same value for each ``inner`` in the inner loop.
Awkward Array's broadcasting manages to have it both ways by applying the
following rules:
* If all dimensions are regular (i.e. :py:obj:`ak.types.RegularType`), like NumPy,
implicit broadcasting aligns to the right, like NumPy.
* If any dimension is variable (i.e. :py:obj:`ak.types.ListType`), which can
never be true of NumPy, implicit broadcasting aligns to the left.
* Explicit broadcasting with a length-1 regular dimension always
broadcasts, like NumPy.
Thus, it is important to be aware of the distinction between a dimension
that is declared to be regular in the type specification and a dimension
that is allowed to be variable (even if it happens to have the same length
for all elements). This distinction is can be accessed through the
:py:obj:`ak.Array.type`, but it is lost when converting an array into JSON or
Python objects.
If arrays have the same depth but different lengths of nested
lists, attempting to broadcast them together is a broadcasting error.
.. code-block:: python
>>> one = ak.Array([[[1, 2, 3], [], [4, 5], [6]], [], [[7, 8]]])
>>> two = ak.Array([[[1.1, 2.2], [3.3], [4.4], [5.5]], [], [[6.6]]])
>>> ak.broadcast_arrays(one, two)
ValueError: while calling
ak.broadcast_arrays(
arrays = (>> this, that = ak.broadcast_arrays(one, two, depth_limit=1)
>>> this.show()
[[[1, 2, 3], [], [4, 5], [6]],
[],
[[7, 8]]]
>>> that.show()
[[[1.1, 2.2], [3.3], [4.4], [5.5]],
[],
[[6.6]]]