Defined in awkward.behaviors.categorical on line 172.

ak.to_categorical(array, highlevel=True)
  • array – Data convertible to an Awkward Array

  • highlevel (bool) – If True, return an ak.Array; otherwise, return a low-level ak.layout.Content subclass.

Creates a categorical dataset, which has the following properties:

  • only distinct values (categories) are stored in their entirety,

  • pointers to those distinct values are represented by integers (an ak.layout.IndexedArray or ak.layout.IndexedOptionArray labeled with parameter "__array__" = "categorical".

This is equivalent to R’s “factor”, Pandas’s “categorical”, and Arrow/Parquet’s “dictionary encoding.” It differs from generic uses of ak.layout.IndexedArray and ak.layout.IndexedOptionArray in Awkward Arrays by the guarantee of no duplicate categories and the "categorical" parameter.

>>> array = ak.Array([["one", "two", "three"], [], ["three", "two"]])
>>> categorical = ak.to_categorical(array)
>>> categorical
<Array [['one', 'two', ... 'three', 'two']] type='3 * var * categorical[type=str...'>
>>> ak.type(categorical)
3 * var * categorical[type=string]
>>> ak.to_list(categorical) == ak.to_list(array)
>>> ak.categories(categorical)
<Array ['one', 'two', 'three'] type='3 * string'>
>>> ak.is_categorical(categorical)
>>> ak.from_categorical(categorical)
<Array [['one', 'two', ... 'three', 'two']] type='3 * var * string'>

This function descends through nested lists, but not into the fields of records, so records can be categories. To make categorical record fields, split up the record, apply this function to each desired field, and ak.zip the results together.

>>> records = ak.Array([
...     {"x": 1.1, "y": "one"},
...     {"x": 2.2, "y": "two"},
...     {"x": 3.3, "y": "three"},
...     {"x": 2.2, "y": "two"},
...     {"x": 1.1, "y": "one"}
... ])
>>> records
<Array [{x: 1.1, y: 'one'}, ... y: 'one'}] type='5 * {"x": float64, "y": string}'>
>>> categorical_records = ak.zip({
...     "x": ak.to_categorical(records["x"]),
...     "y": ak.to_categorical(records["y"]),
... })
>>> categorical_records
<Array [{x: 1.1, y: 'one'}, ... y: 'one'}] type='5 * {"x": categorical[type=floa...'>
>>> ak.type(categorical_records)
5 * {"x": categorical[type=float64], "y": categorical[type=string]}
>>> ak.to_list(categorical_records) == ak.to_list(records)

The check for uniqueness is currently implemented in a Python loop, so conversion to categorical should be regarded as expensive. (This can change, but it would always be an _n log(n)_ operation.)

See also ak.is_categorical, ak.categories, ak.from_categorical.