How to create arrays of strings#

Awkward Arrays can contain strings, although these strings are just a special view of lists of uint8 numbers. As such, the variable-length data are efficiently stored.

NumPy’s strings are padded to have equal width, and Pandas’s strings are Python objects. Awkward Array doesn’t have nearly as many functions for manipulating arrays of strings as NumPy and Pandas, though.

import awkward as ak
import numpy as np

From Python strings#

The ak.Array constructor and ak.from_iter() recognize strings, and strings are returned by ak.to_list().

ak.Array(["one", "two", "three"])

['one',
 'two',
 'three']
----------------
type: 3 * string

They may be nested within anything.

ak.Array([["one", "two"], [], ["three"]])

[['one', 'two'],
 [],
 ['three']]
----------------------
type: 3 * var * string

From NumPy arrays#

NumPy strings are also recognized by ak.from_numpy() and ak.to_numpy().

numpy_array = np.array(["one", "two", "three", "four"])
numpy_array

array(['one', 'two', 'three', 'four'], dtype='<U5')

awkward_array = ak.Array(numpy_array)
awkward_array

['one',
 'two',
 'three',
 'four']
----------------
type: 4 * string

Operations with strings#

Since strings are really just lists, some of the list operations “just work” on strings.

ak.num(awkward_array)

[3,
 3,
 5,
 4]
---------------
type: 4 * int64

awkward_array[:, 1:]

['ne',
 'wo',
 'hree',
 'our']
----------------
type: 4 * string

Others had to be specially overloaded for the string case, such as string-equality. The default meaning for == would be to descend to the lowest level and compare numbers (characters, in this case).

awkward_array == "three"

[False,
 False,
 True,
 False]
--------------
type: 4 * bool

awkward_array == ak.Array(["ONE", "TWO", "three", "four"])

[False,
 False,
 True,
 True]
--------------
type: 4 * bool

Similarly, ak.sort() and ak.argsort() sort strings lexicographically, not individual characters.

ak.sort(awkward_array)

['four',
 'one',
 'three',
 'two']
----------------
type: 4 * string

Still other operations had to be inhibited, since they wouldn’t make sense for strings.

np.sqrt(awkward_array)

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
Cell In[11], line 1
----> 1 np.sqrt(awkward_array)

File ~/micromamba/envs/awkward-docs/lib/python3.11/site-packages/awkward/highlevel.py:1511, in Array.__array_ufunc__(self, ufunc, method, *inputs, **kwargs)
name = f"{type(ufunc).__module__}.{ufunc.__name__}.{method!s}"
with ak._errors.OperationErrorContext(name, inputs, kwargs):
-> 1511     return ak._connect.numpy.array_ufunc(ufunc, method, inputs, kwargs)

File ~/micromamba/envs/awkward-docs/lib/python3.11/site-packages/awkward/_connect/numpy.py:466, in array_ufunc(ufunc, method, inputs, kwargs)
       raise TypeError(
           "no {}.{} overloads for custom types: {}".format(
               type(ufunc).__module__, ufunc.__name__, ", ".join(error_message)
           )
       )
   return None
--> 466 out = ak._broadcasting.broadcast_and_apply(
   inputs, action, allow_records=False, function_name=ufunc.__name__
)
if len(out) == 1:
   return wrap_layout(out[0], behavior=behavior, attrs=attrs)

File ~/micromamba/envs/awkward-docs/lib/python3.11/site-packages/awkward/_broadcasting.py:1140, in broadcast_and_apply(inputs, action, depth_context, lateral_context, allow_records, left_broadcast, right_broadcast, numpy_to_regular, regular_to_jagged, function_name, broadcast_parameters_rule)
backend = backend_of(*inputs, coerce_to_common=False)
isscalar = []
-> 1140 out = apply_step(
   backend,
   broadcast_pack(inputs, isscalar),
   action,
   0,
   depth_context,
   lateral_context,
   {
       "allow_records": allow_records,
       "left_broadcast": left_broadcast,
       "right_broadcast": right_broadcast,
       "numpy_to_regular": numpy_to_regular,
       "regular_to_jagged": regular_to_jagged,
       "function_name": function_name,
       "broadcast_parameters_rule": broadcast_parameters_rule,
   },
)
assert isinstance(out, tuple)
return tuple(broadcast_unpack(x, isscalar) for x in out)

File ~/micromamba/envs/awkward-docs/lib/python3.11/site-packages/awkward/_broadcasting.py:1118, in apply_step(backend, inputs, action, depth, depth_context, lateral_context, options)
   return result
elif result is None:
-> 1118     return continuation()
else:
   raise AssertionError(result)

File ~/micromamba/envs/awkward-docs/lib/python3.11/site-packages/awkward/_broadcasting.py:1087, in apply_step.<locals>.continuation()
# Any non-string list-types?
elif any(x.is_list and not is_string_like(x) for x in contents):
-> 1087     return broadcast_any_list()
# Any RecordArrays?
elif any(x.is_record for x in contents):

File ~/micromamba/envs/awkward-docs/lib/python3.11/site-packages/awkward/_broadcasting.py:623, in apply_step.<locals>.broadcast_any_list()
       nextinputs.append(x)
       nextparameters.append(NO_PARAMETERS)
--> 623 outcontent = apply_step(
   backend,
   nextinputs,
   action,
   depth + 1,
   copy.copy(depth_context),
   lateral_context,
   options,
)
assert isinstance(outcontent, tuple)
parameters = parameters_factory(nextparameters, len(outcontent))

File ~/micromamba/envs/awkward-docs/lib/python3.11/site-packages/awkward/_broadcasting.py:1100, in apply_step(backend, inputs, action, depth, depth_context, lateral_context, options)
   else:
       raise ValueError(
           "cannot broadcast: {}{}".format(
               ", ".join(repr(type(x)) for x in inputs), in_function(options)
           )
       )
-> 1100 result = action(
   inputs,
   depth=depth,
   depth_context=depth_context,
   lateral_context=lateral_context,
   continuation=continuation,
   backend=backend,
   options=options,
)
if isinstance(result, tuple) and all(isinstance(x, Content) for x in result):
   if any(content.backend is not backend for content in result):

File ~/micromamba/envs/awkward-docs/lib/python3.11/site-packages/awkward/_connect/numpy.py:402, in array_ufunc.<locals>.action(inputs, **ignore)
   # Do we have all-strings? If so, we can't proceed
   if all(
       x.is_list and x.parameter("__array__") in ("string", "bytestring")
       for x in contents
   ):
--> 402         raise TypeError(
           f"{type(ufunc).__module__}.{ufunc.__name__} is not implemented for string types. "
           "To register an implementation, add a name to these string(s) and register a behavior overload"
       )
if ufunc is numpy.matmul:
   raise NotImplementedError(
       "matrix multiplication (`@` or `np.matmul`) is not yet implemented for Awkward Arrays"
   )

TypeError: numpy.sqrt is not implemented for string types. To register an implementation, add a name to these string(s) and register a behavior overload

This error occurred while calling

    numpy.sqrt.__call__(
        <Array ['one', 'two', 'three', 'four'] type='4 * string'>
    )

Categorical strings#

A large set of strings with few unique values are more efficiently manipulated as integers than as strings. In Pandas, this is categorical data, in R, it’s called a factor, and in Arrow and Parquet, it’s dictionary encoding.

The ak.str.to_categorical() (requires PyArrow) function makes Awkward Arrays categorical in this sense. ak.to_arrow() and ak.to_parquet() recognize categorical data and convert it to the corresponding Arrow and Parquet types.

uncategorized = ak.Array(["three", "one", "two", "two", "three", "one", "one", "one"])
uncategorized

['three',
 'one',
 'two',
 'two',
 'three',
 'one',
 'one',
 'one']
----------------
type: 8 * string

categorized = ak.str.to_categorical(uncategorized)
categorized

['three',
 'one',
 'two',
 'two',
 'three',
 'one',
 'one',
 'one']
----------------------------------
type: 8 * categorical[type=string]

Internally, the data now have an index that selects from a set of unique strings.

categorized.layout.index

<Index dtype='int64' len='8'>[0 1 2 2 0 1 1 1]</Index>

ak.Array(categorized.layout.content)

['three',
 'one',
 'two']
----------------
type: 3 * string

The main advantage to Awkward categorical data (other than proper conversions to Arrow and Parquet) is that equality is performed using the index integers.

categorized == "one"

[False,
 True,
 False,
 False,
 False,
 True,
 True,
 True]
--------------
type: 8 * bool

With ArrayBuilder#

ak.ArrayBuilder() is described in more detail in this tutorial, but you can add strings by calling the string method or simply appending them.

(This is what ak.from_iter() uses internally to accumulate data.)

builder = ak.ArrayBuilder()

builder.string("one")
builder.append("two")
builder.append("three")

array = builder.snapshot()
array

['one',
 'two',
 'three']
----------------
type: 3 * string