Getting started#

Installation#

If you use pip, rip, pixi, or uv to install packages,
pip install awkward
If you use conda or mamba, it’s in the conda-forge channel.
conda install -c conda-forge awkward

If you’re installing as a developer or testing updates that haven’t been released in a package manager yet, see the developer installation instructions in the Contributor guide.

Tutorials#

See the left side-bar (or bring it into view by clicking on the upper-left ) for tutorials that illustrate the purpose and main concepts behind Awkward Arrays.

Frequently asked questions#

You can test any examples in a new window/tab by clicking on Try It! ⭷.

What is Awkward Array for? How does it compare to other libraries?

Python’s builtin lists, dicts, and classes can be used to analyze arbitrary data structures, but at a cost in speed and memory. Therefore, they can’t be used (easily) with large datasets.

Pandas DataFrames (as well as Polars, cuDF, and Dask DataFrame) are well-suited to tabular data, including tables with relational indexes, but not arbitrary data structures. If a DataFrame is filled with Python’s builtin types, then it offers no speed or memory advantage over Python itself.

NumPy is ideal for rectangular arrays of numbers, but not arbitrary data structures. If a NumPy array is filled with Python’s builtin types, then it offers no speed or memory advantage over Python itself.

Apache Arrow (pyarrow) manages arrays of arbitrary data structures (including those in Polars, cuDF, and to some extent, Pandas), with great language interoperability and interprocess communication, but without manipulation functions oriented toward data analysts.

Awkward Array is a data analyst-friendly extension of NumPy-like idioms for arbitrary data structures. It is intended to be used interchangeably with NumPy and share data with Arrow and DataFrames. Like NumPy, it simplifies and accelerates computations that transform arrays into arrays—all computations over elements in an array are compiled. Also like NumPy, imperative-style computations can be accelerated with Numba.

Note that there is also a ragged array library with simpler (but still non-rectangular) data types that more closely adheres to array APIs.

Where is an Awkward Array’s shape and dtype?

Since Awkward Arrays can contain arbitrary data structures, their type can’t be separated into a shape and a dtype, the way a NumPy array can.

For an array of records like

import awkward as ak

example = ak.Array([
    [{"x": 1.1, "y": [1]}, {"x": 2.2, "y": [1, 2]}, {"x": 3.3, "y": [1, 2, 3]}],
    [],
    [{"x": 4.4, "y": [1, 2, 3, 4]}, {"x": 5.5, "y": [1, 2, 3, 4, 5]}]
])

the x field contains floating point numbers and the y field contains lists of integers. They would have different dtypes, as well as different numbers of dimensions. This array also can’t be separated into x and y columns with different dtypes, as in a DataFrame, since both fields are inside of records in a variable-length list.

Instead, Awkward Arrays have a type, which looks like

3 * var * {x: float64, y: var * int64}

for the above. This combines shape and dtype information in the following way: the length of the array is 3, the first dimension has var or variable length, it contains records with x and y field names in { }, the x field has float64 primitive type and the y field is a var variable length list of int64. You can print(array.type) or array.type.show() to see the type of any array. (For more, see the DataShape language.)

See the ragged array library for variable-length dimensions that are nevertheless separable into a shape and dtype, like a conventional array.

How do I get Awkward Arrays, or read or write files of them?

After importing Awkward Array with

import awkward as ak

the ak.Array constructor takes NumPy arrays, CuPy arrays, pyarrow arrays, or an iterable of Python builtin lists and dicts, such as

example = ak.Array([
    [{"x": 1.1, "y": [1]}, {"x": 2.2, "y": [1, 2]}, {"x": 3.3, "y": [1, 2, 3]}],
    [],
    [{"x": 4.4, "y": [1, 2, 3, 4]}, {"x": 5.5, "y": [1, 2, 3, 4, 5]}]
])

This is a shorthand for functions such as ak.from_numpy(), ak.from_cupy(), ak.from_arrow(), and ak.from_iter(), which you can call explicitly for more control. Similarly, functions like ak.to_numpy(), ak.to_cupy(), ak.to_arrow(), and ak.to_list() convert Awkward Arrays into other types of arrays, or Python lists.

Several file formats have ak.from_* and ak.to_* functions, such as JSON, Parquet, and Feather. To read and write ROOT files, see Uproot.

In addition, there are low-level routines, ak.from_buffers() and ak.to_buffers(), to build new file or line protocol interfaces.

How do I slice Awkward Arrays?

Like NumPy: all NumPy slicing rules are supported, with generalizations to support more data types, as well as slicing rules that have no analog in rectangular arrays; see ak.Array.__getitem__().

Some common examples using

import awkward as ak

example = ak.Array([
    [{"x": 1.1, "y": [1]}, {"x": 2.2, "y": [1, 2]}, {"x": 3.3, "y": [1, 2, 3]}],
    [],
    [{"x": 4.4, "y": [1, 2, 3, 4]}, {"x": 5.5, "y": [1, 2, 3, 4, 5]}]
])

are

  • example[0:-1] to select a range in the first dimension,

  • example[:, 1:] to keep all elements of the first dimension but drop the first element of each nested list,

  • example["x"] or array.x to select the x field of all records,

  • example[array.x > 3] to select all records in which the field x is greater than 3,

  • example.y[:, :, [0, -1]] to select field y and take the first (0) and last (-1) element of each list of field y,

  • and so on.

How do I use NumPy functions with Awkward Arrays?

All NumPy universal functions can be applied to Awkward Arrays that do not contain record structures, as well as any other functions that have Awkward equivalents, such as ak.sum(), ak.argmax(), ak.mean(), ak.sort(), and ak.concatenate().

For example, with

import awkward as ak
import numpy as np

y = ak.Array([
    [[1], [1, 2], [1, 2, 3]],
    [],
    [[1, 2, 3, 4], [1, 2, 3, 4, 5]]
])

you can call

  • np.sqrt(y) to get an array of lists of lists of the square roots of the numbers above,

  • y * 2 or y + y to multiply every value by 2 (which calls np.multiply or np.add, which are NumPy ufuncs),

  • np.sum(y) to get the sum of all values,

  • np.argmax(y, axis=-1) to get the position of the maximum value of each inner list,

  • np.mean(y, axis=0) to get the mean of the first elements of each list, the second elements, and so on,

  • np.sort(y) to sort lists,

  • np.concatenate((y, y)) to concatenate them,

  • and so on.

How do I flatten a ragged array for plotting?

ak.flatten() eliminates one level of nested lists, and ak.ravel() eliminates them all. ak.flatten() also removes missing values (None), which plotting libraries might not recognize.

Depending on what you’re trying to plot, selecting the first element of each list or computing the ak.sum() or ak.mean() of each list might be more meaningful.

How do I make ragged dimensions regular for (ML) algorithms that require it?

The ak.to_regular() function changes the data type from variable-length (var) to fixed-length if all lists in that dimension happen to have the same length anyway.

If you need to change the data to make it conform to a rectangular shape, you can

How can I make or break records in arrays?

Record (struct/class) data structures may come from JSON objects, Arrow Tables, Parquet columns, etc. This small Python dataset produces an array of lists of records:

import awkward as ak

example = ak.Array([
    [{"x": 1.1, "y": [1]}, {"x": 2.2, "y": [1, 2]}, {"x": 3.3, "y": [1, 2, 3]}],
    [],
    [{"x": 4.4, "y": [1, 2, 3, 4]}, {"x": 5.5, "y": [1, 2, 3, 4, 5]}]
])

Individual fields can be extracted by slicing it: example["x"] (example.x) and example["y"] (example.y), and all fields can be extracted at once with ak.unzip():

x, y = ak.unzip(example)

The following is a particularly useful idiom, for turning an array of records into a Python dict of arrays, using both ak.fields() and ak.unzip():

dict_of_arrays = dict(zip(ak.fields(example), ak.unzip(example)))

The opposite, ak.zip(), takes a Python dict of arrays and makes a record array:

ak.zip(dict_of_arrays)

When a set of Awkward Arrays are zipped together, it’s not clear which level of nested lists should be populated with records; ak.zip() attempts to create records at the deepest level, inside of all nested lists (which might not even be possible, if the Awkward Arrays don’t have the same list lengths at all levels). The depth_limit argument of ak.zip() controls this:

ak.zip(dict_of_arrays, depth_limit=2)

reproduces the original example, in which the y field has one more dimension than the x field (scalar x values sit beside y values that are lists).

How do I add a field to an existing record array?

As a shorthand for ak.unzip(), add a field, and ak.zip() (see the question above), new fields can be assigned with ak.Array.__setitem__().

For example, with

import awkward as ak

example = ak.Array([
    [{"x": 1.1, "y": [1]}, {"x": 2.2, "y": [1, 2]}, {"x": 3.3, "y": [1, 2, 3]}],
    [],
    [{"x": 4.4, "y": [1, 2, 3, 4]}, {"x": 5.5, "y": [1, 2, 3, 4, 5]}]
])

you can add a third field, z to the record with

example["z"] = example.x * 10

Note that for assignment, the left-hand side must be expressed with square brackets, not a dot. This is to support assignment into records nested within records.

Why can’t I assign numerical values in an array?

Awkward Arrays are immutable, and almost all operations on them view parts of a data structure and only replace the parts that have changed. Therefore, with an array like

import awkward as ak

example = ak.Array([
    [{"x": 1.1, "y": [1]}, {"x": 2.2, "y": [1, 2]}, {"x": 3.3, "y": [1, 2, 3]}],
    [],
    [{"x": 4.4, "y": [1, 2, 3, 4]}, {"x": 5.5, "y": [1, 2, 3, 4, 5]}]
])

attempting to assign

example[0, "x", 0] = 999

results in an error. (If it were allowed, it could have unpredictable consequences.) Immutability is not enforced at a very low level, so if you know what you’re doing, you can deconstruct the array, view it in NumPy, Arrow, or as raw memory buffers, and change it.

Problems that would be solved by assigning values in place can usually be solved by ak.where().

The only kind of assignment that is allowed is to add new fields, such as

example["z"] = 999

(see the question above). This kind of assignment won’t cause values in another array to change unpredictably.

How do I get rid of missing values (None)?

Some functions, such as ak.min() or ak.max() on empty lists, produce missing values. For example, with

import awkward as ak

x = ak.Array([[1.1, 2.2, 3.3], [], [4.4, 5.5]])

ak.max(x, axis=1) returns

<Array [3.3, None, 5.5] type='3 * ?float64'>

None represents a missing value (distinct from floating-point nan), and the ? or option[float64] in the type means that values could be missing. Such an array can be used in numerical calculations—missing values pass through most functions as missing values in the output—but third-party libraries might not recognize them.

  • ak.drop_none() simply removes the missing values, changing the lengths of lists and the data type to reflect the fact that no values are missing.

  • ak.flatten() removes missing values in the process of flattening nested lists (it treats None like []).

  • ak.fill_none() lets you replace missing values with a specified value.

  • ak.firsts() and ak.singletons() convert between representing option-type data as option[T] and var * T. In the latter, a missing value is an empty list and a non-missing value is a length-1 list.

Why am I getting ValueError or IndexError in mathematical operations?

Most likely, your arrays don’t line up at every level of nested lists. This is a generalization of a shape mismatch in rectangular arrays.

For example, with

import awkward as ak

x = ak.Array([[1.1, 2.2, 3.3], [], [4.4, 5.5]])
y = ak.Array([[1.1, 2.2, 3.3, 999], [], [4.4, 5.5]])

an attempt to add x + y would fail because even though x and y have the same array length (3), the length of the first list differs (3 versus 4).

This type of error is often more subtle than the example above. It won’t happen if two arrays are derived from the same array with shape-preserving operations, but if, for instance, you remove outlier data from from one array and not another, they may fail to line up somewhere in the middle of a large dataset.

One way to avoid that is to introduce missing values (None) instead of removing outliers. Whereas

x[x > 2]

makes an array without values smaller than 2,

<Array [[2.2, 3.3], [], [4.4, 5.5]] type='3 * var * float64'>

a mask,

x.mask[x > 2]

replaces the values smaller than 2 with None:

<Array [[None, 2.2, 3.3], [], [4.4, 5.5]] type='3 * var * ?float64'>

This preserves the shape of the array so that it can continue to be used in mathematical expressions. For instance, x + x.mask[x > 2] returns

<Array [[None, 4.4, 6.6], [], [8.8, 11]] type='3 * var * ?float64'>

(the missing value propagates through to the output).

Missing values can be dropped, using ak.drop_none(), or replaced, using ak.fill_none(), as described in the question above.

How do I use Awkward Array with Numba?

Awkward Arrays can be passed into and out of functions that have been JIT-compiled with Numba. For example, with

import awkward as ak
import numpy as np
import numba as nb

example = ak.Array([
    [{"x": 1.1, "y": [1]}, {"x": 2.2, "y": [1, 2]}, {"x": 3.3, "y": [1, 2, 3]}],
    [],
    [{"x": 4.4, "y": [1, 2, 3, 4]}, {"x": 5.5, "y": [1, 2, 3, 4, 5]}]
])

A function that sums x in each entry (like ak.sum()) can be written in JIT-compiled imperative Python like this:

@nb.jit
def sum_over_x(array):
    output = np.zeros(len(array))
    for i, list_of_records in enumerate(array):
        for record in list_of_records:
            output[i] += record.x
    return output

sum_over_x(example)

Since Numba JIT-compiled the function, it doesn’t suffer the usual slow-down of iterating in Python. On the other hand, all variables in the function must have fixed data type and adhere to Numba’s set of supported Python features and supported NumPy features to be compiled. None of Awkward Array’s ak.* functions can be used—only iteration over values.

A JIT-compiled function can also return a part of the input Awkward Array:

@nb.jit
def record_in_which_y_sums_to_10(array):
    for list_of_records in array:
        for record in list_of_records:
            if np.asarray(record.y).sum() == 10:
                return record

record_in_which_y_sums_to_10(example)

returns

<Record {x: 4.4, y: [1, ..., 4]} type='{x: float64, y: var * int64}'>

which is the record that has np.asarray(record.y).sum() == 10. (One-dimensional Awkward Arrays may be cast as NumPy arrays, to take advantage of NumPy functions.)

Awkward Arrays are immutable inside of JIT-compiled functions, just as they are outside. To create new Awkward Arrays with Numba, use ak.ArrayBuilder.

Awkward Arrays with ak.backend() equal to "cuda" can be passed to Numba functions on GPUs, compiled with @nb.cuda.jit. See How to use Awkward Arrays in Numba’s CUDA target for more.

The choice between computing outside of a Numba JIT-compiled function and outside of one is an either/or choice between imperative style in Numba (only iteration is allowed, no ak.* functions or fancy slices) and array-oriented style outside (iteration is slow in Python; ak.* functions are encouraged).

How do I perform computations on arrays of spatial or momentum vectors?

For 2-D, 3-D, and 4-D space and space-time vectors, see the Vector library. These can be used as momentum vectors for physics with a variety of coordinate transformations, including special relativity. As Awkward Arrays, these each vector is a record whose fields are coordinates, such as x, y, z or rho, phi, theta.

To enable Awkward Arrays of vectors, import Vector as

import vector
vector.register_awkward()

and now any Awkward records with an appropriate name, such as

example = ak.zip({
    "x": ak.Array([[1.1, 2.2, 3.3], [], [4.4, 5.5]]),
    "y": ak.Array([[  1,   2,   3], [], [  4,   5]]),
    "z": ak.Array([[ 10,  20,  30], [], [ 40,  50]])
}, with_name="Momentum3D")

is recognized as an array of vectors with methods like

example.phi

and

(2 * example).is_parallel(example)

These methods also work in Numba JIT-compiled functions. (See the above question.)

Several array-constructing functions accept a with_name argument, including the ak.Array constructor and ak.zip(). There’s also a ak.with_name() function to add a name after an array has already been created.

How would I write my own suite of functions, like Vector?

Add new classes or functions to ak.behavior, which links record names to Python code.

Names are strings that can be saved in files or transferred across networks, but Python code is not always serializable.

How can I delay or distribute a computation?

Use Dask. The dask-awkward library provides a new high-level collection for Awkward Arrays, similar to dask.array and dask.dataframe.

For example, with

import awkward as ak
import dask_awkward as dak

example = ak.Array([
    [{"x": 1.1, "y": [1]}, {"x": 2.2, "y": [1, 2]}, {"x": 3.3, "y": [1, 2, 3]}],
    [],
    [{"x": 4.4, "y": [1, 2, 3, 4]}, {"x": 5.5, "y": [1, 2, 3, 4, 5]}]
])

you can make delayed data with

dak.from_awkward(example, npartitions=1)

although it’s more common to use dak.from_parquet or uproot.dask.

Any operations on this delayed array are collected as a Directed Acyclic Graph (DAG) that is computed when you call dask.compute. The computation may be distributed across multiple CPUs on one computer or across multiple computers in a network.

The dask-awkward project intends to cover the same interface as Awkward Array, though there may be some functions implemented in ak.* that aren’t in dak.* yet. See dask-awkward’s GitHub Issues for Dask-specific issues.

How do I use Awkward Array with ROOT?

Uproot can read and write ROOT files, and works with Awkward Arrays by default.

Also, ak.to_rdataframe() and ak.from_rdataframe() converts Awkward Arrays in memory to and from ROOT’s RDataFrame for computations. See How to convert to/from ROOT RDataFrame for details.

How do I use Awkward Array with C++?

One method is to convert Awkward Arrays to or from ROOT’s RDataFrame using ak.to_rdataframe() and ak.from_rdataframe(). RDataFrame supports computation in JIT-compiled C++.

Another method is to pass Awkward Arrays into JIT-compiled C++ functions defined with cppyy’s cppdef. This interface is similar to Numba, in that the JIT-compiled functions have arbitrary arguments and return values, rather than fitting into a pipeline like RDataFrame, but it also means that you need to set up the loop over entries manually and inside the compiled block. See How to use Awkward Arrays in C++ with cppyy for details.

If you are a library developer wishing to produce and/or consume Awkward Arrays in ahead-of-time compiled code (not JIT), like fastjet, you’ll want to use LayoutBuilder, ak.from_buffers()/ak.to_buffers(), or both. LayoutBuilder constructs an append-only array object like ak.ArrayBuilder, but with statically typed array type in header-only C++ that can be integrated with CMake.

How do I use Awkward Array with Julia?

AwkwardArray.jl is a Julia implementation of Awkward Array, sharing the same memory layout. It can therefore be used as a JIT-compilation target like Numba and C++ (see questions above), but with more flexibility: a single array data type can be used as an ak.Array and as an ak.ArrayBuilder. Whereas Pythonic Awkward Arrays are only borrowed by JIT-compiled Numba or C++ (Python continues to own the memory and decide when it will be deleted), Julia’s JIT-compiled environment is the entire environment, so such decisions don’t need to be made. Julia can act as a producer and/or a consumer of Python Awkward Arrays.

See the AwkwardArray.jl documentation for details.

How do I emulate nested for-loops in Awkward Array (combinatorics)?

In the simplest cases, imperative code like

output = []
for x in awkward_array:
    output.append(compute(x))

can be replaced with

output = compute(awkward_array)

But some problems would be solved by imperative code like

output = []
for x in awkward_array1:
    for y in awkward_array2:
        output.append(compute(x, y))

or even

output = []
for i, x in enumerate(awkward_array):
    for j in range(i + 1, len(awkward_array)):  # avoid repeating x
        y = awkward_array[j]
        output.append(compute(x, y))

These cases involve combinatorics: a Cartesian product and sampling without replacement. To perform such operations at compiled speeds on Awkward Arrays, you may either

  • JIT-compile these for loops with Numba, C++, or Julia (as in the questions above),

  • use Awkward Array’s combinatorics primitives.

ak.cartesian() is Awkward Array’s primitive for Cartesian products: it makes an array of all pairs drawn from two (or more) provided arrays. It emulates nested, unrestricted for loops.

ak.combinations() is Awkward Array’s primitive for sampling without replacement: it makes an array of all pairs drawn from an array and itself without duplicates. It emulates nested for loops that avoid repeating the same element.

These pairs (or triples, etc.) are tuples, which are records without field names. Often, nested=True is a useful argument to avoid flattening the output.

Why don’t my arrays broadcast as in NumPy?

See the last section of How Awkward broadcasting works.

What if I need more help? What if I think I’ve found a bug?

After checking the tutorials on the left-bar of this Getting started guide, the User guide, and the API reference, you can ask questions about how to use a feature or solve a problem on Awkward Array’s GitHub Discussions.

If the behavior you’re seeing looks like a bug, an error in Awkward Array itself, post some simplified code to reproduce it on Awkward Array’s GitHub Issues.