The N-dimensional array (ndarray)

An ndarray is a (usually fixed-size) multidimensional container of items of the same type and size. The number of dimensions and items in an array is defined by its shape, which is a tuple of N non-negative integers that specify the sizes of each dimension. The type of items in the array is specified by a separate data-type object (dtype), one of which is associated with each ndarray.

As with other container objects in Python, the contents of an ndarray can be accessed and modified by indexing or slicing the array (using, for example, N integers), and via the methods and attributes of the ndarray.

Different ndarrays can share the same data, so that changes made in one ndarray may be visible in another. That is, an ndarray can be a “view” to another ndarray, and the data it is referring to is taken care of by the “base” ndarray. ndarrays can also be views to memory owned by Python strings or objects implementing the buffer or array interfaces.

Example

A 2-dimensional array of size 2 x 3, composed of 4-byte integer elements:

>>> x = np.array([[1, 2, 3], [4, 5, 6]], np.int32)
>>> type(x)
<type 'numpy.ndarray'>
>>> x.shape
(2, 3)
>>> x.dtype
dtype('int32')

The array can be indexed using Python container-like syntax:

>>> # The element of x in the *second* row, *third* column, namely, 6.
>>> x[1, 2]

For example slicing can produce views of the array:

>>> y = x[:,1]
>>> y
array([2, 5])
>>> y[0] = 9 # this also changes the corresponding element in x
>>> y
array([9, 5])
>>> x
array([[1, 9, 3],
       [4, 5, 6]])

Constructing arrays

New arrays can be constructed using the routines detailed in Array creation routines, and also by using the low-level ndarray constructor:

Indexing arrays

Arrays can be indexed using an extended Python slicing syntax, array[selection]. Similar syntax is also used for accessing fields in a structured data type.

Internal memory layout of an ndarray

An instance of class ndarray consists of a contiguous one-dimensional segment of computer memory (owned by the array, or by some other object), combined with an indexing scheme that maps N integers into the location of an item in the block. The ranges in which the indices can vary is specified by the shape of the array. How many bytes each item takes and how the bytes are interpreted is defined by the data-type object associated with the array.

A segment of memory is inherently 1-dimensional, and there are many different schemes for arranging the items of an N-dimensional array in a 1-dimensional block. NumPy is flexible, and ndarray objects can accommodate any strided indexing scheme. In a strided scheme, the N-dimensional index corresponds to the offset (in bytes):

from the beginning of the memory block associated with the array. Here, are integers which specify the strides of the array. The column-major order (used, for example, in the Fortran language and in Matlab) and row-major order (used in C) schemes are just specific kinds of strided scheme, and correspond to memory that can be addressed by the strides:

where = self.shape[j].

Both the C and Fortran orders are contiguous, i.e., single-segment, memory layouts, in which every part of the memory block can be accessed by some combination of the indices.

While a C-style and Fortran-style contiguous array, which has the corresponding flags set, can be addressed with the above strides, the actual strides may be different. This can happen in two cases:

1. If self.shape[k] == 1 then for any legal index index[k] == 0. This means that in the formula for the offset and thus and the value of = self.strides[k] is arbitrary.
2. If an array has no elements (self.size == 0) there is no legal index and the strides are never used. Any array with no elements may be considered C-style and Fortran-style contiguous.

Point 1. means that self and self.squeeze() always have the same contiguity and aligned flags value. This also means that even a high dimensional array could be C-style and Fortran-style contiguous at the same time.

An array is considered aligned if the memory offsets for all elements and the base offset itself is a multiple of self.itemsize. Understanding memory-alignment leads to better performance on most hardware.

Data in new ndarrays is in the row-major (C) order, unless otherwise specified, but, for example, basic array slicing often produces views in a different scheme.

Array attributes

Array attributes reflect information that is intrinsic to the array itself. Generally, accessing an array through its attributes allows you to get and sometimes set intrinsic properties of the array without creating a new array. The exposed attributes are the core parts of an array and only some of them can be reset meaningfully without creating a new array. Information on each attribute is given below.

Memory layout

The following attributes contain information about the memory layout of the array:

Data type

The data type object associated with the array can be found in the dtype attribute:

Other attributes

Array interface

ctypes foreign function interface

Array methods

An ndarray object has many methods which operate on or with the array in some fashion, typically returning an array result. These methods are briefly explained below. (Each method’s docstring has a more complete description.)

For the following methods there are also corresponding functions in numpy: all, any, argmax, argmin, argpartition, argsort, choose, clip, compress, copy, cumprod, cumsum, diagonal, imag, max, mean, min, nonzero, partition, prod, ptp, put, ravel, real, repeat, reshape, round, searchsorted, sort, squeeze, std, sum, swapaxes, take, trace, transpose, var.

Array conversion

Shape manipulation

For reshape, resize, and transpose, the single tuple argument may be replaced with n integers which will be interpreted as an n-tuple.

Item selection and manipulation

For array methods that take an axis keyword, it defaults to None. If axis is None, then the array is treated as a 1-D array. Any other value for axis represents the dimension along which the operation should proceed.

Calculation

Many of these methods take an argument named axis. In such cases,

• If axis is None (the default), the array is treated as a 1-D array and the operation is performed over the entire array. This behavior is also the default if self is a 0-dimensional array or array scalar. (An array scalar is an instance of the types/classes float32, float64, etc., whereas a 0-dimensional array is an ndarray instance containing precisely one array scalar.)
• If axis is an integer, then the operation is done over the given axis (for each 1-D subarray that can be created along the given axis).

>>> x
array([[[ 0,  1,  2],
        [ 3,  4,  5],
        [ 6,  7,  8]],
       [[ 9, 10, 11],
        [12, 13, 14],
        [15, 16, 17]],
       [[18, 19, 20],
        [21, 22, 23],
        [24, 25, 26]]])
>>> x.sum(axis=0)
array([[27, 30, 33],
       [36, 39, 42],
       [45, 48, 51]])
>>> # for sum, axis is the first keyword, so we may omit it,
>>> # specifying only its value
>>> x.sum(0), x.sum(1), x.sum(2)
(array([[27, 30, 33],
        [36, 39, 42],
        [45, 48, 51]]),
 array([[ 9, 12, 15],
        [36, 39, 42],
        [63, 66, 69]]),
 array([[ 3, 12, 21],
        [30, 39, 48],
        [57, 66, 75]]))

The parameter dtype specifies the data type over which a reduction operation (like summing) should take place. The default reduce data type is the same as the data type of self. To avoid overflow, it can be useful to perform the reduction using a larger data type.

For several methods, an optional out argument can also be provided and the result will be placed into the output array given. The out argument must be an ndarray and have the same number of elements. It can have a different data type in which case casting will be performed.

Arithmetic, matrix multiplication, and comparison operations

Arithmetic and comparison operations on ndarrays are defined as element-wise operations, and generally yield ndarray objects as results.

Each of the arithmetic operations (+, -, *, /, //, %, divmod(), ** or pow(), <<, >>, &, ^, |, ~) and the comparisons (==, <, >, <=, >=, !=) is equivalent to the corresponding universal function (or ufunc for short) in NumPy. For more information, see the section on Universal Functions.

Comparison operators:

Truth value of an array (bool):

Unary operations:

Arithmetic:

Arithmetic, in-place:

Matrix Multiplication:

Special methods

For standard library functions:

Basic customization:

Container customization: (see Indexing)

Conversion; the operations int, float and complex. . They work only on arrays that have one element in them and return the appropriate scalar.

String representations: