Frequently Asked Questions¶
- General questions about NumPy
- General questions about SciPy
- NumPy vs. SciPy vs. other packages
- Python version support
- Basic NumPy/SciPy usage
- What is the preferred way to check for an empty (zero element) array?
- I want to load data from a text file. How do I make this code more efficient?
- What is the difference between matrices and arrays?
- Why not just have a separate operator for matrix multiplication?
- How do I find the indices of an array where some condition is true?
- How do I count the number of times each value appears in an array of integers?
- Advanced NumPy usage
- Where to get help
NumPy is a Python extension module that provides efficient operation on arrays of homogeneous data. It allows Python to serve as a high-level language for manipulating numerical data, much like for example IDL or MATLAB.
As always, you should choose the programming tools that suit your problem and your environment. Advantages many people cite are that it is open-source, it doesn’t cost anything, it uses a general-purpose programming language (Python) which is very popular and has high-quality libraries for almost any task available, and it is relatively easy to connect existing C and Fortran code to the Python interpreter.
A NumPy array is a multidimensional array of objects all of the same type. In memory, it is an object which points to a block of memory, keeps track of the type of data stored in that memory, keeps track of how many dimensions there are and how large each one is, and - importantly - the spacing between elements along each axis.
For example, you might have a NumPy array that represents the numbers from zero to nine, stored as 32-bit integers, one right after another, in a single block of memory. (for comparison, each Python integer needs to have some type information stored alongside it). You might also have the array of even numbers from zero to eight, stored in the same block of memory, but with a gap of four bytes (one 32-bit integer) between elements. This is called striding, and it means that you can often create a new array referring to a subset of the elements in an array without copying any data. Such subsets are called views. This is an efficiency gain, obviously, but it also allows modification of selected elements of an array in various ways.
An important constraint on NumPy arrays is that for a given axis, all the elements must be spaced by the same number of bytes in memory. NumPy cannot use double-indirection to access array elements, so indexing modes that would require this must produce copies. This constraint makes it possible for all the inner loops in NumPy’s internals to be written in efficient C code.
NumPy arrays offer a number of other possibilities, including using a memory-mapped disk file as the storage space for an array, and record arrays, where each element can have a custom, compound data type.
Python’s lists are efficient general-purpose containers. They support (fairly) efficient insertion, deletion, appending, and concatenation, and Python’s list comprehensions make them easy to construct and manipulate. However, they have certain limitations: they don’t support “vectorized” operations like elementwise addition and multiplication, and the fact that they can contain objects of differing types mean that Python must store type information for every element, and must execute type dispatching code when operating on each element. This also means that very few list operations can be carried out by efficient C loops – each iteration would require type checks and other Python API bookkeeping.
The short version is that Numeric was the original package that provided efficient homogeneous numeric arrays for Python, but some developers felt it lacked certain essential features, so they began developing an independent implementation called numarray. Having two incompatible implementations of array was clearly a disaster in the making, so NumPy was designed to be an improvement on both.
Neither Numeric nor numarray is currently supported. NumPy has been the standard array package for a number of years now. If you use Numeric or numarray, you should upgrade; NumPy is explicitly designed to have all the capabilities of both (and already boasts new features found in neither of its predecessor packages). There are tools available to ease the upgrade process; only C code should require much modification.
SciPy is a set of open source (BSD licensed) scientific and numerical tools for Python. It currently supports special functions, integration, ordinary differential equation (ODE) solvers, gradient optimization, parallel programming tools, an expression-to-C++ compiler for fast execution, and others. A good rule of thumb is that if it’s covered in a general textbook on numerical computing (for example, the well-known Numerical Recipes series), it’s probably implemented in SciPy.
SciPy is freely available. It is distributed as open source software, meaning that you have complete access to the source code and can use it in any way allowed by its liberal BSD license.
SciPy’s license is free for both commercial and non-commercial use, per the terms of the BSD license here.
Actually, the time-critical loops are usually implemented in C, C++ or Fortran. Parts of SciPy are thin layers of code on top of the scientific routines that are freely available at http://www.netlib.org/. Netlib is a huge repository of incredibly valuable and robust scientific algorithms written in C and Fortran. It would be silly to rewrite these algorithms and would take years to debug them. SciPy uses a variety of methods to generate “wrappers” around these algorithms so that they can be used in Python. Some wrappers were generated by hand coding them in C. The rest were generated using either SWIG or f2py. Some of the newer contributions to SciPy are either written entirely or wrapped with Cython.
A second answer is that for difficult problems, a better algorithm can make a tremendous difference in the time it takes to solve a problem. So using SciPy’s built-in algorithms may be much faster than a simple algorithm coded in C.
The SciPy development team works hard to make SciPy as reliable as possible, but, as in any software product, bugs do occur. If you find bugs that affect your software, please tell us by entering a ticket in the SciPy bug tracker, or NumPy bug tracker, as appropriate.
Drop us a mail on the mailing lists. We are keen for more people to help out writing code, unit tests, documentation (including translations into other languages), and helping out with the website.
In an ideal world, NumPy would contain nothing but the array data type and the most basic operations: indexing, sorting, reshaping, basic elementwise functions, et cetera. All numerical code would reside in SciPy. However, one of NumPy’s important goals is compatibility, so NumPy tries to retain all features supported by either of its predecessors. Thus NumPy contains some linear algebra functions and Fourier transforms, even though these more properly belong in SciPy. In any case, SciPy contains more fully-featured versions of the linear algebra modules, as well as many other numerical algorithms. If you are doing scientific computing with Python, you should probably install both NumPy and SciPy. Most new features belong in SciPy rather than NumPy.
Plotting functionality is beyond the scope of NumPy and SciPy, which focus on numerical objects and algorithms. Several packages exist that integrate closely with NumPy and Pandas to produce high quality plots, such as the immensely popular Matplotlib. Other popular options are Bokeh, Plotly, Altair and Chaco.
Like 2D plotting, 3D graphics is beyond the scope of NumPy and SciPy, but
just as in the 2D case, packages exist that integrate with NumPy.
Matplotlib provides basic 3D plotting in the
mplot3d subpackage, whereas
Mayavi provides a wide range
of high-quality 3D visualization features, utilizing the powerful
One of the design goals of NumPy was to make it buildable without a Fortran compiler, and if you don’t have LAPACK available NumPy will use its own implementation. SciPy requires a Fortran compiler to be built, and heavily depends on wrapped Fortran code.
linalg modules in NumPy and SciPy have some common functions but
with different docstrings, and
scipy.linalg contains functions not
numpy.linalg, such as functions related to
LU decomposition and the Schur decomposition, multiple ways
of calculating the pseudoinverse, and matrix transcendentals like the matrix
logarithm. Some functions that exist in both have augmented functionality
scipy.linalg; for example
scipy.linalg.eig() can take a second
matrix argument for solving generalized eigenvalue problems.
The last version of NumPy to support Python 2.7 is numpy 1.16.x. The last SciPy version to do so is 1.2.x. The first release of NumPy to support Python 3.x was NumPy 1.5.0. Python 3 support in SciPy was introduced in 0.9.0.
In general, yes. Recent improvements in PyPy have made the scientific Python stack work with PyPy. The NumPy and SciPy projects run PyPy in continuous integration and aim to further improve support over time. Since much of NumPy and SciPy is implemented as C extension modules, the code may not run any faster (for most cases it’s significantly slower still, however PyPy is actively working on improving this). As always when benchmarking, your experience is the best guide.
No, neither are supported. Jython never worked, because it runs on top of the Java Virtual Machine and has no way to interface with extensions written in C for the standard Python (CPython) interpreter.
Some years ago there was an effort to make NumPy and SciPy compatible with .NET. Some users at the time reported success in using NumPy with Ironclad on 32-bit Windows.
If you are certain a variable is an array, then use the size attribute.
If the variable may be a list or other sequence type, use
The size attribute is preferable to len because:
>>> a = numpy.zeros((1,0)) >>> a.size 0
>>> len(a) 1
numpy.loadtxt(). Even if your text file has header and footer
lines or comments, loadtxt can almost certainly read it; it is convenient and
If you find this still too slow, you can try Pandas (it has a faster csv reader for example). If that doesn’t help, you should consider changing to a more efficient file format than plain text. There are a large number of alternatives, depending on your needs (and on which version of NumPy/SciPy you are using):
- Text files: slow, huge, portable, human-readable; built into NumPy
- pickle: somewhat slow, somewhat portable (may be incompatible with different NumPy versions); built into NumPy
- HDF5: high-powered kitchen-sink format; both PyTables and h5py provide a NumPy friendly interface on top of the core HDF5 library written in C.
- FITS: standard kitchen-sink format in astronomy; the astropy library provides a convenient Python interface through its io.fits package.
- .npy: NumPy native binary data format, simple, efficient, portable; built into NumPy as of 1.0.5.
Note: NumPy matrices will be deprecated, do not use them for new code. The rest of the answer below is kept for historical reasons.
NumPy’s basic data type is the multidimensional array. These can be
one-dimensional (that is, one index, like a list or a vector),
two-dimensional (two indices, like an image), three-dimensional, or more
(zero-dimensional arrays exist and are a slightly strange corner case).
They support various operations, including addition, subtraction,
multiplication, exponentiation, and so on - but all of these are
elementwise operations. If you want matrix multiplication between two
two-dimensional arrays, the function
numpy.dot() or the built-in Python
@ do this. It also works fine for getting the matrix product of
a two-dimensional array and a one-dimensional array, in either direction, or
two one-dimensional arrays. If you want some kind of matrix
multiplication-like operation on higher-dimensional arrays (tensor
contraction), you need to think which indices you want to be contracting
over. Some combination of
rollaxis() should do
what you want.
However, some users find that they are doing so many matrix multiplications
that always having to write
dot as a prefix is too cumbersome, or they
really want to keep row and column vectors separate. For these users, there
is a matrix class. This is simply a transparent wrapper around arrays that
forces arrays to be at least two-dimensional, and that overloads the
multiplication and exponentiation operations. Multiplication becomes matrix
multiplication, and exponentiation becomes matrix exponentiation. If you want
elementwise multiplication, use
asmatrix() converts an array into a matrix (without ever
copying any data);
asarray() converts matrices to arrays.
asanyarray() makes sure that the result is either a matrix or an array
(but not, say, a list). Unfortunately, a few of NumPy’s many functions use
asarray() when they should use
asanyarray(), so from time to time
you may find your matrices accidentally get converted into arrays. Just use
asmatrix() on the output of these operations, and consider filing a bug.
From Python 3.5, the
@ symbol will be defined as a matrix multiplication
operator, and Numpy and Scipy will make use of this. This addition was the
subject of PEP 465. The separate
matrix and array types exist to work around the lack of this operator in earlier
versions of Python.
The prefered idiom for doing this is to use the function
, or the
nonzero() method of an array. Given an array
a > 3 returns a boolean array and since
interpreted as 0 in Python and NumPy,
np.nonzero(a > 3) yields the indices
a where the condition is true.
>>> import numpy as np >>> a = np.array([[1,2,3],[4,5,6],[7,8,9]]) >>> a > 3 array([[False, False, False], [ True, True, True], [ True, True, True]], dtype=bool) >>> np.nonzero(a > 3) (array([1, 1, 1, 2, 2, 2]), array([0, 1, 2, 0, 1, 2]))
nonzero() method of the boolean array can also be called.
>>> (a > 3).nonzero() (array([1, 1, 1, 2, 2, 2]), array([0, 1, 2, 0, 1, 2]))
numpy.bincount(). The resulting array is
>>> arr = numpy.array([0, 5, 4, 0, 4, 4, 3, 0, 0, 5, 2, 1, 1, 9]) >>> numpy.bincount(arr)
The argument to
bincount() must consist of positive integers or booleans.
Negative integers are not supported.
nan, short for “not a number”, is a special floating point value
defined by the IEEE-754 specification along with
and other values and behaviours. In theory, IEEE
specifically designed to address the problem of missing values, but the
reality is that different platforms behave differently, making life more
difficult. On some platforms, the presence of
nan slows calculations
10-100 times. For integer data, no
nan value exists.
Despite all these issues NumPy (and SciPy) endeavor to support IEEE-754 behaviour (based on NumPy’s predecessor numarray). The most significant challenge is a lack of cross-platform support within Python itself. Because NumPy is written to take advantage of C99, which supports IEEE-754, it can side-step such issues internally, but users may still face problems when, for example, comparing values within Python interpreter.
Those wishing to avoid potential headaches will be interested in an
alternative solution which has a long history in NumPy’s predecessors
– masked arrays. Masked arrays are standard arrays with a second
“mask” array of the same shape to indicate whether the value is present
or missing. Masked arrays are the domain of the
and continue the cross-platform Numeric/numarray tradition. See
“Cookbook/Matplotlib/Plotting values with masked arrays” (TODO) for
example, to avoid plotting missing data in Matplotlib. Despite their
additional memory requirement, masked arrays are faster than nans on
many floating point units. See also the NumPy documentation on masked
Another good option is to use Pandas - it uses
nan in a similar way to
NumPy for floating point data, and since pandas 0.25.0 also supports missing
>>> A = numpy.zeros(3) >>> A[[0,1,1,2]] += 1 >>> A array([ 1., 1., 1.])
One might, quite reasonably, have expected A to contain [1,2,1]. Unfortunately this is not what is implemented in NumPy. Moreover, the Python Reference Manual specifies that
>>> x = x + y
>>> x += y
should result in
x having the same value (though not necessarily the same
identity). Moreover, even if the NumPy developers wanted to modify this behaviour,
Python does not provide an overloadable
the code acts like
>>> tmp = A.__getitem__([0,1,1,2]) >>> tmp.__iadd__(1) >>> A.__setitem__([0,1,1,2],tmp)
This leads to other peculiarities sometimes; if the indexing operation is
actually able to provide a view rather than a copy, the
writes to the array, then the view is copied into the array, so that the
array is written to twice.
However, do not despair! NumPy does contain functionality for this type of behaviour, and it can be
obtained by using the ufunc
at(), which is an attribute of the addition (
np.subtract), multiplication (
np.multiply), and division (
np.divide) ufuncs between a matrix
and a scalar:
>>> A = numpy.zeros(3) >>> numpy.add.at(A, [0, 1, 1, 2], 1) >>> A array([ 1., 2., 1.])