Python Scientific

Contents
Python Scientific lecture notes

Release 2013.2 beta (euroscipy 2013)
I One document to learn numerics, science, and data with Python 1
1 Getting started with Python for science 3
1.1 Scientific computing with tools and workflow . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 The Python language . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.3 NumPy: creating and manipulating numerical data . . . . . . . . . . . . . . . . . . . . . . . . . 42
1.4 Matplotlib: plotting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
1.5 Scipy : high-level scientific computing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
1.6 Getting help and finding documentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
EuroScipy tutorial team
2 Advanced topics 139
Editors: Valentin Haenel, Emmanuelle Gouillart, Gal Varoquaux 2.1 Advanced Python Constructs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
2.2 Advanced Numpy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
2.3 Debugging code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
http://scipy-lectures.github.com 2.4 Optimizing code . . . . . . . . . . . . . . . . . . . . . . .
2.5 Sparse Matrices in SciPy . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
197
203
2.6 Image manipulation and processing using Numpy and Scipy . . . . . . . . . . . . . . . . . . . . 222
2.7 Mathematical optimization: finding minima of functions . . . . . . . . . . . . . . . . . . . . . . 239
2.8 Interfacing with C . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252
3 Packages and applications 271

3.1 Statistics in Python . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271
3.2 Sympy : Symbolic Mathematics in Python . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 318
3.3 Scikit-image: image processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 324
3.4 Traits: building interactive dialogs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 338
3.5 3D plotting with Mayavi . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 357
3.6 scikit-learn: machine learning in Python . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 369
Index 379
September 26, 2015 (2013.2-beta-328-g3cd80a5)
i
Python Scientific lecture notes, Release 2013.2 beta (euroscipy 2013)
Part I
One document to learn numerics, science,

and data with Python
1 2
CHAPTER 1 Existing solutions
Which solutions do scientists use to work?

Compiled languages: C, C++, Fortran, etc.
Advantages:
Very fast. Very optimized compilers. For heavy computations, its difficult to outperform these lan-
Getting started with Python for science guages.
Some very optimized scientific libraries have been written for these languages. Example: BLAS
(vector/matrix operations)
Drawbacks:
Painful usage: no interactivity during development, mandatory compilation steps, verbose syntax (&,
::, }}, ; etc.), manual memory management (tricky in C). These are difficult languages for non com-
This part of the Scipy lecture notes is a self-contained introduction to everything that is needed to use Python for
puter scientists.
science, from the language itself, to numerical computing or plotting.
Scripting languages: Matlab
Advantages:
Very rich collection of libraries with numerous algorithms, for many different domains. Fast execution
because these libraries are often written in a compiled language.
1.1 Scientific computing with tools and workflow Pleasant development environment: comprehensive and well organized help, integrated editor, etc.
Commercial support is available.
Authors: Fernando Perez, Emmanuelle Gouillart, Gal Varoquaux, Valentin Haenel
Drawbacks:
Base language is quite poor and can become restrictive for advanced users.
1.1.1 Why Python?
Not free.
The scientists needs Other scripting languages: Scilab, Octave, Igor, R, IDL, etc.
Advantages:
Get data (simulation, experiment control),
Open-source, free, or at least cheaper than Matlab.
Manipulate and process data,
Some features can be very advanced (statistics in R, figures in Igor, etc.)
Visualize results (to understand what we are doing!),
Drawbacks:
Communicate results: produce figures for reports or publications, write presentations.
Fewer available algorithms than in Matlab, and the language is not more advanced.
Specifications Some software are dedicated to one domain. Ex: Gnuplot or xmgrace to draw curves. These programs
are very powerful, but they are restricted to a single type of usage, such as plotting.
Rich collection of already existing bricks corresponding to classical numerical methods or basic actions: we What about Python?
dont want to re-program the plotting of a curve, a Fourier transform or a fitting algorithm. Dont reinvent
the wheel! Advantages:
Easy to learn: computer science is neither our job nor our education. We want to be able to draw a curve, Very rich scientific computing libraries (a bit less than Matlab, though)
smooth a signal, do a Fourier transform in a few minutes. Well thought out language, allowing to write very readable and well structured code: we code what
Easy communication with collaborators, students, customers, to make the code live within a lab or a com- we think.
pany: the code should be as readable as a book. Thus, the language should contain as few syntax symbols or Many libraries for other tasks than scientific computing (web server management, serial port access,
unneeded routines as possible that would divert the reader from the mathematical or scientific understanding etc.)
of the code.
Free and open-source software, widely spread, with a vibrant community.
Efficient code that executes quickly... but needless to say that a very fast code becomes useless if we spend
too much time writing it. So, we need both a quick development time and a quick execution time. Drawbacks:
A single environment/language for everything, if possible, to avoid learning a new software for each new less pleasant development environment than, for example, Matlab. (More geek-oriented).
problem. Not all the algorithms that can be found in more specialized software or toolboxes.
3 1.1. Scientific computing with tools and workflow 4

Python Scientific lecture notes, Release 2013.2 beta (euroscipy 2013) Python Scientific lecture notes, Release 2013.2 beta (euroscipy 2013)
1.1.2 Scientific Python building blocks 1.1.3 The interactive workflow: IPython and a text editor
Unlike Matlab, Scilab or R, Python does not come with a pre-bundled set of modules for scientific computing. Interactive work to test and understand algorithms: In this section, we describe an interactive workflow with
Below are the basic building blocks that can be combined to obtain a scientific computing environment: IPython that is handy to explore and understand algorithms.
Python, a generic and modern computing language Python is a general-purpose language. As such, there is not one blessed environment to work in, and not only one
way of using it. Although this makes it harder for beginners to find their way, it makes it possible for Python to be
Python language: data types (string, int), flow control, data collections (lists, dictionaries), pat-
used to write programs, in web servers, or embedded devices.
terns, etc.
Modules of the standard library.
Reference document for this section:
A large number of specialized modules or applications written in Python: web protocols, web frame-
work, etc. ... and scientific computing. IPython user manual: http://ipython.org/ipython-doc/dev/index.html
Development tools (automatic testing, documentation generation)
Command line interaction
Start ipython:
In [1]: print('Hello world')
Hello world
Getting help by using the ? operator after an object:

IPython, an advanced Python shell http://ipython.org/ In [2]: print?
Numpy : provides powerful numerical arrays objects, and routines to manipulate them. Type: builtin_function_or_method
http://www.numpy.org/ Base Class: <type 'builtin_function_or_method'>
String Form: <built-in function print>
Scipy : high-level data processing routines. Optimization, regression, interpolation, etc Namespace: Python builtin
http://www.scipy.org/ Docstring:
print(value, ..., sep=' ', end='\n', file=sys.stdout)
Prints the values to a stream, or to sys.stdout by default.

Optional keyword arguments:
file: a file-like object (stream); defaults to the current sys.stdout.
sep: string inserted between values, default a space.
end: string appended after the last value, default a newline.
Elaboration of the algorithm in an editor
Matplotlib : 2-D visualization, publication-ready plots http://matplotlib.org/ Create a file my_file.py in a text editor. Under EPD (Enthought Python Distribution), you can use Scite,
available from the start menu. Under Python(x,y), you can use Spyder. Under Ubuntu, if you dont already have
your favorite editor, we would advise installing Stanis Python editor. In the file, add the following
lines:
s = 'Hello world'
print(s)
Now, you can run it in IPython and explore the resulting variables:
In [1]: %run my_file.py
Hello world
In [2]: s
Out[2]: 'Hello world'
In [3]: %whos
Variable Type Data/Info
----------------------------
s str Hello world
Mayavi : 3-D visualization http://code.enthought.com/projects/mayavi/
1.1. Scientific computing with tools and workflow 5 1.1. Scientific computing with tools and workflow 6
In [8]: debug
From a script to functions > /.../IPython/core/compilerop.py (87)ast_parse()
86 and are passed to the built-in compile function."""
While it is tempting to work only with scripts, that is a file full of instructions following each other, do plan ---> 87 return compile(source, filename, symbol, self.flags | PyCF_ONLY_AST, 1)
to progressively evolve the script to a set of functions: 88
A script is not reusable, functions are.
Thinking in terms of functions helps breaking the problem in small blocks. ipdb>locals()
{'source': u'x === 10\n', 'symbol': 'exec', 'self':
<IPython.core.compilerop.CachingCompiler instance at 0x2ad8ef0>,
'filename': '<ipython-input-6-12fd421b5f28>'}
IPython Tips and Tricks
The IPython user manual contains a wealth of information about using IPython, but to get you started we want to IPython help
give you a quick introduction to four useful features: history, magic functions, aliases and tab completion.
The built-in IPython cheat-sheet is accessible via the %quickref magic function.
Like a UNIX shell, IPython supports command history. Type up and down to navigate previously typed commands: A list of all available magic functions is shown when typing %magic.
In [1]: x = 10
Furthermore IPython ships with various aliases which emulate common UNIX command line tools such as ls to
In [2]: <UP> list files, cp to copy files and rm to remove files. A list of aliases is shown when typing alias:
In [2]: x = 10 In [1]: alias
Total number of aliases: 16
IPython supports so called magic functions by prefixing a command with the % character. For example, the run Out[1]:
and whos functions from the previous section are magic functions. Note that, the setting automagic, which is [('cat', 'cat'),
enabled by default, allows you to omit the preceding % sign. Thus, you can just type the magic function and it will ('clear', 'clear'),
work. ('cp', 'cp -i'),
('ldir', 'ls -F -o --color %l | grep /$'),
Other useful magic functions are: ('less', 'less'),
('lf', 'ls -F -o --color %l | grep ^-'),
%cd to change the current directory.
('lk', 'ls -F -o --color %l | grep ^l'),
In [2]: cd /tmp ('ll', 'ls -F -o --color'),
/tmp ('ls', 'ls -F --color'),
('lx', 'ls -F -o --color %l | grep ^-..x'),
%timeit allows you to time the execution of short snippets using the timeit module from the standard ('man', 'man'),
library: ('mkdir', 'mkdir'),
('more', 'more'),
In [3]: timeit x = 10 ('mv', 'mv -i'),
10000000 loops, best of 3: 39 ns per loop ('rm', 'rm -i'),
('rmdir', 'rmdir')]
%cpaste allows you to paste code, especially code from websites which has been prefixed with the stan-
dard Python prompt (e.g. >>>) or with an ipython prompt, (e.g. in [3]): Lastly, we would like to mention the tab completion feature, whose description we cite directly from the IPython
manual:
In [5]: cpaste
Pasting code; enter '--' alone on the line to stop or use Ctrl-D. Tab completion, especially for attributes, is a convenient way to explore the structure of any object youre dealing
:In [3]: timeit x = 10 with. Simply type object_name.<TAB> to view the objects attributes. Besides Python objects and keywords, tab
:-- completion also works on file and directory names.
10000000 loops, best of 3: 85.9 ns per loop
In [6]: cpaste In [1]: x = 10
Pasting code; enter '--' alone on the line to stop or use Ctrl-D.
:>>> timeit x = 10 In [2]: x.<TAB>
:-- x.bit_length x.conjugate x.denominator x.imag x.numerator
10000000 loops, best of 3: 86 ns per loop x.real
%debug allows you to enter post-mortem debugging. That is to say, if the code you try to execute, raises In [3]: x.real.
an exception, using %debug will enter the debugger at the point where the exception was thrown. x.real.bit_length x.real.denominator x.real.numerator
x.real.conjugate x.real.imag x.real.real
In [7]: x === 10
File "<ipython-input-6-12fd421b5f28>", line 1 In [4]: x.real.
x === 10
^
SyntaxError: invalid syntax
1.1. Scientific computing with tools and workflow 7 1.1. Scientific computing with tools and workflow 8
1.2 The Python language >>> print("Hello, world!")

Hello, world!
Authors: Chris Burns, Christophe Combelles, Emmanuelle Gouillart, Gal Varoquaux
Tip: The message Hello, world! is then displayed. You just executed your first Python instruction, congratula-
tions!
Python for scientific computing
We introduce here the Python language. Only the bare minimum necessary for getting started with Numpy To get yourself started, type the following stack of instructions
and Scipy is addressed here. To learn more about the language, consider going through the excellent tutorial
>>> a = 3
http://docs.python.org/tutorial. Dedicated books are also available, such as http://diveintopython.net/. >>> b = 2*a
>>> type(b)
<type 'int'>
>>> print(b)
6
>>> a*b
18
>>> b = 'hello'
>>> type(b)
<type 'str'>
Tip: Python is a programming language, as are C, Fortran, BASIC, PHP, etc. Some specific features of Python >>> b + b
are as follows: 'hellohello'
>>> 2*b
an interpreted (as opposed to compiled) language. Contrary to e.g. C or Fortran, one does not compile
'hellohello'
Python code before executing it. In addition, Python can be used interactively: many Python interpreters
are available, from which commands and scripts can be executed.
Tip: Two variables a and b have been defined above. Note that one does not declare the type of an variable
a free software released under an open-source license: Python can be used and distributed free of charge, before assigning its value. In C, conversely, one should write:
even for building commercial software.
int a = 3;
multi-platform: Python is available for all major operating systems, Windows, Linux/Unix, MacOS X,
most likely your mobile phone OS, etc. In addition, the type of a variable may change, in the sense that at one point in time it can be equal to a value of
a certain type, and a second point in time, it can be equal to a value of a different type. b was first equal to an
a very readable language with clear non-verbose syntax
integer, but it became equal to a string when it was assigned the value hello. Operations on integers (b=2*a)
a language for which a large variety of high-quality packages are available for various applications, from are coded natively in Python, and so are some operations on strings such as additions and multiplications, which
web frameworks to scientific computing. amount respectively to concatenation and repetition.
a language very easy to interface with other languages, in particular C and C++.
Some other features of the language are illustrated just below. For example, Python is an object-oriented
language, with dynamic typing (the same variable can contain objects of different types during the course 1.2.2 Basic types
of a program).
Numerical types
See http://www.python.org/about/ for more information about distinguishing features of Python.
Tip: Python supports the following numerical, scalar types:
Integer
1.2.1 First steps >>> 1 + 1
2
Start the Ipython shell (an enhanced interactive Python shell): >>> a = 4
>>> type(a)
by typing ipython from a Linux/Mac terminal, or from the Windows cmd shell, <type 'int'>
or by starting the program from a menu, e.g. in the Python(x,y) or EPD menu if you have installed one of
these scientific-Python suites. Floats
>>> c = 2.1
Tip: If you dont have Ipython installed on your computer, other Python shells are available, such as the plain >>> type(c)
Python shell started by typing python in a terminal, or the Idle interpreter. However, we advise to use the Ipython <type 'float'>
shell because of its enhanced features, especially for interactive scientific computing.
Complex
Once you have started the interpreter, type >>> a = 1.5 + 0.5j
>>> a.real
1.2. The Python language 9 1.2. The Python language 10

1.5 Warning: Integer division

>>> a.imag In Python 2:
0.5
>>> type(1. + 0j) >>> 3 / 2
<type 'complex'> 1
Booleans In Python 3:
>>> 3 > 4 >>> 3 / 2

False 1.5
>>> test = (3 > 4)
>>> test To be safe: use floats:
False
>>> type(test) >>> 3 / 2.
<type 'bool'> 1.5
>>> a = 3
Tip: A Python shell can therefore replace your pocket calculator, with the basic arithmetic operations +, -, *, /, >>> b = 2
% (modulo) natively implemented >>> a / b # In Python 2
1
>>> 7 * 3. >>> a / float(b)
21.0 1.5
>>> 2**10
Future behavior: to always get the behavior of Python3
1024
>>> 8 % 3
>>> from __future__ import division
2
>>> 3 / 2
1.5
Type conversion (casting):
>>> float(1)
1.0 Tip: If you explicitly want integer division use //:
>>> 3.0 // 2
1.0
Note: The behaviour of the division operator has changed in Python 3.
Containers
Tip: Python provides many efficient types of containers, in which collections of objects can be stored.
Lists
Tip: A list is an ordered collection of objects, that may have different types. For example:
>>> l = ['red', 'blue', 'green', 'black', 'white']

>>> type(l)
<type 'list'>
Indexing: accessing individual objects contained in the list:

>>> l[2]
'green'
Counting from the end with negative indices:

>>> l[-1] >>> L = ['red', 'blue', 'green', 'black', 'white']

'white' >>> L.append('pink')
>>> l[-2] >>> L
'black' ['red', 'blue', 'green', 'black', 'white', 'pink']
>>> L.pop() # removes and returns the last item
Warning: Indexing starts at 0 (as in C), not at 1 (as in Fortran or Matlab)! 'pink'
>>> L
['red', 'blue', 'green', 'black', 'white']
Slicing: obtaining sublists of regularly-spaced elements: >>> L.extend(['pink', 'purple']) # extend L, in-place
>>> L
>>> l ['red', 'blue', 'green', 'black', 'white', 'pink', 'purple']
['red', 'blue', 'green', 'black', 'white'] >>> L = L[:-2]
>>> l[2:4] >>> L
['green', 'black'] ['red', 'blue', 'green', 'black', 'white']
Warning: Note that l[start:stop] contains the elements with indices i such as start<= i < stop Reverse:
(i ranging from start to stop-1). Therefore, l[start:stop] has (stop - start) elements.
>>> r = L[::-1]
>>> r
Slicing syntax: l[start:stop:stride] ['white', 'black', 'green', 'blue', 'red']
>>> r2 = list(L)
Tip: All slicing parameters are optional: >>> r2
>>> l
>>> r2.reverse() # in-place
>>> r2
>>> l[3:]
['white', 'black', 'green', 'blue', 'red']
['black', 'white']
>>> l[:3]
['red', 'blue', 'green']
Concatenate and repeat lists:
>>> l[::2] >>> r + L
['red', 'green', 'white'] ['white', 'black', 'green', 'blue', 'red', 'red', 'blue', 'green', 'black', 'white']
>>> r * 2
['white', 'black', 'green', 'blue', 'red', 'white', 'black', 'green', 'blue', 'red']
Lists are mutable objects and can be modified:
>>> l[0] = 'yellow' Tip: Sort:
>>> l
['yellow', 'blue', 'green', 'black', 'white'] >>> sorted(r) # new object
>>> l[2:4] = ['gray', 'purple'] ['black', 'blue', 'green', 'red', 'white']
>>> l >>> r
['yellow', 'blue', 'gray', 'purple', 'white'] ['white', 'black', 'green', 'blue', 'red']
>>> r.sort() # in-place
>>> r
Note: The elements of a list may have different types: ['black', 'blue', 'green', 'red', 'white']
>>> l = [3, -200, 'hello']
>>> l
[3, -200, 'hello']
>>> l[1], l[2] Methods and Object-Oriented Programming
(-200, 'hello') The notation r.method() (e.g. r.append(3) and L.pop()) is our first example of object-oriented
programming (OOP). Being a list, the object r owns the method function that is called using the
Tip: For collections of numerical data that all have the same type, it is often more efficient to use the array notation .. No further knowledge of OOP than understanding the notation . is necessary for going through
type provided by the numpy module. A NumPy array is a chunk of memory containing fixed-sized items. With this tutorial.
NumPy arrays, operations on elements can be faster because elements are regularly spaced in memory and more
operations are performed through specialized C functions instead of Python loops.
Tip: Python offers a large panel of functions to modify lists, or query them. Here are a few examples; for more
details, see http://docs.python.org/tutorial/datastructures.html#more-on-lists
Add and remove elements:

Discovering methods: >>> a[2:10:2] # Syntax: a[start:stop:step]

'lo o'
Reminder: in Ipython: tab-completion (press tab) >>> a[::3] # every three characters, from beginning to end
In [28]: r.<TAB> 'hl r!'
r.__add__ r.__iadd__ r.__setattr__
r.__class__ r.__imul__ r.__setitem__ Tip: Accents and special characters can also be handled in Unicode strings (see
r.__contains__ r.__init__ r.__setslice__ http://docs.python.org/tutorial/introduction.html#unicode-strings).
r.__delattr__ r.__iter__ r.__sizeof__
r.__delitem__ r.__le__ r.__str__
r.__delslice__ r.__len__ r.__subclasshook__ A string is an immutable object and it is not possible to modify its contents. One may however create new strings
r.__doc__ r.__lt__ r.append from the original one.
r.__eq__ r.__mul__ r.count
In [53]: a = "hello, world!"
r.__format__ r.__ne__ r.extend
In [54]: a[2] = 'z'
r.__ge__ r.__new__ r.index
---------------------------------------------------------------------------
r.__getattribute__ r.__reduce__ r.insert
Traceback (most recent call last):
r.__getitem__ r.__reduce_ex__ r.pop
File "<stdin>", line 1, in <module>
r.__getslice__ r.__repr__ r.remove
TypeError: 'str' object does not support item assignment
r.__gt__ r.__reversed__ r.reverse
r.__hash__ r.__rmul__ r.sort
In [55]: a.replace('l', 'z', 1)
Out[55]: 'hezlo, world!'
In [56]: a.replace('l', 'z')
Out[56]: 'hezzo, worzd!'
Strings
Tip: Strings have many useful methods, such as a.replace as seen above. Remember the a. object-oriented
Different string syntaxes (simple, double or triple quotes): notation and use tab completion or help(str) to search for new methods.
s = 'Hello, how are you?'
s = "Hi, what's up" See also:
s = '''Hello, # tripling the quotes allows the
how are you''' # the string to span more than one line Python offers advanced possibilities for manipulating strings, looking for patterns or formatting.
s = """Hi, The interested reader is referred to http://docs.python.org/library/stdtypes.html#string-methods and
what's up?""" http://docs.python.org/library/string.html#new-string-formatting
String formatting:
In [1]: 'Hi, what's up?'
------------------------------------------------------------ >>> 'An integer: %i; a float: %f; another string: %s' % (1, 0.1, 'string')
File "<ipython console>", line 1 'An integer: 1; a float: 0.100000; another string: string'
'Hi, what's up?'
^ >>> i = 102
SyntaxError: invalid syntax >>> filename = 'processing_of_dataset_%d.txt' % i
>>> filename
The newline character is \n, and the tab character is \t. 'processing_of_dataset_102.txt'
Tip: Strings are collections like lists. Hence they can be indexed and sliced, using the same syntax and rules.
Dictionaries
Indexing:
>>> a = "hello" Tip: A dictionary is basically an efficient table that maps keys to values. It is an unordered container
>>> a[0]
'h'
>>> tel = {'emmanuelle': 5752, 'sebastian': 5578}
>>> a[1]
>>> tel['francis'] = 5915
'e'
>>> tel
>>> a[-1]
{'sebastian': 5578, 'francis': 5915, 'emmanuelle': 5752}
'o'
>>> tel['sebastian']
5578
Tip: (Remember that negative indices correspond to counting from the right end.) >>> tel.keys()
['sebastian', 'francis', 'emmanuelle']
Slicing: >>> tel.values()
[5578, 5915, 5752]
>>> a = "hello, world!" >>> 'francis' in tel
>>> a[3:6] # 3rd to 6th (excluded) elements: elements 3, 4, 5 True
'lo,'

Tip: It can be used to conveniently store and retrieve values associated with a name (a string for a date, a name, In [1]: a = [1, 2, 3]
etc.). See http://docs.python.org/tutorial/datastructures.html#dictionaries for more information. In [3]: a
A dictionary can have keys (resp. values) with different types: Out[3]: [1, 2, 3]
In [4]: a = ['a', 'b', 'c'] # Creates another object.
>>> d = {'a':1, 'b':2, 3:'hello'} In [5]: a
>>> d Out[5]: ['a', 'b', 'c']
{'a': 1, 3: 'hello', 'b': 2} In [6]: id(a)
Out[6]: 138641676
In [7]: a[:] = [1, 2, 3] # Modifies object in place.
In [8]: a
More container types Out[8]: [1, 2, 3]
In [9]: id(a)
Out[9]: 138641676 # Same as in Out[6], yours will differ...
Tuples
Tuples are basically immutable lists. The elements of a tuple are written between parentheses, or just separated the key concept here is mutable vs. immutable
by commas: mutable objects can be changed in place
>>> t = 12345, 54321, 'hello!'
immutable objects cannot be modified once created
>>> t[0]
12345 See also:
>>> t
(12345, 54321, 'hello!') A very good and detailed explanation of the above issues can be found in David M. Beazleys article Types and
>>> u = (0, 2) Objects in Python.
Sets: unordered, unique items:

1.2.3 Control Flow
>>> s = set(('a', 'b', 'c', 'a'))
>>> s
Controls the order in which the code is executed.
set(['a', 'c', 'b'])
>>> s.difference(('a', 'b'))
set(['c']) if/elif/else
>>> if 2**2 == 4:
Assignment operator ... print('Obvious!')
...
Obvious!
Tip: Python library reference says:
Assignment statements are used to (re)bind names to values and to modify attributes or items of Blocks are delimited by indentation
mutable objects.
Tip: Type the following lines in your Python interpreter, and be careful to respect the indentation depth. The
In short, it works as follows (simple assignment): Ipython shell automatically increases the indentation depth after a column : sign; to decrease the indentation
1. an expression on the right hand side is evaluated, the corresponding object is created/obtained depth, go four spaces to the left with the Backspace key. Press the Enter key twice to leave the logical block.
2. a name on the left hand side is assigned, or bound, to the r.h.s. object
>>> a = 10
Things to note: >>> if a == 1:

... print(1)
a single object can have several names bound to it:
... elif a == 2:
In [1]: a = [1, 2, 3] ... print(2)
In [2]: b = a ... else:
In [3]: a ... print('A lot')
Out[3]: [1, 2, 3] A lot
In [4]: b
Out[4]: [1, 2, 3] Indentation is compulsory in scripts as well. As an exercise, re-type the previous lines with the same indentation
In [5]: a is b in a script condition.py, and execute the script with run condition.py in Ipython.
Out[5]: True
In [6]: b[1] = 'hi!'
In [7]: a for/range
Out[7]: [1, 'hi!', 3]
Iterating with an index:
to change a list in place, use indexing/slices:

>>> for i in range(4): >>> 1 == 1.

... print(i) True
0
1 a is b Tests identity: both sides are the same object:
2
3 >>> 1 is 1.
False
But most often, it is more readable to iterate over values:
>>> a = 1
>>> for word in ('cool', 'powerful', 'readable'): >>> b = 1
... print('Python is %s' % word) >>> a is b
Python is cool True
Python is powerful
Python is readable a in b For any collection b: b contains a
>>> b = [1, 2, 3]
>>> 2 in b
while/break/continue True
>>> 5 in b
Typical C-style while loop (Mandelbrot problem): False
>>> z = 1 + 1j If b is a dictionary, this tests that a is a key of b.
>>> while abs(z) < 100:
... z = z**2 + 1
>>> z Advanced iteration
(-134+352j)
Iterate over any sequence
More advanced features
break out of enclosing for/while loop: You can iterate over any sequence (string, list, keys in a dictionary, lines in a file, ...):
>>> z = 1 + 1j >>> vowels = 'aeiouy'
>>> while abs(z) < 100: >>> for i in 'powerful':

... if z.imag == 0: ... if i in vowels:
... break ... print(i)
... z = z**2 + 1 o
e
continue the next iteration of a loop.: u
>>> a = [1, 0, 2, 4]
>>> for element in a: >>> message = "Hello how are you?"
... if element == 0: >>> message.split() # returns a list
... continue ['Hello', 'how', 'are', 'you?']
... print(1. / element) >>> for word in message.split():
1.0 ... print(word)
0.5 ...
0.25 Hello
how
are
you?
Conditional Expressions
if <OBJECT> Tip: Few languages (in particular, languages for scientific computing) allow to loop over anything but inte-
gers/indices. With Python it is possible to loop exactly over the objects of interest without bothering with indices
Evaluates to False: you often dont care about. This feature can often be used to make code more readable.
any number equal to zero (0, 0.0, 0+0j)
an empty container (list, tuple, set, dictionary, ...) Warning: Not safe to modify the sequence you are iterating over.
False, None
Evaluates to True: Keeping track of enumeration number
everything else
a == b Tests equality, with logics: Common task is to iterate over a sequence while keeping track of the item number.

Could use while loop with a counter as above. Or a for loop: Warning: Function blocks must be indented as other control-flow blocks.
>>> words = ('cool', 'powerful', 'readable')
>>> for i in range(0, len(words)):
... print((i, words[i]))
(0, 'cool') Return statement
(1, 'powerful')
(2, 'readable') Functions can optionally return values.
In [6]: def disk_area(radius):
But, Python provides a built-in function - enumerate - for this:
...: return 3.14 * radius * radius
>>> for index, item in enumerate(words): ...:
... print((index, item))
(0, 'cool') In [8]: disk_area(1.5)
(1, 'powerful') Out[8]: 7.0649999999999995
(2, 'readable')
Note: By default, functions return None.
Looping over a dictionary
Note: Note the syntax to define a function:
Use items: the def keyword;
>>> d = {'a': 1, 'b':1.2, 'c':1j} is followed by the functions name, then
>>> for key, val in sorted(d.items()): the arguments of the function are given between parentheses followed by a colon.
... print('Key: %s has value: %s' % (key, val)) the function body;
Key: a has value: 1
Key: b has value: 1.2 and return object for optionally returning values.
Key: c has value: 1j
Note: The ordering of a dictionary in random, thus we use sorted() which will sort on the keys. Parameters
Mandatory parameters (positional arguments)

List Comprehensions In [81]: def double_it(x):
....: return x * 2
>>> [i**2 for i in range(4)] ....:
[0, 1, 4, 9]
In [82]: double_it(3)
Out[82]: 6
Exercise In [83]: double_it()

---------------------------------------------------------------------------
Compute the decimals of Pi using the Wallis formula: Traceback (most recent call last):

42 TypeError: double_it() takes exactly 1 argument (0 given)
=2 21
=1
4
Optional parameters (keyword or named arguments)
In [84]: def double_it(x=2):
....: return x * 2
1.2.4 Defining functions ....:
Function definition In [85]: double_it()

Out[85]: 4
In [56]: def test(): In [86]: double_it(3)
....: print('in test function') Out[86]: 6
....:
....:
Keyword arguments allow you to specify default values.
In [57]: test()
in test function

Warning: Default values are evaluated when the function is defined, not when it is called. This can be In [107]: slicer(rhyme, step=2, start=1, stop=4)
problematic when using mutable types (e.g. dictionary or list) and modifying them in the function body, since Out[107]: ['fish,', 'fish,']
the modifications will be persistent across invocations of the function.
Using an immutable type in a keyword argument: but it is good practice to use the same ordering as the functions definition.
In [124]: bigx = 10 Keyword arguments are a very convenient feature for defining functions with a variable number of arguments,
especially when default values are to be used in most calls to the function.
In [125]: def double_it(x=bigx):
.....: return x * 2
.....: Passing by value
In [126]: bigx = 1e9 # Now really big
Tip: Can you modify the value of a variable inside a function? Most languages (C, Java, ...) distinguish passing
In [128]: double_it() by value and passing by reference. In Python, such a distinction is somewhat artificial, and it is a bit subtle
Out[128]: 20 whether your variables are going to be modified or not. Fortunately, there exist clear rules.
Using an mutable type in a keyword argument (and modifying it inside the function body): Parameters to functions are references to objects, which are passed by value. When you pass a variable to a
function, python passes the reference to the object to which the variable refers (the value). Not the variable itself.
In [2]: def add_to_dict(args={'a': 1, 'b': 2}):
...: for i in args.keys():
If the value passed in a function is immutable, the function does not modify the callers variable. If the value is
...: args[i] += 1
mutable, the function may modify the callers variable in-place:
...: print args
...: >>> def try_to_modify(x, y, z):
... x = 23
In [3]: add_to_dict ... y.append(42)
Out[3]: <function __main__.add_to_dict> ... z = [99] # new reference
... print(x)
In [4]: add_to_dict() ... print(y)
{'a': 2, 'b': 3} ... print(z)
...
In [5]: add_to_dict() >>> a = 77 # immutable variable
{'a': 3, 'b': 4} >>> b = [99] # mutable variable
>>> c = [28]
In [6]: add_to_dict() >>> try_to_modify(a, b, c)
{'a': 4, 'b': 5} 23
[99, 42]
[99]
>>> print(a)
Tip: More involved example implementing pythons slicing:
77
In [98]: def slicer(seq, start=None, stop=None, step=None): >>> print(b)
....: """Implement basic python slicing.""" [99, 42]
....: return seq[start:stop:step] >>> print(c)
....: [28]
In [101]: rhyme = 'one fish, two fish, red fish, blue fish'.split() Functions have a local variable table called a local namespace.
In [102]: rhyme The variable x only exists within the function try_to_modify.
Out[102]: ['one', 'fish,', 'two', 'fish,', 'red', 'fish,', 'blue', 'fish']
In [103]: slicer(rhyme) Global variables

Out[103]: ['one', 'fish,', 'two', 'fish,', 'red', 'fish,', 'blue', 'fish']
Variables declared outside the function can be referenced within the function:
In [104]: slicer(rhyme, step=2)
In [114]: x = 5
Out[104]: ['one', 'two', 'red', 'blue']
In [115]: def addx(y):
In [105]: slicer(rhyme, 1, step=2)
.....: return x + y
Out[105]: ['fish,', 'fish,', 'fish,', 'fish']
.....:
In [106]: slicer(rhyme, start=1, stop=4, step=2)
In [116]: addx(10)
Out[106]: ['fish,', 'fish,']
Out[116]: 15
The order of the keyword arguments does not matter:
But these global variables cannot be modified within the function, unless declared global in the function.

This doesnt work: String Form: <function funcname at 0xeaa0f0>

In [117]: def setx(y): Namespace: Interactive
.....: x = y File: <ipython console>
.....: print('x is %d' % x) Definition: funcname(params)
.....: Docstring:
.....: Concise one-line sentence describing the function.
In [118]: setx(10) Extended summary which can contain multiple paragraphs.

x is 10
Note: Docstring guidelines
In [120]: x
Out[120]: 5 For the sake of standardization, the Docstring Conventions webpage documents the semantics and conventions
associated with Python docstrings.
This works: Also, the Numpy and Scipy modules have defined a precise standard for documenting scientific func-
In [121]: def setx(y): tions, that you may want to follow for your own functions, with a Parameters section, an
.....: global x Examples section, etc. See http://projects.scipy.org/numpy/wiki/CodingStyleGuidelines#docstring-standard
.....: x = y and http://projects.scipy.org/numpy/browser/trunk/doc/example.py#L37
.....: print('x is %d' % x)
.....:
.....:
Functions are objects
In [122]: setx(10)
x is 10 Functions are first-class objects, which means they can be:
In [123]: x assigned to a variable

Out[123]: 10 an item in a list (or any collection)
passed as an argument to another function.
Variable number of parameters In [38]: va = variable_args
Special forms of parameters: In [39]: va('three', x=1, y=2)

args is ('three',)
*args: any number of positional arguments packed into a tuple kwargs is {'y': 2, 'x': 1}
**kwargs: any number of keyword arguments packed into a dictionary
In [35]: def variable_args(*args, **kwargs):
Methods
....: print 'args is', args
....: print 'kwargs is', kwargs
....: Methods are functions attached to objects. Youve seen these in our examples on lists, dictionaries, strings, etc...
In [36]: variable_args('one', 'two', x=1, y=2, z=3)

args is ('one', 'two') Exercises
kwargs is {'y': 2, 'x': 1, 'z': 3}
Exercise: Fibonacci sequence

Docstrings Write a function that displays the n first terms of the Fibonacci sequence, defined by:
u_0 = 1; u_1 = 1
Documentation about what the function does and its parameters. General convention: u_(n+2) = u_(n+1) + u_n
In [67]: def funcname(params):
....: """Concise one-line sentence describing the function.
....:
Exercise: Quicksort
....: Extended summary which can contain multiple paragraphs.
....: """ Implement the quicksort algorithm, as defined by wikipedia:
....: # function body function quicksort(array) var list less, greater if length(array) < 2
....: pass return array
....: select and remove a pivot value pivot from array for each x in array
if x < pivot + 1 then append x to less else append x to greater
In [68]: funcname?
return concatenate(quicksort(less), pivot, quicksort(greater))
Type: function
Base Class: type 'function'>

1.2.5 Reusing code: scripts and modules Warning: Dont implement option parsing yourself. Use modules such as optparse, argparse or
docopt.
For now, we have typed all instructions in the interpreter. For longer sets of instructions we need to change track
and write the code in text files (using a text editor), that we will call either scripts or modules. Use your favorite
text editor (provided it offers syntax highlighting for Python), or the editor that comes with the Scientific Python
Suite you may be using (e.g., Scite with Python(x,y)).
Importing objects from modules
Scripts In [1]: import os
In [2]: os
Tip: Let us first write a script, that is a file with a sequence of instructions that are executed each time the script Out[2]: <module 'os' from '/usr/lib/python2.6/os.pyc'>
is called. Instructions may be e.g. copied-and-pasted from the interpreter (but take care to respect indentation
rules!). In [3]: os.listdir('.')
Out[3]:
The extension for Python files is .py. Write or copy-and-paste the following lines in a file called test.py ['conf.py',
'basic_types.rst',
message = "Hello how are you?" 'control_flow.rst',
for word in message.split(): 'functions.rst',
print word 'python_language.rst',
'reusing.rst',
'file_io.rst',
Tip: Let us now execute the script interactively, that is inside the Ipython interpreter. This is maybe the most 'exceptions.rst',
common use of scripts in scientific computing.
'workflow.rst',
'index.rst']
Note: in Ipython, the syntax to execute a script is %run script.py. For example,
And also:
In [1]: %run test.py In [4]: from os import listdir

Hello
how Importing shorthands:
are
In [5]: import numpy as np
you?
In [2]: message Warning:

Out[2]: 'Hello how are you?'
from os import *
The script has been executed. Moreover the variables defined in the script (such as message) are now available
This is called the star import and please, Use it with caution
inside the interpreters namespace.
Makes the code harder to read and understand: where do symbols come from?
Makes it impossible to guess the functionality by the context and the name (hint: os.name is the name
Tip: Other interpreters also offer the possibility to execute scripts (e.g., execfile in the plain Python inter-
preter, etc.). of the OS), and to profit usefully from tab completion.
Restricts the variable names you can use: os.name might override name, or vise-versa.
Creates possible name clashes between modules.
It is also possible In order to execute this script as a standalone program, by executing the script inside a shell
Makes the code impossible to statically check for undefined symbols.
terminal (Linux/Mac console or cmd Windows console). For example, if we are in the same directory as the test.py
file, we can execute this in a console:
$ python test.py Tip: Modules are thus a good way to organize code in a hierarchical way. Actually, all the scientific computing
Hello tools we are going to use are modules:
how >>> import numpy as np # data arrays
are >>> np.linspace(0, 10, 6)
you? array([ 0., 2., 4., 6., 8., 10.])
>>> import scipy # scientific computing
Tip: Standalone scripts may also take command-line arguments
In file.py: In Python(x,y), Ipython(x,y) executes the following imports at startup:
import sys >>> import numpy
print sys.argv >>> import numpy as np
>>> from pylab import *
$ python file.py test arguments >>> import scipy
['file.py', 'test', 'arguments']
and it is not necessary to re-import these modules.

Creating modules '__file__',

'__name__',
'__package__',
Tip: If we want to write larger and better organized programs (compared to simple scripts), where some objects
are defined, (variables, functions, classes) and that we want to reuse several times, we have to create our own 'c',
'd',
modules.
'print_a',
'print_b']
Let us create a module demo contained in the file demo.py:
"A demo module."
In [8]: demo.
demo.__builtins__ demo.__init__ demo.__str__
def print_b():
demo.__class__ demo.__name__ demo.__subclasshook__
"Prints b."
demo.__delattr__ demo.__new__ demo.c
print 'b'
demo.__dict__ demo.__package__ demo.d
demo.__doc__ demo.__reduce__ demo.print_a
def print_a():
demo.__file__ demo.__reduce_ex__ demo.print_b
"Prints a."
demo.__format__ demo.__repr__ demo.py
print 'a'
demo.__getattribute__ demo.__setattr__ demo.pyc
demo.__hash__ demo.__sizeof__
c = 2
Importing objects from modules into the main namespace
d = 2
In [9]: from demo import print_a, print_b
Tip: In this file, we defined two functions print_a and print_b. Suppose we want to call the print_a
function from the interpreter. We could execute the file as a script, but since we just want to have access to the In [10]: whos
function print_a, we are rather going to import it as a module. The syntax is as follows. Variable Type Data/Info
--------------------------------
demo module <module 'demo' from 'demo.py'>
In [1]: import demo print_a function <function print_a at 0xb7421534>
print_b function <function print_b at 0xb74214c4>
In [2]: demo.print_a() In [11]: print_a()

a a
In [3]: demo.print_b() Warning: Module caching

b
Modules are cached: if you modify demo.py and re-import it in the old session, you will get the
Importing the module gives access to its objects, using the module.object syntax. Dont forget to put the old one.
modules name before the objects name, otherwise Python wont recognize the instruction. Solution:
Introspection In [10]: reload(demo)
In [4]: demo?
Type: module
Base Class: <type 'module'>
__main__ and module loading
String Form: <module 'demo' from 'demo.py'>
Namespace: Interactive
File: /home/varoquau/Projects/Python_talks/scipy_2009_tutorial/source/demo.py Tip: Sometimes we want code to be executed when a module is run directly, but not when it is imported by another
Docstring: module. if __name__ == __main__ allows us to check whether the module is being run directly.
A demo module.
File demo2.py:
In [5]: who def print_b():
demo "Prints b."
print 'b'
In [6]: whos
Variable Type Data/Info def print_a():
------------------------------ "Prints a."
demo module <module 'demo' from 'demo.py'> print 'a'
In [7]: dir(demo) # print_b() runs on import

Out[7]: print_b()
['__builtins__',
'__doc__',

if __name__ == '__main__': or modify the sys.path variable itself within a Python script.
# print_a() is only executed when the module is run directly.
print_a() Tip:
import sys
Importing it: new_path = '/home/emma/user_defined_modules'
In [11]: import demo2 if new_path not in sys.path:
b sys.path.append(new_path)
In [12]: import demo2 This method is not very robust, however, because it makes the code less portable (user-dependent path) and
because you have to add the directory to your sys.path each time you want to import from a module in this
Running it: directory.
In [13]: %run demo2
b See also:
a See http://docs.python.org/tutorial/modules.html for more information about modules.
Scripts or modules? How to organize your code Packages
A directory that contains many modules is called a package. A package is a module with submodules (which can
Note: Rule of thumb
have submodules themselves, etc.). A special file called __init__.py (which may be empty) tells Python that
Sets of instructions that are called several times should be written inside functions for better code reusability. the directory is a Python package, from which modules can be imported.
Functions (or other bits of code) that are called from several scripts should be written inside a module, $ ls
so that only the module is imported in the different scripts (do not copy-and-paste your functions in the cluster/ io/ README.txt@ stsci/
different scripts!). __config__.py@ LATEST.txt@ setup.py@ __svn_version__.py@
__config__.pyc lib/ setup.pyc __svn_version__.pyc
constants/ linalg/ setupscons.py@ THANKS.txt@
fftpack/ linsolve/ setupscons.pyc TOCHANGE.txt@
How modules are found and imported
__init__.py@ maxentropy/ signal/ version.py@
__init__.pyc misc/ sparse/ version.pyc
When the import mymodule statement is executed, the module mymodule is searched in a given list of INSTALL.txt@ ndimage/ spatial/ weave/
directories. This list includes a list of installation-dependent default path (e.g., /usr/lib/python) as well as integrate/ odr/ special/
the list of directories specified by the environment variable PYTHONPATH. interpolate/ optimize/ stats/
$ cd ndimage
The list of directories searched by Python is given by the sys.path variable $ ls
doccer.py@ fourier.pyc interpolation.py@ morphology.pyc setup.pyc
In [1]: import sys
doccer.pyc info.py@ interpolation.pyc _nd_image.so
setupscons.py@
In [2]: sys.path
filters.py@ info.pyc measurements.py@ _ni_support.py@
Out[2]:
setupscons.pyc
['',
filters.pyc __init__.py@ measurements.pyc _ni_support.pyc tests/
'/home/varoquau/.local/bin',
fourier.py@ __init__.pyc morphology.py@ setup.py@
'/usr/lib/python2.7',
'/home/varoquau/.local/lib/python2.7/site-packages',
From Ipython:
'/usr/lib/python2.7/dist-packages',
'/usr/local/lib/python2.7/dist-packages', In [1]: import scipy
...]
In [2]: scipy.__file__
Modules must be located in the search path, therefore you can: Out[2]: '/usr/lib/python2.6/dist-packages/scipy/__init__.pyc'
write your own modules within directories already defined in the search path (e.g.
In [3]: import scipy.version
$HOME/.local/lib/python2.7/dist-packages). You may use symbolic links (on Linux) to
keep the code somewhere else. In [4]: scipy.version.version
modify the environment variable PYTHONPATH to include the directories containing the user-defined mod- Out[4]: '0.7.0'
ules.
In [5]: import scipy.ndimage.morphology
Tip: On Linux/Unix, add the following line to a file read by the shell at startup (e.g. /etc/profile, .profile) In [6]: from scipy.ndimage import morphology
export PYTHONPATH=$PYTHONPATH:/home/emma/user_defined_modules
In [17]: morphology.binary_dilation?
On Windows, http://support.microsoft.com/kb/310519 explains how to handle environment variables. Type: function

Base Class: <type 'function'>

String Form: <function binary_dilation at 0x9bedd84> Quick read
Namespace: Interactive
File: /usr/lib/python2.6/dist-packages/scipy/ndimage/morphology.py
If you want to do a first quick pass through the Scipy lectures to learn the ecosystem, you can directly skip
Definition: morphology.binary_dilation(input, structure=None, to the next chapter: NumPy: creating and manipulating numerical data (page 42).
iterations=1, mask=None, output=None, border_value=0, origin=0, The remainder of this chapter is not necessary to follow the rest of the intro part. But be sure to come back
brute_force=False) and finish this chapter later.
Docstring:
Multi-dimensional binary dilation with the given structure.
An output array can optionally be provided. The origin parameter 1.2.6 Input and Output
controls the placement of the filter. If no structuring element is
provided an element is generated with a squared connectivity equal To be exhaustive, here are some information about input and output in Python. Since we will use the Numpy
to one. The dilation operation is repeated iterations times. If methods to read and write files, you may skip this chapter at first reading.
iterations is less than 1, the dilation is repeated until the
result does not change anymore. If a mask is given, only those We write or read strings to/from files (other types must be converted to strings). To write in a file:
elements with a true value at the corresponding mask element are
>>> f = open('workfile', 'w') # opens the workfile file
modified at each iteration.
>>> type(f)
<type 'file'>
>>> f.write('This is a test \nand another test')
Good practices >>> f.close()
Use meaningful object names To read from a file

Indentation: no choice! In [1]: f = open('workfile', 'r')
Tip: Indenting is compulsory in Python! Every command block following a colon bears an additional In [2]: s = f.read()
indentation level with respect to the previous line with a colon. One must therefore indent after def f():
or while:. At the end of such logical blocks, one decreases the indentation depth (and re-increases it if a In [3]: print(s)
new block is entered, etc.) This is a test
and another test
Strict respect of indentation is the price to pay for getting rid of { or ; characters that delineate logical
blocks in other languages. Improper indentation leads to errors such as In [4]: f.close()
------------------------------------------------------------
IndentationError: unexpected indent (test.py, line 2) See also:
For more details: http://docs.python.org/tutorial/inputoutput.html
All this indentation business can be a bit confusing in the beginning. However, with the clear indentation,
and in the absence of extra characters, the resulting code is very nice to read compared to other languages.
Iterating over a file
Indentation depth: Inside your text editor, you may choose to indent with any positive number of spaces
(1, 2, 3, 4, ...). However, it is considered good practice to indent with 4 spaces. You may configure your In [6]: f = open('workfile', 'r')
editor to map the Tab key to a 4-space indentation. In Python(x,y), the editor is already configured this
way. In [7]: for line in f:
...: print line
Style guidelines ...:
This is a test
Long lines: you should not write very long lines that span over more than (e.g.) 80 characters. Long lines
can be broken with the \ character and another test
>>> long_line = "Here is a very very long line \
... that we break in two parts." In [8]: f.close()
Spaces
File modes
Write well-spaced code: put whitespaces after commas, around arithmetic operators, etc.:
>>> a = 1 # yes Read-only: r
>>> a=1 # too cramped
Write-only: w
A certain number of rules for writing beautiful code (and more importantly using the same conventions Note: Create a new file or overwrite existing file.
as anybody else!) are given in the Style Guide for Python Code.
Append a file: a
Read and Write: r+

Binary mode: b In [44]: fp = open('junk.txt', 'w')

Note: Use for binary files, especially on Windows.
In [45]: fp.close()
In [46]: 'junk.txt' in os.listdir(os.curdir)

1.2.7 Standard Library
Out[46]: True
Note: Reference document for this section: In [47]: os.remove('junk.txt')

The Python Standard Library documentation: http://docs.python.org/library/index.html In [48]: 'junk.txt' in os.listdir(os.curdir)
Python Essential Reference, David Beazley, Addison-Wesley Professional Out[48]: False
os.path: path manipulations

os module: operating system functionality
os.path provides common operations on pathnames.
A portable way of using operating system dependent functionality.
In [70]: fp = open('junk.txt', 'w')
Directory and file manipulation In [71]: fp.close()
Current directory: In [72]: a = os.path.abspath('junk.txt')
In [17]: os.getcwd() In [73]: a

Out[17]: '/Users/cburns/src/scipy2009/scipy_2009_tutorial/source' Out[73]: '/Users/cburns/src/scipy2009/scipy_2009_tutorial/source/junk.txt'
List a directory: In [74]: os.path.split(a)

Out[74]: ('/Users/cburns/src/scipy2009/scipy_2009_tutorial/source',
In [31]: os.listdir(os.curdir)
'junk.txt')
Out[31]:
['.index.rst.swo',
In [78]: os.path.dirname(a)
'.python_language.rst.swp',
Out[78]: '/Users/cburns/src/scipy2009/scipy_2009_tutorial/source'
'.view_array.py.swp',
'_static',
In [79]: os.path.basename(a)
'_templates',
Out[79]: 'junk.txt'
'basic_types.rst',
'conf.py',
In [80]: os.path.splitext(os.path.basename(a))
'control_flow.rst',
Out[80]: ('junk', '.txt')
'debugging.rst',
...
In [84]: os.path.exists('junk.txt')
Out[84]: True
Make a directory:
In [32]: os.mkdir('junkdir') In [86]: os.path.isfile('junk.txt')
Out[86]: True
In [33]: 'junkdir' in os.listdir(os.curdir)
Out[33]: True In [87]: os.path.isdir('junk.txt')
Out[87]: False
Rename the directory:
In [88]: os.path.expanduser('~/local')
In [36]: os.rename('junkdir', 'foodir') Out[88]: '/Users/cburns/local'
In [37]: 'junkdir' in os.listdir(os.curdir) In [92]: os.path.join(os.path.expanduser('~'), 'local', 'bin')

Out[37]: False Out[92]: '/Users/cburns/local/bin'
In [38]: 'foodir' in os.listdir(os.curdir)

Out[38]: True
Running an external command
In [41]: os.rmdir('foodir')
In [8]: os.system('ls')
In [42]: 'foodir' in os.listdir(os.curdir) basic_types.rst demo.py functions.rst python_language.rst standard_library.rst
Out[42]: False control_flow.rst exceptions.rst io.rst python-logo.png
demo2.py first_steps.rst oop.rst reusing_code.rst
Delete a file:

/Library/Frameworks/Python.framework/Versions/2.5/lib/python2.5'
Note: Alternative to os.system
A noteworthy alternative to os.system is the sh module. Which provides much more convenient ways to obtain In [16]: os.getenv('PYTHONPATH')
the output, error stream and exit code of the external command. Out[16]: '.:/Users/cburns/src/utils:/Users/cburns/src/nitools:
/Users/cburns/local/lib/python2.5/site-packages/:
In [20]: import sh /usr/local/lib/python2.5/site-packages/:
In [20]: com = sh.ls() /Library/Frameworks/Python.framework/Versions/2.5/lib/python2.5'
In [21]: print com

basic_types.rst exceptions.rst oop.rst standard_library.rst shutil: high-level file operations
control_flow.rst first_steps.rst python_language.rst
demo2.py functions.rst python-logo.png
demo.py io.rst reusing_code.rst
The shutil provides useful file operations:
shutil.rmtree: Recursively delete a directory tree.
In [22]: print com.exit_code
0 shutil.move: Recursively move a file or directory to another location.
In [23]: type(com)
shutil.copy: Copy files or directories.
Out[23]: sh.RunningCommand
glob: Pattern matching on files

Walking a directory
The glob module provides convenient file pattern matching.
os.path.walk generates a list of filenames in a directory tree. Find all files ending in .txt:
In [18]: import glob
In [10]: for dirpath, dirnames, filenames in os.walk(os.curdir):
....: for fp in filenames:
In [19]: glob.glob('*.txt')
....: print os.path.abspath(fp)
Out[19]: ['holy_grail.txt', 'junk.txt', 'newfile.txt']
....:
....:
/Users/cburns/src/scipy2009/scipy_2009_tutorial/source/.index.rst.swo
/Users/cburns/src/scipy2009/scipy_2009_tutorial/source/.view_array.py.swp sys module: system-specific information
/Users/cburns/src/scipy2009/scipy_2009_tutorial/source/basic_types.rst
/Users/cburns/src/scipy2009/scipy_2009_tutorial/source/conf.py System-specific information related to the Python interpreter.
/Users/cburns/src/scipy2009/scipy_2009_tutorial/source/control_flow.rst
... Which version of python are you running and where is it installed:
In [117]: sys.platform
Out[117]: 'darwin'
Environment variables:
In [118]: sys.version
In [9]: import os Out[118]: '2.5.2 (r252:60911, Feb 22 2008, 07:57:53) \n
[GCC 4.0.1 (Apple Computer, Inc. build 5363)]'
In [11]: os.environ.keys()
Out[11]: In [119]: sys.prefix
['_', Out[119]: '/Library/Frameworks/Python.framework/Versions/2.5'
'FSLDIR',
'TERM_PROGRAM_VERSION', List of command line arguments passed to a Python script:
'FSLREMOTECALL', In [100]: sys.argv
'USER', Out[100]: ['/Users/cburns/local/bin/ipython']
'HOME',
'PATH',
sys.path is a list of strings that specifies the search path for modules. Initialized from PYTHONPATH:
'PS1',
'SHELL', In [121]: sys.path
'EDITOR', Out[121]:
'WORKON_HOME', ['',
'PYTHONPATH', '/Users/cburns/local/bin',
... '/Users/cburns/local/lib/python2.5/site-packages/grin-1.1-py2.5.egg',
'/Users/cburns/local/lib/python2.5/site-packages/argparse-0.8.0-py2.5.egg',
In [12]: os.environ['PYTHONPATH'] '/Users/cburns/local/lib/python2.5/site-packages/urwid-0.9.7.1-py2.5.egg',
Out[12]: '.:/Users/cburns/src/utils:/Users/cburns/src/nitools: '/Users/cburns/local/lib/python2.5/site-packages/yolk-0.4.1-py2.5.egg',
/Users/cburns/local/lib/python2.5/site-packages/: '/Users/cburns/local/lib/python2.5/site-packages/virtualenv-1.2-py2.5.egg',
/usr/local/lib/python2.5/site-packages/: ...

pickle: easy persistence Catching exceptions
Useful to store arbitrary objects to a file. Not safe or fast! try/except
In [1]: import pickle

In [10]: while True:
In [2]: l = [1, None, 'Stan'] ....: try:
....: x = int(raw_input('Please enter a number: '))
In [3]: pickle.dump(l, file('test.pkl', 'w')) ....: break
....: except ValueError:
In [4]: pickle.load(file('test.pkl')) ....: print('That was no valid number. Try again...')
Out[4]: [1, None, 'Stan'] ....:
Please enter a number: a
That was no valid number. Try again...
Exercise Please enter a number: 1
Write a program to search your PYTHONPATH for the module site.py. In [9]: x
Out[9]: 1
path_site
try/finally
1.2.8 Exception handling in Python
In [10]: try:
It is likely that you have raised Exceptions if you have typed all the previous commands of the tutorial. For ....: x = int(raw_input('Please enter a number: '))
example, you may have raised an exception if you entered a command with a typo. ....: finally:
....: print('Thank you for your input')
Exceptions are raised by different kinds of errors arising when executing Python code. In your own code, you ....:
may also catch errors, or define custom error types. You may want to look at the descriptions of the the built-in ....:
Exceptions when looking for the right exception type. Please enter a number: a
Thank you for your input
---------------------------------------------------------------------------
Exceptions ValueError: invalid literal for int() with base 10: 'a'
Exceptions are raised by errors in Python: Important for resource management (e.g. closing a file)
In [1]: 1/0
--------------------------------------------------------------------------- Easier to ask for forgiveness than for permission
ZeroDivisionError: integer division or modulo by zero
In [11]: def print_sorted(collection):
In [2]: 1 + 'e'
....: try:
---------------------------------------------------------------------------
....: collection.sort()
TypeError: unsupported operand type(s) for +: 'int' and 'str'
....: except AttributeError:
....: pass
In [3]: d = {1:1, 2:2}
....: print(collection)
....:
In [4]: d[3]
....:
---------------------------------------------------------------------------
KeyError: 3
In [12]: print_sorted([1, 3, 2])
[1, 2, 3]
In [5]: l = [1, 2, 3]
In [13]: print_sorted(set((1, 3, 2)))
In [6]: l[4]
set([1, 2, 3])
---------------------------------------------------------------------------
IndexError: list index out of range
In [14]: print_sorted('132')
132
In [7]: l.foobar
---------------------------------------------------------------------------
AttributeError: 'list' object has no attribute 'foobar'
Raising exceptions
As you can see, there are different types of exceptions for different errors.
Capturing and reraising an exception:

In [15]: def filter_name(name): >>> anna.set_age(21)

....: try: >>> anna.set_major('physics')
....: name = name.encode('ascii')
....: except UnicodeError, e: In the previous example, the Student class has __init__, set_age and set_major methods. Its at-
....: if name == 'Gal': tributes are name, age and major. We can call these methods and attributes with the following notation:
....: print('OK, Gal') classinstance.method or classinstance.attribute. The __init__ constructor is a special
....: else: method we call with: MyClass(init parameters if any).
....: raise e
....: return name Now, suppose we want to create a new class MasterStudent with the same methods and attributes as the previous
....: one, but with an additional internship attribute. We wont copy the previous class, but inherit from it:
In [16]: filter_name('Gal') >>> class MasterStudent(Student):

OK, Gal ... internship = 'mandatory, from March to June'
Out[16]: 'Ga\xc3\xabl' ...
>>> james = MasterStudent('james')
In [17]: filter_name('Stfan') >>> james.internship
--------------------------------------------------------------------------- 'mandatory, from March to June'
UnicodeDecodeError: 'ascii' codec can't decode byte 0xc3 in position 2: ordinal not in range(128) >>> james.set_age(23)
>>> james.age
Exceptions to pass messages between parts of the code: 23
In [17]: def achilles_arrow(x): The MasterStudent class inherited from the Student attributes and methods.
....: if abs(x - 1) < 1e-3:
....: raise StopIteration Thanks to classes and object-oriented programming, we can organize code with different classes corresponding to
....: x = 1 - (1-x)/2. different objects we encounter (an Experiment class, an Image class, a Flow class, etc.), with their own methods
....: return x and attributes. Then we can use inheritance to consider variations around a base class and re-use code. Ex : from
....: a Flow base class, we can create derived StokesFlow, TurbulentFlow, PotentialFlow, etc.
In [18]: x = 0
In [19]: while True:

1.3 NumPy: creating and manipulating numerical data
....: try:
....: x = achilles_arrow(x) Authors: Emmanuelle Gouillart, Didrik Pinte, Gal Varoquaux, and Pauli Virtanen
....: except StopIteration:
....: break This chapter gives an overview of Numpy, the core tool for performant numerical computing with Python.
....:
....:
In [20]: x
Out[20]: 0.9990234375 1.3.1 The Numpy array object
Use exceptions to notify certain conditions are met (e.g. StopIteration) or not (e.g. custom error raising)
Section contents
1.2.9 Object-oriented programming (OOP) What are Numpy and Numpy arrays? (page 42)
Creating arrays (page 44)
Python supports object-oriented programming (OOP). The goals of OOP are: Basic data types (page 46)
Basic visualization (page 47)
to organize the code, and Indexing and slicing (page 49)
to re-use code in similar contexts. Copies and views (page 52)
Fancy indexing (page 53)
Here is a small example: we create a Student class, which is an object gathering several custom functions (meth-
ods) and variables (attributes), we will be able to use:
>>> class Student(object): What are Numpy and Numpy arrays?
... def __init__(self, name):
... self.name = name
... def set_age(self, age): Numpy arrays
... self.age = age
... def set_major(self, major): Python objects
... self.major = major
... high-level number objects: integers, floating point
>>> anna = Student('anna')
containers: lists (costless insertion and append), dictionaries (fast lookup)
1.2. The Python language 41 1.3. NumPy: creating and manipulating numerical data 42
Numpy provides
In [6]: np.con* ?
extension package to Python for multi-dimensional arrays np.concatenate
np.conj
closer to hardware (efficiency) np.conjugate
designed for scientific computation (convenience) np.convolve
Also known as array oriented computing

Import conventions
The recommended convention to import numpy is:

>>> import numpy as np
>>> a = np.array([0, 1, 2, 3])
>>> a Creating arrays
array([0, 1, 2, 3])
Manual construction of arrays
Tip: For example, An array containing:
values of an experiment/simulation at discrete time steps 1-D:
signal recorded by a measurement device, e.g. sound wave >>> a = np.array([0, 1, 2, 3])
>>> a
pixels of an image, grey-level or colour array([0, 1, 2, 3])
3-D data measured at different X-Y-Z positions, e.g. MRI scan >>> a.ndim
1
... >>> a.shape
(4,)
>>> len(a)
Why it is useful: Memory-efficient container that provides fast numerical operations.
4
In [1]: L = range(1000)
2-D, 3-D, ...:
In [2]: %timeit [i**2 for i in L]
>>> b = np.array([[0, 1, 2], [3, 4, 5]]) # 2 x 3 array
1000 loops, best of 3: 403 us per loop
>>> b
array([[0, 1, 2],
In [3]: a = np.arange(1000)
[3, 4, 5]])
>>> b.ndim
In [4]: %timeit a**2
2
100000 loops, best of 3: 12.7 us per loop
>>> b.shape
(2, 3)
>>> len(b) # returns the size of the first dimension
Numpy Reference documentation 2
On the web: http://docs.scipy.org/ >>> c = np.array([[[1], [2]], [[3], [4]]])

>>> c
Interactive help: array([[[1],
[2]],
In [5]: np.array?
String Form:<built-in function array> [[3],
Docstring: [4]]])
array(object, dtype=None, copy=True, order=None, subok=False, ndmin=0, ... >>> c.shape
(2, 2, 1)
Looking for something:
>>> np.lookfor('create array')
Search results for 'create array' Exercise: Simple arrays
--------------------------------- Create a simple two dimensional array. First, redo the examples from above. And then create your
numpy.array own: how about odd numbers counting backwards on the first row, and even numbers on the second?
Create an array.
Use the functions len(), numpy.shape() on these arrays. How do they relate to each other? And
numpy.memmap
Create a memory-map to an array stored in a *binary* file on disk.
to the ndim attribute of the arrays?
1.3. NumPy: creating and manipulating numerical data 43 1.3. NumPy: creating and manipulating numerical data 44
Functions for creating arrays Exercise: Creating arrays using functions

Experiment with arange, linspace, ones, zeros, eye and diag.
Tip: In practice, we rarely enter items one by one... Create different kinds of arrays with random numbers.
Try setting the seed before creating an array with random values.
Evenly spaced: Look at the function np.empty. What does it do? When might this be useful?
>>> a = np.arange(10) # 0 .. n-1 (!)

>>> a
array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9]) Basic data types
>>> b = np.arange(1, 9, 2) # start, end (exclusive), step
>>> b You may have noticed that, in some instances, array elements are displayed with a trailing dot (e.g. 2. vs 2). This
array([1, 3, 5, 7]) is due to a difference in the data-type used:
or by number of points: >>> a = np.array([1, 2, 3])
>>> a.dtype
>>> c = np.linspace(0, 1, 6) # start, end, num-points dtype('int64')
>>> c
array([ 0. , 0.2, 0.4, 0.6, 0.8, 1. ]) >>> b = np.array([1., 2., 3.])
>>> d = np.linspace(0, 1, 5, endpoint=False) >>> b.dtype
>>> d dtype('float64')
array([ 0. , 0.2, 0.4, 0.6, 0.8])
Common arrays: Tip: Different data-types allow us to store data more compactly in memory, but most of the time we simply work
with floating point numbers. Note that, in the example above, NumPy auto-detects the data-type from the input.
>>> a = np.ones((3, 3)) # reminder: (3, 3) is a tuple
>>> a
array([[ 1., 1., 1.],
[ 1., 1., 1.], You can explicitly specify which data-type you want:
[ 1., 1., 1.]])
>>> b = np.zeros((2, 2)) >>> c = np.array([1, 2, 3], dtype=float)
>>> b >>> c.dtype
array([[ 0., 0.], dtype('float64')
[ 0., 0.]])
>>> c = np.eye(3) The default data type is floating point:
>>> c
>>> a = np.ones((3, 3))
array([[ 1., 0., 0.],
>>> a.dtype
[ 0., 1., 0.],
dtype('float64')
[ 0., 0., 1.]])
>>> d = np.diag(np.array([1, 2, 3, 4]))
>>> d There are also other types:
array([[1, 0, 0, 0], Complex
[0, 2, 0, 0],
[0, 0, 3, 0], >>> d = np.array([1+2j, 3+4j, 5+6*1j])
[0, 0, 0, 4]]) >>> d.dtype
dtype('complex128')
np.random: random numbers (Mersenne Twister PRNG):
Bool
>>> a = np.random.rand(4) # uniform in [0, 1]
>>> a >>> e = np.array([True, False, False, True])
array([ 0.95799151, 0.14222247, 0.08777354, 0.51887998]) >>> e.dtype
dtype('bool')
>>> b = np.random.randn(4) # Gaussian
>>> b Strings
array([ 0.37544699, -0.11425369, -0.47616538, 1.79664113])
>>> f = np.array(['Bonjour', 'Hello', 'Hallo',])
>>> np.random.seed(1234) # Setting the random seed >>> f.dtype # <--- strings containing max. 7 letters
dtype('S7')
Much more
int32
int64
uint32
uint64
Basic visualization
9
Now that we have our first data arrays, we are going to visualize them. 8
Start by launching IPython:
$ ipython
7
Or the notebook: 6
$ ipython notebook
5
Once IPython has started, enable interactive plots:
>>> %matplotlib
4
Or, from the notebook, enable plots in the notebook: 3
>>> %matplotlib inline
2
The inline is important for the notebook, so that plots are displayed in the notebook and not in a new window.
Matplotlib is a 2D plotting package. We can import its functions as below: 1
>>> import matplotlib.pyplot as plt
0
# the tidy way
And then use (note that you have to use show explicitly if you have not enabled interactive plots with 0.0 0.5 1.0 1.5 2.0 2.5 3.0
%matplotlib):
>>> plt.plot(x, y) # line plot
2D arrays (such as images):
>>> plt.show() # <-- shows the plot (not needed with interactive plots)
>>> image = np.random.rand(30, 30)
Or, if you have enabled interactive plots with %matplotlib: >>> plt.imshow(image, cmap=plt.cm.hot)
>>> plt.colorbar()
>>> plot(x, y) # line plot <matplotlib.colorbar.Colorbar instance at ...>
1D plotting:
>>> x = np.linspace(0, 3, 20)
>>> y = np.linspace(0, 9, 20)
>>> plt.plot(x, y) # line plot
[<matplotlib.lines.Line2D object at ...>]
>>> plt.plot(x, y, 'o') # dot plot
>>> a = np.diag(np.arange(3))
>>> a
array([[0, 0, 0],
0 [0, 1, 0],
0.9 [0, 0, 2]])
>>> a[1, 1]
5 0.8 1
>>> a[2, 1] = 10 # third line, second column
>>> a
0.7 array([[ 0, 0, 0],
10 [ 0, 1, 0],
0.6 [ 0, 10, 2]])
>>> a[1]
15 0.5 array([0, 1, 0])
0.4 Note:
In 2D, the first dimension corresponds to rows, the second to columns.
20 0.3 for multidimensional a, a[0] is interpreted by taking all elements in the unspecified dimensions.
25 0.2 Slicing: Arrays, like other Python sequences can also be sliced:
0.1 >>> a = np.arange(10)

>>> a
0 5 10 15 20 25 0.0 array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9])

>>> a[2:9:3] # [start:end:step]
array([2, 5, 8])
Note that the last index is not included! :

See also:
>>> a[:4]
More in the: matplotlib chapter (page 81) array([0, 1, 2, 3])
All three slice components are not required: by default, start is 0, end is the last and step is 1:
Exercise: Simple visualizations
>>> a[1:3]
Plot some simple arrays: a cosine as a function of time and a 2D matrix. array([1, 2])
Try using the gray colormap on the 2D matrix. >>> a[::2]
array([0, 2, 4, 6, 8])
>>> a[3:]
array([3, 4, 5, 6, 7, 8, 9])
Indexing and slicing
A small illustrated summary of Numpy indexing and slicing...
The items of an array can be accessed and assigned to the same way as other Python sequences (e.g. lists):
>>> a = np.arange(10)
>>> a
array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9])
>>> a[0], a[2], a[-1]
(0, 2, 9)
Warning: Indices begin at 0, like other Python sequences (and C/C++). In contrast, in Fortran or Matlab,
indices begin at 1.
The usual python idiom for reversing a sequence is supported:

>>> a[::-1]
array([9, 8, 7, 6, 5, 4, 3, 2, 1, 0])
For multidimensional arrays, indexes are tuples of integers:
Exercise: Array creation

Create the following arrays (with correct data types):
[[1, 1, 1, 1],
[1, 1, 1, 1],
[1, 1, 1, 2],
[1, 6, 1, 1]]
[[0., 0., 0., 0., 0.],

[2., 0., 0., 0., 0.],
[0., 3., 0., 0., 0.],
[0., 0., 4., 0., 0.],
[0., 0., 0., 5., 0.],
[0., 0., 0., 0., 6.]]
Par on course: 3 statements for each
Hint: Individual array elements can be accessed similarly to a list, e.g. a[1] or a[1, 2].
Hint: Examine the docstring for diag.
Exercise: Tiling for array creation

Skim through the documentation for np.tile, and use this function to construct the array:
You can also combine assignment and slicing: [[4, 3, 4, 3, 4, 3],
[2, 1, 2, 1, 2, 1],
[4, 3, 4, 3, 4, 3],
>>> a[5:] = 10
[2, 1, 2, 1, 2, 1]]
>>> a
array([ 0, 1, 2, 3, 4, 10, 10, 10, 10, 10])
>>> b = np.arange(5)
>>> a[5:] = b[::-1]
>>> a Copies and views
array([0, 1, 2, 3, 4, 4, 3, 2, 1, 0])
A slicing operation creates a view on the original array, which is just a way of accessing array data. Thus the
original array is not copied in memory. You can use np.may_share_memory() to check if two arrays share
Exercise: Indexing and slicing the same memory block. Note however, that this uses heuristics and may give you false positives.
Try the different flavours of slicing, using start, end and step: starting from a linspace, try to When modifying the view, the original array is modified as well:
obtain odd numbers counting backwards, and even numbers counting forwards. >>> a = np.arange(10)
Reproduce the slices in the diagram above. You may use the following expression to create the array: >>> a
array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9])
>>> np.arange(6) + np.arange(0, 51, 10)[:, np.newaxis]
>>> b = a[::2]
array([[ 0, 1, 2, 3, 4, 5],
>>> b
[10, 11, 12, 13, 14, 15],
array([0, 2, 4, 6, 8])
[20, 21, 22, 23, 24, 25],
>>> np.may_share_memory(a, b)
[30, 31, 32, 33, 34, 35],
True
[40, 41, 42, 43, 44, 45],
>>> b[0] = 12
[50, 51, 52, 53, 54, 55]])
>>> b
array([12, 2, 4, 6, 8])
>>> a # (!)
array([12, 1, 2, 3, 4, 5, 6, 7, 8, 9])
>>> c = a[::2].copy() # force a copy
>>> c[0] = 12
>>> a
array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9])
>>> np.may_share_memory(a, c)
False
This behavior can be surprising at first sight... but it allows to save both memory and time.
>>> extract_from_a = a[mask] # or, a[a%3==0]

Worked example: Prime number sieve >>> extract_from_a # extract a sub-array with the mask
array([ 3, 0, 9, 6, 0, 12])
Indexing with a mask can be very useful to assign a new value to a sub-array:
>>> a[a % 3 == 0] = -1
>>> a
array([10, -1, 8, -1, 19, 10, 11, -1, 10, -1, -1, 20, -1, 7, 14])
Indexing with an array of integers
>>> a = np.arange(0, 100, 10)

>>> a
array([ 0, 10, 20, 30, 40, 50, 60, 70, 80, 90])
Indexing can be done with an array of integers, where the same index is repeated several time:
>>> a[[2, 3, 2, 4, 2]] # note: [2, 3, 2, 4, 2] is a Python list
array([20, 30, 20, 40, 20])
Compute prime numbers in 099, with a sieve
Construct a shape (100,) boolean array is_prime, filled with True in the beginning: New values can be assigned with this kind of indexing:
>>> is_prime = np.ones((100,), dtype=bool) >>> a[[9, 7]] = -100

>>> a
Cross out 0 and 1 which are not primes: array([ 0, 10, 20, 30, 40, 50, 60, -100, 80, -100])
>>> is_prime[:2] = 0
For each integer j starting from 2, cross out its higher multiples: Tip: When a new array is created by indexing with an array of integers, the new array has the same shape than
the array of integers:
>>> N_max = int(np.sqrt(len(is_prime)))
>>> for j in range(2, N_max):
>>> idx = np.array([[3, 4], [9, 7]])
... is_prime[2*j::j] = False
>>> idx.shape
Skim through help(np.nonzero), and print the prime numbers (2, 2)
Follow-up: >>> a[idx]
Move the above code into a script file named prime_sieve.py array([[3, 4],
Run it to check it works [9, 7]])
Use the optimization suggested in the sieve of Eratosthenes:
1. Skip j which are already known to not be primes
2. The first number to cross out is 2
The image below illustrates various fancy indexing applications
Fancy indexing
Tip: Numpy arrays can be indexed with slices, but also with boolean or integer arrays (masks). This method is
called fancy indexing. It creates copies not views.
Using boolean masks
>>> np.random.seed(3)
>>> a = np.random.random_integers(0, 20, 15)
>>> a
array([10, 3, 8, 0, 19, 10, 11, 9, 10, 6, 0, 20, 12, 7, 14])
>>> (a % 3 == 0)
array([False, True, False, True, False, False, False, True, False,
True, True, False, True, False, False], dtype=bool)
>>> mask = (a % 3 == 0)
Warning: Array multiplication is not matrix multiplication:

Exercise: Fancy indexing
>>> c = np.ones((3, 3))
Again, reproduce the fancy indexing shown in the diagram above.
>>> c * c # NOT matrix multiplication!
Use fancy indexing on the left and array creation on the right to assign values into an array, for instance array([[ 1., 1., 1.],
by setting parts of the array in the diagram above to zero. [ 1., 1., 1.],
[ 1., 1., 1.]])
1.3.2 Numerical operations on arrays

Note: Matrix multiplication:
>>> c.dot(c)
Section contents array([[ 3., 3., 3.],
[ 3., 3., 3.],
Elementwise operations (page 55) [ 3., 3., 3.]])
Basic reductions (page 57)
Broadcasting (page 62)
Array shape manipulation (page 65)
Sorting data (page 67) Exercise: Elementwise operations
Summary (page 68)
Try simple arithmetic elementwise operations: add even elements with odd elements
Time them against their pure python counterparts using %timeit.
Generate:
Elementwise operations [2**0, 2**1, 2**2, 2**3, 2**4]
a_j = 2^(3*j) - j
Basic operations
With scalars: Other operations
>>> a = np.array([1, 2, 3, 4])

>>> a + 1 Comparisons:
array([2, 3, 4, 5])
>>> a = np.array([1, 2, 3, 4])
>>> 2**a
>>> b = np.array([4, 2, 2, 4])
array([ 2, 4, 8, 16])
>>> a == b
array([False, True, False, True], dtype=bool)
All arithmetic operates elementwise: >>> a > b
>>> b = np.ones(4) + 1 array([False, False, True, False], dtype=bool)
>>> a - b
array([-1., 0., 1., 2.]) Tip: Array-wise comparisons:
>>> a * b
array([ 2., 4., 6., 8.]) >>> a = np.array([1, 2, 3, 4])
>>> b = np.array([4, 2, 2, 4])
>>> j = np.arange(5) >>> c = np.array([1, 2, 3, 4])
>>> 2**(j + 1) - j >>> np.array_equal(a, b)
array([ 2, 3, 6, 13, 28]) False
>>> np.array_equal(a, c)
These operations are of course much faster than if you did them in pure python: True
>>> a = np.arange(10000)
>>> %timeit a + 1 Logical operations:
>>> a = np.array([1, 1, 0, 0], dtype=bool)
>>> l = range(10000)
>>> b = np.array([1, 0, 1, 0], dtype=bool)
>>> %timeit [i+1 for i in l]
>>> np.logical_or(a, b)
1000 loops, best of 3: 861 us per loop
array([ True, True, True, False], dtype=bool)
>>> np.logical_and(a, b)
array([ True, False, False, False], dtype=bool)
Transcendental functions:
>>> np.sin(a)
array([ 0. , 0.84147098, 0.90929743, 0.14112001, -0.7568025 ])

>>> np.log(a)
array([ -inf, 0. , 0.69314718, 1.09861229, 1.38629436])
>>> np.exp(a)
array([ 1. , 2.71828183, 7.3890561 , 20.08553692, 54.59815003])
Shape mismatches
>>> a + np.array([1, 2])
ValueError: operands could not be broadcast together with shapes (4) (2)
ValueError: operands could not be broadcast together with shapes (4) (2)
Broadcasting? Well return to that later (page 62). Sum by rows and by columns:
Transposition: >>> x = np.array([[1, 1], [2, 2]])
>>> a = np.triu(np.ones((3, 3)), 1) # see help(np.triu) >>> x
>>> a array([[1, 1],
array([[ 0., 1., 1.], [2, 2]])
[ 0., 0., 1.], >>> x.sum(axis=0) # columns (first dimension)
[ 0., 0., 0.]]) array([3, 3])
>>> a.T >>> x[:, 0].sum(), x[:, 1].sum()
array([[ 0., 0., 0.], (3, 3)
[ 1., 0., 0.], >>> x.sum(axis=1) # rows (second dimension)
[ 1., 1., 0.]]) array([2, 4])
>>> x[0, :].sum(), x[1, :].sum()
(2, 4)
Warning: The transposition is a view
As a results, the following code is wrong and will not make a matrix symmetric:
Tip: Same idea in higher dimensions:
>>> a += a.T
>>> x = np.random.rand(2, 2, 2)
>>> x.sum(axis=2)[0, 1]
1.14764...
Note: Linear algebra >>> x[0, 1, :].sum()
1.14764...
The sub-module numpy.linalg implements basic linear algebra, such as solving linear systems, singular value
decomposition, etc. However, it is not guaranteed to be compiled using efficient routines, and thus we recommend
the use of scipy.linalg, as detailed in section Linear algebra operations: scipy.linalg (page 105)
Other reductions
Exercise other operations works the same way (and take axis=)
Look at the help for np.allclose. When might this be useful? Extrema:
Look at the help for np.triu and np.tril.
>>> x = np.array([1, 3, 2])
>>> x.min()
1
Basic reductions >>> x.max()
3
Computing sums
>>> x.argmin() # index of minimum
0
>>> x = np.array([1, 2, 3, 4]) >>> x.argmax() # index of maximum
>>> np.sum(x) 1
10
>>> x.sum() Logical operations:
10
>>> np.all([True, True, False])
False
>>> np.any([True, True, False]) Worked Example: data statistics

True
Data in populations.txt describes the populations of hares and lynxes (and carrots) in northern Canada
during 20 years.
Note: Can be used for array comparisons: You can view the data in an editor, or alternatively in IPython (both shell and notebook):
>>> a = np.zeros((100, 100))
>>> np.any(a != 0) In [1]: !cat data/populations.txt
False First, load the data into a Numpy array:
>>> np.all(a == a)
True >>> data = np.loadtxt('data/populations.txt')
>>> year, hares, lynxes, carrots = data.T # trick: columns to variables
>>> a = np.array([1, 2, 3, 2])
>>> b = np.array([2, 2, 3, 2])
Then plot it:
>>> c = np.array([6, 4, 4, 5]) >>> from matplotlib import pyplot as plt
>>> ((a <= b) & (b <= c)).all() >>> plt.axes([0.2, 0.1, 0.5, 0.8])
True >>> plt.plot(year, hares, year, lynxes, year, carrots)
>>> plt.legend(('Hare', 'Lynx', 'Carrot'), loc=(1.05, 0.5))
Statistics:
>>> x = np.array([1, 2, 3, 1])
>>> y = np.array([[1, 2, 3], [5, 6, 1]])
80000
>>> x.mean()
1.75
>>> np.median(x)
70000
1.5
>>> np.median(y, axis=-1) # last axis
60000
array([ 2., 5.])
Hare
>>> x.std() # full population standard dev. 50000 Lynx
Carrot
0.82915619758884995
... and many more (best to learn as you go). 40000

Exercise: Reductions
30000
Given there is a sum, what other function might you expect to see?
What is the difference between sum and cumsum?
20000
10000
0
1900 1905 1910 1915 1920
The mean populations over time:
>>> populations = data[:, 1:]
>>> populations.mean(axis=0)
array([ 34080.95238095, 20166.66666667, 42400. ])
The sample standard deviations:
>>> populations.std(axis=0)
array([ 20897.90645809, 16254.59153691, 3322.50622558])
Which species has the highest population each year?:
>>> np.argmax(populations, axis=1)
array([2, 2, 0, 0, 1, 1, 2, 2, 2, 2, 2, 2, 0, 0, 0, 1, 2, 2, 2, 2, 2])
Worked Example: diffusion using a random walk algorithm Broadcasting
Basic operations on numpy arrays (addition, etc.) are elementwise

This works on arrays of the same size.
Tip: Let us consider a simple 1D random walk process: at each time step a walker jumps right or left with Nevertheless, Its also possible to do operations on arrays of different
equal probability. sizes if Numpy can transform these arrays so that they all have
We are interested in finding the typical distance from the origin of a random walker after t left or right the same size: this conversion is called broadcasting.
jumps? We are going to simulate many walkers to find this law, and we are going to do so using array
computing tricks: we are going to create a 2D array with the stories (each walker has a story) in one The image below gives an example of broadcasting:
direction, and the time in the other:
>>> n_stories = 1000 # number of walkers

>>> t_max = 200 # time during which we follow the walker
We randomly choose all the steps 1 or -1 of the walk:
>>> t = np.arange(t_max)
>>> steps = 2 * np.random.random_integers(0, 1, (n_stories, t_max)) - 1
>>> np.unique(steps) # Verification: all steps are 1 or -1
array([-1, 1])
We build the walks by summing steps along the time:
>>> positions = np.cumsum(steps, axis=1) # axis = 1: dimension of time
>>> sq_distance = positions**2
We get the mean in the axis of the stories:
Lets verify:
>>> mean_sq_distance = np.mean(sq_distance, axis=0)
Plot the results: >>> a = np.tile(np.arange(0, 40, 10), (3, 1)).T
>>> a
>>> plt.figure(figsize=(4, 3)) array([[ 0, 0, 0],
<matplotlib.figure.Figure object at ...> [10, 10, 10],
>>> plt.plot(t, np.sqrt(mean_sq_distance), 'g.', t, np.sqrt(t), 'y-') [20, 20, 20],
[<matplotlib.lines.Line2D object at ...>, <matplotlib.lines.Line2D object at ...>] [30, 30, 30]])
>>> plt.xlabel(r"$t$") >>> b = np.array([0, 1, 2])
<matplotlib.text.Text object at ...> >>> a + b
>>> plt.ylabel(r"$\sqrt{\langle (\delta x)^2 \rangle}$") array([[ 0, 1, 2],
<matplotlib.text.Text object at ...> [10, 11, 12],
[20, 21, 22],
16 [30, 31, 32]])
14 We have already used broadcasting without knowing it!:

>>> a = np.ones((4, 5))
12 >>> a[0] = 2 # we assign an array of dimension 0 to an array of dimension 1
>>> a
10 array([[ 2., 2., 2., 2., 2.],
(x)2
[ 1., 1., 1., 1., 1.],

8
6
q
4
2
[ 1., 1., 1., 1., 1.], distance from the origin of points on a 10x10 grid, we can do
[ 1., 1., 1., 1., 1.]]) >>> x, y = np.arange(5), np.arange(5)[:, np.newaxis]
>>> distance = np.sqrt(x ** 2 + y ** 2)
An useful trick: >>> distance
>>> a = np.arange(0, 40, 10) array([[ 0. , 1. , 2. , 3. , 4. ],
>>> a.shape [ 1. , 1.41421356, 2.23606798, 3.16227766, 4.12310563],
(4,) [ 2. , 2.23606798, 2.82842712, 3.60555128, 4.47213595],
>>> a = a[:, np.newaxis] # adds a new axis -> 2D array [ 3. , 3.16227766, 3.60555128, 4.24264069, 5. ],
>>> a.shape [ 4. , 4.12310563, 4.47213595, 5. , 5.65685425]])
(4, 1)
>>> a Or in color:
array([[ 0], >>> plt.pcolor(distance)
[10], >>> plt.colorbar()
[20],
[30]])
>>> a + b
array([[ 0, 1, 2],
[10, 11, 12],
[20, 21, 22],
5 5.4
[30, 31, 32]])
4.8
Tip: Broadcasting seems a bit magical, but it is actually quite natural to use it when we want to solve a problem 4
whose output data is an array with more dimensions than input data. 4.2
3.6
Worked Example: Broadcasting 3
Lets construct an array of distances (in miles) between cities of Route 66: Chicago, Springfield, Saint-Louis, 3.0
Tulsa, Oklahoma City, Amarillo, Santa Fe, Albuquerque, Flagstaff and Los Angeles.
>>> mileposts = np.array([0, 198, 303, 736, 871, 1175, 1475, 1544, 2 2.4
... 1913, 2448])
>>> distance_array = np.abs(mileposts - mileposts[:, np.newaxis])
>>> distance_array
1.8
array([[ 0, 198, 303, 736, 871, 1175, 1475, 1544, 1913, 2448],
[ 198, 0, 105, 538, 673, 977, 1277, 1346, 1715, 2250], 1 1.2
[ 303, 105, 0, 433, 568, 872, 1172, 1241, 1610, 2145],
[ 736, 538, 433,
[ 871, 673, 568, 135,
0, 135, 439, 739, 808, 1177, 1712],
0, 304, 604, 673, 1042, 1577],
0.6
[1175, 977, 872, 439, 304, 0, 300, 369, 738, 1273], 0 0.0
[1475, 1277, 1172, 739, 604, 300,
[1544, 1346, 1241, 808, 673, 369,
0,
69,
69, 438, 973],
0, 369, 904], 0 1 2 3 4 5
[1913, 1715, 1610, 1177, 1042, 738, 438, 369, 0, 535],
[2448, 2250, 2145, 1712, 1577, 1273, 973, 904, 535, 0]])
Remark : the numpy.ogrid function allows to directly create vectors x and y of the previous example, with
two significant dimensions:
>>> x, y = np.ogrid[0:5, 0:5]
>>> x, y
(array([[0],
[1],
[2],
[3],
[4]]), array([[0, 1, 2, 3, 4]]))
>>> x.shape, y.shape
((5, 1), (1, 5))
>>> distance = np.sqrt(x ** 2 + y ** 2)
Tip: So, np.ogrid is very useful as soon as we have to handle computations on a grid. On the other hand,
np.mgrid directly provides matrices full of indices for cases where we cant (or dont want to) benefit from
A lot of grid-based or network-based problems can also use broadcasting. For instance, if we want to compute the broadcasting:
>>> x, y = np.mgrid[0:4, 0:4] >>> a = np.zeros((3, 2))

>>> x >>> b = a.T.reshape(3*2)
array([[0, 0, 0, 0], >>> b[0] = 9
[1, 1, 1, 1], >>> a
[2, 2, 2, 2], array([[ 0., 0.],
[3, 3, 3, 3]]) [ 0., 0.],
>>> y [ 0., 0.]])
array([[0, 1, 2, 3],
[0, 1, 2, 3], To understand this you need to learn more about the memory layout of a numpy array.
[0, 1, 2, 3],
[0, 1, 2, 3]])
Adding a dimension
Array shape manipulation Indexing with the np.newaxis object allows us to add an axis to an array (you have seen this already above in
the broadcasting section):
Flattening
>>> z = np.array([1, 2, 3])
>>> z
>>> a = np.array([[1, 2, 3], [4, 5, 6]]) array([1, 2, 3])
>>> a.ravel()
array([1, 2, 3, 4, 5, 6]) >>> z[:, np.newaxis]
>>> a.T array([[1],
array([[1, 4], [2],
[2, 5], [3]])
[3, 6]])
>>> a.T.ravel() >>> z[np.newaxis, :]
array([1, 4, 2, 5, 3, 6]) array([[1, 2, 3]])
Higher dimensions: last dimensions ravel out first.

Dimension shuffling
Reshaping
>>> a = np.arange(4*3*2).reshape(4, 3, 2)
>>> a.shape
The inverse operation to flattening: (4, 3, 2)
>>> a.shape >>> a[0, 2, 1]
(2, 3) 5
>>> b = a.ravel() >>> b = a.transpose(1, 2, 0)
>>> b = b.reshape((2, 3)) >>> b.shape
>>> b (3, 2, 4)
array([[1, 2, 3], >>> b[2, 1, 0]
[4, 5, 6]]) 5
Or, Also creates a view:
>>> a.reshape((2, -1)) # unspecified (-1) value is inferred >>> b[2, 1, 0] = -1

array([[1, 2, 3], >>> a[0, 2, 1]
[4, 5, 6]]) -1
Warning: ndarray.reshape may return a view (cf help(np.reshape))), or copy

Resizing
Tip: Size of an array can be changed with ndarray.resize:

>>> b[0, 0] = 99 >>> a = np.arange(4)
>>> a >>> a.resize((8,))
array([[99, 2, 3], >>> a
[ 4, 5, 6]]) array([0, 1, 2, 3, 0, 0, 0, 0])
Beware: reshape may also return a copy!: However, it must not be referred to somewhere else:
>>> b = a Exercise: Sorting

>>> a.resize((4,))
Try both in-place and out-of-place sorting.
File "<stdin>", line 1, in <module> Try creating arrays with different dtypes and sorting them.
ValueError: cannot resize an array that has been referenced or is Use all or array_equal to check the results.
referencing another array in this way. Use the resize function Look at np.random.shuffle for a way to create sortable input quicker.
Traceback (most recent call last): Combine ravel, sort and reshape.
File "<stdin>", line 1, in <module> Look at the axis keyword for sort and rewrite the previous exercise.
ValueError: cannot resize an array that has been referenced or is
Exercise: Shape manipulations Summary
Look at the docstring for reshape, especially the notes section which has some more information What do you need to know to get started?
about copies and views.
Use flatten as an alternative to ravel. What is the difference? (Hint: check which one returns a Know how to create arrays : array, arange, ones, zeros.
view and which a copy) Know the shape of the array with array.shape, then use slicing to obtain different views of the array:
Experiment with transpose for dimension shuffling. array[::2], etc. Adjust the shape of the array using reshape or flatten it with ravel.
Obtain a subset of the elements of an array and/or modify their values with masks
Sorting data >>> a[a < 0] = 0
Know miscellaneous operations on arrays, such as finding the mean or max (array.max(),
Sorting along an axis:
array.mean()). No need to retain everything, but have the reflex to search in the documentation (online
>>> a = np.array([[4, 3, 5], [1, 2, 1]]) docs, help(), lookfor())!!
>>> b = np.sort(a, axis=1)
>>> b
For advanced use: master the indexing with arrays of integers, as well as broadcasting. Know more Numpy
array([[3, 4, 5], functions to handle various array operations.
[1, 1, 2]])
Quick read
Note: Sorts each row separately!
If you want to do a first quick pass through the Scipy lectures to learn the ecosystem, you can directly skip
to the next chapter: Matplotlib: plotting (page 81).
In-place sort: The remainder of this chapter is not necessary to follow the rest of the intro part. But be sure to come back
>>> a.sort(axis=1) and finish this chapter, as well as to do some more exercices (page 77).
>>> a
array([[3, 4, 5],
[1, 1, 2]])
1.3.3 More elaborate arrays
Sorting with fancy indexing:
>>> a = np.array([4, 3, 1, 2])
>>> j = np.argsort(a) Section contents
>>> j More data types (page 68)
array([2, 3, 1, 0])
Structured data types (page 70)
>>> a[j]
array([1, 2, 3, 4])
maskedarray: dealing with (propagation of) missing data (page 71)
Finding minima and maxima:

>>> a = np.array([4, 3, 1, 2]) More data types
>>> j_max = np.argmax(a)
>>> j_min = np.argmin(a) Casting
>>> j_max, j_min
(0, 2)
Bigger type wins in mixed-type operations:
>>> np.array([1, 2, 3]) + 1.5
array([ 2.5, 3.5, 4.5])
Assignment never changes the type!
>>> a = np.array([1, 2, 3]) float16 16 bits

>>> a.dtype float32 32 bits
dtype('int64') float64 64 bits (same as float)
>>> a[0] = 1.9 # <-- float is truncated to integer float96 96 bits, platform-dependent (same as np.longdouble)
>>> a float128 128 bits, platform-dependent (same as np.longdouble)
array([1, 2, 3])
>>> np.finfo(np.float32).eps
1.1920929e-07
Forced casts:
>>> np.finfo(np.float64).eps
>>> a = np.array([1.7, 1.2, 1.6]) 2.2204460492503131e-16
>>> b = a.astype(int) # <-- truncates to integer
>>> b >>> np.float32(1e-8) + np.float32(1) == 1
array([1, 1, 1]) True
>>> np.float64(1e-8) + np.float64(1) == 1
Rounding: False
>>> a = np.array([1.2, 1.5, 1.6, 2.5, 3.5, 4.5]) Complex floating-point numbers:
>>> b = np.around(a)
>>> b # still floating-point complex64 two 32-bit floats
array([ 1., 2., 2., 2., 4., 4.]) complex128 two 64-bit floats
>>> c = np.around(a).astype(int) complex192 two 96-bit floats, platform-dependent
>>> c complex256 two 128-bit floats, platform-dependent
array([1, 2, 2, 2, 4, 4])
Smaller data types
Different data type sizes If you dont know you need special data types, then you probably dont.
Comparison on using float32 instead of float64:
Half the size in memory and on disk
Integers (signed):
Half the memory bandwidth required (may be a bit faster in some operations)
int8 8 bits
In [1]: a = np.zeros((1e6,), dtype=np.float64)
int16 16 bits
int32 32 bits (same as int on 32-bit platform) In [2]: b = np.zeros((1e6,), dtype=np.float32)
int64 64 bits (same as int on 64-bit platform)
In [3]: %timeit a*a
>>> np.array([1], dtype=int).dtype
1000 loops, best of 3: 1.78 ms per loop
dtype('int64')
>>> np.iinfo(np.int32).max, 2**31 - 1
In [4]: %timeit b*b
(2147483647, 2147483647)
Unsigned integers: But: bigger rounding errors sometimes in surprising places (i.e., dont use them unless you really
need them)
uint8 8 bits
uint16 16 bits
uint32 32 bits
uint64 64 bits Structured data types
>>> np.iinfo(np.uint32).max, 2**32 - 1
(4294967295, 4294967295) sensor_code (4-character string)
position (float)
value (float)
Long integers >>> samples = np.zeros((6,), dtype=[('sensor_code', 'S4'),
Python 2 has a specific type for long integers, that cannot overflow, represented with an L at the end. In ... ('position', float), ('value', float)])
Python 3, all integers are long, and thus cannot overflow. >>> samples.ndim
1
>>> np.iinfo(np.int64).max, 2**63 - 1 >>> samples.shape
(9223372036854775807, 9223372036854775807L) (6,)
>>> samples.dtype.names
('sensor_code', 'position', 'value')
Floating-point numbers:
>>> samples[:] = [('ALFA', 1, 0.37), ('BETA', 1, 0.11), ('TAU', 1, 0.13),
... ('ALFA', 1.5, 0.37), ('ALFA', 3, 0.11), ('TAU', 1.2, 0.13)]
>>> samples
array([('ALFA', 1.0, 0.37), ('BETA', 1.0, 0.11), ('TAU', 1.0, 0.13),
('ALFA', 1.5, 0.37), ('ALFA', 3.0, 0.11), ('TAU', 1.2, 0.13)],

dtype=[('sensor_code', 'S4'), ('position', '<f8'), ('value', '<f8')]) Good practices
Explicit variable names (no need of a comment to explain what is in the variable)
Field access works by indexing with field names:
Style: spaces after commas, around =, etc.
>>> samples['sensor_code'] A certain number of rules for writing beautiful code (and, more importantly, using the same conven-
array(['ALFA', 'BETA', 'TAU', 'ALFA', 'ALFA', 'TAU'], tions as everybody else!) are given in the Style Guide for Python Code and the Docstring Conventions
dtype='|S4') page (to manage help strings).
>>> samples['value'] Except some rare cases, variable names and comments in English.
array([ 0.37, 0.11, 0.13, 0.37, 0.11, 0.13])
>>> samples[0]
('ALFA', 1.0, 0.37)
1.3.4 Advanced operations
>>> samples[0]['sensor_code'] = 'TAU'
>>> samples[0]
('TAU', 1.0, 0.37)
Section contents
Multiple fields at once:
Polynomials (page 72)
>>> samples[['position', 'value']] Loading data files (page 74)
array([(1.0, 0.37), (1.0, 0.11), (1.0, 0.13), (1.5, 0.37), (3.0, 0.11),
(1.2, 0.13)],
dtype=[('position', '<f8'), ('value', '<f8')])
Polynomials
Fancy indexing works, as usual:
Numpy also contains polynomials in different bases:
>>> samples[samples['sensor_code'] == 'ALFA']
array([('ALFA', 1.5, 0.37), ('ALFA', 3.0, 0.11)], For example, 32 + 2 1:
dtype=[('sensor_code', 'S4'), ('position', '<f8'), ('value', '<f8')])
>>> p = np.poly1d([3, 2, -1])
>>> p(0)
Note: There are a bunch of other syntaxes for constructing structured arrays, see here and here. -1
>>> p.roots
array([-1. , 0.33333333])
>>> p.order
maskedarray: dealing with (propagation of) missing data
2
For floats one could use NaNs, but masks work for all types: >>> x = np.linspace(0, 1, 20)
>>> x = np.ma.array([1, 2, 3, 4], mask=[0, 1, 0, 1]) >>> y = np.cos(x) + 0.3*np.random.rand(20)
>>> x >>> p = np.poly1d(np.polyfit(x, y, 3))
masked_array(data = [1 -- 3 --],
mask = [False True False True], >>> t = np.linspace(0, 1, 200)
fill_value = 999999) >>> plt.plot(x, y, 'o', t, p(t), '-')
[<matplotlib.lines.Line2D object at ...>, <matplotlib.lines.Line2D object at ...>]
>>> y = np.ma.array([1, 2, 3, 4], mask=[0, 1, 1, 1])

>>> x + y
masked_array(data = [2 -- -- --],
mask = [False True True True],
fill_value = 999999)
Masking versions of common functions:

>>> np.ma.sqrt([1, -1, 2, -2])
masked_array(data = [1.0 -- 1.41421356237... --],
mask = [False True False True],
fill_value = 1e+20)
Note: There are other useful array siblings (page 181)
While it is off topic in a chapter on numpy, lets take a moment to recall good coding practice, which really do pay
off in the long run:
1.3 1.3
1.2 1.2
1.1 1.1
1.0
1.0
0.9
0.9
0.8
0.8 0.7
0.7 0.6
0.6 0.5
0.0 0.2 0.4 0.6 0.8 1.0 1.0 0.5 0.0 0.5 1.0
See http://docs.scipy.org/doc/numpy/reference/routines.polynomials.poly1d.html for more. The Chebyshev polynomials have some advantages in interpolation.
More polynomials (with more bases) Loading data files
Numpy also has a more sophisticated polynomial interface, which supports e.g. the Chebyshev basis. Text files
32 + 2 1:
Example: populations.txt:
>>> p = np.polynomial.Polynomial([-1, 2, 3]) # coefs in different order!
>>> p(0) # year hare lynx carrot
-1.0 1900 30e3 4e3 48300
>>> p.roots() 1901 47.2e3 6.1e3 48200
array([-1. , 0.33333333]) 1902 70.2e3 9.8e3 41500
>>> p.degree() # In general polynomials do not always expose 'order' 1903 77.4e3 35.2e3 38200
2
Example using polynomials in Chebyshev basis, for polynomials in range [-1, 1]: >>> data = np.loadtxt('data/populations.txt')
>>> data
>>> x = np.linspace(-1, 1, 2000) array([[ 1900., 30000., 4000., 48300.],
>>> y = np.cos(x) + 0.3*np.random.rand(2000) [ 1901., 47200., 6100., 48200.],
>>> p = np.polynomial.Chebyshev.fit(x, y, 90) [ 1902., 70200., 9800., 41500.],
...
>>> t = np.linspace(-1, 1, 200)
>>> plt.plot(x, y, 'r.') >>> np.savetxt('pop2.txt', data)
[<matplotlib.lines.Line2D object at ...>] >>> data2 = np.loadtxt('pop2.txt')
>>> plt.plot(t, p(t), 'k-', lw=3)
Note: If you have a complicated text file, what you can try are:
np.genfromtxt
Using Pythons I/O functions and e.g. regexps for parsing (Python is quite well suited for this)
Reminder: Navigating the filesystem with IPython
In [1]: pwd # show current directory

'/home/user/stuff/2011-numpy-tutorial'
In [2]: cd ex
0
'/home/user/stuff/2011-numpy-tutorial/ex'
In [3]: ls
populations.txt species.txt
5
10
Images
15
Using Matplotlib:
>>> img = plt.imread('data/elephant.png')
20
>>> img.shape, img.dtype
((200, 300, 3), dtype('float32'))
>>> plt.imshow(img)
25
<matplotlib.image.AxesImage object at ...>
>>> plt.savefig('plot.png') 30
>>> plt.imsave('red_elephant', img[:,:,0], cmap=plt.cm.gray)
0 10 20 30 40
This saved only one channel (of RGB):
>>> plt.imshow(plt.imread('red_elephant.png'))
Other libraries:
Numpys own format
>>> from scipy.misc import imsave
>>> imsave('tiny_elephant.png', img[::6,::6])
>>> plt.imshow(plt.imread('tiny_elephant.png'), interpolation='nearest') Numpy has its own binary format, not portable but with efficient I/O:
>>> data = np.ones((3, 3))
>>> np.save('pop.npy', data)
>>> data3 = np.load('pop.npy')
Well-known (& more obscure) file formats
HDF5: h5py, PyTables

NetCDF: scipy.io.netcdf_file, netcdf4-python, ...
Matlab: scipy.io.loadmat, scipy.io.savemat
MatrixMarket: scipy.io.mmread, scipy.io.mmwrite
IDL: scipy.io.readsav
... if somebody uses it, theres probably also a Python library for it.
Exercise: Text data files

Write a Python script that loads data from populations.txt:: and drop the last column and the first 5
rows. Save the smaller dataset to pop2.txt.
Numpy internals >>> import pylab as plt

>>> lena = misc.lena()
If you are interested in the Numpy internals, there is a good discussion in Advanced Numpy (page 155).
>>> plt.imshow(lena)
<matplotlib.image.AxesImage object at 0x...>
Lena is then displayed in false colors. A colormap must be specified for her to be displayed in grey.
1.3.5 Some exercises
>>> plt.imshow(lena, cmap=plt.cm.gray)
Array manipulations <matplotlib.image.AxesImage object at 0x...>
Create an array of the image with a narrower centering [for example,] remove 30 pixels from all the
1. Form the 2-D array (without typing it in explicitly):
borders of the image. To check the result, display this new array with imshow.
[[1, 6, 11],
[2, 7, 12], >>> crop_lena = lena[30:-30,30:-30]
[3, 8, 13],
[4, 9, 14], We will now frame Lenas face with a black locket. For this, we need to create a mask corresponding
[5, 10, 15]] to the pixels we want to be black. The mask is defined by this condition (y-256)**2 +
(x-256)**2
and generate a new array containing its 2nd and 4th rows.
>>> y, x = np.ogrid[0:512,0:512] # x and y indices of pixels
2. Divide each column of the array: >>> y.shape, x.shape
((512, 1), (1, 512))
>>> import numpy as np >>> centerx, centery = (256, 256) # center of the image
>>> a = np.arange(25).reshape(5, 5) >>> mask = ((y - centery)**2 + (x - centerx)**2) > 230**2 # circle
elementwise with the array b = np.array([1., 5, 10, 15, 20]). (Hint: np.newaxis). then we assign the value 0 to the pixels of the image corresponding to the mask. The syntax is
3. Harder one: Generate a 10 x 3 array of random numbers (in range [0,1]). For each row, pick the number extremely simple and intuitive:
closest to 0.5. >>> lena[mask] = 0
>>> plt.imshow(lena)
Use abs and argsort to find the column j closest for each row.
Use fancy indexing to extract the numbers. (Hint: a[i,j] the array i must contain the row numbers
corresponding to stuff in j.) Follow-up: copy all instructions of this exercise in a script called lena_locket.py then execute
this script in IPython with %run lena_locket.py.
Picture manipulation: Framing Lena Change the circle to an ellipsoid.
Lets do some manipulations on numpy arrays by starting with the famous image of Lena Data statistics
(http://www.cs.cmu.edu/~chuck/lennapg/). scipy provides a 2D array of this image with the scipy.lena
function: The data in populations.txt describes the populations of hares and lynxes (and carrots) in northern Canada
>>> from scipy import misc during 20 years:
>>> data = np.loadtxt('data/populations.txt')
>>> year, hares, lynxes, carrots = data.T # trick: columns to variables
Note: In older versions of scipy, you will find lena under scipy.lena()
Here are a few images we will be able to obtain with our manipulations: use different colormaps, crop the image, >>> import matplotlib.pyplot as plt
change some parts of the image. >>> plt.axes([0.2, 0.1, 0.5, 0.8])
<matplotlib.axes...Axes object at ...>
>>> plt.plot(year, hares, year, lynxes, year, carrots)
[<matplotlib.lines.Line2D object at ...>, ...]
>>> plt.legend(('Hare', 'Lynx', 'Carrot'), loc=(1.05, 0.5))
<matplotlib.legend.Legend object at ...>
Lets use the imshow function of pylab to display the image.
def f(a, b, c):

return some_result
80000
Solution: Python source file
70000
Mandelbrot set
60000
Hare
50000 Lynx 1.5
Carrot
40000 1.0
30000
0.5
20000
10000 0.0
0 0.5
1900 1905 1910 1915 1920
Computes and print, based on the data in populations.txt... 1.0
1. The mean and std of the populations of each species for the years in the period.
2. Which year each species had the largest population.
1.5
3. Which species has the largest population for each year.
np.array([H, L, C]))
(Hint: argsort & fancy indexing of 2.0 1.5 1.0 0.5 0.0 0.5 1.0
4. Which years any of the populations is above 50000. (Hint: comparisons and np.any)
Write a script that computes the Mandelbrot fractal. The Mandelbrot iteration:
5. The top 2 years for each species when they had the lowest populations. (Hint: argsort, fancy indexing)
N_max = 50
6. Compare (plot) the change in hare population (see help(np.gradient)) and the number of lynxes. some_threshold = 50
Check correlation (see help(np.corrcoef)).
... all without for-loops. c = x + 1j*y
Solution: Python source file for j in xrange(N_max):

z = z**2 + c
Crude integral approximations Point (x, y) belongs to the Mandelbrot set if || < some_threshold.
Do this computation by:
Write a function f(a, b, c) that returns . Form a 24x12x6 array containing its values in parameter
ranges [0,1] x [0,1] x [0,1]. 1. Construct a grid of c = x + 1j*y values in range [-2, 1] x [-1.5, 1.5]
Approximate the 3-d integral 2. Do the iteration

1 1 1
3. Form the 2-d boolean mask indicating which points are in the set
( )
0 0 0 4. Save the result to an image with:
over this volume with the mean. The exact result is: ln 2 1
2 0.1931 . . . what is your relative error? >>> import matplotlib.pyplot as plt
(Hints: use elementwise operations and broadcasting. You can make np.ogrid give a number of points in given >>> plt.imshow(mask.T, extent=[-2, 1, -1.5, 1.5])
range with np.ogrid[0:1:20j].) <matplotlib.image.AxesImage object at ...>
>>> plt.gray()
Reminder Python functions: >>> plt.savefig('mandelbrot.png')
Solution: Python source file

Chapters contents
Markov chain Introduction (page 82)
IPython and the pylab mode (page 83)
pylab (page 83)
Simple plot (page 83)
Plotting with default settings (page 84)
Instantiating defaults (page 84)
Changing colors and line widths (page 85)
Setting limits (page 86)
Setting ticks (page 86)
Setting tick labels (page 87)
Moving spines (page 87)
Adding a legend (page 88)
Annotate some points (page 88)
Devil is in the details (page 89)
Figures, Subplots, Axes and Ticks (page 89)
Figures (page 90)
Subplots (page 90)
Axes (page 90)
Ticks (page 91)
Other Types of Plots: examples and exercises (page 91)
Markov chain transition matrix P, and probability distribution on the states p: Regular Plots (page 93)
1. 0 <= P[i,j] <= 1: probability to go from state i to state j Scatter Plots (page 93)
Bar Plots (page 94)
2. Transition rule: = Contour Plots (page 94)
3. all(sum(P, axis=1) == 1), p.sum() == 1: normalization Imshow (page 95)
Pie Charts (page 95)
Write a script that works with 5 states, and: Quiver Plots (page 96)
Constructs a random matrix, and normalizes each row so that it is a transition matrix. Grids (page 96)
Multi Plots (page 96)
Starts from a random (normalized) probability distribution p and takes 50 steps => p_50 Polar Axis (page 97)
Computes the stationary distribution: the eigenvector of P.T with eigenvalue 1 (numerically: closest to 1) 3D Plots (page 97)
=> p_stationary Text (page 98)
Beyond this tutorial (page 98)
Remember to normalize the eigenvector I didnt... Tutorials (page 98)
Checks if p_50 and p_stationary are equal to tolerance 1e-5 Matplotlib documentation (page 99)
Code documentation (page 99)
Toolbox: np.random.rand, .dot(), np.linalg.eig, reductions, abs(), argmin, comparisons, all, Galleries (page 100)
np.linalg.norm, etc. Mailing lists (page 100)
Solution: Python source file Quick references (page 100)
Line properties (page 101)
Line styles (page 102)
Markers (page 102)
1.4 Matplotlib: plotting Colormaps (page 102)
Thanks
1.4.1 Introduction
Many thanks to Bill Wing and Christoph Deil for review and corrections.
Tip: Matplotlib is probably the single most used Python package for 2D-graphics. It provides both a very quick
Authors: Nicolas Rougier, Mike Mller, Gal Varoquaux way to visualize data from Python and publication-quality figures in many formats. We are going to explore
matplotlib in interactive mode covering most common cases.
1.4. Matplotlib: plotting 81 1.4. Matplotlib: plotting 82

IPython and the pylab mode Plotting with default settings
Tip: IPython is an enhanced interactive Python shell that has lots of interesting features including named in-
puts and outputs, access to shell commands, improved debugging and many more. It is central to the scientific-
computing workflow in Python for its use in combination with Matplotlib:
We start IPython with the command line argument -pylab (--pylab since IPython version 0.12), for interactive
matplotlib sessions with Matlab/Mathematica-like functionality.
pylab
Tip: pylab provides a procedural interface to the matplotlib object-oriented plotting library. It is modeled
closely after Matlab. Therefore, the majority of plotting commands in pylab have Matlab analogs with similar Hint: Documentation
arguments. Important commands are explained with interactive examples. plot tutorial
plot() command
1.4.2 Simple plot

Tip: Matplotlib comes with a set of default settings that allow customizing all kinds of properties. You can
control the defaults of almost every property in matplotlib: figure size and dpi, line width, color and style, axes,
Tip: In this section, we want to draw the cosine and sine functions on the same plot. Starting from the default axis and grid properties, text and font properties and so on.
settings, well enrich the figure step by step to make it nicer.
First step is to get the data for the sine and cosine functions: import pylab as pl
import numpy as np
import numpy as np X = np.linspace(-np.pi, np.pi, 256, endpoint=True)
C, S = np.cos(X), np.sin(X)
X = np.linspace(-np.pi, np.pi, 256, endpoint=True)
C, S = np.cos(X), np.sin(X) pl.plot(X, C)
pl.plot(X, S)
X is now a numpy array with 256 values ranging from - to + (included). C is the cosine (256 values) and S is
the sine (256 values). pl.show()
To run the example, you can type them in an IPython interactive session:
$ ipython --pylab Instantiating defaults
This brings us to the IPython prompt:

IPython 0.13 -- An enhanced Interactive Python.
? -> Introduction to IPython's features.
%magic -> Information about IPython's 'magic' % functions.
help -> Python's own help system.
object? -> Details about 'object'. ?object also works, ?? prints more.
Welcome to pylab, a matplotlib-based Python environment.

For more information, type 'help(pylab)'.
Tip: You can also download each of the examples and run it using regular python, but you will loose interactive
data manipulation:
Hint: Documentation
$ python exercice_1.py Customizing matplotlib
You can get source for each step by clicking on the corresponding figure.
In the script below, weve instantiated (and commented) all the figure settings that influence the appearance of the
plot.
Tip: The settings have been explicitly set to their default values, but now you can interactively play with the
values to explore their affect (see Line properties (page 101) and Line styles (page 102) below).

import pylab as pl pl.plot(X, S, color="red", linewidth=2.5, linestyle="-")

import numpy as np ...
# Create a figure of size 8x6 inches, 80 dots per inch

pl.figure(figsize=(8, 6), dpi=80) Setting limits
# Create a new subplot from a grid of 1x1
pl.subplot(1, 1, 1)
X = np.linspace(-np.pi, np.pi, 256, endpoint=True)

C, S = np.cos(X), np.sin(X)
# Plot cosine with a blue continuous line of width 1 (pixels)

pl.plot(X, C, color="blue", linewidth=1.0, linestyle="-")
# Plot sine with a green continuous line of width 1 (pixels)

pl.plot(X, S, color="green", linewidth=1.0, linestyle="-")
Hint: Documentation
# Set x limits xlim() command
pl.xlim(-4.0, 4.0)
ylim() command
# Set x ticks
pl.xticks(np.linspace(-4, 4, 9, endpoint=True))
Tip: Current limits of the figure are a bit too tight and we want to make some space in order to clearly see all
# Set y limits data points.
pl.ylim(-1.0, 1.0)
...
# Set y ticks pl.xlim(X.min() * 1.1, X.max() * 1.1)
pl.yticks(np.linspace(-1, 1, 5, endpoint=True)) pl.ylim(C.min() * 1.1, C.max() * 1.1)
...
# Save figure using 72 dots per inch
# savefig("exercice_2.png", dpi=72)
Setting ticks
# Show result on screen
pl.show()
Changing colors and line widths
Hint: Documentation
xticks() command
yticks() command
Hint: Documentation Tick container

Controlling line properties Tick locating and formatting
Line API
Tip: Current ticks are not ideal because they do not show the interesting values (+/-,+/-/2) for sine and cosine.
Well change them such that they show only these values.
Tip: First step, we want to have the cosine in blue and the sine in red and a slighty thicker line for both of them.
Well also slightly alter the figure size to make it more horizontal. ...
pl.xticks([-np.pi, -np.pi/2, 0, np.pi/2, np.pi])
... pl.yticks([-1, 0, +1])
pl.figure(figsize=(10, 6), dpi=80) ...
pl.plot(X, C, color="blue", linewidth=2.5, linestyle="-")

Setting tick labels ...

ax = pl.gca() # gca stands for 'get current axis'
ax.spines['right'].set_color('none')
ax.spines['top'].set_color('none')
ax.xaxis.set_ticks_position('bottom')
ax.spines['bottom'].set_position(('data',0))
ax.yaxis.set_ticks_position('left')
ax.spines['left'].set_position(('data',0))
...
Adding a legend
Hint: Documentation
Working with text
xticks() command
yticks() command
set_xticklabels()
set_yticklabels()
Tip: Ticks are now properly placed but their label is not very explicit. We could guess that 3.142 is but it would Hint: Documentation
be better to make it explicit. When we set tick values, we can also provide a corresponding label in the second
Legend guide
argument list. Note that well use latex to allow for nice rendering of the label.
legend() command
...
pl.xticks([-np.pi, -np.pi/2, 0, np.pi/2, np.pi], Legend API
[r'$-\pi$', r'$-\pi/2$', r'$0$', r'$+\pi/2$', r'$+\pi$'])
pl.yticks([-1, 0, +1], Tip: Lets add a legend in the upper left corner. This only requires adding the keyword argument label (that will
[r'$-1$', r'$0$', r'$+1$']) be used in the legend box) to the plot commands.
...
...
pl.plot(X, C, color="blue", linewidth=2.5, linestyle="-", label="cosine")
pl.plot(X, S, color="red", linewidth=2.5, linestyle="-", label="sine")
Moving spines
pl.legend(loc='upper left')
...
Annotate some points
Hint: Documentation
Spines
Axis container
Transformations tutorial
Hint: Documentation
Annotating axis
Tip: Spines are the lines connecting the axis tick marks and noting the boundaries of the data area. They can be
placed at arbitrary positions and until now, they were on the border of the axis. Well change that since we want annotate() command
to have them in the middle. Since there are four of them (top/bottom/left/right), well discard the top and right by
setting their color to none and well move the bottom and left ones to coordinate 0 in data space coordinates.
Tip: Lets annotate some interesting points using the annotate command. We chose the 2/3 value and we want

to annotate both the sine and the cosine. Well first draw a marker on the curve as well as a straight dotted line. current axes and gca in turn calls gcf() to get the current figure. If there is none it calls figure() to make
Then, well use the annotate command to display some text with an arrow. one, strictly speaking, to make a subplot(111). Lets look at the details.
...
t = 2 * np.pi / 3 Figures
pl.plot([t, t], [0, np.cos(t)], color='blue', linewidth=2.5, linestyle="--")
pl.scatter([t, ], [np.cos(t), ], 50, color='blue')
Tip: A figure is the windows in the GUI that has Figure # as title. Figures are numbered starting from 1 as
opposed to the normal Python way starting from 0. This is clearly MATLAB-style. There are several parameters
pl.annotate(r'$sin(\frac{2\pi}{3})=\frac{\sqrt{3}}{2}$',
that determine what the figure looks like:
xy=(t, np.sin(t)), xycoords='data',
xytext=(+10, +30), textcoords='offset points', fontsize=16,
arrowprops=dict(arrowstyle="->", connectionstyle="arc3,rad=.2")) Argument Default Description
num 1 number of figure
pl.plot([t, t],[0, np.sin(t)], color='red', linewidth=2.5, linestyle="--") figsize figure.figsize figure size in in inches (width, height)
pl.scatter([t, ],[np.sin(t), ], 50, color='red') dpi figure.dpi resolution in dots per inch
facecolor figure.facecolor color of the drawing background
pl.annotate(r'$cos(\frac{2\pi}{3})=-\frac{1}{2}$', edgecolor figure.edgecolor color of edge around the drawing background
xy=(t, np.cos(t)), xycoords='data',
frameon True draw figure frame or not
xytext=(-90, -50), textcoords='offset points', fontsize=16,
arrowprops=dict(arrowstyle="->", connectionstyle="arc3,rad=.2")) Tip: The defaults can be specified in the resource file and will be used most of the time. Only the number of the
... figure is frequently changed.
As with other objects, you can set figure properties also setp or with the set_something methods.
Devil is in the details When you work with the GUI you can close a figure by clicking on the x in the upper right corner. But you
can close a figure programmatically by calling close. Depending on the argument it closes (1) the current figure
(no argument), (2) a specific figure (figure number or figure instance as argument), or (3) all figures ("all" as
argument).
pl.close(1) # Closes figure 1
Subplots
Tip: With subplot you can arrange plots in a regular grid. You need to specify the number of rows and columns
Hint: Documentation and the number of the plot. Note that the gridspec command is a more powerful alternative.
Artists
BBox
Tip: The tick labels are now hardly visible because of the blue and red lines. We can make them bigger and we
can also adjust their properties such that theyll be rendered on a semi-transparent white background. This will
allow us to see both the data and the labels.
...
for label in ax.get_xticklabels() + ax.get_yticklabels():
label.set_fontsize(16) Axes
label.set_bbox(dict(facecolor='white', edgecolor='None', alpha=0.65))
... Axes are very similar to subplots but allow placement of plots at any location in the fig-
ure. So if we want to put a smaller plot inside a bigger one we do so with axes.
1.4.3 Figures, Subplots, Axes and Ticks
A figure in matplotlib means the whole window in the user interface. Within this figure there can be subplots.
Tip: So far we have used implicit figure and axes creation. This is handy for fast plots. We can have more control
over the display using figure, subplot, and axes explicitly. While subplot positions the plots in a regular grid, axes
allows free placement within the figure. Both can be useful depending on your intention. Weve already worked
with figures and subplots without explicitly calling them. When we call plot, matplotlib calls gca() to get the

(page 93) (page 94)

Ticks
Well formatted ticks are an important part of publishing-ready figures. Matplotlib provides a totally configurable
system for ticks. There are tick locators to specify where ticks should appear and tick formatters to give ticks the
appearance you want. Major and minor ticks can be located and formatted independently from each other. Per
default minor ticks are not shown, i.e. there is only an empty list for them because it is as NullLocator (see
below).
Tick Locators
(page 94) (page 95)
Tick locators control the positions of the ticks. They are set as follows:
ax = pl.gca()
ax.xaxis.set_major_locator(eval(locator))
(page 96) (page 95)
There are several locators for different kind of requirements:

All of these locators derive from the base class matplotlib.ticker.Locator. You can make your own
locator deriving from it. Handling dates as ticks can be especially tricky. Therefore, matplotlib provides special
locators in matplotlib.dates.
1.4.4 Other Types of Plots: examples and exercises (page 96) (page 96)
(page 97) (page 97)

(page 93) (page 98)

Regular Plots Bar Plots
Hint: You need to use the fill_between command. Hint: You need to take care of text alignment.
Starting from the code below, try to reproduce the graphic on the right taking care of filled areas: Starting from the code below, try to reproduce the graphic on the right by adding labels for red bars.
n = 256 n = 12
X = np.linspace(-np.pi, np.pi, n, endpoint=True) X = np.arange(n)
Y = np.sin(2 * X) Y1 = (1 - X / float(n)) * np.random.uniform(0.5, 1.0, n)
Y2 = (1 - X / float(n)) * np.random.uniform(0.5, 1.0, n)
pl.plot(X, Y + 1, color='blue', alpha=1.00)
pl.plot(X, Y - 1, color='blue', alpha=1.00) pl.bar(X, +Y1, facecolor='#9999ff', edgecolor='white')
pl.bar(X, -Y2, facecolor='#ff9999', edgecolor='white')
Click on the figure for solution.
for x, y in zip(X, Y1):
pl.text(x + 0.4, y + 0.05, '%.2f' % y, ha='center', va='bottom')
Scatter Plots
pl.ylim(-1.25, +1.25)
Click on figure for solution.
Contour Plots
Hint: Color is given by angle of (X,Y).
Starting from the code below, try to reproduce the graphic on the right taking care of marker size, color and
transparency.
n = 1024
X = np.random.normal(0,1,n) Hint: You need to use the clabel command.
Y = np.random.normal(0,1,n)
Starting from the code below, try to reproduce the graphic on the right taking care of the colormap (see Colormaps
pl.scatter(X,Y) (page 102) below).
Click on figure for solution. def f(x, y):
return (1 - x / 2 + x ** 5 + y ** 3) * np.exp(-x ** 2 -y ** 2)
n = 256
x = np.linspace(-3, 3, n)
y = np.linspace(-3, 3, n)
X, Y = np.meshgrid(x, y)

pl.contourf(X, Y, f(X, Y), 8, alpha=.75, cmap='jet') Quiver Plots

C = pl.contour(X, Y, f(X, Y), 8, colors='black', linewidth=.5)
Imshow
Hint: You need to draw arrows twice.
Starting from the code above, try to reproduce the graphic on the right taking care of colors and orientations.
n = 8
X, Y = np.mgrid[0:n, 0:n]
Hint: You need to take care of the origin of the image in the imshow command and use a colorbar pl.quiver(X, Y)

Starting from the code below, try to reproduce the graphic on the right taking care of colormap, image interpolation
and origin.
def f(x, y): Grids
return (1 - x / 2 + x ** 5 + y ** 3) * np.exp(-x ** 2 - y ** 2)
n = 10
x = np.linspace(-3, 3, 4 * n)
y = np.linspace(-3, 3, 3 * n)
X, Y = np.meshgrid(x, y)
pl.imshow(f(X, Y))
Click on the figure for the solution.
Pie Charts Starting from the code below, try to reproduce the graphic on the right
taking care of line styles.
axes = pl.gca()
axes.set_xlim(0, 4)
axes.set_ylim(0, 3)
axes.set_xticklabels([])
axes.set_yticklabels([])
Multi Plots
Hint: You need to modify Z.
Starting from the code below, try to reproduce the graphic on the right taking care of colors and slices size.
Z = np.random.uniform(0, 1, 20)
pl.pie(Z)
Click on the figure for the solution.

from mpl_toolkits.mplot3d import Axes3D

Hint: You can use several subplots with different partition.
fig = pl.figure()
Starting from the code below, try to reproduce the graphic on the right. ax = Axes3D(fig)
X = np.arange(-4, 4, 0.25)
pl.subplot(2, 2, 1) Y = np.arange(-4, 4, 0.25)
pl.subplot(2, 2, 3) X, Y = np.meshgrid(X, Y)
pl.subplot(2, 2, 4) R = np.sqrt(X**2 + Y**2)
Z = np.sin(R)
ax.plot_surface(X, Y, Z, rstride=1, cstride=1, cmap='hot')
Polar Axis Click on figure for solution.

See also:
3D plotting with Mayavi (page 357)
Text
Hint: You only need to modify the axes line
Starting from the code below, try to reproduce the graphic on the right.
pl.axes([0, 0, 1, 1])
N = 20
theta = np.arange(0., 2 * np.pi, 2 * np.pi / N) Hint: Have a look at the matplotlib logo.
radii = 10 * np.random.rand(N)
width = np.pi / 4 * np.random.rand(N) Try to do the same from scratch !
bars = pl.bar(theta, radii, width=width, bottom=0.0)
for r, bar in zip(radii, bars):
bar.set_facecolor(cm.jet(r / 10.))
bar.set_alpha(0.5)
Quick read
Click on figure for solution. If you want to do a first quick pass through the Scipy lectures to learn the ecosystem, you can directly skip
to the next chapter: Scipy : high-level scientific computing (page 103).
3D Plots The remainder of this chapter is not necessary to follow the rest of the intro part. But be sure to come back
and finish this chapter later.
1.4.5 Beyond this tutorial
Matplotlib benefits from extensive documentation as well as a large community of users and developers. Here are
some links of interest:
Tutorials
Pyplot tutorial
Hint: You need to use contourf Introduction
Controlling line properties
Starting from the code below, try to reproduce the graphic on the right. Working with multiple figures and axes

Working with text argument, allowing for multiple *x*, *y* pairs with an
Image tutorial optional format string. For example, each of the following is
Startup commands legal::
Importing image data into Numpy arrays
Plotting numpy arrays as images plot(x, y) # plot x and y using default line style and color
Text tutorial plot(x, y, 'bo') # plot x and y using blue circle markers
Text introduction plot(y) # plot y using x as index array 0..N-1
Basic text commands plot(y, 'r+') # ditto, but with red plusses
Text properties and layout
If *x* and/or *y* is 2-dimensional, then the corresponding columns
Writing mathematical expressions
will be plotted.
Text rendering With LaTeX ...
Annotating text
Artist tutorial
Introduction Galleries
Customizing your objects
Object containers The matplotlib gallery is also incredibly useful when you search how to render a given graphic. Each example
Figure container comes with its source.
Axes container
Axis containers A smaller gallery is also available here.
Tick containers
Path tutorial
Mailing lists
Introduction
Bzier example
Finally, there is a user mailing list where you can ask for help and a developers mailing list that is more technical.
Compound paths
Transforms tutorial
Introduction 1.4.6 Quick references
Data coordinates
Axes coordinates
Here is a set of tables that show main properties and styles.
Blended transformations
Using offset transforms to create a shadow effect
The transformation pipeline
Matplotlib documentation
User guide
FAQ
Installation
Usage
How-To
Troubleshooting
Environment Variables
Screenshots
Code documentation
The code is well documented and you can quickly access a specific command from within a python session:
>>> import pylab as pl
>>> help(pl.plot)
Help on function plot in module matplotlib.pyplot:
plot(*args, **kwargs)
Plot lines and/or markers to the
:class:`~matplotlib.axes.Axes`. *args* is a variable length

Line properties Line styles
Property Description Appearance

alpha (or a) alpha transparency on 0-1 scale
antialiased True or False - use antialised

rendering
color (or c) matplotlib color arg
linestyle (or ls) see Line properties (page 101)

linewidth (or lw) float, the line width in points
solid_capstyle Cap style for solid lines
solid_joinstyle Join style for solid lines
dash_capstyle Cap style for dashes
dash_joinstyle Join style for dashes
marker see Markers (page 102)

Markers
markeredgewidth line width around the marker
(mew) symbol
markeredgecolor edge color if a marker is used
(mec)
markerfacecolor face color if a marker is used
(mfc)
markersize (ms) size of the marker in points
Colormaps
All colormaps can be reversed by appending _r. For instance, gray_r is the reverse of gray.
If you want to know more about colormaps, checks Documenting the matplotlib colormaps.

scipy.cluster Vector quantization / Kmeans

scipy.constants Physical and mathematical constants
scipy.fftpack Fourier transform
scipy.integrate Integration routines
scipy.interpolate Interpolation
scipy.io Data input and output
scipy.linalg Linear algebra routines
scipy.ndimage n-dimensional image package
scipy.odr Orthogonal distance regression
scipy.optimize Optimization
scipy.signal Signal processing
scipy.sparse Sparse matrices
scipy.spatial Spatial data structures and algorithms
scipy.special Any special mathematical functions
scipy.stats Statistics
Tip: They all depend on numpy, but are mostly independent of each other. The standard way of importing
Numpy and these Scipy modules is:
1.5 Scipy : high-level scientific computing
>>> from scipy import stats # same for other sub-modules
Authors: Adrien Chauve, Andre Espaze, Emmanuelle Gouillart, Gal Varoquaux, Ralf Gommers
The main scipy namespace mostly contains functions that are really numpy functions (try scipy.cos is
np.cos). Those are exposed for historical reasons only; theres usually no reason to use import scipy in
Scipy your code.
The scipy package contains various toolboxes dedicated to common issues in scientific computing. Its
different submodules correspond to different applications, such as interpolation, integration, optimization,
image processing, statistics, special functions, etc.
1.5.1 File input/output: scipy.io
scipy can be compared to other standard scientific-computing libraries, such as the GSL (GNU Scientific
Library for C and C++), or Matlabs toolboxes. scipy is the core package for scientific routines in Python;
Loading and saving matlab files:
it is meant to operate efficiently on numpy arrays, so that numpy and scipy work hand in hand.
Before implementing a routine, it is worth checking if the desired data processing is not already imple- >>> from scipy import io as spio
mented in Scipy. As non-professional programmers, scientists often tend to re-invent the wheel, which >>> a = np.ones((3, 3))
leads to buggy, non-optimal, difficult-to-share and unmaintainable code. By contrast, Scipys routines are >>> spio.savemat('file.mat', {'a': a}) # savemat expects a dictionary
optimized and tested, and should therefore be used when possible. >>> data = spio.loadmat('file.mat', struct_as_record=True)
>>> data['a']
array([[ 1., 1., 1.],
[ 1., 1., 1.],
Chapters contents [ 1., 1., 1.]])
File input/output: scipy.io (page 104) Reading images:

Special functions: scipy.special (page 104)
Linear algebra operations: scipy.linalg (page 105) >>> from scipy import misc
>>> misc.imread('fname.png')
Fast Fourier transforms: scipy.fftpack (page 106)
array(...)
Optimization and fit: scipy.optimize (page 110)
>>> # Matplotlib also has a similar function
Statistics and random numbers: scipy.stats (page 114) >>> import matplotlib.pyplot as plt
Interpolation: scipy.interpolate (page 116) >>> plt.imread('fname.png')
Numerical integration: scipy.integrate (page 117) array(...)
Signal processing: scipy.signal (page 119)
Image processing: scipy.ndimage (page 121) See also:
Summary exercises on scientific computing (page 125)
Load text files: numpy.loadtxt()/numpy.savetxt()
Clever loading of text/csv files: numpy.genfromtxt()/numpy.recfromcsv()
Warning: This tutorial is far from an introduction to numerical computing. As enumerating the different
submodules and functions in scipy would be very boring, we concentrate instead on a few examples to give a Fast and efficient, but numpy-specific, binary format: numpy.save()/numpy.load()
general idea of how to use scipy for scientific computing.
1.5.2 Special functions: scipy.special

scipy is composed of task-specific sub-modules:
Special functions are transcendental functions. The docstring of the scipy.special module is well-written,
so we wont list all functions here. Frequently used ones are:
1.5. Scipy : high-level scientific computing 103 1.5. Scipy : high-level scientific computing 104
Bessel function, such as scipy.special.jn() (nth integer order Bessel function) >>> spec
array([ 14.88982544, 0.45294236, 0.29654967])
Elliptic function (scipy.special.ellipj() for the Jacobian elliptic function, ...)
Gamma function: scipy.special.gamma(), also note scipy.special.gammaln() which will The original matrix can be re-composed by matrix multiplication of the outputs of svd with np.dot:
give the log of Gamma to a higher numerical precision.
>>> sarr = np.diag(spec)
Erf, the area under a Gaussian curve: scipy.special.erf() >>> svd_mat = uarr.dot(sarr).dot(vharr)
>>> np.allclose(svd_mat, arr)
True
1.5.3 Linear algebra operations: scipy.linalg
SVD is commonly used in statistics and signal processing. Many other standard decompositions (QR, LU,
The scipy.linalg module provides standard linear algebra operations, relying on an underlying efficient Cholesky, Schur), as well as solvers for linear systems, are available in scipy.linalg.
implementation (BLAS, LAPACK).
The scipy.linalg.det() function computes the determinant of a square matrix: 1.5.4 Fast Fourier transforms: scipy.fftpack
>>> from scipy import linalg
The scipy.fftpack module allows to compute fast Fourier transforms. As an illustration, a (noisy) input
>>> arr = np.array([[1, 2],
... [3, 4]]) signal may look like:
>>> linalg.det(arr) >>> time_step = 0.02
-2.0 >>> period = 5.
>>> arr = np.array([[3, 2], >>> time_vec = np.arange(0, 20, time_step)
... [6, 4]]) >>> sig = np.sin(2 * np.pi / period * time_vec) + \
>>> linalg.det(arr) ... 0.5 * np.random.randn(time_vec.size)
0.0
>>> linalg.det(np.ones((3, 4))) The observer doesnt know the signal frequency, only the sampling time step of the signal sig.
The signal is supposed to come from a real function so the Fourier transform will be sym-
...
metric. The scipy.fftpack.fftfreq() function will generate the sampling frequencies and
ValueError: expected square matrix
Traceback (most recent call last): scipy.fftpack.fft() will compute the fast Fourier transform:
... >>> from scipy import fftpack
ValueError: expected square matrix >>> sample_freq = fftpack.fftfreq(sig.size, d=time_step)
>>> sig_fft = fftpack.fft(sig)
The scipy.linalg.inv() function computes the inverse of a square matrix:
>>> arr = np.array([[1, 2], Because the resulting power is symmetric, only the positive part of the spectrum needs to be used for finding the
... [3, 4]]) frequency:
>>> iarr = linalg.inv(arr) >>> pidxs = np.where(sample_freq > 0)
>>> iarr >>> freqs = sample_freq[pidxs]
array([[-2. , 1. ], >>> power = np.abs(sig_fft)[pidxs]
[ 1.5, -0.5]])
>>> np.allclose(np.dot(arr, iarr), np.eye(2))
True
600
Finally computing the inverse of a singular matrix (its determinant is zero) will raise LinAlgError: Peak frequency
>>> arr = np.array([[3, 2], 500
... [6, 4]])
>>> linalg.inv(arr)
Traceback (most recent call last): 400
...
plower
...LinAlgError: singular matrix
300
...
...LinAlgError: singular matrix 200
More advanced operations are available, for example singular-value decomposition (SVD): 0.050.100.150.200.250.300.350.400.45
100
>>> arr = np.arange(9).reshape((3, 3)) + np.diag([1, 0, 1])
>>> uarr, spec, vharr = linalg.svd(arr)
0
The resulting array spectrum is:
0 5 10 15 20 25
Frequency [Hz]
The signal frequency can be found by:
>>> freq = freqs[power.argmax()] Worked example: Crude periodicity finding

>>> np.allclose(freq, 1./period) # check that correct freq is found
True
Now the high-frequency noise will be removed from the Fourier transformed signal: 80
>>> sig_fft[np.abs(sample_freq) > freq] = 0 hare
The resulting filtered signal can be computed by the scipy.fftpack.ifft() function:
70 lynx
60 carrot
Population number ( 103 )

>>> main_sig = fftpack.ifft(sig_fft)
The result can be viewed with:

>>> import pylab as plt 50
>>> plt.figure()
<matplotlib.figure.Figure object at 0x...>
>>> plt.plot(time_vec, sig) 40
[<matplotlib.lines.Line2D object at 0x...>]
>>> plt.plot(time_vec, main_sig, linewidth=3)
[<matplotlib.lines.Line2D object at 0x...>] 30
>>> plt.xlabel('Time [s]')
<matplotlib.text.Text object at 0x...>
>>> plt.ylabel('Amplitude') 20
<matplotlib.text.Text object at 0x...>
10
3
0
2
1900 1905 1910 1915 1920
Year
1
300
Amplitude
1 250
2
200
Power ( 103 )
3
0 5 10 15 20
Time [s]
150
numpy.fft
Numpy also has an implementation of FFT (numpy.fft). However, in general the scipy one should be 100
preferred, as it uses more efficient underlying implementations.
50
0
0 5 10 15 20
Period
Worked example: Gaussian image blur Exercise: Denoise moon landing image
Convolution:

1 () = ( )0 ( )
1 () = ()
0 ()
50
1. Examine the provided image moonlanding.png, which is heavily contaminated with periodic noise. In
100 this exercise, we aim to clean up the noise using the Fast Fourier Transform.
2. Load the image using pylab.imread().
3. Find and use the 2-D FFT function in scipy.fftpack, and plot the spectrum (Fourier transform
of) the image. Do you have any trouble visualising the spectrum? If so, why?
150 4. The spectrum consists of high and low frequency components. The noise is contained in the high-
frequency part of the spectrum, so set some of those components to zero (use array slicing).
5. Apply the inverse Fourier transform to see the resulting image.
0 50 100 150 200 250 1.5.5 Optimization and fit: scipy.optimize
Optimization is the problem of finding a numerical solution to a minimization or equality.

The scipy.optimize module provides useful algorithms for function minimization (scalar or multi-
dimensional), curve fitting and root finding.
>>> from scipy import optimize
Finding the minimum of a scalar function

Lets define the following function:
>>> def f(x):
... return x**2 + 10*np.sin(x)
and plot it:

>>> x = np.arange(-10, 10, 0.1)
>>> plt.plot(x, f(x))
>>> plt.show()
Another available (but much less efficient) global optimizer is scipy.optimize.brute() (brute force opti-
mization on a grid). More efficient algorithms for different classes of global optimization problems exist, but this
120 is out of the scope of scipy. Some useful packages for global optimization are OpenOpt, IPOPT, PyGMO and
PyEvolve.
100 Note: scipy used to contain the routine anneal, it has been deprecated since SciPy 0.14.0 and removed in
SciPy 0.16.0.
80 To find the local minimum, lets

scipy.optimize.fminbound():
constraint the variable to the interval (0, 10) using
60 >>> xmin_local = optimize.fminbound(f, 0, 10)

>>> xmin_local
3.8374671...
40 Note: Finding minima of function is discussed in more details in the advanced chapter: Mathematical optimiza-
tion: finding minima of functions (page 239).
20 Finding the roots of a scalar function

To find a root, i.e. a point where f(x) = 0, of the function f above we can use for example
0 scipy.optimize.fsolve():
>>> root = optimize.fsolve(f, 1) # our initial guess is 1
>>> root
20 array([ 0.])
10 5 0 5 10
Note that only one root is found. Inspecting the plot of f reveals that there is a second root around -2.5. We find
the exact value of it by adjusting our initial guess:
This function has a global minimum around -1.3 and a local minimum around 3.8. >>> root2 = optimize.fsolve(f, -2.5)
The general and efficient way to find a minimum for this function is to conduct a gradient descent starting from a >>> root2
given initial point. The BFGS algorithm is a good way of doing this: array([-2.47948183])
>>> optimize.fmin_bfgs(f, 0) Curve fitting

Optimization terminated successfully.
Current function value: -7.945823 Suppose we have data sampled from f with some noise:
Iterations: 5 .. Comment to make doctest pass
Function evaluations: 24
Gradient evaluations: 8
array([-1.30644003])
>>> xdata = np.linspace(-10, 10, num=20)
A possible issue with this approach is that, if the function has local minima the algorithm may find these local
>>> ydata = f(xdata) + np.random.randn(xdata.size)
minima instead of the global minimum depending on the initial point:
>>> optimize.fmin_bfgs(f, 3, disp=0) Now if we know the functional form of the function from which the samples were drawn (x^2 + sin(x) in
array([ 3.83746663]) this case) but not the amplitudes of the terms, we can find those by least squares curve fitting. First we have to
define the function to fit:
If we dont know the neighborhood of the global minimum to choose the initial point, we need to resort to costlier
>>> def f2(x, a, b):
global optimization. To find the global minimum, we use scipy.optimize.basinhopping() (which
... return a*x**2 + b*np.sin(x)
combines a local optimizer with stochastic sampling of starting points for the local optimizer):
New in version 0.12.0: basinhopping was added in version 0.12.0 of Scipy Then we can use scipy.optimize.curve_fit() to find a and b:
>>> optimize.basinhopping(f, 0) >>> guess = [2, 2]
nfev: 1725 >>> params, params_covariance = optimize.curve_fit(f2, xdata, ydata, guess)
minimization_failures: 0 >>> params
fun: -7.9458233756152845 array([ 0.99667386, 10.17808313])
x: array([-1.30644001])
message: ['requested number of basinhopping iterations completed successfully'] Now we have found the minima and roots of f and used curve fitting on it, we put all those resuls together in a
njev: 575 single plot:
nit: 100
Exercise: 2-D minimization
120
f(x) Six-hump Camelback function
100 Curve fit result
Minima 6
80 Roots 5
4
60 3
f(x, y)
2
f(x)
40 1
0
1
20 2
1.0
0 0.5
2.01.51.0 0.0y
0.50.0 0.5
20 x 0.51.01.5
10 5 0 5 10 2.0 1.0
x
The six-hump camelback function
Note: In Scipy >= 0.11 unified interfaces to all minimization and root finding algorithms
are available: scipy.optimize.minimize(), scipy.optimize.minimize_scalar() and 4 2
scipy.optimize.root(). They allow comparing various algorithms easily through the method keyword. (, ) = (4 2.12 + ) + + (4 2 4) 2
3
You can find algorithms with the same functionalities for multi-dimensional problems in scipy.optimize. has multiple global and local minima. Find the global minima of this function.
Hints:
Variables can be restricted to -2 < x < 2 and -1 < y < 1.
Exercise: Curve fitting of temperature data Use numpy.meshgrid() and pylab.imshow() to find visually the regions.
The temperature extremes in Alaska for each month, starting in January, are given by (in degrees Use scipy.optimize.fmin_bfgs() or another multi-dimensional minimizer.
Celcius): How many global minima are there, and what is the function value at those points? What
happens for an initial guess of (x, y) = (0, 0)?
max: 17, 19, 21, 28, 33, 38, 37, 37, 31, 23, 19, 18
min: -62, -59, -56, -46, -32, -18, -9, -13, -25, -46, -52, -58
See the summary exercise on Non linear least squares curve fitting: application to point extraction in topograph-
1. Plot these temperature extremes. ical lidar data (page 129) for another, more advanced example.
2. Define a function that can describe min and max temperatures. Hint: this function has to
have a period of 1 year. Hint: include a time offset.
3. Fit this function to the data with scipy.optimize.curve_fit(). 1.5.6 Statistics and random numbers: scipy.stats
4. Plot the result. Is the fit reasonable? If not, why?
5. Is the time offset for min and max temperatures the same within the fit accuracy? The module scipy.stats contains statistical tools and probabilistic descriptions of random processes. Random
number generators for various random process can be found in numpy.random.
Histogram and probability density function
Given observations of a random process, their histogram is an estimator of the random processs PDF (probability
density function):
>>> a = np.random.normal(size=1000)
>>> bins = np.arange(-4, 5)
>>> bins
array([-4, -3, -2, -1, 0, 1, 2, 3, 4]) >>> stats.scoreatpercentile(a, 50)

>>> histogram = np.histogram(a, bins=bins, normed=True)[0] 0.0404176959...
>>> bins = 0.5*(bins[1:] + bins[:-1])
>>> bins Similarly, we can calculate the percentile 90:
array([-3.5, -2.5, -1.5, -0.5, 0.5, 1.5, 2.5, 3.5])
>>> from scipy import stats >>> stats.scoreatpercentile(a, 90)
>>> b = stats.norm.pdf(bins) # norm is a distribution 1.3185699120...
>>> plt.plot(bins, histogram)

The percentile is an estimator of the CDF: cumulative distribution function.
>>> plt.plot(bins, b)
Statistical tests
A statistical test is a decision indicator. For instance, if we have two sets of observations, that we assume are gen-
erated from Gaussian processes, we can use a T-test to decide whether the two sets of observations are significantly
0.40 different:
0.35 >>> a = np.random.normal(0, 1, size=100)
>>> b = np.random.normal(1, 1, size=10)
0.30 >>> stats.ttest_ind(a, b)
(array(-3.177574054...), 0.0019370639...)
0.25
Tip: The resulting output is composed of:
0.20
The T statistic value: it is a number the sign of which is proportional to the difference between the two
0.15 random processes and the magnitude is related to the significance of this difference.
the p value: the probability of both processes being identical. If it is close to 1, the two process are almost
0.10
certainly identical. The closer it is to zero, the more likely it is that the processes have different means.
0.05
See also:
0.00
6 4 2 0 2 4 6 The chapter on statistics (page 271) introduces much more elaborate tools for statistical testing and statistical data
loading and visualization outside of scipy.
If we know that the random process belongs to a given family of random processes, such as normal processes,
we can do a maximum-likelihood fit of the observations to estimate the parameters of the underlying distribution. 1.5.7 Interpolation: scipy.interpolate
Here we fit a normal process to the observed data:
>>> loc, std = stats.norm.fit(a) The scipy.interpolate is useful for fitting a function from experimental data and thus evaluating points
>>> loc where no measure exists. The module is based on the FITPACK Fortran subroutines from the netlib project.
0.0314345570...
By imagining experimental data close to a sine function:
>>> std
0.9778613090... >>> measured_time = np.linspace(0, 1, 10)
>>> noise = (np.random.random(10)*2 - 1) * 1e-1
>>> measures = np.sin(2 * np.pi * measured_time) + noise
Exercise: Probability distributions
The scipy.interpolate.interp1d class can build a linear interpolation function:
Generate 1000 random variates from a gamma distribution with a shape parameter of 1, then plot a histogram
from those samples. Can you plot the pdf on top (it should match)? >>> from scipy.interpolate import interp1d
Extra: the distributions have a number of useful methods. Explore them by reading the docstring or by using >>> linear_interp = interp1d(measured_time, measures)
IPython tab completion. Can you find the shape parameter of 1 back by using the fit method on your
random variates? Then the scipy.interpolate.linear_interp instance needs to be evaluated at the time of interest:
>>> computed_time = np.linspace(0, 1, 50)
>>> linear_results = linear_interp(computed_time)
Percentiles
A cubic interpolation can also be selected by providing the kind optional keyword argument:
The median is the value with half of the observations below, and half above: >>> cubic_interp = interp1d(measured_time, measures, kind='cubic')
>>> cubic_results = cubic_interp(computed_time)
>>> np.median(a)
0.04041769593...
The results are now gathered on the following Matplotlib figure:
It is also called the percentile 50, because 50% of the observation are below it:
... return -2 * ypos

...
1.0
measures An extra argument counter_arr has been added to illustrate that the function may be called several times for
a single time step, until solver convergence. The counter array is defined as:
0.5
linear interp >>> counter = np.zeros((1,), dtype=np.uint16)
cubic interp The trajectory will now be computed:

>>> from scipy.integrate import odeint
>>> time_vec = np.linspace(0, 4, 40)
0.0 >>> yvec, info = odeint(calc_derivative, 1, time_vec,
... args=(counter,), full_output=True)
Thus the derivative function has been called more than 40 times (which was the number of time steps):
0.5 >>> counter
array([129], dtype=uint16)
and the cumulative number of iterations for each of the 10 first time steps can be obtained by:
1.0 >>> info['nfe'][:10]
array([31, 35, 43, 49, 53, 57, 59, 63, 65, 69], dtype=int32)
Note that the solver requires more iterations for the first time step. The solution yvec for the trajectory can now
1.5 be plotted:
0.0 0.2 0.4 0.6 0.8 1.0
1.0
scipy.interpolate.interp2d is similar to scipy.interpolate.interp1d, but for 2-D arrays.
Note that for the interp family, the computed time must stay within the measured time range. See the sum- 0.8
mary exercise on Maximum wind speed prediction at the Sprog station (page 125) for a more advance spline
interpolation example.
y position [m]
0.6
1.5.8 Numerical integration: scipy.integrate
The most generic integration routine is scipy.integrate.quad():

0.4
>>> from scipy.integrate import quad
>>> res, err = quad(np.sin, 0, np.pi/2) 0.2
>>> np.allclose(res, 1)
True
>>> np.allclose(err, 1 - res) 0.0
True 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0
Time [s]
Others integration schemes are available with fixed_quad, quadrature, romberg.
Another example with scipy.integrate.odeint() will be a damped spring-mass oscillator (2nd order
scipy.integrate also features routines for integrating Ordinary Differential Equations (ODE). In particular, oscillator). The position of a mass attached to a spring obeys the 2nd order ODE y + 2 eps wo y +
scipy.integrate.odeint() is a general-purpose integrator using LSODA (Livermore Solver for Ordinary wo^2 y = 0 with wo^2 = k/m with k the spring constant, m the mass and eps=c/(2 m wo) with c the
Differential equations with Automatic method switching for stiff and non-stiff problems), see the ODEPACK damping coefficient. For this example, we choose the parameters as:
Fortran library for more details.
>>> mass = 0.5 # kg
odeint solves first-order ODE systems of the form: >>> kspring = 4 # N/m
dy/dt = rhs(y1, y2, .., t0,...) >>> cviscous = 0.4 # N s/m
As an introduction, let us solve the ODE dy/dt = -2y between t = 0..4, with the initial condition y(t=0) so the system will be underdamped, because:
= 1. First the function computing the derivative of the position needs to be defined: >>> eps = cviscous / (2 * mass * np.sqrt(kspring/mass))
>>> eps < 1
>>> def calc_derivative(ypos, time, counter_arr):
True
... counter_arr += 1
For the scipy.integrate.odeint() solver the 2nd order equation needs to be transformed in a system of
two first-order equations for the vector Y=(y, y). It will be convenient to define nu = 2 eps * wo = c 8
/ m and om = wo^2 = k/m:
>>> nu_coef = cviscous / mass 6
>>> om_coef = kspring / mass
Thus the function will calculate the velocity and acceleration by: 4
>>> def calc_deri(yvec, time, nuc, omc):
... return (yvec[1], -nuc * yvec[1] - omc * yvec[0]) 2
...
>>> time_vec = np.linspace(0, 10, 100)
>>> yarr = odeint(calc_deri, (1, 0), time_vec, args=(nu_coef, om_coef)) 0
The final position and velocity are shown on the following Matplotlib figure:
2
1.5 4
y 0 1 2 3 4 5
1.0 y'
scipy.signal.resample(): resample a signal to n points using FFT.
0.5
>>> t = np.linspace(0, 5, 100)
0.0 >>> x = np.sin(t)
0.5 >>> plt.plot(t, x, linewidth=3)

1.0 >>> plt.plot(t[::2], signal.resample(x, 50), 'ko')
1.5
2.0 1.5
2.5
0 2 4 6 8 10 1.0
There is no Partial Differential Equations (PDE) solver in Scipy. Some Python packages for solving PDEs are 0.5
available, such as fipy or SfePy.
0.0
1.5.9 Signal processing: scipy.signal
0.5
>>> from scipy import signal
scipy.signal.detrend(): remove linear trend from signal: 1.0

>>> t = np.linspace(0, 5, 100)
>>> x = t + np.random.normal(size=100) 1.5
0 1 2 3 4 5
>>> plt.plot(t, x, linewidth=3)
Notice how on the side of the window the resampling is less accurate and has a rippling effect.
>>> plt.plot(t, signal.detrend(x), linewidth=3)
[<matplotlib.lines.Line2D object at ...>] scipy.signal has many window functions: scipy.signal.hamming(),
scipy.signal.bartlett(), scipy.signal.blackman()...
scipy.signal has filtering (median filter scipy.signal.medfilt(), Wiener
scipy.signal.wiener()), but we will discuss this in the image section.
1.5.10 Image processing: scipy.ndimage
The submodule dedicated to image processing in scipy is scipy.ndimage.

>>> from scipy import ndimage
Image processing routines may be sorted according to the category of processing they perform.
Geometrical transformations on images
Changing orientation, resolution, .. Mathematical morphology
>>> from scipy import misc Mathematical morphology is a mathematical theory that stems from set theory. It characterizes and transforms
>>> lena = misc.lena() geometrical structures. Binary (black and white) images, in particular, can be transformed using this theory:
>>> shifted_lena = ndimage.shift(lena, (50, 50)) the sets to be transformed are the sets of neighboring non-zero-valued pixels. The theory was also extended to
>>> shifted_lena2 = ndimage.shift(lena, (50, 50), mode='nearest')
gray-valued images.
>>> rotated_lena = ndimage.rotate(lena, 30)
>>> cropped_lena = lena[50:-50, 50:-50]
>>> zoomed_lena = ndimage.zoom(lena, 2)
>>> zoomed_lena.shape
(1024, 1024)
>>> plt.subplot(151)
<matplotlib.axes._subplots.AxesSubplot object at 0x...> Elementary mathematical-morphology operations use a structuring element in order to modify other geometrical
structures.
>>> plt.imshow(shifted_lena, cmap=plt.cm.gray)
<matplotlib.image.AxesImage object at 0x...> Let us first generate a structuring element
>>> plt.axis('off') >>> el = ndimage.generate_binary_structure(2, 1)

(-0.5, 511.5, 511.5, -0.5) >>> el
array([[False, True, False],
>>> # etc. [...True, True, True],
[False, True, False]], dtype=bool)
>>> el.astype(np.int)
array([[0, 1, 0],
Image filtering
[1, 1, 1],
[0, 1, 0]])
>>> from scipy import misc
>>> lena = misc.lena() Erosion
>>> noisy_lena = np.copy(lena).astype(np.float) >>> a = np.zeros((7,7), dtype=np.int)
>>> noisy_lena += lena.std()*0.5*np.random.standard_normal(lena.shape) >>> a[1:6, 2:5] = 1
>>> blurred_lena = ndimage.gaussian_filter(noisy_lena, sigma=3) >>> a
>>> median_lena = ndimage.median_filter(blurred_lena, size=5) array([[0, 0, 0, 0, 0, 0, 0],
>>> from scipy import signal [0, 0, 1, 1, 1, 0, 0],
>>> wiener_lena = signal.wiener(blurred_lena, (5,5)) [0, 0, 1, 1, 1, 0, 0],
[0, 0, 1, 1, 1, 0, 0],
Many other filters in scipy.ndimage.filters and scipy.signal can be applied to images. [0, 0, 1, 1, 1, 0, 0],
[0, 0, 1, 1, 1, 0, 0],
[0, 0, 0, 0, 0, 0, 0]])
Exercise >>> ndimage.binary_erosion(a).astype(a.dtype)
array([[0, 0, 0, 0, 0, 0, 0],
Compare histograms for the different filtered images. [0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 1, 0, 0, 0],
[0, 0, 0, 1, 0, 0, 0],
[0, 0, 0, 1, 0, 0, 0], >>> a = np.zeros((50, 50))

[0, 0, 0, 0, 0, 0, 0], >>> a[10:-10, 10:-10] = 1
[0, 0, 0, 0, 0, 0, 0]]) >>> a += 0.25*np.random.standard_normal(a.shape)
>>> #Erosion removes objects smaller than the structure >>> mask = a>=0.5
>>> ndimage.binary_erosion(a, structure=np.ones((5,5))).astype(a.dtype) >>> opened_mask = ndimage.binary_opening(mask)
array([[0, 0, 0, 0, 0, 0, 0], >>> closed_mask = ndimage.binary_closing(opened_mask)
[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0]])
Dilation
>>> a = np.zeros((5, 5))
>>> a[2, 2] = 1
>>> a
array([[ 0., 0., 0., 0., 0.], Exercise
[ 0., 0., 0., 0., 0.],
[ 0., 0., 1., 0., 0.], Check that the area of the reconstructed square is smaller than the area of the initial square. (The
[ 0., 0., 0., 0., 0.], opposite would occur if the closing step was performed before the opening).
[ 0., 0., 0., 0., 0.]])
>>> ndimage.binary_dilation(a).astype(a.dtype)
array([[ 0., 0., 0., 0., 0.], For gray-valued images, eroding (resp. dilating) amounts to replacing a pixel by the minimal (resp. maximal)
[ 0., 0., 1., 0., 0.], value among pixels covered by the structuring element centered on the pixel of interest.
[ 0., 1., 1., 1., 0.],
>>> a = np.zeros((7,7), dtype=np.int)
[ 0., 0., 1., 0., 0.],
>>> a[1:6, 1:6] = 3
[ 0., 0., 0., 0., 0.]])
>>> a[4,4] = 2; a[2,3] = 1
>>> a
Opening
array([[0, 0, 0, 0, 0, 0, 0],
>>> a = np.zeros((5,5), dtype=np.int) [0, 3, 3, 3, 3, 3, 0],
>>> a[1:4, 1:4] = 1; a[4, 4] = 1 [0, 3, 3, 1, 3, 3, 0],
>>> a [0, 3, 3, 3, 3, 3, 0],
array([[0, 0, 0, 0, 0], [0, 3, 3, 3, 2, 3, 0],
[0, 1, 1, 1, 0], [0, 3, 3, 3, 3, 3, 0],
[0, 1, 1, 1, 0], [0, 0, 0, 0, 0, 0, 0]])
[0, 1, 1, 1, 0], >>> ndimage.grey_erosion(a, size=(3,3))
[0, 0, 0, 0, 1]]) array([[0, 0, 0, 0, 0, 0, 0],
>>> # Opening removes small objects [0, 0, 0, 0, 0, 0, 0],
>>> ndimage.binary_opening(a, structure=np.ones((3,3))).astype(np.int) [0, 0, 1, 1, 1, 0, 0],
array([[0, 0, 0, 0, 0], [0, 0, 1, 1, 1, 0, 0],
[0, 1, 1, 1, 0], [0, 0, 3, 2, 2, 0, 0],
[0, 1, 1, 1, 0], [0, 0, 0, 0, 0, 0, 0],
[0, 1, 1, 1, 0], [0, 0, 0, 0, 0, 0, 0]])
[0, 0, 0, 0, 0]])
>>> # Opening can also smooth corners
>>> ndimage.binary_opening(a).astype(np.int) Measurements on images
array([[0, 0, 0, 0, 0],
[0, 0, 1, 0, 0],
Let us first generate a nice synthetic binary image.
[0, 1, 1, 1, 0],
[0, 0, 1, 0, 0], >>> x, y = np.indices((100, 100))
[0, 0, 0, 0, 0]]) >>> sig = np.sin(2*np.pi*x/50.)*np.sin(2*np.pi*y/50.)*(1+x*y/50.**2)**2
>>> mask = sig > 1
Closing: ndimage.binary_closing
Now we look for various information about the objects in the image:
Exercise >>> labels, nb = ndimage.label(mask)
>>> nb
Check that opening amounts to eroding, then dilating. 8
>>> areas = ndimage.sum(mask, labels, range(1, labels.max()+1))
An opening operation removes small structures, while a closing operation fills small holes. Such operations can >>> areas
array([ 190., 45., 424., 278., 459., 190., 549., 424.])
therefore be used to clean an image.
>>> maxima = ndimage.maximum(sig, labels, range(1, labels.max()+1)) Computing the cumulative probabilities
>>> maxima
array([ 1.80238238, 1.13527605, 5.51954079, 2.49611818, The annual wind speeds maxima have already been computed and saved in the numpy format in the file
6.71673619, 1.80238238, 16.76547217, 5.51954079]) examples/max-speeds.npy, thus they will be loaded by using numpy:
>>> ndimage.find_objects(labels==4)
[(slice(30L, 48L, None), slice(30L, 48L, None))] >>> import numpy as np
>>> sl = ndimage.find_objects(labels==4) >>> max_speeds = np.load('intro/summary-exercises/examples/max-speeds.npy')
>>> import pylab as pl >>> years_nb = max_speeds.shape[0]
>>> pl.imshow(sig[sl[0]])
<matplotlib.image.AxesImage object at ...> Following the cumulative probability definition p_i from the previous section, the corresponding values will be:
>>> cprob = (np.arange(years_nb, dtype=np.float32) + 1)/(years_nb + 1)
and they are assumed to fit the given wind speeds:

>>> sorted_max_speeds = np.sort(max_speeds)
Prediction with UnivariateSpline
In this section the quantile function will be estimated by using the UnivariateSpline class which can
represent a spline from points. The default behavior is to build a spline of degree 3 and points can have
different weights according to their reliability. Variants are InterpolatedUnivariateSpline and
LSQUnivariateSpline on which errors checking is going to change. In case a 2D spline is wanted, the
See the summary exercise on Image processing application: counting bubbles and unmolten grains (page 133) for
BivariateSpline class family is provided. All those classes for 1D and 2D splines use the FITPACK For-
a more advanced example.
tran subroutines, thats why a lower library access is available through the splrep and splev functions for
respectively representing and evaluating a spline. Moreover interpolation functions without the use of FITPACK
1.5.11 Summary exercises on scientific computing parameters are also provided for simpler use (see interp1d, interp2d, barycentric_interpolate
and so on).
The summary exercises use mainly Numpy, Scipy and Matplotlib. They provide some real-life examples of For the Sprog maxima wind speeds, the UnivariateSpline will be used because a spline of degree 3 seems
scientific computing with Python. Now that the basics of working with Numpy and Scipy have been introduced, to correctly fit the data:
the interested user is invited to try these exercises.
>>> from scipy.interpolate import UnivariateSpline
>>> quantile_func = UnivariateSpline(cprob, sorted_max_speeds)
Maximum wind speed prediction at the Sprog station
The quantile function is now going to be evaluated from the full range of probabilities:
The exercise goal is to predict the maximum wind speed occurring every 50 years even if no measure exists for >>> nprob = np.linspace(0, 1, 1e2)
such a period. The available data are only measured over 21 years at the Sprog meteorological station located >>> fitted_max_speeds = quantile_func(nprob)
in Denmark. First, the statistical steps will be given and then illustrated with functions from the scipy.interpolate
module. At the end the interested readers are invited to compute results from raw data and in a slightly different In the current model, the maximum wind speed occurring every 50 years is defined as the upper 2% quantile. As
approach. a result, the cumulative probability value will be:
>>> fifty_prob = 1. - 0.02
Statistical approach
So the storm wind speed occurring every 50 years can be guessed by:
The annual maxima are supposed to fit a normal probability density function. However such function is not going >>> fifty_wind = quantile_func(fifty_prob)
to be estimated because it gives a probability from a wind speed maxima. Finding the maximum wind speed >>> fifty_wind
occurring every 50 years requires the opposite approach, the result needs to be found from a defined probability. array(32.97989825...)
That is the quantile function role and the exercise goal will be to find it. In the current model, it is supposed that
the maximum wind speed occurring every 50 years is defined as the upper 2% quantile. The results are now gathered on a Matplotlib figure:
By definition, the quantile function is the inverse of the cumulative distribution function. The latter describes the
probability distribution of an annual maxima. In the exercise, the cumulative probability p_i for a given year Exercise with the Gumbell distribution
i is defined as p_i = i/(N+1) with N = 21, the number of measured years. Thus it will be possible to
calculate the cumulative probability of every measured wind speed maxima. From those experimental points, the The interested readers are now invited to make an exercise by using the wind speeds measured over 21 years.
scipy.interpolate module will be very useful for fitting the quantile function. Finally the 50 years maxima is going The measurement period is around 90 minutes (the original period was around 10 minutes but the file size
to be evaluated from the cumulative probability of the 2% quantile. has been reduced for making the exercise setup easier). The data are stored in numpy format inside the file
examples/sprog-windspeeds.npy. Do not look at the source code for the plots until you have completed
the exercise.
Figure 1.1: Solution: Python source file Figure 1.2: Solution: Python source file
The first step will be to find the annual maxima by using numpy and plot them as a matplotlib bar figure. In this tutorial, the goal is to analyze the waveform recorded by the lidar system 2 . Such a signal contains peaks
whose center and amplitude permit to compute the position and some characteristics of the hit target. When the
The second step will be to use the Gumbell distribution on cumulative probabilities p_i defined as -log(
footprint of the laser beam is around 1m on the Earth surface, the beam can hit multiple targets during the two-way
-log(p_i) ) for fitting a linear quantile function (remember that you can define the degree of the
propagation (for example the ground and the top of a tree or building). The sum of the contributions of each target
UnivariateSpline). Plotting the annual maxima versus the Gumbell distribution should give you
hit by the laser beam then produces a complex signal with multiple peaks, each one containing information about
the following figure.
one target.
One state of the art method to extract information from these data is to decompose them in a sum of Gaussian
functions where each function represents the contribution of a target hit by the laser beam.
Therefore, we use the scipy.optimize module to fit a waveform to one or a sum of Gaussian functions.
Loading and visualization
Load the first waveform using:

>>> waveform_1 = np.load('data/waveform_1.npy')
and visualize it:

>>> t = np.arange(len(waveform_1))
>>> plt.plot(t, waveform_1)
>>> plt.show()
Figure 1.3: Solution: Python source file
The last step will be to find 34.23 m/s for the maximum wind speed occurring every 50 years.
Non linear least squares curve fitting: application to point extraction in topographical lidar data
The goal of this exercise is to fit a model to some data. The data used in this tutorial are lidar data and are described
in details in the following introductory paragraph. If youre impatient and want to practice now, please skip it and
go directly to Loading and visualization (page 130).
Introduction
Lidars systems are optical rangefinders that analyze property of scattered light to measure distances. Most of them
emit a short light impulsion towards a target and record the reflected signal. This signal is then processed to extract
the distance between the lidar system and the target.
Topographical lidar systems are such systems embedded in airborne platforms. They measure distances between
the platform and the Earth, so as to deliver information on the Earths topography (see 1 for more details). As you can notice, this waveform is a 80-bin-length signal with a single peak.
1 Mallet, C. and Bretar, F. Full-Waveform Topographic Lidar: State-of-the-Art. ISPRS Journal of Photogrammetry and Remote Sensing 2 The data used for this tutorial are part of the demonstration data available for the FullAnalyze software and were kindly provided by the
64(1), pp.1-16, January 2009 http://dx.doi.org/10.1016/j.isprsjprs.2008.09.007 GIS DRAIX.
Fitting a waveform with a simple Gaussian model Going further
The signal is very simple and can be modeled as a single Gaussian function and an offset corresponding to the Try with a more complex waveform (for instance data/waveform_2.npy) that contains three signif-
background noise. To fit the signal with the function, we must: icant peaks. You must adapt the model which is now a sum of Gaussian functions instead of only one
Gaussian peak.
define the model
propose an initial solution
call scipy.optimize.leastsq
Model A Gaussian function defined by

{ ( )2 }

+ exp

can be defined in python by:

>>> def model(t, coeffs):
... return coeffs[0] + coeffs[1] * np.exp( - ((t-coeffs[2])/coeffs[3])**2 )
where
coeffs[0] is (noise)
coeffs[1] is (amplitude)
coeffs[2] is (center)
coeffs[3] is (width)
Initial solution An approximative initial solution that we can find from looking at the graph is for instance:
>>> x0 = np.array([3, 30, 15, 1], dtype=float)
Fit scipy.optimize.leastsq minimizes the sum of squares of the function given as an argument. Basi-
cally, the function to minimize is the residuals (the difference between the data and the model):
In some cases, writing an explicit function to compute the Jacobian is faster than letting leastsq esti-
>>> def residuals(coeffs, y, t): mate it numerically. Create a function to compute the Jacobian of the residuals and use it as an input for
... return y - model(t, coeffs) leastsq.
So lets get our solution by calling scipy.optimize.leastsq() with the following arguments: When we want to detect very small peaks in the signal, or when the initial guess is too far from a good
solution, the result given by the algorithm is often not satisfying. Adding constraints to the parameters of
the function to minimize
the model enables to overcome such limitations. An example of a priori knowledge we can add is the sign
an initial solution of our variables (which are all positive).
the additional arguments to pass to the function With the following initial solution:
>>> from scipy.optimize import leastsq >>> x0 = np.array([3, 50, 20, 1], dtype=float)
>>> x, flag = leastsq(residuals, x0, args=(waveform_1, t))
>>> print(x) compare the result of scipy.optimize.leastsq() and what you can get with
[ 2.70363341 27.82020742 15.47924562 3.05636228] scipy.optimize.fmin_slsqp() when adding boundary constraints.
And visualize the solution:

>>> plt.plot(t, waveform_1, t, model(t, x))
[<matplotlib.lines.Line2D object at ...>, <matplotlib.lines.Line2D object at ...>]
>>> plt.legend(['waveform', 'model'])
<matplotlib.legend.Legend object at ...>
>>> plt.show()
Remark: from scipy v0.8 and above, you should rather use scipy.optimize.curve_fit() which takes
the model and the data as arguments, so you dont need to define the residuals any more.
Image processing application: counting bubbles and unmolten grains Example of solution for the image processing exercise: unmolten grains in glass
1. Open the image file MV_HFV_012.jpg and display it. Browse through the keyword arguments in the
docstring of imshow to display the image with the right orientation (origin in the bottom left corner, and
not the upper left corner as for standard arrays).
Statement of the problem >>> dat = pl.imread('data/MV_HFV_012.jpg')
2. Crop the image to remove the lower panel with measure information.
1. Open the image file MV_HFV_012.jpg and display it. Browse through the keyword arguments in the docstring
of imshow to display the image with the right orientation (origin in the bottom left corner, and not the upper >>> dat = dat[:-60]
left corner as for standard arrays).
3. Slightly filter the image with a median filter in order to refine its histogram. Check how the histogram
This Scanning Element Microscopy image shows a glass sample (light gray matrix) with some bubbles (on black) changes.
and unmolten sand grains (dark gray). We wish to determine the fraction of the sample covered by these three
phases, and to estimate the typical size of sand grains and bubbles, their sizes, etc. >>> filtdat = ndimage.median_filter(dat, size=(7,7))
>>> hi_dat = np.histogram(dat, bins=np.arange(256))
2. Crop the image to remove the lower panel with measure information. >>> hi_filtdat = np.histogram(filtdat, bins=np.arange(256))
3. Slightly filter the image with a median filter in order to refine its histogram. Check how the histogram changes.
4. Using the histogram of the filtered image, determine thresholds that allow to define masks for sand pixels, glass
pixels and bubble pixels. Other option (homework): write a function that determines automatically the thresholds
from the minima of the histogram.
5. Display an image in which the three phases are colored with three different colors.
6. Use mathematical morphology to clean the different phases.
7. Attribute labels to all bubbles and sand grains, and remove from the sand mask grains that are smaller than 10
pixels. To do so, use ndimage.sum or np.bincount to compute the grain sizes.
8. Compute the mean size of bubbles.
Proposed solution

4. Using the histogram of the filtered image, determine thresholds that allow to define masks for sand pixels, 6. Use mathematical morphology to clean the different phases.
glass pixels and bubble pixels. Other option (homework): write a function that determines automatically
>>> sand_op = ndimage.binary_opening(sand, iterations=2)
the thresholds from the minima of the histogram.
>>> void = filtdat <= 50 7. Attribute labels to all bubbles and sand grains, and remove from the sand mask grains that are smaller than
>>> sand = np.logical_and(filtdat > 50, filtdat <= 114) 10 pixels. To do so, use ndimage.sum or np.bincount to compute the grain sizes.
>>> glass = filtdat > 114
>>> sand_labels, sand_nb = ndimage.label(sand_op)
>>> sand_areas = np.array(ndimage.sum(sand_op, sand_labels, np.arange(sand_labels.max()+1)))
5. Display an image in which the three phases are colored with three different colors.
>>> mask = sand_areas > 100
>>> phases = void.astype(np.int) + 2*glass.astype(np.int) + 3*sand.astype(np.int) >>> remove_small_sand = mask[sand_labels.ravel()].reshape(sand_labels.shape)
8. Compute the mean size of bubbles.

>>> bubbles_labels, bubbles_nb = ndimage.label(void)
>>> bubbles_areas = np.bincount(bubbles_labels.ravel())[1:]
>>> mean_bubble_size = bubbles_areas.mean()
>>> median_bubble_size = np.median(bubbles_areas)
>>> mean_bubble_size, median_bubble_size
(1699.875, 65.0)
1.6 Getting help and finding documentation
Author: Emmanuelle Gouillart

Rather than knowing all functions in Numpy and Scipy, it is important to find rapidly information throughout the
documentation and the available help. Here are some ways to get information: countered, such as fitting data points, solving ODE, etc.
In Ipython, help function opens the docstring of the function. Only type the beginning of the func- Matplotlibs website http://matplotlib.sourceforge.net/ features a very nice gallery with a large number of
tions name and use tab completion to display the matching functions. plots, each of them shows both the source code and the resulting plot. This is very useful for learning by
In [204]: help np.v example. More standard documentation is also available.
np.vander np.vdot np.version np.void0 np.vstack
Finally, two more technical possibilities are useful as well:
np.var np.vectorize np.void np.vsplit
In Ipython, the magical function %psearch search for objects matching patterns. This is useful if, for
In [204]: help np.vander example, one does not know the exact name of a function.
In Ipython it is not possible to open a separated window for help and documentation; however one can always open In [3]: import numpy as np
In [4]: %psearch np.diag*
np.diag
np.diagflat
np.diagonal
numpy.lookfor looks for keywords inside the docstrings of specified modules.

In [45]: numpy.lookfor('convolution')
Search results for 'convolution'
--------------------------------
numpy.convolve
a second Ipython shell just to display help and docstrings... Returns the discrete, linear convolution of two one-dimensional
sequences.
Numpys and Scipys documentations can be browsed online on http://docs.scipy.org/doc. numpy.bartlett
The search button is quite useful inside the reference documentation of the two packages Return the Bartlett window.
(http://docs.scipy.org/doc/numpy/reference/ and http://docs.scipy.org/doc/scipy/reference/). numpy.correlate
Discrete, linear correlation of two 1-dimensional sequences.
Tutorials on various topics as well as the complete API with all docstrings are found on this website. In [46]: numpy.lookfor('remove', module='os')
Search results for 'remove'
---------------------------
os.remove
remove(path)
os.removedirs
removedirs(path)
os.rmdir
rmdir(path)
os.unlink
unlink(path)
os.walk
Directory tree generator.
Numpys and Scipys documentation is enriched and updated on a regular basis by users on a wiki
http://docs.scipy.org/numpy/. As a result, some docstrings are clearer or more detailed on the wiki, and If everything listed above fails (and Google doesnt have the answer)... dont despair! Write to the mailing-
you may want to read directly the documentation on the wiki instead of the official documentation website. list suited to your problem: you should have a quick answer if you describe your problem well. Experts on
Note that anyone can create an account on the wiki and write better documentation; this is an easy way to scientific python often give very enlightening explanations on the mailing-list.
contribute to an open-source project and improve the tools you are using! Numpy discussion (numpy-discussion@scipy.org): all about numpy arrays, manipulating them, in-
Scipys cookbook http://www.scipy.org/Cookbook gives recipes on many common problems frequently en- dexation questions, etc.
SciPy Users List (scipy-user@scipy.org): scientific computing with Python, high-level data process-
ing, in particular with the scipy package.
matplotlib-users@lists.sourceforge.net for plotting with matplotlib.
1.6. Getting help and finding documentation 137 1.6. Getting help and finding documentation 138
CHAPTER 2 2.1.1 Iterators, generator expressions and generators
Iterators
Simplicity
Duplication of effort is wasteful, and replacing the various home-grown approaches with a standard feature
Advanced topics usually ends up making things more readable, and interoperable as well.
Guido van Rossum Adding Optional Static Typing to Python
An iterator is an object adhering to the iterator protocol basically this means that it has a next method,
which, when called, returns the next item in the sequence, and when theres nothing to return, raises the
StopIteration exception.
This part of the Scipy lecture notes is dedicated to advanced usage. It strives to educate the proficient Python coder An iterator object allows to loop just once. It holds the state (position) of a single iteration, or from the other side,
to be an expert and tackles various specific topics. each loop over a sequence requires a single iterator object. This means that we can iterate over the same sequence
more than once concurrently. Separating the iteration logic from the sequence allows us to have more than one
way of iteration.
Calling the __iter__ method on a container to create an iterator object is the most straightforward way to get
hold of an iterator. The iter function does that for us, saving a few keystrokes.
2.1 Advanced Python Constructs >>> nums = [1, 2, 3] # note that ... varies: these are different objects
>>> iter(nums)
<...iterator object at ...>
Author Zbigniew Jedrzejewski-Szmek >>> nums.__iter__()
<...iterator object at ...>
This section covers some features of the Python language which can be considered advanced in the sense that >>> nums.__reversed__()
not every language has them, and also in the sense that they are more useful in more complicated programs or <...reverseiterator object at ...>
libraries, but not in the sense of being particularly specialized, or particularly complicated.
>>> it = iter(nums)
It is important to underline that this chapter is purely about the language itself about features supported through
>>> next(it)
special syntax complemented by functionality of the Python stdlib, which could not be implemented through
1
clever external modules. >>> next(it)
The process of developing the Python programming language, its syntax, is very transparent; proposed changes 2
are evaluated from various angles and discussed via Python Enhancement Proposals PEPs. As a result, features >>> next(it)
3
described in this chapter were added after it was shown that they indeed solve real problems and that their use is
>>> next(it)
as simple as possible.
Chapters contents StopIteration
Iterators, generator expressions and generators (page 140) File "<stdin>", line 1, in <module>
Iterators (page 140) StopIteration
Generator expressions (page 141)
Generators (page 141) When used in a loop, StopIteration is swallowed and causes the loop to finish. But with explicit invocation,
Bidirectional communication (page 142) we can see that once the iterator is exhausted, accessing it raises an exception.
Chaining generators (page 144) Using the for..in loop also uses the __iter__ method. This allows us to transparently start the iteration over a
Decorators (page 144) sequence. But if we already have the iterator, we want to be able to use it in an for loop in the same way. In order
Replacing or tweaking the original object (page 145) to achieve this, iterators in addition to next are also required to have a method called __iter__ which returns
Decorators implemented as classes and as functions (page 145) the iterator (self).
Copying the docstring and other attributes of the original function (page 147)
Examples in the standard library (page 148) Support for iteration is pervasive in Python: all sequences and unordered containers in the standard library allow
Deprecation of functions (page 150) this. The concept is also stretched to other things: e.g. file objects support iteration over lines.
A while-loop removing decorator (page 151) >>> f = open('/etc/fstab')
A plugin registration system (page 151) >>> f is f.__iter__()
Context managers (page 152) True
Catching exceptions (page 153)
Using generators to define context managers (page 154) The file is an iterator itself and its __iter__ method doesnt create a separate object: only a single thread of
sequential access is allowed.
139 2.1. Advanced Python Constructs 140

Generator expressions Lets go over the life of the single invocation of the generator function.
>>> def f():
A second way in which iterator objects are created is through generator expressions, the basis for list compre- ... print("-- start --")
hensions. To increase clarity, a generator expression must always be enclosed in parentheses or an expression. If ... yield 3
round parentheses are used, then a generator iterator is created. If rectangular parentheses are used, the process is ... print("-- middle --")
short-circuited and we get a list. ... yield 4
... print("-- finished --")
>>> (i for i in nums)
>>> gen = f()
<generator object <genexpr> at 0x...>
>>> next(gen)
>>> [i for i in nums]
-- start --
[1, 2, 3]
3
>>> list(i for i in nums)
>>> next(gen)
[1, 2, 3]
-- middle --
4
The list comprehension syntax also extends to dictionary and set comprehensions. A set is created when >>> next(gen)
the generator expression is enclosed in curly braces. A dict is created when the generator expression contains -- finished --
pairs of the form key:value: Traceback (most recent call last):
>>> {i for i in range(3)} ...
set([0, 1, 2]) StopIteration
>>> {i:i**2 for i in range(3)} Traceback (most recent call last):
{0: 0, 1: 1, 2: 4} ...
StopIteration
One gotcha should be mentioned: in old Pythons the index variable (i) would leak, and in versions >= 3 this is
Contrary to a normal function, where executing f() would immediately cause the first print to be executed,
fixed.
gen is assigned without executing any statements in the function body. Only when gen.next() is invoked
by next, the statements up to the first yield are executed. The second next prints -- middle -- and
Generators execution halts on the second yield. The third next prints -- finished -- and falls of the end of the
function. Since no yield was reached, an exception is raised.
What happens with the function after a yield, when the control passes to the caller? The state of each generator
Generators is stored in the generator object. From the point of view of the generator function, is looks almost as if it was
A generator is a function that produces a sequence of results instead of a single value. running in a separate thread, but this is just an illusion: execution is strictly single-threaded, but the interpreter
David Beazley A Curious Course on Coroutines and Concurrency keeps and restores the state in between the requests for the next value.
Why are generators useful? As noted in the parts about iterators, a generator function is just a different way to
A third way to create iterator objects is to call a generator function. A generator is a function containing the create an iterator object. Everything that can be done with yield statements, could also be done with next
keyword yield. It must be noted that the mere presence of this keyword completely changes the nature of the methods. Nevertheless, using a function and having the interpreter perform its magic to create an iterator has
function: this yield statement doesnt have to be invoked, or even reachable, but causes the function to be advantages. A function can be much shorter than the definition of a class with the required next and __iter__
marked as a generator. When a normal function is called, the instructions contained in the body start to be methods. What is more important, it is easier for the author of the generator to understand the state which is
executed. When a generator is called, the execution stops before the first instruction in the body. An invocation kept in local variables, as opposed to instance attributes, which have to be used to pass data between consecutive
of a generator function creates a generator object, adhering to the iterator protocol. As with normal function invocations of next on an iterator object.
invocations, concurrent and recursive invocations are allowed. A broader question is why are iterators useful? When an iterator is used to power a loop, the loop becomes very
When next is called, the function is executed until the first yield. Each encountered yield statement gives a simple. The code to initialise the state, to decide if the loop is finished, and to find the next value is extracted into
value becomes the return value of next. After executing the yield statement, the execution of this function is a separate place. This highlights the body of the loop the interesting part. In addition, it is possible to reuse the
suspended. iterator code in other places.
>>> def f():

... yield 1 Bidirectional communication
... yield 2
>>> f() Each yield statement causes a value to be passed to the caller. This is the reason for the introduction of generators
<generator object f at 0x...> by PEP 255 (implemented in Python 2.2). But communication in the reverse direction is also useful. One obvious
>>> gen = f()
way would be some external state, either a global variable or a shared mutable object. Direct communication is
>>> next(gen)
possible thanks to PEP 342 (implemented in 2.5). It is achieved by turning the previously boring yield statement
1
>>> next(gen) into an expression. When the generator resumes execution after a yield statement, the caller can call a method
2 on the generator object to either pass a value into the generator, which then is returned by the yield statement,
>>> next(gen) or a different method to inject an exception into the generator.
The first of the new methods is send(value), which is similar to next(), but passes value into the generator
to be used for the value of the yield expression. In fact, g.next() and g.send(None) are equivalent.
StopIteration
2.1. Advanced Python Constructs 141 2.1. Advanced Python Constructs 142
The second of the new methods is throw(type, value=None, traceback=None) which is equivalent next or __next__?
to:
In Python 2.x, the iterator method to retrieve the next value is called next. It is invoked implicitly through
raise type, value, traceback the global function next, which means that it should be called __next__. Just like the global func-
tion iter calls __iter__. This inconsistency is corrected in Python 3.x, where it.next becomes
at the point of the yield statement. it.__next__. For other generator methods send and throw the situation is more complicated,
Unlike raise (which immediately raises an exception from the current execution point), throw() first resumes because they are not called implicitly by the interpreter. Nevertheless, theres a proposed syntax extension
the generator, and only then raises the exception. The word throw was picked because it is suggestive of putting to allow continue to take an argument which will be passed to send of the loops iterator. If this exten-
the exception in another location, and is associated with exceptions in other languages. sion is accepted, its likely that gen.send will become gen.__send__. The last of generator methods,
close, is pretty obviously named incorrectly, because it is already invoked implicitly.
What happens when an exception is raised inside the generator? It can be either raised explicitly or when executing
some statements or it can be injected at the point of a yield statement by means of the throw() method.
In either case, such an exception propagates in the standard manner: it can be intercepted by an except or
finally clause, or otherwise it causes the execution of the generator function to be aborted and propagates in Chaining generators
the caller.
For completeness sake, its worth mentioning that generator iterators also have a close() method, which can Note: This is a preview of PEP 380 (not yet implemented, but accepted for Python 3.3).
be used to force a generator that would otherwise be able to provide more values to finish immediately. It allows
the generator __del__ method to destroy objects holding the state of generator. Lets say we are writing a generator and we want to yield a number of values generated by a second generator,
a subgenerator. If yielding of values is the only concern, this can be performed without much difficulty using a
Lets define a generator which just prints what is passed in through send and throw.
loop such as
>>> import itertools
subgen = some_other_generator()
>>> def g():
for v in subgen:
... print('--start--')
yield v
... for i in itertools.count():
... print('--yielding %i--' % i)
... try: However, if the subgenerator is to interact properly with the caller in the case of calls to send(), throw()
... ans = yield i and close(), things become considerably more difficult. The yield statement has to be guarded by a
... except GeneratorExit: try..except..finally structure similar to the one defined in the previous section to debug the generator function.
... print('--closing--') Such code is provided in PEP 380, here it suffices to say that new syntax to properly yield from a subgenerator is
... raise being introduced in Python 3.3:
... except Exception as e:
... print('--yield raised %r--' % e) yield from some_other_generator()
... else:
... print('--yield returned %s--' % ans) This behaves like the explicit loop above, repeatedly yielding values from some_other_generator until it is
exhausted, but also forwards send, throw and close to the subgenerator.
>>> it = g()
>>> next(it)
--start-- 2.1.2 Decorators
--yielding 0--
0
>>> it.send(11) Summary
--yield returned 11--
--yielding 1-- This amazing feature appeared in the language almost apologetically and with concern that it might not be
1 that useful.
>>> it.throw(IndexError) Bruce Eckel An Introduction to Python Decorators
--yield raised IndexError()--
--yielding 2--
2 Since a function or a class are objects, they can be passed around. Since they are mutable objects, they can be
>>> it.close() modified. The act of altering a function or class object after it has been constructed but before is is bound to its
--closing-- name is called decorating.
There are two things hiding behind the name decorator one is the function which does the work of decorating,
i.e. performs the real work, and the other one is the expression adhering to the decorator syntax, i.e. an at-symbol
and the name of the decorating function.
Function can be decorated by using the decorator syntax for functions:
@decorator #
def function(): #
pass
A function is defined in the standard way.
An expression starting with @ placed before the function definition is the decorator . The part after @ must doing decoration
be a simple expression, usually this is just the name of a function or class. This part is evaluated first, and >>> function()
after the function defined below is ready, the decorator is called with the newly defined function object as inside function
the single argument. The value returned by the decorator is attached to the original name of the function.
>>> def decorator_with_arguments(arg):
Decorators can be applied to functions and to classes. For classes the semantics are identical the original class ... print("defining the decorator")
definition is used as an argument to call the decorator and whatever is returned is assigned under the original name. ... def _decorator(function):
Before the decorator syntax was implemented (PEP 318), it was possible to achieve the same effect by assigning ... # in this inner function, arg is available too
... print("doing decoration, %r" % arg)
the function or class object to a temporary variable and then invoking the decorator explicitly and then assigning
... return function
the return value to the name of the function. This sounds like more typing, and it is, and also the name of the
... return _decorator
decorated function doubling as a temporary variable must be used at least three times, which is prone to errors. >>> @decorator_with_arguments("abc")
Nevertheless, the example above is equivalent to: ... def function():
... print("inside function")
def function(): #
defining the decorator
pass
doing decoration, 'abc'
function = decorator(function) #
>>> function()
inside function
Decorators can be stacked the order of application is bottom-to-top, or inside-out. The semantics are such
that the originally defined function is used as an argument for the first decorator, whatever is returned by the first The two trivial decorators above fall into the category of decorators which return the original function. If they
decorator is used as an argument for the second decorator, ..., and whatever is returned by the last decorator is were to return a new function, an extra level of nestedness would be required. In the worst case, three levels of
attached under the name of the original function. nested functions.
The decorator syntax was chosen for its readability. Since the decorator is specified before the header of the >>> def replacing_decorator_with_args(arg):
function, it is obvious that its is not a part of the function body and its clear that it can only operate on the whole ... print("defining the decorator")
function. Because the expression is prefixed with @ is stands out and is hard to miss (in your face, according to ... def _decorator(function):
the PEP :) ). When more than one decorator is applied, each one is placed on a separate line in an easy to read ... # in this inner function, arg is available too
way. ... print("doing decoration, %r" % arg)
... def _wrapper(*args, **kwargs):
... print("inside wrapper, %r %r" % (args, kwargs))
Replacing or tweaking the original object ... return function(*args, **kwargs)
... return _wrapper
Decorators can either return the same function or class object or they can return a completely different object. ... return _decorator
In the first case, the decorator can exploit the fact that function and class objects are mutable and add attributes, >>> @replacing_decorator_with_args("abc")
e.g. add a docstring to a class. A decorator might do something useful even without modifying the object, for ... def function(*args, **kwargs):
example register the decorated class in a global registry. In the second case, virtually anything is possible: when ... print("inside function, %r %r" % (args, kwargs))
... return 14
something different is substituted for the original function or class, the new object can be completely different.
Nevertheless, such behaviour is not the purpose of decorators: they are intended to tweak the decorated object, not
do something unpredictable. Therefore, when a function is decorated by replacing it with a different function, >>> function(11, 12)
the new function usually calls the original function, after doing some preparatory work. Likewise, when a class is inside wrapper, (11, 12) {}
decorated by replacing if with a new class, the new class is usually derived from the original class. When the inside function, (11, 12) {}
purpose of the decorator is to do something every time, like to log every call to a decorated function, only the 14
second type of decorators can be used. On the other hand, if the first type is sufficient, it is better to use it, because
it is simpler. The _wrapper function is defined to accept all positional and keyword arguments. In general we cannot know
what arguments the decorated function is supposed to accept, so the wrapper function just passes everything to the
wrapped function. One unfortunate consequence is that the apparent argument list is misleading.
Decorators implemented as classes and as functions
Compared to decorators defined as functions, complex decorators defined as classes are simpler. When an object
The only requirement on decorators is that they can be called with a single argument. This means that decorators is created, the __init__ method is only allowed to return None, and the type of the created object cannot be
can be implemented as normal functions, or as classes with a __call__ method, or in theory, even as lambda changed. This means that when a decorator is defined as a class, it doesnt make much sense to use the argument-
functions. less form: the final decorated object would just be an instance of the decorating class, returned by the constructor
call, which is not very useful. Therefore its enough to discuss class-based decorators where arguments are given
Lets compare the function and class approaches. The decorator expression (the part after @) can be either just a in the decorator expression and the decorator __init__ method is used for decorator construction.
name, or a call. The bare-name approach is nice (less to type, looks cleaner, etc.), but is only possible when no
arguments are needed to customise the decorator. Decorators written as functions can be used in those two cases: >>> class decorator_class(object):
... def __init__(self, arg):
>>> def simple_decorator(function): ... # this method is called in the decorator expression
... print("doing decoration") ... print("in decorator init, %s" % arg)
... return function ... self.arg = arg
>>> @simple_decorator ... def __call__(self, function):
... def function(): ... # this method is called to do the job
... print("inside function") ... print("in decorator call, %s" % self.arg)
... return function functools.update_wrapper(wrapper, wrapped)

>>> deco_instance = decorator_class('foo')
Update a wrapper function to look like the wrapped function.
in decorator init, foo
>>> @deco_instance >>> import functools
... def function(*args, **kwargs): >>> def better_replacing_decorator_with_args(arg):
... print("in function, %s %s" % (args, kwargs)) ... print("defining the decorator")
in decorator call, foo ... def _decorator(function):
>>> function() ... print("doing decoration, %r" % arg)
in function, () {} ... def _wrapper(*args, **kwargs):
... print("inside wrapper, %r %r" % (args, kwargs))
Contrary to normal rules (PEP 8) decorators written as classes behave more like functions and therefore their ... return function(*args, **kwargs)
name often starts with a lowercase letter. ... return functools.update_wrapper(_wrapper, function)
... return _decorator
In reality, it doesnt make much sense to create a new class just to have a decorator which returns the original >>> @better_replacing_decorator_with_args("abc")
function. Objects are supposed to hold state, and such decorators are more useful when the decorator returns a ... def function():
new object. ... "extensive documentation"
... print("inside function")
>>> class replacing_decorator_class(object):
... return 14
... def __init__(self, arg):
... # this method is called in the decorator expression
... print("in decorator init, %s" % arg)
>>> function
... self.arg = arg
<function function at 0x...>
... def __call__(self, function):
>>> print(function.__doc__)
... # this method is called to do the job
extensive documentation
... print("in decorator call, %s" % self.arg)
... self.function = function
... return self._wrapper One important thing is missing from the list of attributes which can be copied to the replacement func-
... def _wrapper(self, *args, **kwargs):
tion: the argument list. The default values for arguments can be modified through the __defaults__,
... print("in the wrapper, %s %s" % (args, kwargs))
... return self.function(*args, **kwargs)
__kwdefaults__ attributes, but unfortunately the argument list itself cannot be set as an attribute. This means
>>> deco_instance = replacing_decorator_class('foo') that help(function) will display a useless argument list which will be confusing for the user of the function.
in decorator init, foo An effective but ugly way around this problem is to create the wrapper dynamically, using eval. This can be
>>> @deco_instance automated by using the external decorator module. It provides support for the decorator decorator, which
... def function(*args, **kwargs): takes a wrapper and turns it into a decorator which preserves the function signature.
... print("in function, %s %s" % (args, kwargs))
in decorator call, foo
To sum things up, decorators should always use functools.update_wrapper or some other means of
>>> function(11, 12) copying function attributes.
in the wrapper, (11, 12) {}
in function, (11, 12) {}
Examples in the standard library
A decorator like this can do pretty much anything, since it can modify the original function object and mangle the
arguments, call the original function or not, and afterwards mangle the return value. First, it should be mentioned that theres a number of useful decorators available in the standard library. There are
three decorators which really form a part of the language:
classmethod causes a method to become a class method, which means that it can be invoked without
Copying the docstring and other attributes of the original function
creating an instance of the class. When a normal method is invoked, the interpreter inserts the instance
object as the first positional parameter, self. When a class method is invoked, the class itself is given as
When a new function is returned by the decorator to replace the original function, an unfortunate consequence
the first parameter, often called cls.
is that the original function name, the original docstring, the original argument list are lost. Those attributes of
the original function can partially be transplanted to the new function by setting __doc__ (the docstring), Class methods are still accessible through the class namespace, so they dont pollute the modules names-
__module__ and __name__ (the full name of the function), and __annotations__ (extra information pace. Class methods can be used to provide alternative constructors:
about arguments and the return value of the function available in Python 3). This can be done automatically by
class Array(object):
using functools.update_wrapper. def __init__(self, data):
self.data = data
@classmethod
def fromfile(cls, file):
data = numpy.load(file)
return cls(data)
This is cleaner then using a multitude of flags to __init__.

staticmethod is applied to methods to make them static, i.e. basically a normal function, but acces-
sible through the class namespace. This can be useful when the function is only needed inside this class (its
name would then be prefixed with _), or when we want the user to think of the method as connected to the <property object at 0x...>
class, despite an implementation which doesnt require this. >>> D.a.fget
<function ...>
property is the pythonic answer to the problem of getters and setters. A method decorated with
>>> D.a.fset
property becomes a getter which is automatically called on attribute access.
<function ...>
>>> class A(object): >>> D.a.fdel
... @property <function ...>
... def a(self): >>> d = D() # ... varies, this is not the same à` function
... "an important attribute" >>> d.a
... return "a value" getting 1
>>> A.a 1
<property object at 0x...> >>> d.a = 2
>>> A().a setting 2
'a value' >>> del d.a
deleting
In this example, A.a is an read-only attribute. It is also documented: help(A) includes the docstring for >>> d.a
attribute a taken from the getter method. Defining a as a property allows it to be a calculated on the fly, and getting 1
has the side effect of making it read-only, because no setter is defined. 1
To have a setter and a getter, two methods are required, obviously. Since Python 2.6 the following syntax is Properties are a bit of a stretch for the decorator syntax. One of the premises of the decorator syntax that
preferred: the name is not duplicated is violated, but nothing better has been invented so far. It is just good style to
use the same name for the getter, setter, and deleter methods.
class Rectangle(object):
def __init__(self, edge): Some newer examples include:
self.edge = edge
functools.lru_cache memoizes an arbitrary function maintaining a limited cache of argu-
@property ments:answer pairs (Python 3.2)
def area(self):
functools.total_ordering is a class decorator which fills in missing ordering methods (__lt__,
"""Computed area.
__gt__, __le__, ...) based on a single available one (Python 2.7).
Setting this updates the edge length to the proper value.
"""
Deprecation of functions
return self.edge**2
@area.setter Lets say we want to print a deprecation warning on stderr on the first invocation of a function we dont like
def area(self, area): anymore. If we dont want to modify the function, we can use a decorator:
self.edge = area ** 0.5 class deprecated(object):
"""Print a deprecation warning once on first use of the function.
The way that this works, is that the property decorator replaces the getter method with a property object.
This object in turn has three methods, getter, setter, and deleter, which can be used as decorators. >>> @deprecated() # doctest: +SKIP
Their job is to set the getter, setter and deleter of the property object (stored as attributes fget, fset, and ... def f():
fdel). The getter can be set like in the example above, when creating the object. When defining the setter, ... pass
we already have the property object under area, and we add the setter to it by using the setter method. >>> f() # doctest: +SKIP
All this happens when we are creating the class. f is deprecated
"""
Afterwards, when an instance of the class has been created, the property object is special. When the inter- def __call__(self, func):
preter executes attribute access, assignment, or deletion, the job is delegated to the methods of the property self.func = func
object. self.count = 0
return self._wrapper
To make everything crystal clear, lets define a debug example: def _wrapper(self, *args, **kwargs):
>>> class D(object): self.count += 1
... @property if self.count == 1:
... def a(self): print self.func.__name__, 'is deprecated'
... print("getting 1") return self.func(*args, **kwargs)
... return 1
... @a.setter It can also be implemented as a function:
... def a(self, value):
def deprecated(func):
... print("setting %r" % value)
"""Print a deprecation warning once on first use of the function.
... @a.deleter
... def a(self):
>>> @deprecated # doctest: +SKIP
... print("deleting")
... def f():
>>> D.a
... pass
>>> f() # doctest: +SKIP

@WordProcessor.plugin
f is deprecated
class CleanMdashesExtension(object):
"""
def cleanup(self, text):
count = [0]
return text.replace('—', u'\N{em dash}')
def wrapper(*args, **kwargs):
count[0] += 1
if count[0] == 1:
Here we use a decorator to decentralise the registration of plugins. We call our decorator with a noun, instead of
print func.__name__, 'is deprecated' a verb, because we use it to declare that our class is a plugin for WordProcessor. Method plugin simply
return func(*args, **kwargs) appends the class to the list of plugins.
return wrapper A word about the plugin itself: it replaces HTML entity for em-dash with a real Unicode em-dash character. It
exploits the unicode literal notation to insert a character by using its name in the unicode database (EM DASH).
If the Unicode character was inserted directly, it would be impossible to distinguish it from an en-dash in the
A while-loop removing decorator source of a program.
Lets say we have function which returns a lists of things, and this list created by running a loop. If we dont know See also:
how many objects will be needed, the standard way to do this is something like: More examples and reading
def find_answers(): PEP 318 (function and method decorator syntax)
answers = []
while True: PEP 3129 (class decorator syntax)
ans = look_for_next_answer()
http://wiki.python.org/moin/PythonDecoratorLibrary
if ans is None:
break http://docs.python.org/dev/library/functools.html
answers.append(ans)
return answers http://pypi.python.org/pypi/decorator
Bruce Eckel
This is fine, as long as the body of the loop is fairly compact. Once it becomes more complicated, as often happens
in real code, this becomes pretty unreadable. We could simplify this by using yield statements, but then the user Decorators I: Introduction to Python Decorators
would have to explicitly call list(find_answers()).
Python Decorators II: Decorator Arguments
We can define a decorator which constructs the list for us:
Python Decorators III: A Decorator-Based Build System
def vectorized(generator_func):
def wrapper(*args, **kwargs):
return list(generator_func(*args, **kwargs)) 2.1.3 Context managers
return functools.update_wrapper(wrapper, generator_func)
A context manager is an object with __enter__ and __exit__ methods which can be used in the with state-
Our function then becomes: ment:
@vectorized with manager as var:
def find_answers(): do_something(var)
while True:
ans = look_for_next_answer() is in the simplest case equivalent to
if ans is None:
break var = manager.__enter__()
yield ans try:
do_something(var)
finally:
A plugin registration system manager.__exit__()
In other words, the context manager protocol defined in PEP 343 permits the extraction of the boring part of a
This is a class decorator which doesnt modify the class, but just puts it in a global registry. It falls into the category
try..except..finally structure into a separate class leaving only the interesting do_something block.
of decorators returning the original object:
1. The __enter__ method is called first. It can return a value which will be assigned to var. The as-part
class WordProcessor(object):
PLUGINS = []
is optional: if it isnt present, the value returned by __enter__ is simply ignored.
def process(self, text): 2. The block of code underneath with is executed. Just like with try clauses, it can either execute success-
for plugin in self.PLUGINS: fully to the end, or it can break, continue or return, or it can throw an exception. Either way, after the
text = plugin().cleanup(text) block is finished, the __exit__ method is called. If an exception was thrown, the information about the
return text
exception is passed to __exit__, which is described below in the next subsection. In the normal case,
@classmethod
exceptions can be ignored, just like in a finally clause, and will be rethrown after __exit__ is finished.
def plugin(cls, plugin): Lets say we want to make sure that a file is closed immediately after we are done writing to it:
cls.PLUGINS.append(plugin)
>>> class closing(object): class assert_raises(object):

... def __init__(self, obj): # based on pytest and unittest.TestCase
... self.obj = obj def __init__(self, type):
... def __enter__(self): self.type = type
... return self.obj def __enter__(self):
... def __exit__(self, *args): pass
... self.obj.close() def __exit__(self, type, value, traceback):
>>> with closing(open('/tmp/file', 'w')) as f: if type is None:
... f.write('the contents\n') raise AssertionError('exception expected')
if issubclass(type, self.type):
Here we have made sure that the f.close() is called when the with block is exited. Since closing files is such return True # swallow the expected exception
a common operation, the support for this is already present in the file class. It has an __exit__ method which raise AssertionError('wrong exception type')
calls close and can be used as a context manager itself:
with assert_raises(KeyError):
>>> with open('/tmp/file', 'a') as f: {}['foo']
... f.write('more contents\n')
The common use for try..finally is releasing resources. Various different cases are implemented similarly: Using generators to define context managers
in the __enter__ phase the resource is acquired, in the __exit__ phase it is released, and the exception, if
thrown, is propagated. As with files, theres often a natural operation to perform after the object has been used and When discussing generators (page 141), it was said that we prefer generators to iterators implemented as classes
it is most convenient to have the support built in. With each release, Python provides support in more places: because they are shorter, sweeter, and the state is stored as local, not instance, variables. On the other hand, as
all file-like objects: described in Bidirectional communication (page 142), the flow of data between the generator and its caller can
be bidirectional. This includes exceptions, which can be thrown into the generator. We would like to implement
file automatically closed context managers as special generator functions. In fact, the generator protocol was designed to support this use
fileinput, tempfile (py >= 3.2) case.
bz2.BZ2File, gzip.GzipFile, tarfile.TarFile, zipfile.ZipFile @contextlib.contextmanager

def some_generator(<arguments>):
ftplib, nntplib close connection (py >= 3.2 or 3.3) <setup>
try:
locks
yield <value>
multiprocessing.RLock lock and unlock finally:
<cleanup>
multiprocessing.Semaphore
memoryview automatically release (py >= 3.2 and 2.7) The contextlib.contextmanager helper takes a generator and turns it into a context manager. The gen-
erator has to obey some rules which are enforced by the wrapper function most importantly it must yield
decimal.localcontext modify precision of computations temporarily exactly once. The part before the yield is executed from __enter__, the block of code protected by the con-
_winreg.PyHKEY open and close hive key text manager is executed when the generator is suspended in yield, and the rest is executed in __exit__. If
an exception is thrown, the interpreter hands it to the wrapper through __exit__ arguments, and the wrapper
warnings.catch_warnings kill warnings temporarily function then throws it at the point of the yield statement. Through the use of generators, the context manager
contextlib.closing the same as the example above, call close is shorter and simpler.
parallel programming Lets rewrite the closing example as a generator:
concurrent.futures.ThreadPoolExecutor invoke in parallel then kill thread pool (py @contextlib.contextmanager

def closing(obj):
>= 3.2)
try:
concurrent.futures.ProcessPoolExecutor invoke in parallel then kill process pool yield obj
(py >= 3.2) finally:
obj.close()
nogil solve the GIL problem temporarily (cython only :( )
Lets rewrite the assert_raises example as a generator:
Catching exceptions @contextlib.contextmanager
def assert_raises(type):
When an exception is thrown in the with-block, it is passed as arguments to __exit__. Three arguments are try:
used, the same as returned by sys.exc_info(): type, value, traceback. When no exception is thrown, None yield
except type:
is used for all three arguments. The context manager can swallow the exception by returning a true value from
return
__exit__. Exceptions can be easily ignored, because if __exit__ doesnt use return and just falls of the except Exception as value:
end, None is returned, a false value, and therefore the exception is rethrown after __exit__ is finished. raise AssertionError('wrong exception type')
The ability to catch exceptions opens interesting possibilities. A classic example comes from unit-tests we else:
want to make sure that some code throws the right kind of exception: raise AssertionError('exception expected')
Chapter contents
Here we use a decorator to turn generator functions into context managers! Life of ndarray (page 156)
Its... (page 156)
Block of memory (page 157)
2.2 Advanced Numpy Data types (page 158)
Indexing scheme: strides (page 163)
Findings in dissection (page 169)
Author: Pauli Virtanen Universal functions (page 169)
Numpy is at the base of Pythons scientific stack of tools. Its purpose to implement efficient operations on many What they are? (page 169)
items in a block of memory. Understanding how it works in detail helps in making efficient use of its flexibility, Exercise: building an ufunc from scratch (page 170)
taking useful shortcuts. Solution: building an ufunc from scratch (page 174)
Generalized ufuncs (page 177)
This section covers: Interoperability features (page 178)
Anatomy of Numpy arrays, and its consequences. Tips and tricks. Sharing multidimensional, typed data (page 178)
The old buffer protocol (page 178)
Universal functions: what, why, and what to do if you want a new one. The old buffer protocol (page 179)
Integration with other tools: Numpy offers several ways to wrap any data in an ndarray, without unnecessary Array interface protocol (page 180)
copies. Array siblings: chararray, maskedarray, matrix (page 181)
chararray: vectorized string operations (page 181)
Recently added features, and whats in them: PEP 3118 buffers, generalized ufuncs, ... masked_array missing data (page 181)
recarray: purely convenience (page 184)
Prerequisites matrix: convenience? (page 184)
Summary (page 184)
Numpy Contributing to Numpy/Scipy (page 184)
Cython Why (page 184)
Pillow (Python imaging library, used in a couple of examples) Reporting bugs (page 184)
Contributing to documentation (page 185)
Contributing features (page 186)
How to help, in general (page 186)
Tip: In this section, numpy will be imported as follows:

2.2.1 Life of ndarray
Its...
ndarray =
block of memory + indexing scheme + data type descriptor
raw data
how to locate an element
how to interpret an element
2.2. Advanced Numpy 155 2.2. Advanced Numpy 156

>>> x[0] = 9
>>> y
array([9, 2, 3])
Memory does not need to be owned by an ndarray:

>>> x = b'1234' # The 'b' is for "bytes", necessary in Python 3
x is a string (in Python 3 a bytes), we can represent its data as an array of ints:
>>> y = np.frombuffer(x, dtype=np.int8)
>>> y.data
<... at ...>
>>> y.base is x
True
typedef struct PyArrayObject { >>> y.flags

PyObject_HEAD C_CONTIGUOUS : True
F_CONTIGUOUS : True
/* Block of memory */ OWNDATA : False
char *data; WRITEABLE : False
ALIGNED : True
/* Data type descriptor */ UPDATEIFCOPY : False
PyArray_Descr *descr;
The owndata and writeable flags indicate status of the memory block.
/* Indexing scheme */
int nd;
npy_intp *dimensions; Data types
npy_intp *strides;
The descriptor
/* Other stuff */
PyObject *base;
int flags; dtype describes a single item in the array:
PyObject *weakreflist;
type scalar type of the data, one of:
} PyArrayObject;
int8, int16, float64, et al. (fixed size)
str, unicode, void (flexible size)
itemsize size of the data block
Block of memory
byte- byte order: big-endian > / little-endian < / not applicable |
order
>>> x = np.array([1, 2, 3], dtype=np.int32)
>>> x.data
fields sub-dtypes, if its a structured data type
<... at ...> shape shape of the array, if its a sub-array
>>> str(x.data) >>> np.dtype(int).type
'\x01\x00\x00\x00\x02\x00\x00\x00\x03\x00\x00\x00' <type 'numpy.int64'>
>>> np.dtype(int).itemsize
Memory address of the data: 8
>>> x.__array_interface__['data'][0] >>> np.dtype(int).byteorder
64803824 '='
The whole __array_interface__:

Example: reading .wav files
>>> x.__array_interface__
{'data': (35828928, False),
'descr': [('', '<i4')], The .wav file header:
'shape': (4,),
'strides': None,
'typestr': '<i4',
'version': 3}
Reminder: two ndarrays may share the same memory:

>>> x = np.array([1, 2, 3, 4])
>>> y = x[:-1]

chunk_id "RIFF" Exercise

chunk_size 4-byte unsigned little-endian integer
format "WAVE" Mini-exercise, make a sparse dtype by using offsets, and only some of the fields:
fmt_id "fmt " >>> wav_header_dtype = np.dtype(dict(
fmt_size 4-byte unsigned little-endian integer ... names=['format', 'sample_rate', 'data_id'],
audio_fmt 2-byte unsigned little-endian integer ... offsets=[offset_1, offset_2, offset_3], # counted from start of structure in bytes
num_channels 2-byte unsigned little-endian integer ... formats=list of dtypes for each of the fields,
sample_rate 4-byte unsigned little-endian integer ... ))
byte_rate 4-byte unsigned little-endian integer and use that to read the sample rate, and data_id (as sub-array).
block_align 2-byte unsigned little-endian integer
bits_per_sample 2-byte unsigned little-endian integer
>>> f = open('data/test.wav', 'r')
data_id "data" >>> wav_header = np.fromfile(f, dtype=wav_header_dtype, count=1)
data_size 4-byte unsigned little-endian integer >>> f.close()
44-byte block of raw data (in the beginning of the file) >>> print(wav_header)
[ ('RIFF', 17402L, 'WAVE', 'fmt ', 16L, 1, 1, 16000L, 32000L, 2, 16, [['d', 'a'], ['t', 'a']], 173
... followed by data_size bytes of actual sound data. >>> wav_header['sample_rate']
array([16000], dtype=uint32)
The .wav file header as a Numpy structured data type:
>>> wav_header_dtype = np.dtype([ Lets try accessing the sub-array:
... ("chunk_id", (bytes, 4)), # flexible-sized scalar type, item size 4
>>> wav_header['data_id']
... ("chunk_size", "<u4"), # little-endian unsigned 32-bit integer
array([[['d', 'a'],
... ("format", "S4"), # 4-byte string
['t', 'a']]],
... ("fmt_id", "S4"),
dtype='|S1')
... ("fmt_size", "<u4"),
>>> wav_header.shape
... ("audio_fmt", "<u2"), #
(1,)
... ("num_channels", "<u2"), # .. more of the same ...
>>> wav_header['data_id'].shape
... ("sample_rate", "<u4"), #
(1, 2, 2)
... ("byte_rate", "<u4"),
... ("block_align", "<u2"),
When accessing sub-arrays, the dimensions get added to the end!
... ("bits_per_sample", "<u2"),
... ("data_id", ("S1", (2, 2))), # sub-array, just for fun!
... ("data_size", "u4"), Note: There are existing modules such as wavfile, audiolab, etc. for loading sound data...
... #
... # the sound data itself cannot be represented here:
... # it does not have a fixed size Casting and re-interpretation/views
... ])
See also: casting
wavreader.py on assignment
>>> wav_header_dtype['format'] on array construction

dtype('S4') on arithmetic
>>> wav_header_dtype.fields
etc. 'fmt_id': (dtype('S4'), 12), 'byte_rate': (dtype('uint32'), 28), 'chunk_id': (dtype('S4'),
dict_proxy({'block_align': (dtype('uint16'), 32), 'format': (dtype('S4'), 8), 'data_id': (dtype(('S1', (2, 2))), 36),
>>> wav_header_dtype.fields['format']
(dtype('S4'), 8)
and manually: .astype(dtype)
data re-interpretation
The first element is the sub-dtype in the structured data, corresponding to the name format
manually: .view(dtype)
The second one is its offset (in bytes) from the beginning of the item
Casting
Casting in arithmetic, in nutshell:
only type (not value!) of operands matters
largest safe type able to represent both is picked
scalars can lose to arrays in some situations
Casting in general copies data:

>>> x = np.array([1, 2, 3, 4], dtype=np.float)

>>> x Note:
array([ 1., 2., 3., 4.]) .view() makes views, does not copy (or alter) the memory block
>>> y = x.astype(np.int8)
>>> y only changes the dtype (and adjusts array shape):
array([1, 2, 3, 4], dtype=int8)
>>> x[1] = 5
>>> y + 1
>>> y
array([328193], dtype=int32)
>>> y + 256
>>> y.base is x
array([257, 258, 259, 260], dtype=int16)
True
>>> y + 256.0
array([ 257., 258., 259., 260.])
>>> y + np.array([256], dtype=np.int32)
array([257, 258, 259, 260], dtype=int32)
Mini-exercise: data re-interpretation
Casting on setitem: dtype of the array is not changed on item assignment:
See also:
>>> y[:] = y + 1.5
>>> y view-colors.py
You have RGBA data in an array:
Note: Exact rules: see documentation: http://docs.scipy.org/doc/numpy/reference/ufuncs.html#casting-rules >>> x = np.zeros((10, 10, 4), dtype=np.int8)
>>> x[:, :, 0] = 1
>>> x[:, :, 1] = 2
>>> x[:, :, 2] = 3
Re-interpretation / viewing >>> x[:, :, 3] = 4
Data block in memory (4 bytes)
where the last three dimensions are the R, B, and G, and alpha channels.
0x01 || 0x02 || 0x03 || 0x04
How to make a (10, 10) structured array with field names r, g, b, a without copying data?
4 of uint8, OR, >>> y = ...
4 of int8, OR,
>>> assert (y['r'] == 1).all()
2 of int16, OR, >>> assert (y['g'] == 2).all()
>>> assert (y['b'] == 3).all()
1 of int32, OR,
>>> assert (y['a'] == 4).all()
1 of float32, OR,
Solution
...
>>> y = x.view([('r', 'i1'),
How to switch from one to another? ... ('g', 'i1'),
1. Switch the dtype: ... ('b', 'i1'),
... ('a', 'i1')]
>>> x = np.array([1, 2, 3, 4], dtype=np.uint8) ... )[:, :, 0]
>>> x.dtype = "<i2"
>>> x
array([ 513, 1027], dtype=int16)
>>> 0x0201, 0x0403
(513, 1027)
0x01 0x02 || 0x03 0x04
Note: little-endian: least significant byte is on the left in memory
2. Create a new view:

>>> y = x.view("<i4")
>>> y
array([67305985], dtype=int32)
>>> 0x04030201
67305985
0x01 0x02 0x03 0x04

Warning: Another array taking exactly 4 bytes of memory: >>> y = np.array(x, order='F')
>>> y.strides
>>> y = np.array([[1, 3], [2, 4]], dtype=np.uint8).transpose() (2, 4)
>>> x = y.copy() >>> str(y.data)
>>> x '\x01\x00\x04\x00\x02\x00\x05\x00\x03\x00\x06\x00'
array([[1, 2],
[3, 4]], dtype=uint8) Need to jump 2 bytes to find the next row
>>> y
array([[1, 2], Need to jump 4 bytes to find the next column
[3, 4]], dtype=uint8) Similarly to higher dimensions:
>>> x.view(np.int16)
array([[ 513], C: last dimensions vary fastest (= smaller strides)
[1027]], dtype=int16)
F: first dimensions vary fastest
>>> 0x0201, 0x0403
(513, 1027)
>>> y.view(np.int16) shape = (1 , 2 , ..., )
array([[ 769, 1026]], dtype=int16) strides = (1 , 2 , ..., )
What happened?
= +1 +2 ... itemsize
... we need to look into what x[0,1] actually means

= 1 2 ...1 itemsize
>>> 0x0301, 0x0402

(769, 1026) Note: Now we can understand the behavior of .view():
>>> y = np.array([[1, 3], [2, 4]], dtype=np.uint8).transpose()
>>> x = y.copy()
Indexing scheme: strides Transposition does not affect the memory layout of the data, only strides
>>> x.strides
Main point
(2, 1)
>>> y.strides
The question: (1, 2)
>>> x = np.array([[1, 2, 3], >>> str(x.data)

... [4, 5, 6], '\x01\x02\x03\x04'
... [7, 8, 9]], dtype=np.int8) >>> str(y.data)
>>> str(x.data) '\x01\x03\x02\x04'
'\x01\x02\x03\x04\x05\x06\x07\x08\t'
the results are different when interpreted as 2 of int16
At which byte in ``x.data`` does the item ``x[1, 2]`` begin?
.copy() creates new arrays in the C order (by default)
The answer (in Numpy)
strides: the number of bytes to jump to find the next element
Slicing with integers
1 stride per dimension:
Everything can be represented by changing only shape, strides, and possibly adjusting the data
>>> x.strides
pointer!
(3, 1) >>> byte_offset = 3*1 + 1*2 # to find x[1, 2] >>> x.flat[byte_offset] 6 >>> x[1, 2] 6 Never makes copies of the data
simple, flexible >>> x = np.array([1, 2, 3, 4, 5, 6], dtype=np.int32)
>>> y = x[::-1]
C and Fortran order >>> y
array([6, 5, 4, 3, 2, 1], dtype=int32)
>>> x = np.array([[1, 2, 3], >>> y.strides
... [4, 5, 6]], dtype=np.int16, order='C') (-4,)
>>> x.strides
(6, 2) >>> y = x[2:]
>>> str(x.data) >>> y.__array_interface__['data'][0] - x.__array_interface__['data'][0]
'\x01\x00\x02\x00\x03\x00\x04\x00\x05\x00\x06\x00' 8
Need to jump 6 bytes to find the next row >>> x = np.zeros((10, 10, 10), dtype=np.float)
>>> x.strides
Need to jump 2 bytes to find the next column

(800, 80, 8) using only as_strided.:

>>> x[::2,::3,::4].strides Hint: byte_offset = stride[0]*index[0] + stride[1]*index[1] + ...
(1600, 240, 32)
Spoiler
Similarly, transposes never make copies (it just swaps strides):
Stride can also be 0:
>>> x = np.zeros((10, 10, 10), dtype=np.float)
>>> x.strides >>> x = np.array([1, 2, 3, 4], dtype=np.int8)
(800, 80, 8) >>> y = as_strided(x, strides=(0, 1), shape=(3, 4))
>>> x.T.strides >>> y
(8, 80, 800) array([[1, 2, 3, 4],
[1, 2, 3, 4],
But: not all reshaping operations can be represented by playing with strides: [1, 2, 3, 4]], dtype=int8)
>>> y.base.base is x
>>> a = np.arange(6, dtype=np.int8).reshape(3, 2) True
>>> b = a.T
>>> b.strides
(1, 2)
Broadcasting
So far, so good. However:
Doing something useful with it: outer product of [1, 2, 3, 4] and [5, 6, 7]
>>> str(a.data)
'\x00\x01\x02\x03\x04\x05' >>> x = np.array([1, 2, 3, 4], dtype=np.int16)
>>> b >>> x2 = as_strided(x, strides=(0, 1*2), shape=(3, 4))
array([[0, 2, 4], >>> x2
[1, 3, 5]], dtype=int8) array([[1, 2, 3, 4],
>>> c = b.reshape(3*2) [1, 2, 3, 4],
>>> c [1, 2, 3, 4]], dtype=int16)
array([0, 2, 4, 1, 3, 5], dtype=int8)
>>> y = np.array([5, 6, 7], dtype=np.int16)
Here, there is no way to represent the array c given one stride and the block of memory for a. Therefore, the >>> y2 = as_strided(y, strides=(1*2, 0), shape=(3, 4))
reshape operation needs to make a copy here. >>> y2
array([[5, 5, 5, 5],
[6, 6, 6, 6],
Example: fake dimensions with strides [7, 7, 7, 7]], dtype=int16)
Stride manipulation >>> x2 * y2

array([[ 5, 10, 15, 20],
[ 6, 12, 18, 24],
>>> from numpy.lib.stride_tricks import as_strided [ 7, 14, 21, 28]], dtype=int16)
>>> help(as_strided)
as_strided(x, shape=None, strides=None)
Make an ndarray from the given array with the given shape and strides ... seems somehow familiar ...
Warning: as_strided does not check that you stay inside the memory block bounds...
>>> x = np.array([1, 2, 3, 4], dtype=np.int16)
>>> y = np.array([5, 6, 7], dtype=np.int16)
>>> x = np.array([1, 2, 3, 4], dtype=np.int16) >>> x[np.newaxis,:] * y[:,np.newaxis]
>>> as_strided(x, strides=(2*2, ), shape=(2, )) array([[ 5, 10, 15, 20],
array([1, 3], dtype=int16) [ 6, 12, 18, 24],
>>> x[::2] [ 7, 14, 21, 28]], dtype=int16)
array([1, 3], dtype=int16)
Internally, array broadcasting is indeed implemented using 0-strides.
See also:
stride-fakedims.py More tricks: diagonals
Exercise
See also:
array([1, 2, 3, 4], dtype=np.int8)
stride-diagonals.py
-> array([[1, 2, 3, 4],
[1, 2, 3, 4],
Challenge
[1, 2, 3, 4]], dtype=np.int8) Pick diagonal entries of the matrix: (assume C memory order):

>>> x = np.array([[1, 2, 3], CPU cache effects

... [4, 5, 6],
... [7, 8, 9]], dtype=np.int32) Memory layout can affect performance:
>>> x_diag = as_strided(x, shape=(3,), strides=(???,)) In [1]: x = np.zeros((20000,))
Pick the first super-diagonal entries [2, 6]. In [2]: y = np.zeros((20000*67,))[::67]
And the sub-diagonals? In [3]: x.shape, y.shape

((20000,), (20000,))
(Hint to the last two: slicing first moves the point where striding starts from.)
Solution In [4]: %timeit x.sum()
Pick diagonals:
>>> x_diag = as_strided(x, shape=(3, ), strides=((3+1)*x.itemsize, )) In [5]: %timeit y.sum()
>>> x_diag 100000 loops, best of 3: 2.34 ms per loop
array([1, 5, 9], dtype=int32)
In [6]: x.strides, y.strides
((8,), (536,))
Slice first, to adjust the data pointer:
>>> as_strided(x[0, 1:], shape=(2, ), strides=((3+1)*x.itemsize, ))
array([2, 6], dtype=int32) Smaller strides are faster?
>>> as_strided(x[1:, 0], shape=(2, ), strides=((3+1)*x.itemsize, ))

Note: Using np.diag

>>> y = np.diag(x, k=1)
>>> y
However,
>>> y.flags.owndata CPU pulls data from main memory to its cache in blocks
False
If many array items consecutively operated on fit in a single block (small stride):
Note This behavior has changed: before numpy 1.9, np.diag would make a copy.
fewer transfers needed
See also: faster
stride-diagonals.py See also:
Challenge numexpr is designed to mitigate cache effects in array computing.
Compute the tensor trace:

Example: inplace operations (caveat emptor)
>>> x = np.arange(5*5*5*5).reshape(5, 5, 5, 5)
>>> s = 0
>>> for i in range(5): Sometimes,
... for j in range(5):
>>> a -= b
... s += x[j, i, j, i]
is not the same as
by striding, and using sum() on the result.
>>> a -= b.copy()
>>> y = as_strided(x, shape=(5, 5), strides=(TODO, TODO))
>>> s2 = ...
>>> x = np.array([[1, 2], [3, 4]])
>>> assert s == s2
>>> x -= x.transpose()
>>> x
Solution array([[ 0, -1],
>>> y = as_strided(x, shape=(5, 5), strides=((5*5*5 + 5)*x.itemsize, [ 4, 0]])
... (5*5 + 1)*x.itemsize))
>>> s2 = y.sum()

>>> y = np.array([[1, 2], [3, 4]]) * This function must compute the ufunc for many values at once,
>>> y -= y.T.copy() * in the way shown below.
>>> y */
array([[ 0, -1], char *input_1 = (char*)args[0];
[ 1, 0]]) char *input_2 = (char*)args[1];
char *output = (char*)args[2];
x and x.transpose() share data int i;
x -= x.transpose() modifies the data element-by-element... for (i = 0; i < dimensions[0]; ++i) {

because x and x.transpose() have different striding, modified data re-appears on the RHS *output = elementwise_function(*input_1, *input_2);
input_1 += steps[0];
input_2 += steps[1];
Findings in dissection output += steps[2];
}
}
2. The Numpy part, built by

char types[3]
types[0] = NPY_BYTE /* type of first input arg */

types[1] = NPY_BYTE /* type of second input arg */
types[2] = NPY_BYTE /* type of third input arg */
PyObject *python_ufunc = PyUFunc_FromFuncAndData(

ufunc_loop,
NULL,
types,
1, /* ntypes */
2, /* num_inputs */
1, /* num_outputs */
memory block: may be shared, .base, .data
identity_element,
data type descriptor: structured data, sub-arrays, byte order, casting, viewing, .astype(), .view() name,
docstring,
strided indexing: strides, C/F-order, slicing w/ integers, as_strided, broadcasting, stride tricks, diag, unused)
CPU cache coherence
A ufunc can also support multiple different input-output type combinations.
2.2.2 Universal functions

Making it easier
What they are?
3. ufunc_loop is of very generic form, and Numpy provides pre-made ones
Ufunc performs and elementwise operation on all elements of an array. PyUfunc_f_f float elementwise_func(float input_1)
Examples: PyUfunc_ff_ffloat elementwise_func(float input_1, float input_2)
PyUfunc_d_d double elementwise_func(double input_1)
np.add, np.subtract, scipy.special.*, ...
PyUfunc_dd_ddouble elementwise_func(double input_1, double input_2)
Automatically support: broadcasting, casting, ... PyUfunc_D_D elementwise_func(npy_cdouble *input, npy_cdouble* output)
PyUfunc_DD_Delementwise_func(npy_cdouble *in1, npy_cdouble *in2,
The author of an ufunc only has to supply the elementwise operation, Numpy takes care of the rest. npy_cdouble* out)
The elementwise operation needs to be implemented in C (or, e.g., Cython) Only elementwise_func needs to be supplied
... except when your elementwise function is not in one of the above forms
Parts of an Ufunc
Exercise: building an ufunc from scratch

1. Provided by user
void ufunc_loop(void **args, int *dimensions, int *steps, void *data) The Mandelbrot fractal is defined by the iteration
{
/* 2 +
* int8 output = elementwise_function(int8 input_1, int8 input_2) where = + is a complex number. This iteration is repeated if stays finite no matter how long the iteration
*
runs, belongs to the Mandelbrot set.

Make ufunc called mandel(z0, c) that computes: # Say, use 100 as the maximum number of iterations, and 1000
z = z0 # as the cutoff for z.real**2 + z.imag**2.
for k in range(iterations): #
z = z*z + c
TODO: mandelbrot iteration should go here
say, 100 iterations or until z.real**2 + z.imag**2 > 1000. Use it to determine which c are in the
# Return the answer for this point
Mandelbrot set.
z_out[0] = z
Our function is a simple one, so make use of the PyUFunc_* helpers.
Write it in Cython # Boilerplate Cython definitions
See also: #
# The litany below is particularly long, but you don't really need to
mandel.pyx, mandelplot.py # read this part; it just pulls in stuff from the Numpy C headers.
# ----------------------------------------------------------
#
# Fix the parts marked by TODO cdef extern from "numpy/arrayobject.h":
# void import_array()
ctypedef int npy_intp
# cdef enum NPY_TYPES:
# Compile this file by (Cython >= 0.12 required because of the complex vars) NPY_DOUBLE
# NPY_CDOUBLE
# cython mandel.pyx NPY_LONG
# python setup.py build_ext -i
# cdef extern from "numpy/ufuncobject.h":
# and try it out with, in this directory, void import_ufunc()
# ctypedef void (*PyUFuncGenericFunction)(char**, npy_intp*, npy_intp*, void*)
# >>> import mandel object PyUFunc_FromFuncAndData(PyUFuncGenericFunction* func, void** data,
# >>> mandel.mandel(0, 1 + 2j) char* types, int ntypes, int nin, int nout,
# int identity, char* name, char* doc, int c)
#
# List of pre-defined loop functions
# The elementwise function
# ------------------------ void PyUFunc_f_f_As_d_d(char** args, npy_intp* dimensions, npy_intp* steps, void* func)
void PyUFunc_d_d(char** args, npy_intp* dimensions, npy_intp* steps, void* func)
cdef void mandel_single_point(double complex *z_in, void PyUFunc_f_f(char** args, npy_intp* dimensions, npy_intp* steps, void* func)
double complex *c_in, void PyUFunc_g_g(char** args, npy_intp* dimensions, npy_intp* steps, void* func)
double complex *z_out) nogil: void PyUFunc_F_F_As_D_D(char** args, npy_intp* dimensions, npy_intp* steps, void* func)
# void PyUFunc_F_F(char** args, npy_intp* dimensions, npy_intp* steps, void* func)
# The Mandelbrot iteration void PyUFunc_D_D(char** args, npy_intp* dimensions, npy_intp* steps, void* func)
# void PyUFunc_G_G(char** args, npy_intp* dimensions, npy_intp* steps, void* func)
void PyUFunc_ff_f_As_dd_d(char** args, npy_intp* dimensions, npy_intp* steps, void* func)
# void PyUFunc_ff_f(char** args, npy_intp* dimensions, npy_intp* steps, void* func)
# Some points of note: void PyUFunc_dd_d(char** args, npy_intp* dimensions, npy_intp* steps, void* func)
# void PyUFunc_gg_g(char** args, npy_intp* dimensions, npy_intp* steps, void* func)
# - It's *NOT* allowed to call any Python functions here. void PyUFunc_FF_F_As_DD_D(char** args, npy_intp* dimensions, npy_intp* steps, void* func)
# void PyUFunc_DD_D(char** args, npy_intp* dimensions, npy_intp* steps, void* func)
# The Ufunc loop runs with the Python Global Interpreter Lock released. void PyUFunc_FF_F(char** args, npy_intp* dimensions, npy_intp* steps, void* func)
# Hence, the ``nogil``. void PyUFunc_GG_G(char** args, npy_intp* dimensions, npy_intp* steps, void* func)
#
# - And so all local variables must be declared with ``cdef``
# # Required module initialization
# - Note also that this function receives *pointers* to the data # ------------------------------
#
import_array()
cdef double complex z = z_in[0] import_ufunc()
cdef double complex c = c_in[0]
cdef int k # the integer we use in the for loop
# The actual ufunc declaration
# # ----------------------------
# TODO: write the Mandelbrot iteration for one point here,
# as you would write it in Python. cdef PyUFuncGenericFunction loop_func[1]
# cdef char input_output_types[3]

cdef void *elementwise_funcs[1] NPY_LONGDOUBLE, NPY_CFLOAT, NPY_CDOUBLE, NPY_CLONGDOUBLE, NPY_DATETIME,

NPY_TIMEDELTA, NPY_OBJECT, NPY_STRING, NPY_UNICODE, NPY_VOID
#
# Reminder: some pre-made Ufunc loops:
# Solution: building an ufunc from scratch
# ================ =======================================================
# ``PyUfunc_f_f`` ``float elementwise_func(float input_1)``
# The elementwise function
# ``PyUfunc_ff_f`` ``float elementwise_func(float input_1, float input_2)``
# ------------------------
# ``PyUfunc_d_d`` ``double elementwise_func(double input_1)``
# ``PyUfunc_dd_d`` ``double elementwise_func(double input_1, double input_2)``
cdef void mandel_single_point(double complex *z_in,
# ``PyUfunc_D_D`` `èlementwise_func(complex_double *input, complex_double* complex_double)``
double complex *c_in,
# ``PyUfunc_DD_D`` `èlementwise_func(complex_double *in1, complex_double *in2, complex_double* out)``
double complex *z_out) nogil:
# ================ =======================================================
#
#
# The Mandelbrot iteration
# The full list is above.
#
#
#
#
# Type codes:
# Some points of note:
#
#
# NPY_BOOL, NPY_BYTE, NPY_UBYTE, NPY_SHORT, NPY_USHORT, NPY_INT, NPY_UINT,
# - It's *NOT* allowed to call any Python functions here.
# NPY_LONG, NPY_ULONG, NPY_LONGLONG, NPY_ULONGLONG, NPY_FLOAT, NPY_DOUBLE,
#
# NPY_LONGDOUBLE, NPY_CFLOAT, NPY_CDOUBLE, NPY_CLONGDOUBLE, NPY_DATETIME,
# The Ufunc loop runs with the Python Global Interpreter Lock released.
# NPY_TIMEDELTA, NPY_OBJECT, NPY_STRING, NPY_UNICODE, NPY_VOID
# Hence, the ``nogil``.
#
#
# - And so all local variables must be declared with ``cdef``
loop_func[0] = ... TODO: suitable PyUFunc_* ...
#
input_output_types[0] = ... TODO ...
# - Note also that this function receives *pointers* to the data;
... TODO: fill in rest of input_output_types ...
# the "traditional" solution to passing complex variables around
#
# This thing is passed as the ``data`` parameter for the generic
# PyUFunc_* loop, to let it know which function it should call.
cdef double complex z = z_in[0]
elementwise_funcs[0] = <void*>mandel_single_point
cdef double complex c = c_in[0]
cdef int k # the integer we use in the for loop
# Construct the ufunc:
# Straightforward iteration
mandel = PyUFunc_FromFuncAndData(
loop_func,
for k in range(100):
elementwise_funcs,
z = z*z + c
input_output_types,
if z.real**2 + z.imag**2 > 1000:
1, # number of supported input types
break
TODO, # number of input args
TODO, # number of output args
# Return the answer for this point
0, # ìdentity` element, never mind this
z_out[0] = z
"mandel", # function name
"mandel(z, c) -> computes z*z + c", # docstring
0 # unused
# Boilerplate Cython definitions
)
#
# You don't really need to read this part, it just pulls in
Reminder: some pre-made Ufunc loops:
# stuff from the Numpy C headers.
PyUfunc_f_f float elementwise_func(float input_1) # ----------------------------------------------------------
PyUfunc_ff_ffloat elementwise_func(float input_1, float input_2)
PyUfunc_d_d double elementwise_func(double input_1) cdef extern from "numpy/arrayobject.h":
void import_array()
PyUfunc_dd_ddouble elementwise_func(double input_1, double input_2)
ctypedef int npy_intp
PyUfunc_D_D elementwise_func(complex_double *input, complex_double* cdef enum NPY_TYPES:
output) NPY_CDOUBLE
PyUfunc_DD_Delementwise_func(complex_double *in1, complex_double *in2,
complex_double* out) cdef extern from "numpy/ufuncobject.h":
void import_ufunc()
Type codes: ctypedef void (*PyUFuncGenericFunction)(char**, npy_intp*, npy_intp*, void*)
NPY_BOOL, NPY_BYTE, NPY_UBYTE, NPY_SHORT, NPY_USHORT, NPY_INT, NPY_UINT, object PyUFunc_FromFuncAndData(PyUFuncGenericFunction* func, void** data,
NPY_LONG, NPY_ULONG, NPY_LONGLONG, NPY_ULONGLONG, NPY_FLOAT, NPY_DOUBLE, char* types, int ntypes, int nin, int nout,

int identity, char* name, char* doc, int c)
void PyUFunc_DD_D(char**, npy_intp*, npy_intp*, void*)
# Required module initialization

# ------------------------------
import_array()
import_ufunc()
# The actual ufunc declaration

# ----------------------------
cdef PyUFuncGenericFunction loop_func[1]

cdef char input_output_types[3]
Note: Most of the boilerplate could be automated by these Cython modules:
cdef void *elementwise_funcs[1]
http://wiki.cython.org/MarkLodato/CreatingUfuncs
loop_func[0] = PyUFunc_DD_D
input_output_types[0] = NPY_CDOUBLE
input_output_types[1] = NPY_CDOUBLE Several accepted input types
E.g. supporting both single- and double-precision versions
elementwise_funcs[0] = <void*>mandel_single_point
cdef void mandel_single_point(double complex *z_in,
mandel = PyUFunc_FromFuncAndData( double complex *c_in,
loop_func, double complex *z_out) nogil:
elementwise_funcs, ...
input_output_types,
1, # number of supported input types cdef void mandel_single_point_singleprec(float complex *z_in,
2, # number of input args float complex *c_in,
1, # number of output args float complex *z_out) nogil:
0, # ìdentity` element, never mind this ...
"mandel(z, c) -> computes iterated z*z + c", # docstring cdef PyUFuncGenericFunction loop_funcs[2]
0 # unused cdef char input_output_types[3*2]
) cdef void *elementwise_funcs[1*2]
""" loop_funcs[0] = PyUFunc_DD_D

Plot Mandelbrot input_output_types[0] = NPY_CDOUBLE
================ input_output_types[1] = NPY_CDOUBLE
Plot the Mandelbrot ensemble. elementwise_funcs[0] = <void*>mandel_single_point
""" loop_funcs[1] = PyUFunc_FF_F

input_output_types[3] = NPY_CFLOAT
import numpy as np input_output_types[4] = NPY_CFLOAT
import mandel input_output_types[5] = NPY_CFLOAT
x = np.linspace(-1.7, 0.6, 1000) elementwise_funcs[1] = <void*>mandel_single_point_singleprec
y = np.linspace(-1.4, 1.4, 1000)
c = x[None,:] + 1j*y[:,None] mandel = PyUFunc_FromFuncAndData(
z = mandel.mandel(c, c) loop_func,
elementwise_funcs,
import matplotlib.pyplot as plt input_output_types,
plt.imshow(abs(z)**2 < 1000, extent=[-1.7, 0.6, -1.4, 1.4]) 2, # number of supported input types <----------------
plt.gray() 2, # number of input args
plt.show() 1, # number of output args
0, # ìdentity` element, never mind this
"mandel(z, c) -> computes iterated z*z + c", # docstring
0 # unused

) void gufunc_loop(void **args, int *dimensions, int *steps, void *data)

{
char *input_1 = (char*)args[0]; /* these are as previously */
Generalized ufuncs char *input_2 = (char*)args[1];
char *output = (char*)args[2];
ufunc
int input_1_stride_m = steps[3]; /* strides for the core dimensions */
output = elementwise_function(input) int input_1_stride_n = steps[4]; /* are added after the non-core */
int input_2_strides_n = steps[5]; /* steps */
Both output and input can be a single array element only. int input_2_strides_p = steps[6];
int output_strides_n = steps[7];
generalized ufunc
int output_strides_p = steps[8];
output and input can be arrays with a fixed number of dimensions
int m = dimension[1]; /* core dimensions are added after */
For example, matrix trace (sum of diag elements): int n = dimension[2]; /* the main dimension; order as in */
input shape = (n, n) int p = dimension[3]; /* signature */
output shape = () i.e. scalar
int i;
(n, n) -> ()
for (i = 0; i < dimensions[0]; ++i) {
Matrix product: matmul_for_strided_matrices(input_1, input_2, output,
strides for each array...);
input_1 shape = (m, n)
input_2 shape = (n, p) input_1 += steps[0];
output shape = (m, p) input_2 += steps[1];
output += steps[2];
(m, n), (n, p) -> (m, p) }
}
This is called the signature of the generalized ufunc
The dimensions on which the g-ufunc acts, are core dimensions
2.2.3 Interoperability features
Status in Numpy Sharing multidimensional, typed data
g-ufuncs are in Numpy already ... Suppose you

new ones can be created with PyUFunc_FromFuncAndDataAndSignature 1. Write a library than handles (multidimensional) binary data,
... but we dont ship with public g-ufuncs, except for testing, ATM 2. Want to make it easy to manipulate the data with Numpy, or whatever other library,
>>> import numpy.core.umath_tests as ut 3. ... but would not like to have Numpy as a dependency.
>>> ut.matrix_multiply.signature
'(m,n),(n,p)->(m,p)'
Currently, 3 solutions:
1. the old buffer interface
>>> x = np.ones((10, 2, 4))
>>> y = np.ones((10, 4, 5)) 2. the array interface
>>> ut.matrix_multiply(x, y).shape 3. the new buffer interface (PEP 3118)
(10, 2, 5)
the last two dimensions became core dimensions, and are modified as per the signature The old buffer protocol
otherwise, the g-ufunc operates elementwise
Only 1-D buffers
matrix multiplication this way could be useful for operating on many small matrices at once
No data type information
Generalized ufunc loop C-level interface; PyBufferProcs tp_as_buffer in the type object
But its integrated into Python (e.g. strings support it)
Matrix multiplication (m,n),(n,p) -> (m,p) Mini-exercise using Pillow (Python Imaging Library):
See also:
pilbuffer.py

>>> from PIL import Image

>>> data = np.zeros((200, 200, 4), dtype=np.int8)
>>> data[:, :] = [255, 0, 0, 255] # Red
>>> # In PIL, RGBA images consist of 32-bit integers whose bytes are [RR,GG,BB,AA]
>>> data = data.view(np.int32).squeeze()
>>> img = Image.frombuffer("RGBA", (200, 200), data, "raw", "RGBA", 0, 1)
>>> img.save('test.png')
Q:
Check what happens if data is now modified, and img saved again.
The old buffer protocol
"""
From buffer
============
Show how to exchange data between numpy and a library that only knows
the buffer interface.
"""
import numpy as np
import Image
# Let's make a sample image, RGBA format
x = np.zeros((200, 200, 4), dtype=np.int8)
x[:,:,0] = 254 # red

x[:,:,3] = 255 # opaque
data = x.view(np.int32) # Check that you understand why this is OK!
img = Image.frombuffer("RGBA", (200, 200), data)

img.save('test.png')
#
# Modify the original data, and save again. Array interface protocol
#
# It turns out that PIL, which knows next to nothing about Numpy, Multidimensional buffers
# happily shares the same data.
# Data type information present
Numpy-specific approach; slowly deprecated (but not going away)
x[:,:,1] = 254
img.save('test2.png') Not integrated in Python otherwise
See also:
Documentation: http://docs.scipy.org/doc/numpy/reference/arrays.interface.html
>>> x = np.array([[1, 2], [3, 4]])
>>> x.__array_interface__
{'data': (171694552, False), # memory address of data, is readonly?
'descr': [('', '<i4')], # data type descriptor
'typestr': '<i4', # same, in another form
'strides': None, # strides; or None if in C-order
'shape': (2, 2),
'version': 3,
}
::

>>> from PIL import Image The mask

>>> img = Image.open('data/test.png')
>>> img.__array_interface__ You can modify the mask by assigning:
{'data': ...,
'shape': (200, 200, 4), >>> mx[1] = np.ma.masked
'typestr': '|u1'} >>> mx
>>> x = np.asarray(img) masked_array(data = [1 -- 3 -- 5],
>>> x.shape mask = [False True False True False],
(200, 200, 4) fill_value = 999999)
The mask is cleared on assignment:

Note: A more C-friendly variant of the array interface is also defined.
>>> mx[1] = 9
>>> mx
masked_array(data = [1 9 3 -- 5],
2.2.4 Array siblings: chararray, maskedarray, matrix mask = [False False False True False],
fill_value = 999999)
chararray: vectorized string operations
The mask is also available directly:
>>> x = np.array(['a', ' bbb', ' ccc']).view(np.chararray) >>> mx.mask
>>> x.lstrip(' ') array([False, False, False, True, False], dtype=bool)
chararray(['a', 'bbb', 'ccc'],
dtype='...') The masked entries can be filled with a given value to get an usual array back:
>>> x.upper()
chararray(['A', ' BBB', ' CCC'], >>> x2 = mx.filled(-1)
dtype='...') >>> x2
array([ 1, 9, 3, -1, 5])
Note: .view() has a second meaning: it can make an ndarray an instance of a specialized ndarray subclass
The mask can also be cleared:
>>> mx.mask = np.ma.nomask
>>> mx
masked_array missing data
masked_array(data = [1 9 3 -99 5],
mask = [False False False False False],
Masked arrays are arrays that may have missing or invalid entries. fill_value = 999999)
For example, suppose we have an array where the fourth entry is invalid:
>>> x = np.array([1, 2, 3, -99, 5]) Domain-aware functions
One way to describe this is to create a masked array:

The masked array package also contains domain-aware functions:
>>> mx = np.ma.masked_array(x, mask=[0, 0, 0, 1, 0])
>>> mx >>> np.ma.log(np.array([1, 2, -1, -2, 3, -5]))
masked_array(data = [1 2 3 -- 5], masked_array(data = [0.0 0.6931471805599453 -- -- 1.0986122886681098 --],
mask = [False False False True False], mask = [False False True True False True],
fill_value = 999999) fill_value = 1e+20)
Masked mean ignores masked data: Note: Streamlined and more seamless support for dealing with missing data in arrays is making its way into
>>> mx.mean()
Numpy 1.7. Stay tuned!
2.75
>>> np.mean(mx)
2.75
Warning: Not all Numpy functions respect masks, for instance np.dot, so check the return types.
The masked_array returns a view to the original array:

>>> mx[1] = 9
>>> x
array([ 1, 9, 3, -99, 5])

Example: Masked statistics recarray: purely convenience
Canadian rangers were distracted when counting hares and lynxes in 1903-1910 and 1917-1918, and got >>> arr = np.array([('a', 1), ('b', 2)], dtype=[('x', 'S1'), ('y', int)])
the numbers are wrong. (Carrot farmers stayed alert, though.) Compute the mean populations over time, >>> arr2 = arr.view(np.recarray)
ignoring the invalid numbers. >>> arr2.x
chararray(['a', 'b'],
>>> data = np.loadtxt('data/populations.txt')
dtype='|S1')
>>> populations = np.ma.masked_array(data[:,1:])
>>> arr2.y
>>> year = data[:, 0]
array([1, 2])
>>> bad_years = (((year >= 1903) & (year <= 1910))
... | ((year >= 1917) & (year <= 1918)))
>>> # '&' means 'and' and '|' means 'or' matrix: convenience?
>>> populations[bad_years, 0] = np.ma.masked
>>> populations[bad_years, 1] = np.ma.masked always 2-D
>>> populations.mean(axis=0) * is the matrix product, not the elementwise one

masked_array(data = [40472.72727272727 18627.272727272728 42400.0],
>>> np.matrix([[1, 0], [0, 1]]) * np.matrix([[1, 2], [3, 4]])
mask = [False False False],
matrix([[1, 2],
fill_value = 1e+20)
[3, 4]])
>>> populations.std(axis=0)
masked_array(data = [21087.656489006717 15625.799814240254 3322.5062255844787],
mask = [False False False], 2.2.5 Summary
fill_value = 1e+20)
Note that Matplotlib knows about masked arrays: Anatomy of the ndarray: data, dtype, strides.
>>> plt.plot(year, populations, 'o-') Universal functions: elementwise operations, how to make new ones
[<matplotlib.lines.Line2D object at ...>, ...] Ndarray subclasses
Various buffer interfaces for integration with other tools
80000 Recent additions: PEP 3118, generalized ufuncs
70000 2.2.6 Contributing to Numpy/Scipy
Get this tutorial: http://www.euroscipy.org/talk/882

60000
Why
50000
Theres a bug?
40000 I dont understand what this is supposed to do?
I have this fancy code. Would you like to have it?
30000 Id like to help! What can I do?
20000 Reporting bugs
10000 Bug tracker (prefer this)

http://projects.scipy.org/numpy
0 http://projects.scipy.org/scipy
1900 1905 1910 1915 1920 Click the Register link to get an account
Mailing lists ( scipy.org/Mailing_Lists )
If youre unsure
No replies in a week or so? Just file a bug ticket.

Good bug report * change your subscription options, at the bottom of

http://mail.scipy.org/mailman/listinfo/scipy-dev
Title: numpy.random.permutations fails for non-integer arguments
Send a mail @ scipy-dev mailing list; ask for activation:
I'm trying to generate random permutations, using numpy.random.permutations
To: scipy-dev@scipy.org
When calling numpy.random.permutation with non-integer arguments
it fails with a cryptic error message:: Hi,
>>> np.random.permutation(12) I'd like to edit Numpy/Scipy docstrings. My account is XXXXX

array([ 6, 11, 4, 10, 2, 8, 1, 7, 9, 3, 0, 5])
>>> np.random.permutation(12.) Cheers,
Traceback (most recent call last): N. N.
File "mtrand.pyx", line 3311, in mtrand.RandomState.permutation Check the style guide:
File "mtrand.pyx", line 3254, in mtrand.RandomState.shuffle
http://docs.scipy.org/numpy/
TypeError: len() of unsized object
Dont be intimidated; to fix a small thing, just fix it
This also happens with long arguments, and so
np.random.permutation(X.shape[0]) where X is an array fails on 64 Edit
bit windows (where shape is a tuple of longs).
2. Edit sources and send patches (as for bugs)
It would be great if it could cast to integer or at least raise a 3. Complain on the mailing list
proper error for non-integer types.
I'm using Numpy 1.4.1, built from the official tarball, on Windows Contributing features
64 with Visual studio 2008, on Python.org 64-bit Python.
0. Ask on mailing list, if unsure where it should go
0. What are you trying to do?
1. Write a patch, add an enhancement ticket on the bug tracket
1. Small code snippet reproducing the bug (if possible)
2. OR, create a Git branch implementing the feature + add enhancement ticket.
What actually happens
Especially for big/invasive additions
What youd expect
http://projects.scipy.org/numpy/wiki/GitMirror
2. Platform (Windows / Linux / OSX, 32/64 bits, x86/PPC, ...)
http://www.spheredev.org/wiki/Git_for_the_lazy
3. Version of Numpy/Scipy
# Clone numpy repository
>>> print(np.__version__) git clone --origin svn http://projects.scipy.org/git/numpy.git numpy
1... cd numpy
Check that the following is what you expect # Create a feature branch
git checkout -b name-of-my-feature-branch svn/trunk
>>> print(np.__file__)
/...
<edit stuff>
In case you have old/broken Numpy installations lying around. git commit -a
If unsure, try to remove existing Numpy installations, and reinstall...
Create account on http://github.com (or anywhere)
Contributing to documentation Create a new repository @ Github

Push your work to github
1. Documentation editor
git remote add github git@github:YOURUSERNAME/YOURREPOSITORYNAME.git
http://docs.scipy.org/numpy git push github name-of-my-feature-branch
Registration
Register an account How to help, in general
Subscribe to scipy-dev mailing list (subscribers-only)
Bug fixes always welcome!
Problem with mailing lists: you get mail
What irks you most
* But: you can turn mail delivery off

Browse the tracker

Documentation work Brian Kernighan
API docs: improvements to docstrings Everyone knows that debugging is twice as hard as writing a program in the first place. So if youre as
clever as you can be when you write it, how will you ever debug it?
* Know some Scipy module well?
User guide We all write buggy code. Accept it. Deal with it.
* Needs to be done eventually. Write your code with testing and debugging in mind.
* Want to think? Come up with a Table of Contents Keep It Simple, Stupid (KISS).
http://scipy.org/Developer_Zone/UG_Toc What is the simplest thing that could possibly work?
Ask on communication channels: Dont Repeat Yourself (DRY).
numpy-discussion list Every piece of knowledge must have a single, unambiguous, authoritative representation within a
system.
scipy-dev list
Constants, algorithms, etc...
Try to limit interdependencies of your code. (Loose Coupling)
2.3 Debugging code
Give your variables, functions and modules meaningful names (not mathematics names)
Author: Gal Varoquaux
pyflakes: fast static analysis
This section explores tools to understand better your code base: debugging, to find and fix bugs.
It is not specific to the scientific Python community, but the strategies that we will employ are tailored to its needs. They are several static analysis tools in Python; to name a few:
pylint
Prerequisites
pychecker
Numpy
pyflakes
IPython
nosetests (http://readthedocs.org/docs/nose/en/latest/) pep8
pyflakes (http://pypi.python.org/pypi/pyflakes)
flake8
gdb for the C-debugging part.
Here we focus on pyflakes, which is the simplest tool.
Fast, simple
Chapters contents Detects syntax errors, missing imports, typos on names.
Avoiding bugs (page 187) Another good recommendation is the flake8 tool which is a combination of pyflakes and pep8. Thus, in addition
Coding best practices to avoid getting in trouble (page 187) to the types of errors that pyflakes catches, flake8 detects violations of the recommendation in PEP8 style guide.
pyflakes: fast static analysis (page 188)
* Running pyflakes on the current edited file (page 188) Integrating pyflakes (or flake8) in your editor or IDE is highly recommended, it does yield productivity gains.
* A type-as-go spell-checker like integration (page 189)
Debugging workflow (page 190)
Running pyflakes on the current edited file
Using the Python debugger (page 190)
Invoking the debugger (page 191)
* Postmortem (page 191) You can bind a key to run pyflakes in the current buffer.
* Step-by-step execution (page 192) In kate Menu: settings -> configure kate
* Other ways of starting a debugger (page 194)
Debugger commands and interaction (page 195) In plugins enable external tools
* Getting help when in the debugger (page 195) In external Tools, add pyflakes:
Debugging segmentation faults using gdb (page 195)
kdialog --title "pyflakes %filename" --msgbox "$(pyflakes %filename)"
In TextMate
2.3.1 Avoiding bugs Menu: TextMate -> Preferences -> Advanced -> Shell variables, add a shell variable:
Coding best practices to avoid getting in trouble TM_PYCHECKER=/Library/Frameworks/Python.framework/Versions/Current/bin/pyflakes
Then Ctrl-Shift-V is binded to a pyflakes report
2.3. Debugging code 187 2.3. Debugging code 188

In vim In your vimrc (binds F5 to pyflakes): In emacs Use the flymake mode with pyflakes, documented on
http://www.plope.com/Members/chrism/flymake-mode : add the following to your .emacs file:
autocmd FileType python let &mp = 'echo "*** running % ***" ; pyflakes %'
autocmd FileType tex,mp,rst,python imap <Esc>[15~ <C-O>:make!^M (when (load "flymake" t)
autocmd FileType tex,mp,rst,python map <Esc>[15~ :make!^M (defun flymake-pyflakes-init ()
autocmd FileType tex,mp,rst,python set autowrite (let* ((temp-file (flymake-init-create-temp-buffer-copy
'flymake-create-temp-inplace))
In emacs In your emacs (binds F5 to pyflakes): (local-file (file-relative-name
temp-file
(defun pyflakes-thisfile () (interactive) (file-name-directory buffer-file-name))))
(compile (format "pyflakes %s" (buffer-file-name))) (list "pyflakes" (list local-file))))
)
(add-to-list 'flymake-allowed-file-name-masks
(define-minor-mode pyflakes-mode '("\\.py\\'" flymake-pyflakes-init)))
"Toggle pyflakes mode.
With no argument, this command toggles the mode. (add-hook 'find-file-hook 'flymake-find-file-hook)
Non-null prefix argument turns on the mode.
Null prefix argument turns off the mode."
;; The initial value.
nil 2.3.2 Debugging workflow
;; The indicator for the mode line.
" Pyflakes" If you do have a non trivial bug, this is when debugging strategies kick in. There is no silver bullet. Yet, strategies
;; The minor mode bindings. help:
'( ([f5] . pyflakes-thisfile) )
) For debugging a given problem, the favorable situation is when the problem is isolated in a
small number of lines of code, outside framework or application code, with short modify-run-
(add-hook 'python-mode-hook (lambda () (pyflakes-mode t))) fail cycles
1. Make it fail reliably. Find a test case that makes the code fail every time.
A type-as-go spell-checker like integration 2. Divide and Conquer. Once you have a failing test case, isolate the failing code.
Which module.
In vim
Which function.
Use the pyflakes.vim plugin:
Which line of code.
1. download the zip file from http://www.vim.org/scripts/script.php?script_id=2441
=> isolate a small reproducible failure: a test case
2. extract the files in ~/.vim/ftplugin/python
3. Change one thing at a time and re-run the failing test case.
3. make sure your vimrc has filetype plugin indent on
4. Use the debugger to understand what is going wrong.
5. Take notes and be patient. It may take a while.
Note: Once you have gone through this process: isolated a tight piece of code reproducing the bug and fix the
bug using this piece of code, add the corresponding code to your test suite.
Alternatively: use the syntastic plugin. This can be configured to use flake8 too and also handles
on-the-fly checking for many other languages. 2.3.3 Using the Python debugger
The python debugger, pdb: http://docs.python.org/library/pdb.html, allows you to inspect your code interactively.
Specifically it allows you to:
View the source code.
Walk up and down the call stack.
Inspect values of variables.
Modify values of variables.
Set breakpoints.

r
print ipdb> quit
Yes, print statements do work as a debugging tool. However to inspect runtime, it is often more efficient In [3]:
to use the debugger.
Post-mortem debugging without IPython

Invoking the debugger In some situations you cannot use IPython, for instance to debug a script that wants to be called from the
command line. In this case, you can call the script with python -m pdb script.py:
Ways to launch the debugger:
$ python -m pdb index_error.py
1. Postmortem, launch debugger after module errors. > /home/varoquau/dev/scipy-lecture-notes/advanced/debugging_optimizing/index_error.py(1)<module>(
-> """Small snippet to raise an IndexError."""
2. Launch the module with the debugger. (Pdb) continue
3. Call the debugger inside the module Traceback (most recent call last):
File "/usr/lib/python2.6/pdb.py", line 1296, in main
pdb._runscript(mainpyfile)
Postmortem File "/usr/lib/python2.6/pdb.py", line 1215, in _runscript
self.run(statement)
File "/usr/lib/python2.6/bdb.py", line 372, in run
Situation: Youre working in IPython and you get a traceback. exec cmd in globals, locals
Here we debug the file index_error.py. When running it, an IndexError is raised. Type %debug and File "<string>", line 1, in <module>
drop into the debugger. File "index_error.py", line 8, in <module>
index_error()
In [1]: %run index_error.py File "index_error.py", line 5, in index_error
--------------------------------------------------------------------------- print lst[len(lst)]
IndexError Traceback (most recent call last) IndexError: list index out of range
/home/varoquau/dev/scipy-lecture-notes/advanced/debugging_optimizing/index_error.py in <module>() Uncaught exception. Entering post mortem debugging
6 Running 'cont' or 'step' will restart the program
7 if __name__ == '__main__': > /home/varoquau/dev/scipy-lecture-notes/advanced/debugging_optimizing/index_error.py(5)index_err
----> 8 index_error() -> print lst[len(lst)]
9 (Pdb)
/home/varoquau/dev/scipy-lecture-notes/advanced/debugging_optimizing/index_error.py in index_error()
3 def index_error():
4 lst = list('foobar') Step-by-step execution
----> 5 print lst[len(lst)]
6
7 if __name__ == '__main__': Situation: You believe a bug exists in a module but are not sure where.
For instance we are trying to debug wiener_filtering.py. Indeed the code runs, but the filtering does not
IndexError: list index out of range work well.
In [2]: %debug Run the script in IPython with the debugger using %run -d wiener_filtering.p :
> /home/varoquau/dev/scipy-lecture-notes/advanced/debugging_optimizing/index_error.py(5)index_error()
4 lst = list('foobar') In [1]: %run -d wiener_filtering.py
----> 5 print lst[len(lst)] *** Blank or comment
6 *** Blank or comment
*** Blank or comment
ipdb> list Breakpoint 1 at /home/varoquau/dev/scipy-lecture-notes/advanced/debugging_optimizing/wiener_f
1 """Small snippet to raise an IndexError.""" NOTE: Enter 'c' at the ipdb> prompt to start your script.
2 > <string>(1)<module>()
3 def index_error():
4 lst = list('foobar') Set a break point at line 34 using b 34:
----> 5 print lst[len(lst)] ipdb> n
6 > /home/varoquau/dev/scipy-lecture-notes/advanced/debugging_optimizing/wiener_filtering.py(4)
7 if __name__ == '__main__': 3
8 index_error() 1---> 4 import numpy as np
9 5 import scipy as sp
ipdb> len(lst) ipdb> b 34

6 Breakpoint 2 at /home/varoquau/dev/scipy-lecture-notes/advanced/debugging_optimizing/wiener_f
ipdb> print lst[len(lst)-1]

Continue execution to next breakpoint with c(ont(inue)): Raising exception on numerical errors
ipdb> c When we run the wiener_filtering.py file, the following warnings are raised:
> /home/varoquau/dev/scipy-lecture-notes/advanced/debugging_optimizing/wiener_filtering.py(34)iterated_wiener()
33 """ In [2]: %run wiener_filtering.py
2--> 34 noisy_img = noisy_img wiener_filtering.py:40: RuntimeWarning: divide by zero encountered in divide
35 denoised_img = local_mean(noisy_img, size=size) noise_level = (1 - noise/l_var )
We can turn these warnings in exception, which enables us to do post-mortem debugging on them, and find
Step into code with n(ext) and s(tep): next jumps to the next statement in the current execution our problem more quickly:
context, while step will go across execution contexts, i.e. enable exploring inside function calls:
In [3]: np.seterr(all='raise')
ipdb> s Out[3]: {'divide': 'print', 'invalid': 'print', 'over': 'print', 'under': 'ignore'}
In [4]: %run wiener_filtering.py
2 34 noisy_img = noisy_img ---------------------------------------------------------------------------
---> 35 denoised_img = local_mean(noisy_img, size=size) FloatingPointError Traceback (most recent call last)
36 l_var = local_var(noisy_img, size=size) /home/esc/anaconda/lib/python2.7/site-packages/IPython/utils/py3compat.pyc in execfile(fname, *wh
176 else:
ipdb> n 177 filename = fname
--> 178 __builtin__.execfile(filename, *where)
35 denoised_img = local_mean(noisy_img, size=size)
---> 36 l_var = local_var(noisy_img, size=size) /home/esc/physique-cuso-python-2013/scipy-lecture-notes/advanced/debugging/wiener_filtering.py in
37 for i in range(3): 55 pl.matshow(noisy_lena[cut], cmap=pl.cm.gray)
56
Step a few lines and explore the local variables: ---> 57 denoised_lena = iterated_wiener(noisy_lena)
58 pl.matshow(denoised_lena[cut], cmap=pl.cm.gray)
ipdb> n
59
36 l_var = local_var(noisy_img, size=size)
/home/esc/physique-cuso-python-2013/scipy-lecture-notes/advanced/debugging/wiener_filtering.py in
---> 37 for i in range(3):
38 res = noisy_img - denoised_img
38 res = noisy_img - denoised_img
39 noise = (res**2).sum()/res.size
ipdb> print l_var
---> 40 noise_level = (1 - noise/l_var )
[[5868 5379 5316 ..., 5071 4799 5149]
41 noise_level[noise_level<0] = 0
[5013 363 437 ..., 346 262 4355]
42 denoised_img += noise_level*res
[5379 410 344 ..., 392 604 3377]
...,
FloatingPointError: divide by zero encountered in divide
[ 435 362 308 ..., 275 198 1632]
[ 548 392 290 ..., 248 263 1653]
[ 466 789 736 ..., 1835 1725 1940]]
ipdb> print l_var.min()
0 Other ways of starting a debugger
Oh dear, nothing but integers, and 0 variation. Here is our bug, we are doing integer arithmetic. Raising an exception as a poor man break point
If you find it tedious to note the line number to set a break point, you can simply raise an exception at the
point that you want to inspect and use IPythons %debug. Note that in this case you cannot step or continue
the execution.
Debugging test failures using nosetests
You can run nosetests --pdb to drop in post-mortem debugging on exceptions, and nosetests
--pdb-failure to inspect test failures using the debugger.
In addition, you can use the IPython interface for the debugger in nose by installing the nose plugin ipdb-
plugin. You can than pass --ipdb and --ipdb-failure options to nosetests.
Calling the debugger explicitly
Insert the following line where you want to drop in the debugger:
import pdb; pdb.set_trace()
Warning: When running nosetests, the output is captured, and thus it seems that the debugger does not
work. Simply run the nosetests with the -s flag.

To debug with gdb the Python script segfault.py, we can run the script in gdb as follows
Graphical debuggers and alternatives
$ gdb python
For stepping through code and inspecting variables, you might find it more convenient to use a graph- ...
ical debugger such as winpdb. (gdb) run segfault.py
Alternatively, pudb is a good semi-graphical debugger with a text user interface in the console. Starting program: /usr/bin/python segfault.py
[Thread debugging using libthread_db enabled]
Also, the pydbgr project is probably worth looking at.
Program received signal SIGSEGV, Segmentation fault.
_strided_byte_copy (dst=0x8537478 "\360\343G", outstrides=4, src=
Debugger commands and interaction 0x86c0690 <Address 0x86c0690 out of bounds>, instrides=32, N=3,
elsize=4)
at numpy/core/src/multiarray/ctors.c:365
l(list) Lists the code at the current position 365 _FAST_MOVE(Int32);
u(p) Walk up the call stack (gdb)
d(own) Walk down the call stack
n(ext) Execute the next line (does not go down in new functions) We get a segfault, and gdb captures it for post-mortem debugging in the C level stack (not the Python call stack).
s(tep) Execute the next statement (goes down in new functions) We can debug the C call stack using gdbs commands:
bt Print the call stack (gdb) up
a Print the local variables #1 0x004af4f5 in _copy_from_same_shape (dest=<value optimized out>,
!command Execute the given Python command (by opposition to pdb commands src=<value optimized out>, myfunc=0x496780 <_strided_byte_copy>,
swap=0)
Warning: Debugger commands are not Python code at numpy/core/src/multiarray/ctors.c:748
You cannot name the variables the way you want. For instance, if in you cannot override the variables in the 748 myfunc(dit->dataptr, dest->strides[maxaxis],
current frame with the same name: use different names than your local variable when typing code in the
debugger. As you can see, right now, we are in the C code of numpy. We would like to know what is the Python code that
triggers this segfault, so we go up the stack until we hit the Python execution loop:
(gdb) up
Getting help when in the debugger #8 0x080ddd23 in call_function (f=
Frame 0x85371ec, for file /home/varoquau/usr/lib/python2.6/site-packages/numpy/core/arrayprint
at ../Python/ceval.c:3750
Type h or help to access the interactive help: 3750 ../Python/ceval.c: No such file or directory.
in ../Python/ceval.c
ipdb> help
(gdb) up
Documented commands (type help <topic>):
#9 PyEval_EvalFrameEx (f=
========================================
Frame 0x85371ec, for file /home/varoquau/usr/lib/python2.6/site-packages/numpy/core/arrayprint
EOF bt cont enable jump pdef r tbreak w
at ../Python/ceval.c:2412
a c continue exit l pdoc restart u whatis
2412 in ../Python/ceval.c
alias cl d h list pinfo return unalias where
(gdb)
args clear debug help n pp run unt
b commands disable ignore next q s until
break condition down j p quit step up Once we are in the Python execution loop, we can use our special Python helper function. For instance we can
find the corresponding Python code:
Miscellaneous help topics: (gdb) pyframe
========================== /home/varoquau/usr/lib/python2.6/site-packages/numpy/core/arrayprint.py (158): _leading_trailing
exec pdb (gdb)
Undocumented commands: This is numpy code, we need to go up until we find code that we have written:
======================
retval rv (gdb) up
...
(gdb) up
2.3.4 Debugging segmentation faults using gdb #34 0x080dc97a in PyEval_EvalFrameEx (f=
Frame 0x82f064c, for file segfault.py, line 11, in print_big_array (small_array=<numpy.ndarray
1630 ../Python/ceval.c: No such file or directory.
If you have a segmentation fault, you cannot debug it with pdb, as it crashes the Python interpreter before it can in ../Python/ceval.c
drop in the debugger. Similarly, if you have a bug in C code embedded in Python, pdb is useless. For this we turn (gdb) pyframe
to the gnu debugger, gdb, available on Linux. segfault.py (12): print_big_array
Before we start with gdb, let us add a few Python-specific tools to it. For this we add a few macros to our
~/.gbdinit. The optimal choice of macro depends on your Python version and your gdb version. I have added The corresponding code is:
a simplified version in gdbinit, but feel free to read DebuggingWithGdb.

def make_big_array(small_array): 3. Optimize the code by profiling simple use-cases to find the bottlenecks and speeding up these bottleneck,
big_array = stride_tricks.as_strided(small_array, finding a better algorithm or implementation. Keep in mind that a trade off should be found between
shape=(2e6, 2e6), strides=(32, 32)) profiling on a realistic example and the simplicity and speed of execution of the code. For efficient work, it
return big_array is best to work with profiling runs lasting around 10s.
def print_big_array(small_array):
big_array = make_big_array(small_array) 2.4.2 Profiling Python code
Thus the segfault happens when printing big_array[-10:]. The reason is simply that big_array has been
allocated with its end outside the program memory. No optimization without measuring!
Note: For a list of Python-specific commands defined in the gdbinit, read the source of this file. Measure: profiling, timing
Youll have surprises: the fastest code is not always what you think
Wrap up exercise
Timeit
The following script is well documented and hopefully legible. It seeks to answer a problem of actual interest
for numerical computing, but it does not work... Can you debug it? In IPython, use timeit (http://docs.python.org/library/timeit.html) to time elementary operations:
Python source code: to_debug.py
In [1]: import numpy as np
In [2]: a = np.arange(1000)
2.4 Optimizing code In [3]: %timeit a ** 2

In [4]: %timeit a ** 2.1

Donald Knuth 1000 loops, best of 3: 154 us per loop
Premature optimization is the root of all evil
In [5]: %timeit a * a
This chapter deals with strategies to make Python code go faster. Use this to guide your choice between strategies.
Note: For long running calls, using %time instead of %timeit; it is less precise but faster
Prerequisites
line_profiler
Profiler
Useful when you have a large program to profile, for example the following file:
Chapters contents
# For this example to run, you also need the 'ica.py' file
Optimization workflow (page 197)
Profiling Python code (page 198) import numpy as np
Timeit (page 198) from scipy import linalg
Profiler (page 198)
Line-profiler (page 200) from ica import fastica
Making code go faster (page 200)
Algorithmic optimization (page 200)
def test():
* Example of the SVD (page 200) data = np.random.random((5000, 100))
Writing faster numerical code (page 201)
u, s, v = linalg.svd(data)
Additional Links (page 203) pca = np.dot(u[:, :10].T, data)
results = fastica(pca.T, whiten=False)
if __name__ == '__main__':
2.4.1 Optimization workflow test()
1. Make it work: write the code in a simple legible ways.

Note: This is a combination of two unsupervised learning techniques, principal component analysis (PCA)
2. Make it work reliably: write automated test cases, make really sure that your algorithm is right and that if and independent component analysis (ICA). PCA is a technique for dimensionality reduction, i.e. an algorithm
you break it, the tests will capture the breakage. to explain the observed variance in your data using less dimensions. ICA is a source seperation technique, for
2.4. Optimizing code 197 2.4. Optimizing code 198

example to unmix multiple signals that have been recorded through multiple sensors. Doing a PCA first and then Line-profiler
an ICA can be useful if you have more sensors than signals. For more information see: the FastICA example from
scikits-learn. The profiler tells us which function takes most of the time, but not where it is called.
For this, we use the line_profiler: in the source file, we decorate a few functions that we want to inspect with
To run it, you also need to download the ica module. In IPython we can time the script: @profile (no need to import it)
In [1]: %run -t demo.py @profile
def test():
IPython CPU timings (estimated): data = np.random.random((5000, 100))
User : 14.3929 s. u, s, v = linalg.svd(data)
System: 0.256016 s. pca = np.dot(u[:, :10], data)
results = fastica(pca.T, whiten=False)
and profile it:
In [2]: %run -p demo.py Then we run the script using the kernprof.py program, with switches -l, --line-by-line and -v,
--view to use the line-by-line profiler and view the results in addition to saving them:
916 function calls in 14.551 CPU seconds $ kernprof.py -l -v demo.py
Ordered by: internal time Wrote profile results to demo.py.lprof
Timer unit: 1e-06 s
ncalls tottime percall cumtime percall filename:lineno (function)
1 14.457 14.457 14.479 14.479 decomp.py:849 (svd) File: demo.py
1 0.054 0.054 0.054 0.054 {method 'random_sample' of 'mtrand.RandomState' objects} Function: test at line 5
1 0.017 0.017 0.021 0.021 function_base.py:645 (asarray_chkfinite) Total time: 14.2793 s
54 0.011 0.000 0.011 0.000 {numpy.core._dotblas.dot}
2 0.005 0.002 0.005 0.002 {method 'any' of 'numpy.ndarray' objects} Line # Hits Time Per Hit % Time Line Contents
6 0.001 0.000 0.001 0.000 ica.py:195 (gprime) =========== ============ ===== ========= ======= ==== ========
6 0.001 0.000 0.001 0.000 ica.py:192 (g) 5 @profile
14 0.001 0.000 0.001 0.000 {numpy.linalg.lapack_lite.dsyevd} 6 def test():
19 0.001 0.000 0.001 0.000 twodim_base.py:204 (diag) 7 1 19015 19015.0 0.1 data = np.random.random((5000, 100))
1 0.001 0.001 0.008 0.008 ica.py:69 (_ica_par) 8 1 14242163 14242163.0 99.7 u, s, v = linalg.svd(data)
1 0.001 0.001 14.551 14.551 {execfile} 9 1 10282 10282.0 0.1 pca = np.dot(u[:10, :], data)
107 0.000 0.000 0.001 0.000 defmatrix.py:239 (__array_finalize__) 10 1 7799 7799.0 0.1 results = fastica(pca.T, whiten=False)
7 0.000 0.000 0.004 0.001 ica.py:58 (_sym_decorrelation)
7 0.000 0.000 0.002 0.000 linalg.py:841 (eigh)
The SVD is taking all the time. We need to optimise this line.
172 0.000 0.000 0.000 0.000 {isinstance}
1 0.000 0.000 14.551 14.551 demo.py:1 (<module>)
29 0.000 0.000 0.000 0.000 numeric.py:180 (asarray)
35 0.000 0.000 0.000 0.000 defmatrix.py:193 (__new__)
2.4.3 Making code go faster
35 0.000 0.000 0.001 0.000 defmatrix.py:43 (asmatrix)
21 0.000 0.000 0.001 0.000 defmatrix.py:287 (__mul__) Once we have identified the bottlenecks, we need to make the corresponding code go faster.
41 0.000 0.000 0.000 0.000 {numpy.core.multiarray.zeros}
28 0.000 0.000 0.000 0.000 {method 'transpose' of 'numpy.ndarray' objects}
1 0.000 0.000 0.008 0.008 ica.py:97 (fastica) Algorithmic optimization
...
The first thing to look for is algorithmic optimization: are there ways to compute less, or better?
Clearly the svd (in decomp.py) is what takes most of our time, a.k.a. the bottleneck. We have to find a way to
For a high-level view of the problem, a good understanding of the maths behind the algorithm helps. However, it
make this step go faster, or to avoid this step (algorithmic optimization). Spending time on the rest of the code is
is not uncommon to find simple changes, like moving computation or memory allocation outside a for loop,
useless.
that bring in big gains.
Profiling outside of IPython, running cProfile

Example of the SVD
Similar profiling can be done outside of IPython, simply calling the built-in Python profilers cProfile
and profile. In both examples above, the SVD - Singular Value Decomposition - is what takes most of the time. Indeed, the
$ python -m cProfile -o demo.prof demo.py computational cost of this algorithm is roughly 3 in the size of the input matrix.
Using the -o switch will output the profiler results to the file demo.prof to view with an external tool. However, in both of these example, we are not using all the output of the SVD, but only the first few rows of its
This can be useful if you wish to process the profiler output with a visualization tool. first return argument. If we use the svd implementation of scipy, we can ask for an incomplete version of the
SVD. Note that implementations of linear algebra in scipy are richer then those in numpy and should be preferred.

In [3]: %timeit np.linalg.svd(data) Use broadcasting (page 62) to do operations on arrays as small as possible before combining them.
1 loops, best of 3: 14.5 s per loop In place operations
In [4]: from scipy import linalg In [1]: a = np.zeros(1e7)
In [5]: %timeit linalg.svd(data) In [2]: %timeit global a ; a = 0*a

1 loops, best of 3: 14.2 s per loop 10 loops, best of 3: 111 ms per loop
In [6]: %timeit linalg.svd(data, full_matrices=False) In [3]: %timeit global a ; a *= 0

1 loops, best of 3: 295 ms per loop 10 loops, best of 3: 48.4 ms per loop
In [7]: %timeit np.linalg.svd(data, full_matrices=False) note: we need global a in the timeit so that it work, as it is assigning to a, and thus considers it as a
1 loops, best of 3: 293 ms per loop local variable.
We can then use this insight to optimize the previous code: Be easy on the memory: use views, and not copies
def test(): Copying big arrays is as costly as making simple numerical operations on them:
data = np.random.random((5000, 100)) In [1]: a = np.zeros(1e7)
u, s, v = linalg.svd(data, full_matrices=False)
pca = np.dot(u[:, :10].T, data) In [2]: %timeit a.copy()
results = fastica(pca.T, whiten=False) 10 loops, best of 3: 124 ms per loop
In [1]: import demo In [3]: %timeit a + 1

10 loops, best of 3: 112 ms per loop
In [2]: %timeit demo.
demo.fastica demo.np demo.prof.pdf demo.py demo.pyc Beware of cache effects
demo.linalg demo.prof demo.prof.png demo.py.lprof demo.test
Memory access is cheaper when it is grouped: accessing a big array in a continuous way is much faster than
In [2]: %timeit demo.test() random access. This implies amongst other things that smaller strides are faster (see CPU cache effects
ica.py:65: RuntimeWarning: invalid value encountered in sqrt (page 168)):
W = (u * np.diag(1.0/np.sqrt(s)) * u.T) * W # W = (W * W.T) ^{-1/2} * W
1 loops, best of 3: 17.5 s per loop In [1]: c = np.zeros((1e4, 1e4), order='C')
In [3]: import demo_opt In [2]: %timeit c.sum(axis=0)

1 loops, best of 3: 3.89 s per loop
In [4]: %timeit demo_opt.test()
1 loops, best of 3: 208 ms per loop In [3]: %timeit c.sum(axis=1)
Real incomplete SVDs, e.g. computing only the first 10 eigenvectors, can be computed with arpack, available in
scipy.sparse.linalg.eigsh. In [4]: c.strides
Out[4]: (80000, 8)
Computational linear algebra This is the reason why Fortran ordering or C ordering may make a big difference on operations:
For certain algorithms, many of the bottlenecks will be linear algebra computations. In this case, using the In [5]: a = np.random.rand(20, 2**18)
right function to solve the right problem is key. For instance, an eigenvalue problem with a symmetric matrix
is easier to solve than with a general matrix. Also, most often, you can avoid inverting a matrix and use a In [6]: b = np.random.rand(20, 2**18)
less costly (and more numerically stable) operation.
In [7]: %timeit np.dot(b, a.T)
Know your computational linear algebra. When in doubt, explore scipy.linalg, and use %timeit to
try out different alternatives on your data.
In [8]: c = np.ascontiguousarray(a.T)
In [9]: %timeit np.dot(b, c)

2.4.4 Writing faster numerical code 10 loops, best of 3: 84.2 ms per loop
A complete discussion on advanced use of numpy is found in chapter Advanced Numpy (page 155), or in the Note that copying the data to work around this effect may not be worth it:
article The NumPy array: a structure for efficient numerical computation by van der Walt et al. Here we discuss
In [10]: %timeit c = np.ascontiguousarray(a.T)
only some commonly encountered tricks to make code faster.
Vectorizing for loops
Using numexpr can be useful to automatically optimize code for such effects.
Find tricks to avoid for loops using numpy arrays. For this, masks and indices arrays can be useful.
Use compiled code
Broadcasting

The last resort, once you are sure that all the high-level optimizations have been explored, is to transfer the Sparse Matrices vs. Sparse Matrix Storage Schemes
hot spots, i.e. the few lines or functions in which most of the time is spent, to compiled code. For compiled
code, the preferred option is to use Cython: it is easy to transform exiting Python code in compiled code, sparse matrix is a matrix, which is almost empty
and with a good use of the numpy support yields efficient code on numpy arrays, for instance by unrolling
storing all the zeros is wasteful -> store only nonzero items
loops.
think compression
Warning: For all the above: profile and time your choices. Dont base your optimization on theoretical
considerations. pros: huge memory savings
cons: depends on actual storage scheme, (*) usually does not hold
Additional Links
Typical Applications
If you need to profile memory usage, you could try the memory_profiler
solution of partial differential equations (PDEs)
If you need to profile down into C extensions, you could try using gperftools from Python with yep.
the finite element method
If you would like to track performace of your code across time, i.e. as you make new commits to your
mechanical engineering, electrotechnics, physics, ...
repository, you could try: vbench
graph theory
If you need some interactive visualization why not try RunSnakeRun
nonzero at (i, j) means that node i is connected to node j
...
2.5 Sparse Matrices in SciPy
Prerequisites
Author: Robert Cimrman
recent versions of
numpy
scipy
2.5.1 Introduction matplotlib (optional)

ipython (the enhancements come handy)
(dense) matrix is:
mathematical object
Sparsity Structure Visualization
data structure for storing a 2D array of values
spy() from matplotlib
important features:
example plots:
memory allocated once for all items
usually a contiguous chunk, think NumPy ndarray
fast access to individual items (*)
Why Sparse Matrices?
the memory, that grows like n**2

small example (double precision matrix):
>>> x = np.linspace(0, 1e6, 10)
>>> plt.plot(x, 8.0 * (x**2) / 1e6, lw=5)
>>> plt.xlabel('size n')
<matplotlib.text.Text object at ...>
>>> plt.ylabel('memory [MB]')
<matplotlib.text.Text object at ...>
2.5. Sparse Matrices in SciPy 203 2.5. Sparse Matrices in SciPy 204
Common Methods
all scipy.sparse classes are subclasses of spmatrix

default implementation of arithmetic operations
* always converts to CSR

* subclasses override for efficiency
shape, data type set/get
nonzero indices
format conversion, interaction with NumPy (toarray(), todense())
...
attributes:
mtx.A - same as mtx.toarray()
mtx.T - transpose (same as mtx.transpose())
mtx.H - Hermitian (conjugate) transpose
mtx.real - real part of complex matrix
mtx.imag - imaginary part of complex matrix
mtx.size - the number of nonzeros (same as self.getnnz())
mtx.shape - the number of rows and columns (tuple)
data usually stored in NumPy arrays
Sparse Matrix Classes

2.5.2 Storage Schemes
Diagonal Format (DIA)
seven sparse matrix types in scipy.sparse:
1. csc_matrix: Compressed Sparse Column format very simple scheme
2. csr_matrix: Compressed Sparse Row format diagonals in dense NumPy array of shape (n_diag, length)
3. bsr_matrix: Block Sparse Row format fixed length -> waste space a bit when far from main diagonal
4. lil_matrix: List of Lists format subclass of _data_matrix (sparse matrix classes with data attribute)
5. dok_matrix: Dictionary of Keys format offset for each diagonal
6. coo_matrix: COOrdinate format (aka IJV, triplet format) 0 is the main diagonal
7. dia_matrix: DIAgonal format negative offset = below
each suitable for some tasks positive offset = above
many employ sparsetools C++ module by Nathan Bell fast matrix * vector (sparsetools)
assume the following is imported: fast and easy item-wise operations
>>> import numpy as np manipulate data array directly (fast NumPy machinery)
>>> import scipy.sparse as sps
constructor accepts:
dense matrix (array)
warning for NumPy users:
sparse matrix
the multiplication with * is the matrix multiplication (dot product)
shape tuple (create empty matrix)
not part of NumPy!
(data, offsets) tuple
* passing a sparse matrix object to NumPy functions expecting ndarray/matrix does not work
no slicing, no individual item access
use: -2: . 6 3 .
rather specialized -3: . . 7 4
---------8
solving PDEs by finite differences
matrix-vector multiplication
with an iterative solver
>>> vec = np.ones((4, ))
>>> vec
Examples array([ 1., 1., 1., 1.])
>>> mtx * vec
create some DIA matrices: array([ 12., 19., 9., 11.])
>>> data = np.array([[1, 2, 3, 4]]).repeat(3, axis=0) >>> mtx.toarray() * vec
>>> data array([[ 1., 0., 11., 0.],
array([[1, 2, 3, 4], [ 5., 2., 0., 12.],
[1, 2, 3, 4], [ 0., 6., 3., 0.],
[1, 2, 3, 4]]) [ 0., 0., 7., 4.]])
>>> offsets = np.array([0, -1, 2])
>>> mtx = sparse.dia_matrix((data, offsets), shape=(4, 4))
>>> mtx List of Lists Format (LIL)
<4x4 sparse matrix of type '<... 'numpy.int64'>'
with 9 stored elements (3 diagonals) in DIAgonal format>
row-based linked list
>>> mtx.todense()
matrix([[1, 0, 3, 0], each row is a Python list (sorted) of column indices of non-zero elements
[1, 2, 0, 4],
[0, 2, 3, 0], rows stored in a NumPy array (dtype=np.object)
[0, 0, 3, 4]])
non-zero values data stored analogously
>>> data = np.arange(12).reshape((3, 4)) + 1 efficient for constructing sparse matrices incrementally
>>> data
array([[ 1, 2, 3, 4], constructor accepts:
[ 5, 6, 7, 8], dense matrix (array)
[ 9, 10, 11, 12]])
>>> mtx = sparse.dia_matrix((data, offsets), shape=(4, 4)) sparse matrix
>>> mtx.data
array([[ 1, 2, 3, 4], shape tuple (create empty matrix)
[ 5, 6, 7, 8], flexible slicing, changing sparsity structure is efficient
[ 9, 10, 11, 12]]...)
>>> mtx.offsets slow arithmetics, slow column slicing due to being row-based
array([ 0, -1, 2], dtype=int32)
use:
>>> print(mtx)
(0, 0) 1 when sparsity pattern is not known apriori or changes
(1, 1) 2
(2, 2) 3 example: reading a sparse matrix from a text file
(3, 3) 4
(1, 0) 5
(2, 1) 6 Examples
(3, 2) 7 create an empty LIL matrix:
(0, 2) 11
(1, 3) 12 >>> mtx = sparse.lil_matrix((4, 5))
>>> mtx.todense()
matrix([[ 1, 0, 11, 0], prepare random data:
[ 5, 2, 0, 12],
[ 0, 6, 3, 0], >>> from numpy.random import rand
[ 0, 0, 7, 4]]) >>> data = np.round(rand(2, 3))
>>> data
explanation with a scheme: array([[ 1., 1., 1.],
[ 1., 0., 1.]])
offset: row
assign the data using fancy indexing:
2: 9
1: --10------ >>> mtx[:2, [1, 2, 3]] = data
0: 1 . 11 . >>> mtx
-1: 5 2 . 12 <4x5 sparse matrix of type '<... 'numpy.float64'>'
with 5 stored elements in LInked List format>
>>> print(mtx) can be efficiently converted to a coo_matrix once constructed

(0, 1) 1.0 slow arithmetics (for loops with dict.iteritems())
(0, 2) 1.0
(0, 3) 1.0 use:
(1, 1) 1.0
(1, 3) 1.0 when sparsity pattern is not known apriori or changes
>>> mtx.todense()
matrix([[ 0., 1., 1., 1., 0.],
Examples
[ 0., 1., 0., 1., 0.],
[ 0., 0., 0., 0., 0.], create a DOK matrix element by element:
[ 0., 0., 0., 0., 0.]])
>>> mtx.toarray() >>> mtx = sparse.dok_matrix((5, 5), dtype=np.float64)
array([[ 0., 1., 1., 1., 0.], >>> mtx
[ 0., 1., 0., 1., 0.], <5x5 sparse matrix of type '<... 'numpy.float64'>'
[ 0., 0., 0., 0., 0.], with 0 stored elements in Dictionary Of Keys format>
[ 0., 0., 0., 0., 0.]]) >>> for ir in range(5):
... for ic in range(5):
more slicing and indexing: ... mtx[ir, ic] = 1.0 * (ir != ic)
>>> mtx
>>> mtx = sparse.lil_matrix([[0, 1, 2, 0], [3, 0, 1, 0], [1, 0, 0, 1]]) <5x5 sparse matrix of type '<... 'numpy.float64'>'
>>> mtx.todense() with 20 stored elements in Dictionary Of Keys format>
matrix([[0, 1, 2, 0], >>> mtx.todense()
[3, 0, 1, 0], matrix([[ 0., 1., 1., 1., 1.],
[1, 0, 0, 1]]...) [ 1., 0., 1., 1., 1.],
>>> print(mtx) [ 1., 1., 0., 1., 1.],
(0, 1) 1 [ 1., 1., 1., 0., 1.],
(0, 2) 2 [ 1., 1., 1., 1., 0.]])
(1, 0) 3
(1, 2) 1 slicing and indexing:
(2, 0) 1
(2, 3) 1 >>> mtx[1, 1]
>>> mtx[:2, :] 0.0
<2x4 sparse matrix of type '<... 'numpy.int64'>' >>> mtx[1, 1:3]
with 4 stored elements in LInked List format> <1x2 sparse matrix of type '<... 'numpy.float64'>'
>>> mtx[:2, :].todense() with 1 stored elements in Dictionary Of Keys format>
matrix([[0, 1, 2, 0], >>> mtx[1, 1:3].todense()
[3, 0, 1, 0]]...) matrix([[ 0., 1.]])
>>> mtx[1:2, [0,2]].todense() >>> mtx[[2,1], 1:3].todense()
matrix([[3, 1]]...) matrix([[ 1., 0.],
>>> mtx.todense() [ 0., 1.]])
matrix([[0, 1, 2, 0],
[3, 0, 1, 0],
[1, 0, 0, 1]]...) Coordinate Format (COO)
also known as the ijv or triplet format

Dictionary of Keys Format (DOK)
three NumPy arrays: row, col, data
subclass of Python dict data[i] is value at (row[i], col[i]) position
keys are (row, column) index tuples (no duplicate entries allowed) permits duplicate entries
values are corresponding non-zero values subclass of _data_matrix (sparse matrix classes with data attribute)
efficient for constructing sparse matrices incrementally fast format for constructing sparse matrices
constructor accepts: constructor accepts:
dense matrix (array) dense matrix (array)
sparse matrix sparse matrix
shape tuple (create empty matrix) shape tuple (create empty matrix)
efficient O(1) access to individual elements (data, ij) tuple
flexible slicing, changing sparsity structure is efficient very fast conversion to and from CSR/CSC formats
fast matrix * vector (sparsetools) Compressed Sparse Row Format (CSR)
fast and easy item-wise operations

row oriented
manipulate data array directly (fast NumPy machinery)
three NumPy arrays: indices, indptr, data
no slicing, no arithmetics (directly)
* indices is array of column indices
use:
* data is array of corresponding nonzero values
facilitates fast conversion among sparse formats
* indptr points to row starts in indices and data
when converting to other format (usually CSR or CSC), duplicate entries are summed together
* length is n_row + 1, last item = number of values = length of both indices and data
* facilitates efficient construction of finite element matrices
* nonzero values of the i-th row are data[indptr[i]:indptr[i+1]] with column
indices indices[indptr[i]:indptr[i+1]]
Examples
* item (i, j) can be accessed as data[indptr[i]+k], where k is position of j in
create empty COO matrix: indices[indptr[i]:indptr[i+1]]
>>> mtx = sparse.coo_matrix((3, 4), dtype=np.int8) subclass of _cs_matrix (common CSR/CSC functionality)
>>> mtx.todense()
matrix([[0, 0, 0, 0], * subclass of _data_matrix (sparse matrix classes with data attribute)
[0, 0, 0, 0],
fast matrix vector products and other arithmetics (sparsetools)
[0, 0, 0, 0]], dtype=int8)
create using (data, ij) tuple:
>>> row = np.array([0, 3, 1, 0])
>>> col = np.array([0, 3, 1, 2]) sparse matrix
>>> data = np.array([4, 5, 7, 9]) shape tuple (create empty matrix)
>>> mtx = sparse.coo_matrix((data, (row, col)), shape=(4, 4))
>>> mtx (data, ij) tuple
(data, indices, indptr) tuple
with 4 stored elements in COOrdinate format>
>>> mtx.todense() efficient row slicing, row-oriented operations
matrix([[4, 0, 9, 0],
[0, 7, 0, 0], slow column slicing, expensive changes to the sparsity structure
[0, 0, 0, 0],
use:
[0, 0, 0, 5]])
actual computations (most linear solvers support this format)
duplicates entries are summed together:
>>> row = np.array([0, 0, 1, 3, 1, 0, 0]) Examples
>>> col = np.array([0, 2, 1, 3, 1, 0, 0])
>>> data = np.array([1, 1, 1, 1, 1, 1, 1]) create empty CSR matrix:
>>> mtx = sparse.coo_matrix((data, (row, col)), shape=(4, 4))
>>> mtx.todense() >>> mtx = sparse.csr_matrix((3, 4), dtype=np.int8)
matrix([[3, 0, 1, 0], >>> mtx.todense()
[0, 2, 0, 0], matrix([[0, 0, 0, 0],
[0, 0, 0, 0], [0, 0, 0, 0],
[0, 0, 0, 1]]) [0, 0, 0, 0]], dtype=int8)
no slicing...: create using (data, ij) tuple:
>>> mtx[2, 3] >>> row = np.array([0, 0, 1, 2, 2, 2])

Traceback (most recent call last): >>> col = np.array([0, 2, 2, 0, 1, 2])
... >>> data = np.array([1, 2, 3, 4, 5, 6])
TypeError: 'coo_matrix' object ... >>> mtx = sparse.csr_matrix((data, (row, col)), shape=(3, 3))
Traceback (most recent call last): >>> mtx
... <3x3 sparse matrix of type '<... 'numpy.int64'>'
TypeError: 'coo_matrix' object ... with 6 stored elements in Compressed Sparse Row format>
>>> mtx.todense()
matrix([[1, 0, 2],
[0, 0, 3],
[4, 5, 6]]...)
>>> mtx.data
array([1, 2, 3, 4, 5, 6]...) >>> mtx = sparse.csc_matrix((3, 4), dtype=np.int8)

>>> mtx.indices >>> mtx.todense()
array([0, 2, 2, 0, 1, 2], dtype=int32) matrix([[0, 0, 0, 0],
>>> mtx.indptr [0, 0, 0, 0],
array([0, 2, 3, 6], dtype=int32) [0, 0, 0, 0]], dtype=int8)
create using (data, indices, indptr) tuple: create using (data, ij) tuple:
>>> data = np.array([1, 2, 3, 4, 5, 6]) >>> row = np.array([0, 0, 1, 2, 2, 2])
>>> indices = np.array([0, 2, 2, 0, 1, 2]) >>> col = np.array([0, 2, 2, 0, 1, 2])
>>> indptr = np.array([0, 2, 3, 6]) >>> data = np.array([1, 2, 3, 4, 5, 6])
>>> mtx = sparse.csr_matrix((data, indices, indptr), shape=(3, 3)) >>> mtx = sparse.csc_matrix((data, (row, col)), shape=(3, 3))
>>> mtx.todense() >>> mtx
matrix([[1, 0, 2], <3x3 sparse matrix of type '<... 'numpy.int64'>'
[0, 0, 3], with 6 stored elements in Compressed Sparse Column format>
[4, 5, 6]]) >>> mtx.todense()
matrix([[1, 0, 2],
[0, 0, 3],
Compressed Sparse Column Format (CSC) [4, 5, 6]]...)
>>> mtx.data
array([1, 4, 5, 2, 3, 6]...)
column oriented >>> mtx.indices
three NumPy arrays: indices, indptr, data array([0, 2, 2, 0, 1, 2], dtype=int32)
>>> mtx.indptr
* indices is array of row indices array([0, 2, 3, 6], dtype=int32)
* data is array of corresponding nonzero values create using (data, indices, indptr) tuple:
* indptr points to column starts in indices and data >>> data = np.array([1, 4, 5, 2, 3, 6])
>>> indices = np.array([0, 2, 2, 0, 1, 2])
* length is n_col + 1, last item = number of values = length of both indices and data
>>> indptr = np.array([0, 2, 3, 6])
* nonzero values of the i-th column are data[indptr[i]:indptr[i+1]] with row >>> mtx = sparse.csc_matrix((data, indices, indptr), shape=(3, 3))
indices indices[indptr[i]:indptr[i+1]] >>> mtx.todense()
matrix([[1, 0, 2],
* item (i, j) can be accessed as data[indptr[j]+k], where k is position of i in [0, 0, 3],
indices[indptr[j]:indptr[j+1]] [4, 5, 6]])
subclass of _cs_matrix (common CSR/CSC functionality)
* subclass of _data_matrix (sparse matrix classes with data attribute) Block Compressed Row Format (BSR)
basically a CSR with dense sub-matrices of fixed shape instead of scalar items
block size (R, C) must evenly divide the shape of the matrix (M, N)
three NumPy arrays: indices, indptr, data
sparse matrix
shape tuple (create empty matrix) * indices is array of column indices for each block
(data, ij) tuple * data is array of corresponding nonzero values of shape (nnz, R, C)
(data, indices, indptr) tuple * ...

subclass of _cs_matrix (common CSR/CSC functionality)
efficient column slicing, column-oriented operations
slow row slicing, expensive changes to the sparsity structure * subclass of _data_matrix (sparse matrix classes with data attribute)
use:
actual computations (most linear solvers support this format)
Examples sparse matrix
create empty CSC matrix: shape tuple (create empty matrix)
(data, ij) tuple
(data, indices, indptr) tuple >>> mtx.indptr

many arithmetic operations considerably more efficient than CSR for sparse matrices with dense sub- array([0, 2, 3, 6], dtype=int32)
matrices
create using (data, indices, indptr) tuple with (2, 2) block size:
use:
>>> indptr = np.array([0, 2, 3, 6])
like CSR >>> indices = np.array([0, 2, 2, 0, 1, 2])
>>> data = np.array([1, 2, 3, 4, 5, 6]).repeat(4).reshape(6, 2, 2)
vector-valued finite element discretizations >>> mtx = sparse.bsr_matrix((data, indices, indptr), shape=(6, 6))
>>> mtx.todense()
matrix([[1, 1, 0, 0, 2, 2],
Examples
[1, 1, 0, 0, 2, 2],
create empty BSR matrix with (1, 1) block size (like CSR...): [0, 0, 0, 0, 3, 3],
[0, 0, 0, 0, 3, 3],
>>> mtx = sparse.bsr_matrix((3, 4), dtype=np.int8) [4, 4, 5, 5, 6, 6],
>>> mtx [4, 4, 5, 5, 6, 6]])
<3x4 sparse matrix of type '<... 'numpy.int8'>' >>> data
with 0 stored elements (blocksize = 1x1) in Block Sparse Row format> array([[[1, 1],
>>> mtx.todense() [1, 1]],
matrix([[0, 0, 0, 0],
[0, 0, 0, 0], [[2, 2],
[0, 0, 0, 0]], dtype=int8) [2, 2]],
create empty BSR matrix with (3, 2) block size: [[3, 3],
[3, 3]],
>>> mtx = sparse.bsr_matrix((3, 4), blocksize=(3, 2), dtype=np.int8)
>>> mtx
[[4, 4],
[4, 4]],
with 0 stored elements (blocksize = 3x2) in Block Sparse Row format>
>>> mtx.todense()
[[5, 5],
matrix([[0, 0, 0, 0],
[5, 5]],
[0, 0, 0, 0],
[0, 0, 0, 0]], dtype=int8)
[[6, 6],
[6, 6]]])
a bug?
create using (data, ij) tuple with (1, 1) block size (like CSR...):
Summary
>>> row = np.array([0, 0, 1, 2, 2, 2])
>>> col = np.array([0, 2, 2, 0, 1, 2])
>>> data = np.array([1, 2, 3, 4, 5, 6]) Table 2.1: Summary of storage schemes.
>>> mtx = sparse.bsr_matrix((data, (row, col)), shape=(3, 3))
>>> mtx for- matrix * get fancy set fancy solvers note
<3x3 sparse matrix of type '<... 'numpy.int64'>' mat vector item get item set
with 6 stored elements (blocksize = 1x1) in Block Sparse Row format> DIA sparsetools . . . . itera- has data array, specialized
>>> mtx.todense() tive
matrix([[1, 0, 2], LIL via CSR yes yes yes yes itera- arithmetics via CSR,
[0, 0, 3], tive incremental construction
[4, 5, 6]]...)
DOK python yes one axis yes yes itera- O(1) item access, incremental
>>> mtx.data
array([[[1]],
only tive construction
COO sparsetools . . . . itera- has data array, facilitates fast
[[2]], tive conversion
CSR sparsetools yes yes slow . any has data array, fast row-wise ops
[[3]], CSC sparsetools yes yes slow . any has data array, fast column-wise
ops
[[4]], BSR sparsetools . . . . special- has data array, specialized
ized
[[5]],
[[6]]]...)
>>> mtx.indices
2.5.3 Linear System Solvers
array([0, 2, 2, 0, 1, 2], dtype=int32)
sparse matrix/eigenvalue problem solvers live in scipy.sparse.linalg
the submodules: >>> mtx1 = mtx.astype(np.float32)

dsolve: direct factorization methods for solving linear systems >>> x = dsolve.spsolve(mtx1, rhs, use_umfpack=False)
>>> print(x)
isolve: iterative methods for solving linear systems [ 106. -21. 5.5 -1.5 1. ]
>>> print("Error: %s" % (mtx1 * x - rhs))
eigen: sparse eigenvalue problem solvers Error: [ 0. 0. 0. 0. 0.]
all solvers are accessible from:
solve as double precision real:
>>> import scipy.sparse.linalg as spla
>>> spla.__all__ >>> mtx2 = mtx.astype(np.float64)
['LinearOperator', 'Tester', 'arpack', 'aslinearoperator', 'bicg', >>> x = dsolve.spsolve(mtx2, rhs, use_umfpack=True)
'bicgstab', 'cg', 'cgs', 'csc_matrix', 'csr_matrix', 'dsolve', >>> print(x)
'eigen', 'eigen_symmetric', 'factorized', 'gmres', 'interface', [ 106. -21. 5.5 -1.5 1. ]
'isolve', 'iterative', 'lgmres', 'linsolve', 'lobpcg', 'lsqr', >>> print("Error: %s" % (mtx2 * x - rhs))
'minres', 'np', 'qmr', 'speigs', 'spilu', 'splu', 'spsolve', 'svd', Error: [ 0. 0. 0. 0. 0.]
'test', 'umfpack', 'use_solver', 'utils', 'warnings']
solve as single precision complex:
>>> mtx1 = mtx.astype(np.complex64)
Sparse Direct Solvers >>> x = dsolve.spsolve(mtx1, rhs, use_umfpack=False)
>>> print(x)
default solver: SuperLU 4.0 [ 106.0+0.j -21.0+0.j 5.5+0.j -1.5+0.j 1.0+0.j]
>>> print("Error: %s" % (mtx1 * x - rhs))
included in SciPy Error: [ 0.+0.j 0.+0.j 0.+0.j 0.+0.j 0.+0.j]
real and complex systems
solve as double precision complex:
both single and double precision
>>> mtx2 = mtx.astype(np.complex128)
optional: umfpack >>> x = dsolve.spsolve(mtx2, rhs, use_umfpack=True)
>>> print(x)
real and complex systems [ 106.0+0.j -21.0+0.j 5.5+0.j -1.5+0.j 1.0+0.j]
double precision only >>> print("Error: %s" % (mtx2 * x - rhs))
Error: [ 0.+0.j 0.+0.j 0.+0.j 0.+0.j 0.+0.j]
recommended for performance
"""
wrappers now live in scikits.umfpack
Solve a linear system
check-out the new scikits.suitesparse by Nathaniel Smith =======================
Construct a 1000x1000 lil_matrix and add some values to it, convert it

Examples to CSR format and solve A x = b for x:and solve a linear system with a
direct solver.
"""
import the whole module, and see its docstring:
import numpy as np
>>> from scipy.sparse.linalg import dsolve import scipy.sparse as sps
>>> help(dsolve) from matplotlib import pyplot as plt
from scipy.sparse.linalg.dsolve import linsolve
both superlu and umfpack can be used (if the latter is installed) as follows:
rand = np.random.rand
prepare a linear system:
mtx = sps.lil_matrix((1000, 1000), dtype=np.float64)
mtx[0, :100] = rand(100)
>>> from scipy import sparse
mtx[1, 100:200] = mtx[0, :100]
>>> mtx = sparse.spdiags([[1, 2, 3, 4, 5], [6, 5, 8, 9, 10]], [0, 1], 5, 5)
mtx.setdiag(rand(1000))
>>> mtx.todense()
matrix([[ 1, 5, 0, 0, 0],
plt.clf()
[ 0, 2, 8, 0, 0],
plt.spy(mtx, marker='.', markersize=2)
[ 0, 0, 3, 9, 0],
plt.show()
[ 0, 0, 0, 4, 10],
[ 0, 0, 0, 0, 5]])
mtx = mtx.tocsr()
>>> rhs = np.array([1, 2, 3, 4, 5], dtype=np.float32)
rhs = rand(1000)
solve as single precision real: x = linsolve.spsolve(mtx, rhs)
print('rezidual: %r' % np.linalg.norm(mtx * x - rhs))
examples/direct_solve.py >>> A.matvec(np.ones(2))

array([ 2., 3.])
>>> A * np.ones(2)
Iterative Solvers array([ 2., 3.])
the isolve module contains the following solvers:

bicg (BIConjugate Gradient) A Few Notes on Preconditioning
bicgstab (BIConjugate Gradient STABilized)

problem specific
cg (Conjugate Gradient) - symmetric positive definite matrices only
often hard to develop
cgs (Conjugate Gradient Squared)
if not sure, try ILU
gmres (Generalized Minimal RESidual)
available in dsolve as spilu()
minres (MINimum RESidual)
qmr (Quasi-Minimal Residual) Eigenvalue Problem Solvers
Common Parameters The eigen module
mandatory: arpack * a collection of Fortran77 subroutines designed to solve large scale eigenvalue problems
A [{sparse matrix, dense matrix, LinearOperator}] The N-by-N matrix of the linear system. lobpcg (Locally Optimal Block Preconditioned Conjugate Gradient Method) * works very well in com-
bination with PyAMG * example by Nathan Bell:
b [{array, matrix}] Right hand side of the linear system. Has shape (N,) or (N,1).
"""
optional: Compute eigenvectors and eigenvalues using a preconditioned eigensolver
========================================================================
x0 [{array, matrix}] Starting guess for the solution.
tol [float] Relative tolerance to achieve before terminating. In this example Smoothed Aggregation (SA) is used to precondition
the LOBPCG eigensolver on a two-dimensional Poisson problem with
maxiter [integer] Maximum number of iterations. Iteration will stop after maxiter steps even if the specified Dirichlet boundary conditions.
tolerance has not been achieved. """
M [{sparse matrix, dense matrix, LinearOperator}] Preconditioner for A. The preconditioner should ap-
import scipy
proximate the inverse of A. Effective preconditioning dramatically improves the rate of convergence, from scipy.sparse.linalg import lobpcg
which implies that fewer iterations are needed to reach a given error tolerance.
callback [function] User-supplied function to call after each iteration. It is called as callback(xk), where from pyamg import smoothed_aggregation_solver
xk is the current solution vector. from pyamg.gallery import poisson
N = 100
LinearOperator Class K = 9
A = poisson((N,N), format='csr')
from scipy.sparse.linalg.interface import LinearOperator # create the AMG hierarchy

ml = smoothed_aggregation_solver(A)
common interface for performing matrix vector products
# initial approximation to the K eigenvectors
useful abstraction that enables using dense and sparse matrices within the solvers, as well as matrix-free
X = scipy.rand(A.shape[0], K)
solutions
has shape and matvec() (+ some optional parameters) # preconditioner based on ml
M = ml.aspreconditioner()
example:
# compute eigenvalues and eigenvectors with LOBPCG
W,V = lobpcg(A, X, M=M, tol=1e-8, largest=False)
>>> from scipy.sparse.linalg import LinearOperator
>>> def mv(v):
... return np.array([2*v[0], 3*v[1]])
#plot the eigenvectors
...
import pylab
>>> A = LinearOperator((2, 2), matvec=mv)
>>> A
pylab.figure(figsize=(9,9))
<2x2 LinearOperator with unspecified dtype>
for i in range(K): http://code.google.com/p/pyamg

pylab.subplot(3, 3, i+1) Pysparse
pylab.title('Eigenvector %d' % i)
pylab.pcolor(V[:,i].reshape(N,N)) own sparse matrix classes
pylab.axis('equal')
pylab.axis('off') matrix and eigenvalue problem solvers
pylab.show() http://pysparse.sourceforge.net/
examples/pyamg_with_lobpcg.py
example by Nils Wagner: 2.6 Image manipulation and processing using Numpy and Scipy
examples/lobpcg_sakurai.py
Authors: Emmanuelle Gouillart, Gal Varoquaux
output:
This section addresses basic image manipulation and processing using the core scientific modules NumPy and
$ python examples/lobpcg_sakurai.py
SciPy. Some of the operations covered by this tutorial may be useful for other kinds of multidimensional array
Results by LOBPCG for n=2500
processing than image processing. In particular, the submodule scipy.ndimage provides functions operating
[ 0.06250083 0.06250028 0.06250007] on n-dimensional NumPy arrays.
See also:
Exact eigenvalues
For more advanced image processing and image-specific routines, see the tutorial Scikit-image: image processing
[ 0.06250005 0.0625002 0.06250044] (page 324), dedicated to the skimage module.
Elapsed time 7.01

Image = 2-D numerical array
(or 3-D: CT, MRI, 2D + time; 4-D, ...)
Here, image == Numpy array np.array
Tools used in this tutorial:

numpy: basic array manipulation
scipy: scipy.ndimage submodule dedicated to image processing (n-dimensional images). See
http://docs.scipy.org/doc/scipy/reference/tutorial/ndimage.html
Common tasks in image processing:

Input/Output, displaying images
Basic manipulations: cropping, flipping, rotating, ...
Image filtering: denoising, sharpening
Image segmentation: labeling pixels corresponding to different objects
Classification
Feature extraction
Registration
...
2.5.4 Other Interesting Packages
PyAMG
algebraic multigrid solvers
2.5. Sparse Matrices in SciPy 221 2.6. Image manipulation and processing using Numpy and Scipy 222
Chapters contents >>> misc.imsave('lena.png', lena) # First we need to create the PNG file
Opening and writing to image files (page 223) >>> lena = misc.imread('lena.png')
Displaying images (page 224) >>> type(lena)
Basic manipulations (page 225) <... 'numpy.ndarray'>
Statistical information (page 226) >>> lena.shape, lena.dtype
Geometrical transformations (page 227) ((512, 512), dtype('uint8'))
Image filtering (page 227)
Blurring/smoothing (page 227) dtype is uint8 for 8-bit images (0-255)
Sharpening (page 228)
Opening raw files (camera, 3-D images)
Denoising (page 228)
Mathematical morphology (page 229) >>> lena.tofile('lena.raw') # Create raw file
Feature extraction (page 232) >>> lena_from_raw = np.fromfile('lena.raw', dtype=np.uint8)
Edge detection (page 232) >>> lena_from_raw.shape
Segmentation (page 232) (262144,)
Measuring objects properties: ndimage.measurements (page 234) >>> lena_from_raw.shape = (512, 512)
Need to know the shape and dtype of the image (how to separate data bytes).
For large data, use np.memmap for memory mapping:
2.6.1 Opening and writing to image files
>>> lena_memmap = np.memmap('lena.raw', dtype=np.uint8, shape=(512, 512))
Writing an array to a file:
(data are read from the file, and not loaded into memory)
Displaying Lena
================ Working on a list of image files
>>> for i in range(10):
Small example to plot lena. ... im = np.random.random_integers(0, 255, 10000).reshape((100, 100))
""" ... misc.imsave('random_%02d.png' % i, im)
>>> from glob import glob
from scipy import misc >>> filelist = glob('random*.png')
l = misc.lena() >>> filelist.sort()
misc.imsave('lena.png', l) # uses the Image module (PIL)
import matplotlib.pyplot as plt

plt.imshow(l) 2.6.2 Displaying images
plt.show()
Use matplotlib and imshow to display an image inside a matplotlib figure:
>>> l = misc.lena()
>>> plt.imshow(l, cmap=plt.cm.gray)
Increase contrast by setting min and max values:

>>> plt.imshow(l, cmap=plt.cm.gray, vmin=30, vmax=200)
>>> # Remove axes and ticks
>>> plt.axis('off')
(-0.5, 511.5, 511.5, -0.5)
Draw contour lines:

>>> plt.contour(l, [60, 211])
<matplotlib.contour.QuadContourSet ...>
For fine inspection of intensity variations, use interpolation=nearest:

>>> plt.imshow(l[200:220, 200:220], cmap=plt.cm.gray)
Creating a numpy array from an image file:
>>> plt.imshow(l[200:220, 200:220], cmap=plt.cm.gray, interpolation='nearest')
>>> from scipy import misc <matplotlib.image.AxesImage object at 0x...>
2.6. Image manipulation and processing using Numpy and Scipy 223 2.6. Image manipulation and processing using Numpy and Scipy 224
>>> # Slicing
>>> lena[10:13, 20:23]
array([[158, 156, 157],
[157, 155, 155],
[157, 157, 158]])
>>> lena[100:120] = 255
>>>
>>> lx, ly = lena.shape
>>> X, Y = np.ogrid[0:lx, 0:ly]
>>> mask = (X - lx / 2) ** 2 + (Y - ly / 2) ** 2 > lx * ly / 4
>>> # Masks
>>> lena[mask] = 0
>>> # Fancy indexing
>>> lena[range(400), range(400)] = 255
3-D visualization: Mayavi

See 3D plotting with Mayavi (page 357).
Image plane widgets
Isosurfaces
...
Statistical information

2.6.3 Basic manipulations >>> lena.mean()
124.04678344726562
Images are arrays: use the whole numpy machinery. >>> lena.max(), lena.min()
(245, 25)
np.histogram

>>> lena[0, 40]
166
Exercise >>> from scipy import misc

Open as an array the scikit-image logo (http://scikit-image.org/_static/img/logo.png), or an im-
>>> blurred_lena = ndimage.gaussian_filter(lena, sigma=3)
age that you have on your computer. >>> very_blurred = ndimage.gaussian_filter(lena, sigma=5)
Crop a meaningful part of the image, for example the python circle in the logo.
Display the image array using matplotlib. Change the interpolation method and zoom to see the Uniform filter
difference.
Transform your image to greyscale >>> local_mean = ndimage.uniform_filter(lena, size=11)
Increase the contrast of the image by changing its minimum and maximum values. Optional: use
scipy.stats.scoreatpercentile (read the docstring!) to saturate 5% of the darkest pixels
and 5% of the lightest pixels.
Save the array to two different file formats (png, jpg, tiff)
Sharpening
Geometrical transformations
Sharpen a blurred image:
>>> lx, ly = lena.shape >>> from scipy import misc
>>> # Cropping >>> lena = misc.lena()
>>> crop_lena = lena[lx / 4: - lx / 4, ly / 4: - ly / 4] >>> blurred_l = ndimage.gaussian_filter(lena, 3)
>>> # up <-> down flip
>>> flip_ud_lena = np.flipud(lena) increase the weight of edges by adding an approximation of the Laplacian:
>>> # rotation
>>> rotate_lena = ndimage.rotate(lena, 45) >>> filter_blurred_l = ndimage.gaussian_filter(blurred_l, 1)
>>> rotate_lena_noreshape = ndimage.rotate(lena, 45, reshape=False) >>> alpha = 30
>>> sharpened = blurred_l + alpha * (blurred_l - filter_blurred_l)
2.6.4 Image filtering
Local filters: replace the value of pixels by a function of the values of neighboring pixels.
Neighbourhood: square (choose size), disk, or more complicated structuring element. Denoising
Noisy lena:
>>> l = misc.lena()
>>> l = l[230:310, 210:350]
>>> noisy = l + 0.4 * l.std() * np.random.random(l.shape)
A Gaussian filter smoothes the noise out... and the edges as well:
Blurring/smoothing >>> gauss_denoised = ndimage.gaussian_filter(noisy, 2)
Gaussian filter from scipy.ndimage:
Most local linear isotropic filters blur the image (ndimage.uniform_filter) array([[0, 1, 0],
A median filter preserves better the edges: [1, 1, 1],
[0, 1, 0]])
>>> med_denoised = ndimage.median_filter(noisy, 3)
Median filter: better result for straight boundaries (low curvature):

>>> im = np.zeros((20, 20))
>>> im[5:-5, 5:-5] = 1
>>> im = ndimage.distance_transform_bf(im)
>>> im_noise = im + 0.2 * np.random.randn(*im.shape)
>>> im_med = ndimage.median_filter(im_noise, 3)
Erosion = minimum filter. Replace the value of a pixel by the minimal value covered by the structuring element.:
>>> a[1:6, 2:5] = 1
>>> a
array([[0, 0, 0, 0, 0, 0, 0],
[0, 0, 1, 1, 1, 0, 0],
[0, 0, 1, 1, 1, 0, 0],
[0, 0, 1, 1, 1, 0, 0],
[0, 0, 1, 1, 1, 0, 0],
Other rank filter: ndimage.maximum_filter, ndimage.percentile_filter [0, 0, 1, 1, 1, 0, 0],
[0, 0, 0, 0, 0, 0, 0]])
Other local non-linear filters: Wiener (scipy.signal.wiener), etc.
>>> ndimage.binary_erosion(a).astype(a.dtype)
Non-local filters array([[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 1, 0, 0, 0],
Exercise: denoising [0, 0, 0, 1, 0, 0, 0],
[0, 0, 0, 1, 0, 0, 0],
Create a binary image (of 0s and 1s) with several objects (circles, ellipses, squares, or random shapes). [0, 0, 0, 0, 0, 0, 0],
Add some noise (e.g., 20% of noise) [0, 0, 0, 0, 0, 0, 0]])
Try two different denoising methods for denoising the image: gaussian filtering and median filtering. >>> #Erosion removes objects smaller than the structure
Compare the histograms of the two different denoised images. Which one is the closest to the his- >>> ndimage.binary_erosion(a, structure=np.ones((5,5))).astype(a.dtype)
togram of the original (noise-free) image? array([[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0],
See also: [0, 0, 0, 0, 0, 0, 0],
More denoising filters are available in skimage.denoising, see the Scikit-image: image processing [0, 0, 0, 0, 0, 0, 0],
(page 324) tutorial. [0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0]])
Mathematical morphology
See http://en.wikipedia.org/wiki/Mathematical_morphology
Probe an image with a simple shape (a structuring element), and modify this image according to how the shape
locally fits or misses the image.
Structuring element:
>>> el = ndimage.generate_binary_structure(2, 1)
>>> el
array([[False, True, False],
[ True, True, True],
[False, True, False]], dtype=bool)
>>> el.astype(np.int) Dilation: maximum filter:
>>> a = np.zeros((5, 5)) >>> square = np.zeros((32, 32))

>>> a[2, 2] = 1 >>> square[10:-10, 10:-10] = 1
>>> a >>> np.random.seed(2)
array([[ 0., 0., 0., 0., 0.], >>> x, y = (32*np.random.random((2, 20))).astype(np.int)
[ 0., 0., 0., 0., 0.], >>> square[x, y] = 1
[ 0., 0., 1., 0., 0.],
[ 0., 0., 0., 0., 0.], >>> open_square = ndimage.binary_opening(square)
[ 0., 0., 0., 0., 0.]])
>>> ndimage.binary_dilation(a).astype(a.dtype) >>> eroded_square = ndimage.binary_erosion(square)
array([[ 0., 0., 0., 0., 0.], >>> reconstruction = ndimage.binary_propagation(eroded_square, mask=square)
[ 0., 0., 1., 0., 0.],
[ 0., 1., 1., 1., 0.],
[ 0., 0., 1., 0., 0.],
[ 0., 0., 0., 0., 0.]])
Also works for grey-valued images:

>>> im = np.zeros((64, 64)) Closing: dilation + erosion
>>> x, y = (63*np.random.random((2, 8))).astype(np.int)
>>> im[x, y] = np.arange(8) Many other mathematical morphology operations: hit and miss transform, tophat, etc.
>>> bigger_points = ndimage.grey_dilation(im, size=(5, 5), structure=np.ones((5, 5)))

2.6.5 Feature extraction
>>> square = np.zeros((16, 16))
>>> square[4:-4, 4:-4] = 1 Edge detection
>>> dist = ndimage.distance_transform_bf(square)
>>> dilate_dist = ndimage.grey_dilation(dist, size=(3, 3), \
Synthetic data:
... structure=np.ones((3, 3)))
>>> im = np.zeros((256, 256))
>>> im[64:-64, 64:-64] = 1
>>>
>>> im = ndimage.rotate(im, 15, mode='constant')
>>> im = ndimage.gaussian_filter(im, 8)
Opening: erosion + dilation: Use a gradient operator (Sobel) to find high intensity variations:
>>> a = np.zeros((5,5), dtype=np.int) >>> sx = ndimage.sobel(im, axis=0, mode='constant')

>>> a[1:4, 1:4] = 1; a[4, 4] = 1 >>> sy = ndimage.sobel(im, axis=1, mode='constant')
>>> a >>> sob = np.hypot(sx, sy)
array([[0, 0, 0, 0, 0],
[0, 1, 1, 1, 0],
[0, 1, 1, 1, 0],
[0, 1, 1, 1, 0],
[0, 0, 0, 0, 1]])
>>> # Opening removes small objects
>>> ndimage.binary_opening(a, structure=np.ones((3,3))).astype(np.int)
array([[0, 0, 0, 0, 0],
[0, 1, 1, 1, 0],
[0, 1, 1, 1, 0], Segmentation
[0, 1, 1, 1, 0],
[0, 0, 0, 0, 0]]) Histogram-based segmentation (no spatial information)
>>> # Opening can also smooth corners
>>> ndimage.binary_opening(a).astype(np.int) >>> n = 10
array([[0, 0, 0, 0, 0], >>> l = 256
[0, 0, 1, 0, 0], >>> im = np.zeros((l, l))
[0, 1, 1, 1, 0], >>> np.random.seed(1)
[0, 0, 1, 0, 0], >>> points = l*np.random.random((2, n**2))
[0, 0, 0, 0, 0]]) >>> im[(points[0]).astype(np.int), (points[1]).astype(np.int)] = 1
>>> im = ndimage.gaussian_filter(im, sigma=l/(4.*n))
Application: remove noise:
>>> mask = (im > im.mean()).astype(np.float)
>>> mask += 0.1 * im
>>> img = mask + 0.2*np.random.randn(*mask.shape) >>> from sklearn.feature_extraction import image

>>> from sklearn.cluster import spectral_clustering
>>> hist, bin_edges = np.histogram(img, bins=60)
>>> bin_centers = 0.5*(bin_edges[:-1] + bin_edges[1:]) >>> l = 100
>>> x, y = np.indices((l, l))
>>> binary_img = img > 0.5
>>> center1 = (28, 24)
>>> center2 = (40, 50)
>>> center3 = (67, 58)
>>> center4 = (24, 70)
>>> radius1, radius2, radius3, radius4 = 16, 14, 15, 14
>>> circle1 = (x - center1[0])**2 + (y - center1[1])**2 < radius1**2

>>> # 4 circles
Use mathematical morphology to clean up the result: >>> img = circle1 + circle2 + circle3 + circle4
>>> mask = img.astype(bool)
>>> # Remove small white regions >>> img = img.astype(float)
>>> open_img = ndimage.binary_opening(binary_img)
>>> # Remove small black hole >>> img += 1 + 0.2*np.random.randn(*img.shape)
>>> close_img = ndimage.binary_closing(open_img) >>> # Convert the image into a graph with the value of the gradient on
>>> # the edges.
>>> graph = image.img_to_graph(img, mask=mask)
>>> # Take a decreasing function of the gradient: we take it weakly

>>> # dependant from the gradient the segmentation is close to a voronoi
>>> graph.data = np.exp(-graph.data/graph.data.std())
>>> labels = spectral_clustering(graph, n_clusters=4, eigen_solver='arpack')

>>> label_im = -np.ones(mask.shape)
>>> label_im[mask] = labels
Exercise
Check that reconstruction operations (erosion + propagation) produce a better result than opening/closing:
>>> eroded_img = ndimage.binary_erosion(binary_img)
>>> reconstruct_img = ndimage.binary_propagation(eroded_img, mask=binary_img)
>>> tmp = np.logical_not(reconstruct_img)
>>> eroded_tmp = ndimage.binary_erosion(tmp)
>>> reconstruct_final = np.logical_not(ndimage.binary_propagation(eroded_tmp, mask=tmp))
>>> np.abs(mask - close_img).mean()
0.00727836...
>>> np.abs(mask - reconstruct_final).mean()
0.00059502...
Exercise
Check how a first denoising step (e.g. with a median filter) modifies the histogram, and check that the
resulting histogram-based segmentation is more accurate.
2.6.6 Measuring objects properties: ndimage.measurements
See also:
Synthetic data:
More advanced segmentation algorithms are found in the scikit-image: see Scikit-image: image processing
(page 324). >>> n = 10
>>> l = 256
See also: >>> im = np.zeros((l, l))
>>> points = l*np.random.random((2, n**2))
Other Scientific Packages provide algorithms that can be useful for image processing. In this example, we use the
>>> im[(points[0]).astype(np.int), (points[1]).astype(np.int)] = 1
spectral clustering function of the scikit-learn in order to segment glued objects.

>>> mask = im > im.mean()
Analysis of connected components

Label connected components: ndimage.label:
>>> label_im, nb_labels = ndimage.label(mask)
>>> nb_labels # how many regions?
16
>>> plt.imshow(label_im)
Compute size, mean_value, etc. of each region:

>>> sizes = ndimage.sum(mask, label_im, range(nb_labels + 1))
>>> mean_vals = ndimage.sum(im, label_im, range(1, nb_labels + 1))
Clean up small connect components:

>>> mask_size = sizes < 1000
>>> remove_pixel = mask_size[label_im]
>>> remove_pixel.shape
(256, 256)
>>> label_im[remove_pixel] = 0
>>> plt.imshow(label_im)
Now reassign labels with np.searchsorted:

>>> labels = np.unique(label_im)
>>> label_im = np.searchsorted(labels, label_im)
Find region of interest enclosing object:

>>> slice_x, slice_y = ndimage.find_objects(label_im==4)[0]
>>> roi = im[slice_x, slice_y]
>>> plt.imshow(roi)
Other spatial measures: ndimage.center_of_mass, ndimage.maximum_position, etc.

Can be used outside the limited scope of segmentation applications.
Example: block mean:
>>> l = misc.lena()
>>> sx, sy = l.shape
>>> X, Y = np.ogrid[0:sx, 0:sy]
>>> regions = (sy//6) * (X//4) + (Y//6) # note that we use broadcasting
>>> block_mean = ndimage.mean(l, labels=regions, index=np.arange(1,
... regions.max() +1)) Correlation function, Fourier/wavelet spectrum, etc.

>>> block_mean.shape = (sx // 4, sy // 6) One example with mathematical morphology: granulometry (http://en.wikipedia.org/wiki/Granulometry_%28morphology%29)
>>> def disk_structure(n):
... struct = np.zeros((2 * n + 1, 2 * n + 1))
... x, y = np.indices((2 * n + 1, 2 * n + 1))
... mask = (x - n)**2 + (y - n)**2 <= n**2
... struct[mask] = 1
... return struct.astype(np.bool)
...
>>>
>>> def granulometry(data, sizes=None):
... s = max(data.shape)
... if sizes == None:
... sizes = range(1, s/2, 2)
... granulo = [ndimage.binary_opening(data, \
... structure=disk_structure(n)).sum() for n in sizes]
... return granulo
...
>>>
>>> n = 10
>>> l = 256
>>> im = np.zeros((l, l))
>>> points = l*np.random.random((2, n**2))
>>>
When regions are regular blocks, it is more efficient to use stride tricks (Example: fake dimensions with strides >>> mask = im > im.mean()
(page 165)). >>>
Non-regularly-spaced blocks: radial mean: >>> granulo = granulometry(mask, sizes=np.arange(2, 19, 4))
>>> sx, sy = l.shape

>>> X, Y = np.ogrid[0:sx, 0:sy]
>>> r = np.hypot(X - sx/2, Y - sy/2)
>>> rbin = (20* r/r.max()).astype(np.int)
>>> radial_mean = ndimage.mean(l, labels=rbin, index=np.arange(1, rbin.max() +1))
See also:
More on image-processing:
The chapter on Scikit-image (page 324)
Other, more powerful and complete modules: OpenCV (Python bindings), CellProfiler, ITK with Python
bindings
Other measures
2.7 Mathematical optimization: finding minima of functions

Dimensionality of the problem
Authors: Gal Varoquaux The scale of an optimization problem is pretty much set by the dimensionality of the problem, i.e. the number
of scalar variables on which the search is performed.
Mathematical optimization deals with the problem of finding numerically minimums (or maximums or zeros) of
a function. In this context, the function is called cost function, or objective function, or energy.
Here, we are interested in using scipy.optimize for black-box optimization: we do not rely on the mathe-
matical expression of the function that we are optimizing. Note that this expression can often be used for more Convex versus non-convex optimization
efficient, non black-box, optimization.
Prerequisites
Numpy, Scipy
IPython
matplotlib
References
Mathematical optimization is very ... mathematical. If you want performance, it really pays to read the
books:
Convex Optimization by Boyd and Vandenberghe (pdf available free online).
Numerical Optimization, by Nocedal and Wright. Detailed reference on gradient descent methods.
Practical Methods of Optimization by Fletcher: good at hand-waving explainations.
A convex function: A non-convex function

f is above all its tangents.
Chapters contents equivalently, for two point A, B, f(C) lies below
Knowing your problem (page 239) the segment [f(A), f(B])], if A < C < B
Convex versus non-convex optimization (page 240)
Smooth and non-smooth problems (page 240) Optimizing convex functions is easy. Optimizing non-convex functions can be very hard.
Noisy versus exact cost functions (page 241)
Constraints (page 241) Note: It can be proven that for a convex function a local minimum is also a global minimum. Then, in some
A review of the different optimizers (page 241) sense, the minimum is unique.
Getting started: 1D optimization (page 241)
Gradient based methods (page 242)
Newton and quasi-newton methods (page 244)
Smooth and non-smooth problems
Gradient-less methods (page 246)
Global optimizers (page 248)
Practical guide to optimization with scipy (page 248)
Choosing a method (page 248)
Making your optimizer faster (page 249)
Computing gradients (page 249)
Synthetic exercices (page 249)
Special case: non-linear least-squares (page 250)
Minimizing the norm of a vector function (page 250)
Curve fitting (page 251)
Optimization with constraints (page 251)
Box bounds (page 251)
General constraints (page 252)
2.7.1 Knowing your problem
Not all optimization problems are equal. Knowing your problem enables you to choose the right tool. A smooth function: A non-smooth function
The gradient is defined everywhere, and is a continuous function
2.7. Mathematical optimization: finding minima of functions 239 2.7. Mathematical optimization: finding minima of functions 240
Optimizing smooth functions is easier (true in the context of black-box optimization, otherwise Linear Program-
ming is an example of methods which deal very efficiently with piece-wise linear functions).
Noisy versus exact cost functions

Brents method on a quadratic function: it converges
in 3 iterations, as the quadratic approximation is then
exact.
Brents method on a non-convex function: note that

the fact that the optimizer avoided the local minimum is a
matter of luck.
>>> from scipy import optimize
>>> def f(x):
... return -np.exp(-(x - .7)**2)
>>> x_min = optimize.brent(f) # It actually converges in 9 iterations!
>>> x_min
Noisy (blue) and non-noisy (green) functions 0.699999999...
>>> x_min - .7
Noisy gradients -2.1605...e-10
Many optimization methods rely on gradients of the objective function. If the gradient function is not given,
they are computed numerically, which induces errors. In such situation, even if the objective function is not Note: Brents method can also be used for optimization constrained to an interval using
noisy, scipy.optimize.fminbound()
Note: In scipy 0.11, scipy.optimize.minimize_scalar() gives a generic interface to 1D scalar mini-

Constraints mization
Gradient based methods
Some intuitions about gradient descent
Here we focus on intuitions, not code. Code will follow.

Gradient descent basically consists in taking small steps in the direction of the gradient, that is the direction of the
steepest descent.
Table 2.2: Fixed step gradient descent
Optimizations under constraints

Here:
1 < 1 < 1
1 < 2 < 1 A well-conditionned quadratic function.
2.7.2 A review of the different optimizers
Getting started: 1D optimization An ill-conditionned quadratic function.

The core problem of gradient-methods on ill-conditioned
Use scipy.optimize.brent() to minimize 1D functions. It combines a bracketing strategy with a parabolic problems is that the gradient tends not to point in the direction
approximation. of the minimum.
We can see that very anisotropic (ill-conditionned) functions are harder to optimize.
The conjugate gradient solves this problem by adding a friction term: each step depends on the two last values of
Take home message: conditioning number and preconditioning the gradient and sharp turns are reduced.
If you know natural scaling for your variables, prescale them so that they behave similarly. This is related to
preconditioning. Table 2.4: Conjugate gradient descent
Also, it clearly can be advantageous to take bigger steps. This is done in gradient descent code using a line search.
Table 2.3: Adaptive step gradient descent
An ill-conditionned
non-quadratic function.
A well-conditionned quadratic
function.
An ill-conditionned very
Methods based on conjugate gradient are named with cg in scipy. The simple conjugate gradient method to
minimize a function is scipy.optimize.fmin_cg():
An ill-conditionned quadratic >>> def f(x): # The rosenbrock function
function. ... return .5*(1 - x[0])**2 + (x[1] - x[0]**2)**2
>>> optimize.fmin_cg(f, [2, 2])
Current function value: 0.000000
Iterations: 13
array([ 0.99998968, 0.99997855])
These methods need the gradient of the function. They can compute it, but will perform better if you can pass
An ill-conditionned
them the gradient:
>>> def fprime(x):
... return np.array((-2*.5*(1 - x[0]) - 4*x[0]*(x[1] - x[0]**2), 2*(x[1] - x[0]**2)))
>>> optimize.fmin_cg(f, [2, 2], fprime=fprime)
Iterations: 13
An ill-conditionned very array([ 0.99999199, 0.99998336])
Note that the function has only been evaluated 30 times, compared to 120 without the gradient.
The more a function looks like a quadratic function (elliptic iso-curves), the easier it is to optimize.
Newton and quasi-newton methods

Conjugate gradient descent
Newton methods: using the Hessian (2nd differential)
The gradient descent algorithms above are toys not to be used on real problems.
As can be seen from the above experiments, one of the problems of the simple gradient descent algorithms, is that Newton methods use a local quadratic approximation to compute the jump direction. For this purpose, they rely
it tends to oscillate across a valley, each time following the direction of the gradient, that makes it cross the valley. on the 2 first derivative of the function: the gradient and the Hessian.
Quasi-Newton methods: approximating the Hessian on the fly
BFGS: BFGS (Broyden-Fletcher-Goldfarb-Shanno algorithm) refines at each step an approximation of the Hes-
sian.
An ill-conditionned quadratic function:
Note that, as the quadratic approximation is exact, the Newton
method is blazing fast

On a exactly quadratic function, BFGS is not as fast as
An ill-conditionned non-quadratic function: Newtons method, but still very fast.
Here we are optimizing a Gaussian, which is always below its
quadratic approximation. As a result, the Newton method
overshoots and leads to oscillations.
An ill-conditionned non-quadratic function:

Here BFGS does better than Newton, as its empirical estimate
of the curvature is better than that given by the Hessian.
An ill-conditionned very non-quadratic function:
In scipy, the Newton method for optimization is implemented in scipy.optimize.fmin_ncg() (cg here
refers to that fact that an inner operation, the inversion of the Hessian, is performed by conjugate gradient).
scipy.optimize.fmin_tnc() can be use for constraint problems, although it is less versatile:
An ill-conditionned very non-quadratic function:
>>> def f(x): # The rosenbrock function >>> def f(x): # The rosenbrock function
... return .5*(1 - x[0])**2 + (x[1] - x[0]**2)**2 ... return .5*(1 - x[0])**2 + (x[1] - x[0]**2)**2
>>> def fprime(x): >>> def fprime(x):
... return np.array((-2*.5*(1 - x[0]) - 4*x[0]*(x[1] - x[0]**2), 2*(x[1] - x[0]**2))) ... return np.array((-2*.5*(1 - x[0]) - 4*x[0]*(x[1] - x[0]**2), 2*(x[1] - x[0]**2)))
>>> optimize.fmin_ncg(f, [2, 2], fprime=fprime) >>> optimize.fmin_bfgs(f, [2, 2], fprime=fprime)
Optimization terminated successfully. Optimization terminated successfully.
Current function value: 0.000000 Current function value: 0.000000
Iterations: 9 Iterations: 16
Function evaluations: 11 Function evaluations: 24
Gradient evaluations: 51 Gradient evaluations: 24
Hessian evaluations: 0 array([ 1.00000017, 1.00000026])
array([ 1., 1.])
L-BFGS: Limited-memory BFGS Sits between BFGS and conjugate gradient: in very high dimensions (> 250)
Note that compared to a conjugate gradient (above), Newtons method has required less function evaluations, but
the Hessian matrix is too costly to compute and invert. L-BFGS keeps a low-rank version. In addition, the scipy
more gradient evaluations, as it uses it to approximate the Hessian. Lets compute the Hessian and pass it to the
version, scipy.optimize.fmin_l_bfgs_b(), includes box bounds:
algorithm:
>>> def f(x): # The rosenbrock function
>>> def hessian(x): # Computed with sympy ... return .5*(1 - x[0])**2 + (x[1] - x[0]**2)**2
... return np.array(((1 - 4*x[1] + 12*x[0]**2, -4*x[0]), (-4*x[0], 2))) >>> def fprime(x):
>>> optimize.fmin_ncg(f, [2, 2], fprime=fprime, fhess=hessian) ... return np.array((-2*.5*(1 - x[0]) - 4*x[0]*(x[1] - x[0]**2), 2*(x[1] - x[0]**2)))
Optimization terminated successfully. >>> optimize.fmin_l_bfgs_b(f, [2, 2], fprime=fprime)
Current function value: 0.000000 (array([ 1.00000005, 1.00000009]), 1.4417677473011859e-15, {...})
Iterations: 9
Gradient evaluations: 19 Note: If you do not specify the gradient to the L-BFGS solver, you need to add approx_grad=1
Hessian evaluations: 9
array([ 1., 1.])
Gradient-less methods
Note: At very high-dimension, the inversion of the Hessian can be costly and unstable (large scale > 250).
A shooting method: the Powell algorithm
Note: Newton optimizers should not to be confused with Newtons root finding method, based on the same
principles, scipy.optimize.newton(). Almost a gradient approach
Global optimizers
If your problem does not admit a unique local minimum (which can be hard to test unless the function is convex),
and you do not have prior information to initialize the optimization close to the solution, you may need a global
optimizer.
Powells method isnt too sensitive to local
ill-conditionning in low dimensions Brute force: a grid search
scipy.optimize.brute() evaluates the function on a given grid of parameters and returns the parameters
corresponding to the minimum value. The parameters are specified with ranges given to numpy.mgrid. By
default, 20 steps are taken in each direction:
An ill-conditionned very non-quadratic function: >>> def f(x): # The rosenbrock function
... return .5*(1 - x[0])**2 + (x[1] - x[0]**2)**2
>>> optimize.brute(f, ((-1, 2), (-1, 2)))
Simplex method: the Nelder-Mead array([ 1.00001462, 1.00001547])
The Nelder-Mead algorithms is a generalization of dichotomy approaches to high-dimensional spaces. The al-
gorithm works by refining a simplex, the generalization of intervals and triangles to high-dimensional spaces, to 2.7.3 Practical guide to optimization with scipy
bracket the minimum.
Choosing a method
Strong points: it is robust to noise, as it does not rely on computing gradients. Thus it can work on functions that
are not locally smooth such as experimental data points, as long as they display a large-scale bell-shape behavior.
However it is slower than gradient-based methods on smooth, non-noisy functions.
An ill-conditionned
non-quadratic function:
Without knowledge of the gradient

In general, prefer BFGS (scipy.optimize.fmin_bfgs()) or L-BFGS
(scipy.optimize.fmin_l_bfgs_b()), even if you have to approximate nu-
An ill-conditionned very merically gradients
non-quadratic function: On well-conditioned problems, Powell (scipy.optimize.fmin_powell()) and
In scipy, scipy.optimize.fmin() implements the Nelder-Mead approach: Nelder-Mead (scipy.optimize.fmin()), both gradient-free methods, work well in
high dimension, but they collapse for ill-conditioned problems.
>>> def f(x): # The rosenbrock function
... return .5*(1 - x[0])**2 + (x[1] - x[0]**2)**2 With knowledge of the gradient
>>> optimize.fmin(f, [2, 2]) BFGS (scipy.optimize.fmin_bfgs()) or L-BFGS
Optimization terminated successfully. (scipy.optimize.fmin_l_bfgs_b()).
Iterations: 46 Computational overhead of BFGS is larger than that L-BFGS, itself larger than that of
Function evaluations: 91 conjugate gradient. On the other side, BFGS usually needs less function evaluations than
array([ 0.99998568, 0.99996682]) CG. Thus conjugate gradient method is better than BFGS at optimizing computationally
cheap functions.
With the Hessian
If you can compute the Hessian, prefer the Newton method
(scipy.optimize.fmin_ncg()).
If you have noisy measurements

Use Nelder-Mead (scipy.optimize.fmin()) or Powell Exercice: A locally flat minimum
(scipy.optimize.fmin_powell()). Consider the function exp(-1/(.1*x**2 + y**2). This function admits a minimum in (0, 0). Starting
from an initialization at (1, 1), try to get within 1e-8 of this minimum point.
Making your optimizer faster
Choose the right method (see above), do compute analytically the gradient and Hessian, if you can.
Use preconditionning when possible.
Choose your initialization points wisely. For instance, if you are running many similar optimizations, warm-
restart one with the results of another.
Relax the tolerance if you dont need precision
Computing gradients
Computing gradients, and even more Hessians, is very tedious but worth the effort. Symbolic computation with
Sympy (page 318) may come in handy.
Warning: A very common source of optimization not converging well is human error in the computation of 2.7.4 Special case: non-linear least-squares
the gradient. You can use scipy.optimize.check_grad() to check that your gradient is correct. It
returns the norm of the different between the gradient given, and a gradient computed numerically: Minimizing the norm of a vector function
>>> optimize.check_grad(f, fprime, [2, 2])
Least square problems, minimizing the norm of a vector function, have a specific structure that can be used in the
2.384185791015625e-07
LevenbergMarquardt algorithm implemented in scipy.optimize.leastsq().
See also scipy.optimize.approx_fprime() to find your errors.
Lets try to minimize the norm of the following vectorial function:
>>> def f(x):
... return np.arctan(x) - np.arctan(np.linspace(0, 1, len(x)))
Synthetic exercices
>>> x0 = np.zeros(10)
>>> optimize.leastsq(f, x0)
(array([ 0. , 0.11111111, 0.22222222, 0.33333333, 0.44444444,
0.55555556, 0.66666667, 0.77777778, 0.88888889, 1. ]), 2)
This took 67 function evaluations (check it with full_output=1). What if we compute the norm ourselves and use
a good generic optimizer (BFGS):
>>> def g(x):
... return np.sum(f(x)**2)
>>> optimize.fmin_bfgs(g, x0)
Iterations: 11
Exercice: A simple (?) quadratic function Function evaluations: 144
Optimize the following function, using K[0] as a starting point: array([ -7.4...-09, 1.1...e-01, 2.2...e-01,
3.3...e-01, 4.4...e-01, 5.5...e-01,
np.random.seed(0)
6.6...e-01, 7.7...e-01, 8.8...e-01,
K = np.random.normal(size=(100, 100))
1.0...e+00])
def f(x):
return np.sum((np.dot(K, x - 1))**2) + np.sum(x**2)**2
BFGS needs more function calls, and gives a less precise result.
Time your approach. Find the fastest approach. Why is BFGS not working well? Note: leastsq is interesting compared to BFGS only if the dimensionality of the output vector is large, and
larger than the number of parameters to optimize.
Warning: If the function is linear, this is a linear-algebra problem, and should be solved with
scipy.linalg.lstsq().
Curve fitting scipy.optimize.fmin_l_bfgs_b() a quasi-Newton (page 246) method with bound constraints:
>>> def f(x):
... return np.sqrt((x[0] - 3)**2 + (x[1] - 2)**2)
>>> optimize.fmin_l_bfgs_b(f, np.array([0, 0]), approx_grad=1, bounds=((-1.5, 1.5), (-1.5, 1.
(array([ 1.5, 1.5]), 1.5811388300841898, {...})
General constraints
Equality and inequality constraints specified as functions: f(x) = 0 and g(x)< 0.

scipy.optimize.fmin_slsqp() Sequential least square programming: equality and inequality con-
Least square problems occur often when fitting a non-

linear to data. While it is possible to construct our optimization problem ourselves, scipy provides a helper
function for this purpose: scipy.optimize.curve_fit():
.. Comment to make doctest pass
>>> def f(t, omega, phi): straints:

... return np.cos(omega * t + phi)
>>> def f(x):
>>> x = np.linspace(0, 3, 50) ... return np.sqrt((x[0] - 3)**2 + (x[1] - 2)**2)
>>> y = f(x, 1.5, 1) + .1*np.random.normal(size=50)
>>> def constraint(x):
... return np.atleast_1d(1.5 - np.sum(np.abs(x)))
>>> optimize.curve_fit(f, x, y)
(array([ 1.51854577, 0.92665541]), array([[ 0.00037994, -0.00056796],
>>> optimize.fmin_slsqp(f, np.array([0, 0]), ieqcons=[constraint, ])
[-0.00056796, 0.00123978]]))
Optimization terminated successfully. (Exit mode 0)
Iterations: 5
Exercise Function evaluations: 20
Do the same with omega = 3. What is the difficulty? Gradient evaluations: 5
array([ 1.25004696, 0.24995304])
scipy.optimize.fmin_cobyla() Constraints optimization by linear approximation: inequality

2.7.5 Optimization with constraints constraints only:
>>> optimize.fmin_cobyla(f, np.array([0, 0]), cons=constraint)
Box bounds array([ 1.25009622, 0.24990378])
Box bounds correspond to limiting each of the individual parameters of the optimization. Note that some Warning: The above problem is known as the Lasso problem in statistics, and there exists very efficient
problems that are not originally written as box bounds can be rewritten as such via change of variables. solvers for it (for instance in scikit-learn). In general do not use generic solvers when specific ones exist.
Lagrange multipliers
If you are ready to do a bit of math, many constrained optimization problems can be converted to non-
constrained optimization problems using a mathematical trick known as Lagrange multipliers.
2.8 Interfacing with C
scipy.optimize.fminbound() for 1D-optimization Author: Valentin Haenel
2.7. Mathematical optimization: finding minima of functions 251 2.8. Interfacing with C 252
This chapter contains an introduction to the many different routes for making your native code (primarily C/C++) You will need a C compiler for most of the examples.
available from Python, a process commonly referred to wrapping. The goal of this chapter is to give you a
flavour of what technologies exist and what their respective merits and shortcomings are, so that you can select
the appropriate one for your specific needs. In any case, once you do start wrapping, you almost certainly will 2.8.2 Python-C-Api
want to consult the respective documentation for your selected technique.
The Python-C-API is the backbone of the standard Python interpreter (a.k.a CPython). Using this API it is possible
to write Python extension module in C and C++. Obviously, these extension modules can, by virtue of language
Chapters contents compatibility, call any function written in C or C++.
Introduction (page 253) When using the Python-C-API, one usually writes much boilerplate code, first to parse the arguments that were
Python-C-Api (page 254) given to a function, and later to construct the return type.
Ctypes (page 258)
SWIG (page 261) Advantages
Cython (page 265) Requires no additional libraries
Summary (page 269)
Further Reading and References (page 269) Lots of low-level control
Exercises (page 269) Entirely usable from C++
Disadvantages
May require a substantial amount of effort
2.8.1 Introduction
Much overhead in the code
This chapter covers the following techniques:
Must be compiled
Python-C-Api
High maintenance cost
Ctypes
No forward compatibility across Python versions as C-API changes
SWIG (Simplified Wrapper and Interface Generator)
Reference count bugs are easy to create and very hard to track down.
Cython
Note: The Python-C-Api example here serves mainly for didactic reasons. Many of the other techniques actually
These four techniques are perhaps the most well known ones, of which Cython is probably the most advanced one depend on this, so it is good to have a high-level understanding of how it works. In 99% of the use-cases you will
and the one you should consider using first. The others are also important, if you want to understand the wrapping be better off, using an alternative technique.
problem from different angles. Having said that, there are other alternatives out there, but having understood the
basics of the ones above, you will be in a position to evaluate the technique of your choice to see if it fits your
needs. Note: Since refernce counting bugs are easy to create and hard to track down, anyone really needing to use
the Python C-API should read the section about objects, types and reference counts from the official python
The following criteria may be useful when evaluating a technology: documentation. Additionally, there is a tool by the name of cpychecker which can help discover common errors
Are additional libraries required? with reference counting.
Is the code autogenerated?
Does it need to be compiled? Example
Is there good support for interacting with Numpy arrays?
The following C-extension module, make the cos function from the standard math library available to Python:
Does it support C++?
/* Example of wrapping cos function from math.h with the Python-C-API. */
Before you set out, you should consider your use case. When interfacing with native code, there are usually two
use-cases that come up: #include <Python.h>
Existing code in C/C++ that needs to be leveraged, either because it already exists, or because it is faster. #include <math.h>
Python code too slow, push inner loops to native code /* wrapped cosine function */
static PyObject* cos_func(PyObject* self, PyObject* args)
Each technology is demonstrated by wrapping the cos function from math.h. While this is a mostly a trivial {
example, it should serve us well to demonstrate the basics of the wrapping solution. Since each technique also double value;
includes some form of Numpy support, this is also demonstrated using an example where the cosine is computed double answer;
on some kind of array.
/* parse the input, from python float to c double */
Last but not least, two small warnings: if (!PyArg_ParseTuple(args, "d", &value))
All of these techniques may crash (segmentation fault) the Python interpreter, which is (usually) due to bugs return NULL;
in the C code. /* if the above function returns -1, an appropriate Python exception will
* have been set, and the function simply returns NULL
All the examples have been done on Linux, they should be possible on other operating systems. */
2.8. Interfacing with C 253 2.8. Interfacing with C 254

Type: module
/* call cos from libm */
String Form:<module 'cos_module' from 'cos_module.so'>
answer = cos(value);
File: /home/esc/git-working/scipy-lecture-notes/advanced/interfacing_with_c/python_c_api/cos
Docstring: <no docstring>
/* construct the output from cos, from c double to python float */
return Py_BuildValue("f", answer);
In [3]: dir(cos_module)
}
Out[3]: ['__doc__', '__file__', '__name__', '__package__', 'cos_func']
/* define functions in module */
In [4]: cos_module.cos_func(1.0)
static PyMethodDef CosMethods[] =
Out[4]: 0.5403023058681398
{
{"cos_func", cos_func, METH_VARARGS, "evaluate the cosine"},
{NULL, NULL, 0, NULL}
Out[5]: 1.0
};
/* module initialization */
Out[7]: -1.0
PyMODINIT_FUNC
initcos_module(void)
Now lets see how robust this is:
{ In [10]: cos_module.cos_func('foo')
(void) Py_InitModule("cos_module", CosMethods); ---------------------------------------------------------------------------
} TypeError Traceback (most recent call last)
<ipython-input-10-11bee483665d> in <module>()
As you can see, there is much boilerplate, both to massage the arguments and return types into place and for ----> 1 cos_module.cos_func('foo')
the module initialisation. Although some of this is amortised, as the extension grows, the boilerplate required for
each function(s) remains. TypeError: a float is required
The standard python build system distutils supports compiling C-extensions from a setup.py, which is
rather convenient:
Numpy Support
from distutils.core import setup, Extension
Analog to the Python-C-API, Numpy, which is itself implemented as a C-extension, comes with the Numpy-C-
# define the extension module API. This API can be used to create and manipulate Numpy arrays from C, when writing a custom C-extension.
cos_module = Extension('cos_module', sources=['cos_module.c']) See also: :ref:advanced_numpy_.
# run the setup
Note: If you do ever need to use the Numpy C-API refer to the documentation about Arrays and Iterators.
setup(ext_modules=[cos_module])
This can be compiled: The following example shows how to pass Numpy arrays as arguments to functions and how to iterate over Numpy
arrays using the (old) Numpy-C-API. It simply takes an array as argument applies the cosine function from the
$ cd advanced/interfacing_with_c/python_c_api math.h and returns a resulting new array.
$ ls /* Example of wrapping the cos function from math.h using the Numpy-C-API. */
cos_module.c setup.py
#include <Python.h>
$ python setup.py build_ext --inplace #include <numpy/arrayobject.h>
running build_ext #include <math.h>
building 'cos_module' extension
creating build /* wrapped cosine function */
creating build/temp.linux-x86_64-2.7 static PyObject* cos_func_np(PyObject* self, PyObject* args)
{
gcc -pthread -fno-strict-aliasing -g -O2 -DNDEBUG -g -fwrapv -O3 -Wall -Wstrict-prototypes -fPIC -I/home/esc/anaconda/include/python2.7 -c cos_module.c -o build/temp.linux-x86_64-2.7/cos_module.o
gcc -pthread -shared build/temp.linux-x86_64-2.7/cos_module.o -L/home/esc/anaconda/lib -lpython2.7 -o /home/esc/git-working/scipy-lecture-notes/advanced/interfacing_with_c/python_c_api/cos_module.so
PyArrayObject *in_array;
$ ls PyObject *out_array;
build/ cos_module.c cos_module.so setup.py NpyIter *in_iter;
NpyIter *out_iter;
build_ext is to build extension modules NpyIter_IterNextFunc *in_iternext;
NpyIter_IterNextFunc *out_iternext;
--inplace will output the compiled extension module into the current directory
The file cos_module.so contains the compiled extension, which we can now load in the IPython interpreter: /* parse single numpy array argument */
if (!PyArg_ParseTuple(args, "O!", &PyArray_Type, &in_array))
In [1]: import cos_module return NULL;
In [2]: cos_module? /* construct the output array, like the input array */

out_array = PyArray_NewLikeArray(in_array, NPY_ANYORDER, NULL, 0); from distutils.core import setup, Extension
if (out_array == NULL) import numpy
return NULL;
# define the extension module
/* create the iterators */ cos_module_np = Extension('cos_module_np', sources=['cos_module_np.c'],
in_iter = NpyIter_New(in_array, NPY_ITER_READONLY, NPY_KEEPORDER, include_dirs=[numpy.get_include()])
NPY_NO_CASTING, NULL);
if (in_iter == NULL) # run the setup
goto fail; setup(ext_modules=[cos_module_np])
out_iter = NpyIter_New((PyArrayObject *)out_array, NPY_ITER_READWRITE, To convince ourselves if this does actually works, we run the following test script:
NPY_KEEPORDER, NPY_NO_CASTING, NULL);
if (out_iter == NULL) { import cos_module_np
NpyIter_Deallocate(in_iter); import numpy as np
goto fail; import pylab
}
x = np.arange(0, 2 * np.pi, 0.1)
in_iternext = NpyIter_GetIterNext(in_iter, NULL); y = cos_module_np.cos_func_np(x)
out_iternext = NpyIter_GetIterNext(out_iter, NULL); pylab.plot(x, y)
if (in_iternext == NULL || out_iternext == NULL) { pylab.show()
NpyIter_Deallocate(in_iter);
NpyIter_Deallocate(out_iter); And this should result in the following figure:
goto fail;
}
double ** in_dataptr = (double **) NpyIter_GetDataPtrArray(in_iter);
double ** out_dataptr = (double **) NpyIter_GetDataPtrArray(out_iter);
/* iterate over the arrays */

do {
**out_dataptr = cos(**in_dataptr);
} while(in_iternext(in_iter) && out_iternext(out_iter));
2.8.3 Ctypes
/* clean up and return the result */
NpyIter_Deallocate(in_iter); Ctypes is a foreign function library for Python. It provides C compatible data types, and allows calling functions
NpyIter_Deallocate(out_iter); in DLLs or shared libraries. It can be used to wrap these libraries in pure Python.
Py_INCREF(out_array);
return out_array; Advantages
Part of the Python standard library
/* in case bad things happen */
fail: Does not need to be compiled
Py_XDECREF(out_array);
return NULL; Wrapping code entirely in Python
} Disadvantages
/* define functions in module */ Requires code to be wrapped to be available as a shared library (roughly speaking *.dll in Windows
static PyMethodDef CosMethods[] = *.so in Linux and *.dylib in Mac OSX.)
{
{"cos_func_np", cos_func_np, METH_VARARGS, No good support for C++
"evaluate the cosine on a numpy array"},
{NULL, NULL, 0, NULL}
}; Example
/* module initialization */ As advertised, the wrapper code is in pure Python.

PyMODINIT_FUNC
""" Example of wrapping cos function from math.h using ctypes. """
initcos_module_np(void)
{
import ctypes
(void) Py_InitModule("cos_module_np", CosMethods);
from ctypes.util import find_library
/* IMPORTANT: this must be called */
import_array();
# find and load the library
}
libm = ctypes.cdll.LoadLibrary(find_library('m'))
# set the argument type
To compile this we can use distutils again. However we need to be sure to include the Numpy headers by using
libm.cos.argtypes = [ctypes.c_double]
:func:numpy.get_include.

# set the return type Numpy Support

libm.cos.restype = ctypes.c_double
Numpy contains some support for interfacing with ctypes. In particular there is support for exporting certain
attributes of a Numpy array as ctypes data-types and there are functions to convert from C arrays to Numpy arrays
def cos_func(arg): and back.
''' Wrapper for cos from math.h '''
return libm.cos(arg) For more information, consult the corresponding section in the Numpy Cookbook and the API documentation for
numpy.ndarray.ctypes and numpy.ctypeslib.
Finding and loading the library may vary depending on your operating system, check the documentation for For the following example, lets consider a C function in a library that takes an input and an output array, computes
details the cosine of the input array and stores the result in the output array.
This may be somewhat deceptive, since the math library exists in compiled form on the system already. If The library consists of the following header file (although this is not strictly needed for this example, we list it for
you were to wrap a in-house library, you would have to compile it first, which may or may not require some completeness):
additional effort.
void cos_doubles(double * in_array, double * out_array, int size);
We may now use this, as before:
In [1]: import cos_module The function implementation resides in the following C source file:
#include <math.h>
In [2]: cos_module?
Type: module /* Compute the cosine of each element in in_array, storing the result in
String Form:<module 'cos_module' from 'cos_module.py'>
* out_array. */
File: /home/esc/git-working/scipy-lecture-notes/advanced/interfacing_with_c/ctypes/cos_module.py void cos_doubles(double * in_array, double * out_array, int size){
Docstring: <no docstring> int i;
for(i=0;i<size;i++){
In [3]: dir(cos_module) out_array[i] = cos(in_array[i]);
Out[3]: }
['__builtins__', }
'__doc__',
'__file__',
And since the library is pure C, we cant use distutils to compile it, but must use a combination of make and
'__name__',
gcc:
'__package__',
'cos_func', m.PHONY : clean
'ctypes',
'find_library', libcos_doubles.so : cos_doubles.o
'libm'] gcc -shared -Wl,-soname,libcos_doubles.so -o libcos_doubles.so cos_doubles.o
In [4]: cos_module.cos_func(1.0) cos_doubles.o : cos_doubles.c

Out[4]: 0.5403023058681398 gcc -c -fPIC cos_doubles.c -o cos_doubles.o
In [5]: cos_module.cos_func(0.0) clean :

Out[5]: 1.0 -rm -vf libcos_doubles.so cos_doubles.o cos_doubles.pyc
In [6]: cos_module.cos_func(3.14159265359) We can then compile this (on Linux) into the shared library libcos_doubles.so:
Out[6]: -1.0
$ ls
As with the previous example, this code is somewhat robust, although the error message is not quite as helpful, cos_doubles.c cos_doubles.h cos_doubles.py makefile test_cos_doubles.py
since it does not tell us what the type should be. $ make
gcc -c -fPIC cos_doubles.c -o cos_doubles.o
In [7]: cos_module.cos_func('foo') gcc -shared -Wl,-soname,libcos_doubles.so -o libcos_doubles.so cos_doubles.o
--------------------------------------------------------------------------- $ ls
ArgumentError Traceback (most recent call last) cos_doubles.c cos_doubles.o libcos_doubles.so* test_cos_doubles.py
<ipython-input-7-11bee483665d> in <module>() cos_doubles.h cos_doubles.py makefile
----> 1 cos_module.cos_func('foo')
Now we can proceed to wrap this library via ctypes with direct support for (certain kinds of) Numpy arrays:
/home/esc/git-working/scipy-lecture-notes/advanced/interfacing_with_c/ctypes/cos_module.py in cos_func(arg)
12 def cos_func(arg): """ Example of wrapping a C library function that accepts a C double array as
13 ''' Wrapper for cos from math.h ''' input using the numpy.ctypeslib. """
---> 14 return libm.cos(arg)
import numpy as np
ArgumentError: argument 1: <type 'exceptions.TypeError'>: wrong type import numpy.ctypeslib as npct
from ctypes import c_int
# input type for the cos_doubles function

# must be a double array, with single dimension that is contiguous Steep learning curve
array_1d_double = npct.ndpointer(dtype=np.double, ndim=1, flags='CONTIGUOUS')
# load the library, using numpy mechanisms

Example
libcd = npct.load_library("libcos_doubles", ".")
Lets imagine that our cos function lives in a cos_module which has been written in c and consists of the
# setup the return types and argument types source file cos_module.c:
libcd.cos_doubles.restype = None
#include <math.h>
libcd.cos_doubles.argtypes = [array_1d_double, array_1d_double, c_int]
double cos_func(double arg){
return cos(arg);
def cos_doubles_func(in_array, out_array):
}
return libcd.cos_doubles(in_array, out_array, len(in_array))
and the header file cos_module.h:
Note the inherent limitation of contiguous single dimensional Numpy arrays, since the C functions requires
this kind of buffer. double cos_func(double arg);
Also note that the output array must be preallocated, for example with numpy.zeros() and the function And our goal is to expose the cos_func to Python. To achieve this with SWIG, we must write an interface file
will write into its buffer. which contains the instructions for SWIG.
Although the original signature of the cos_doubles function is ARRAY, ARRAY, int the final /* Example of wrapping cos function from math.h using SWIG. */
cos_doubles_func takes only two Numpy arrays as arguments.
And, as before, we convince ourselves that it worked: %module cos_module
%{
import numpy as np /* the resulting C file should be built as a python extension */
import pylab #define SWIG_FILE_WITH_INIT
import cos_doubles /* Includes the header in the wrapper code */
#include "cos_module.h"
x = np.arange(0, 2 * np.pi, 0.1) %}
y = np.empty_like(x) /* Parse the header file to generate wrappers */
%include "cos_module.h"
cos_doubles.cos_doubles_func(x, y)
pylab.plot(x, y) As you can see, not too much code is needed here. For this simple example it is enough to simply include the
pylab.show() header file in the interface file, to expose the function to Python. However, SWIG does allow for more fine grained
inclusion/exclusion of functions found in header files, check the documentation for details.
Generating the compiled wrappers is a two stage process:
1. Run the swig executable on the interface file to generate the files cos_module_wrap.c, which is the
source file for the autogenerated Python C-extension and cos_module.py, which is the autogenerated
pure python module.
2. Compile the cos_module_wrap.c into the _cos_module.so. Luckily, distutils knows how to
handle SWIG interface files, so that our setup.py is simply:
2.8.4 SWIG
SWIG, the Simplified Wrapper Interface Generator, is a software development tool that connects programs written
setup(ext_modules=[Extension("_cos_module",
in C and C++ with a variety of high-level programming languages, including Python. The important thing with
sources=["cos_module.c", "cos_module.i"])])
SWIG is, that it can autogenerate the wrapper code for you. While this is an advantage in terms of development
time, it can also be a burden. The generated file tend to be quite large and may not be too human readable and the $ cd advanced/interfacing_with_c/swig
multiple levels of indirection which are a result of the wrapping process, may be a bit tricky to understand.
$ ls
Note: The autogenerated C code uses the Python-C-Api. cos_module.c cos_module.h cos_module.i setup.py
Advantages $ python setup.py build_ext --inplace

running build_ext
Can automatically wrap entire libraries given the headers building '_cos_module' extension
swigging cos_module.i to cos_module_wrap.c
Works nicely with C++
swig -python -o cos_module_wrap.c cos_module.i
Disadvantages creating build
creating build/temp.linux-x86_64-2.7
Autogenerates enormous files gcc -pthread -fno-strict-aliasing -g -O2 -DNDEBUG -g -fwrapv -O3 -Wall -Wstrict-prototypes -fPIC -
gcc -pthread -fno-strict-aliasing -g -O2 -DNDEBUG -g -fwrapv -O3 -Wall -Wstrict-prototypes -fPIC -
Hard to debug if something goes wrong

gcc -pthread -shared build/temp.linux-x86_64-2.7/cos_module.o build/temp.linux-x86_64-2.7/cos_module_wrap.o -L/home/esc/anaconda/lib

#include <math.h> -lpython2.7 -o /home/esc/git-working/scipy-lecture-notes/advanced/interfaci
$ ls /* Compute the cosine of each element in in_array, storing the result in

build/ cos_module.c cos_module.h cos_module.i cos_module.py _cos_module.so* cos_module_wrap.c setup.py * out_array. */
void cos_doubles(double * in_array, double * out_array, int size){
We can now load and execute the cos_module as we have done in the previous examples: int i;
In [1]: import cos_module out_array[i] = cos(in_array[i]);
}
In [2]: cos_module? }
Type: module
String Form:<module 'cos_module' from 'cos_module.py'> This is wrapped as cos_doubles_func using the following SWIG interface file:
File: /home/esc/git-working/scipy-lecture-notes/advanced/interfacing_with_c/swig/cos_module.py
Docstring: <no docstring> /* Example of wrapping a C function that takes a C double array as input using
* numpy typemaps for SWIG. */
In [3]: dir(cos_module)
Out[3]: %module cos_doubles
['__builtins__', %{
'__doc__', /* the resulting C file should be built as a python extension */
'__file__', #define SWIG_FILE_WITH_INIT
'__name__', /* Includes the header in the wrapper code */
'__package__', #include "cos_doubles.h"
'_cos_module', %}
'_newclass',
'_object', /* include the numpy typemaps */
'_swig_getattr', %include "numpy.i"
'_swig_property', /* need this for correct module initialization */
'_swig_repr', %init %{
'_swig_setattr', import_array();
'_swig_setattr_nondynamic', %}
'cos_func']
/* typemaps for the two arrays, the second will be modified in-place */
In [4]: cos_module.cos_func(1.0) %apply (double* IN_ARRAY1, int DIM1) {(double * in_array, int size_in)}
Out[4]: 0.5403023058681398 %apply (double* INPLACE_ARRAY1, int DIM1) {(double * out_array, int size_out)}
In [5]: cos_module.cos_func(0.0) /* Wrapper for cos_doubles that massages the types */

Out[5]: 1.0 %inline %{
/* takes as input two numpy arrays */
In [6]: cos_module.cos_func(3.14159265359) void cos_doubles_func(double * in_array, int size_in, double * out_array, int size_out) {
Out[6]: -1.0 /* calls the original funcion, providing only the size of the first */
cos_doubles(in_array, out_array, size_in);
Again we test for robustness, and we see that we get a better error message (although, strictly speaking in Python }
there is no double type): %}
In [7]: cos_module.cos_func('foo') To use the Numpy typemaps, we need include the numpy.i file.
---------------------------------------------------------------------------
TypeError Traceback (most recent call last) Observe the call to import_array() which we encountered already in the Numpy-C-API example.
<ipython-input-7-11bee483665d> in <module>()
Since the type maps only support the signature ARRAY, SIZE we need to wrap the cos_doubles as
----> 1 cos_module.cos_func('foo')
cos_doubles_func which takes two arrays including sizes as input.
TypeError: in method 'cos_func', argument 1 of type 'double' As opposed to the simple SWIG example, we dont include the cos_doubles.h header, There is nothing
there that we wish to expose to Python since we expose the functionality through cos_doubles_func.
Numpy Support And, as before we can use distutils to wrap this:

Numpy provides support for SWIG with the numpy.i file. This interface file defines various so-called typemaps import numpy
which support conversion between Numpy arrays and C-Arrays. In the following example we will take a quick
look at how such typemaps work in practice. setup(ext_modules=[Extension("_cos_doubles",
sources=["cos_doubles.c", "cos_doubles.i"],
We have the same cos_doubles function as in the ctypes example: include_dirs=[numpy.get_include()])])
void cos_doubles(double * in_array, double * out_array, int size);
As previously, we need to use include_dirs to specify the location.

$ ls Advantages
cos_doubles.c cos_doubles.h cos_doubles.i numpy.i setup.py test_cos_doubles.py Python like language for writing C-extensions
$ python setup.py build_ext -i
running build_ext Autogenerated code
building '_cos_doubles' extension
swigging cos_doubles.i to cos_doubles_wrap.c Supports incremental optimization
swig -python -o cos_doubles_wrap.c cos_doubles.i Includes a GNU debugger extension
cos_doubles.i:24: Warning(490): Fragment 'NumPy_Backward_Compatibility' not found.
cos_doubles.i:24: Warning(490): Fragment 'NumPy_Backward_Compatibility' not found. Support for C++ (Since version 0.13)
cos_doubles.i:24: Warning(490): Fragment 'NumPy_Backward_Compatibility' not found.
Disadvantages
creating build
creating build/temp.linux-x86_64-2.7 Must be compiled
gcc -pthread -fno-strict-aliasing -g -O2 -DNDEBUG -g -fwrapv -O3 -Wall -Wstrict-prototypes -fPIC -I/home/esc/anaconda/lib/python2.7/site-packages/numpy/core/include -I/home/esc/anaconda/include/python2.7 -c c
Requires an additional library ( but only at build time, at this problem
gcc -pthread -fno-strict-aliasing -g -O2 -DNDEBUG -g -fwrapv -O3 -Wall -Wstrict-prototypes -fPIC -I/home/esc/anaconda/lib/python2.7/site-packages/numpy/core/include can be overcome by shipping the
-I/home/esc/anaconda/include/python2.7 -c c
In file included from /home/esc/anaconda/lib/python2.7/site-packages/numpy/core/include/numpy/ndarraytypes.h:1722, generated C files)
from /home/esc/anaconda/lib/python2.7/site-packages/numpy/core/include/numpy/ndarrayobject.h:17,
from /home/esc/anaconda/lib/python2.7/site-packages/numpy/core/include/numpy/arrayobject.h:15,
from cos_doubles_wrap.c:2706: Example
/home/esc/anaconda/lib/python2.7/site-packages/numpy/core/include/numpy/npy_deprecated_api.h:11:2: warning: #warning "Using deprecated NumPy API, disable it by #defining NPY_NO_DEPRECATED_API NPY_1_7_API_VERS
The main Cython code for our cos_module
gcc -pthread -shared build/temp.linux-x86_64-2.7/cos_doubles.o build/temp.linux-x86_64-2.7/cos_doubles_wrap.o -L/home/esc/anaconda/lib -lpython2.7 is contained in the file cos_module.pyx:
-o /home/esc/git-working/scipy-lecture-notes/advanced/interfa
$ ls
build/ cos_doubles.h cos_doubles.py cos_doubles_wrap.c setup.py """ Example of wrapping cos function from math.h using Cython. """
cos_doubles.c cos_doubles.i _cos_doubles.so* numpy.i test_cos_doubles.py
cdef extern from "math.h":
double cos(double arg)
And, as before, we convince ourselves that it worked:
import numpy as np def cos_func(arg):
import pylab return cos(arg)
import cos_doubles
Note the additional keywords such as cdef and extern. Also the cos_func is then pure Python.
x = np.arange(0, 2 * np.pi, 0.1)
y = np.empty_like(x) Again we can use the standard distutils module, but this time we need some additional pieces from the
Cython.Distutils:
cos_doubles.cos_doubles_func(x, y) from distutils.core import setup, Extension
pylab.plot(x, y) from Cython.Distutils import build_ext
pylab.show()
setup(
cmdclass={'build_ext': build_ext},
ext_modules=[Extension("cos_module", ["cos_module.pyx"])]
)
Compiling this:
$ cd advanced/interfacing_with_c/cython
$ ls
2.8.5 Cython cos_module.pyx setup.py
$ python setup.py build_ext --inplace
running build_ext
Cython is both a Python-like language for writing C-extensions and an advanced compiler for this language. The
cythoning cos_module.pyx to cos_module.c
Cython language is a superset of Python, which comes with additional constructs that allow you call C functions
building 'cos_module' extension
and annotate variables and class attributes with c types. In this sense one could also call it a Python with types. creating build
In addition to the basic use case of wrapping native code, Cython supports an additional use-case, namely interac- creating build/temp.linux-x86_64-2.7
tive optimization. Basically, one starts out with a pure-Python script and incrementally adds Cython types to the gcc -pthread -fno-strict-aliasing -g -O2 -DNDEBUG -g -fwrapv -O3 -Wall -Wstrict-prototypes -fPIC -
gcc -pthread -shared build/temp.linux-x86_64-2.7/cos_module.o -L/home/esc/anaconda/lib -lpython2.7
bottleneck code to optimize only those code paths that really matter.
$ ls
In this sense it is quite similar to SWIG, since the code can be autogenerated but in a sense it also quite similar to build/ cos_module.c cos_module.pyx cos_module.so* setup.py
ctypes since the wrapping code can (almost) be written in Python.
And running it:
While others solutions that autogenerate code can be quite difficult to debug (for example SWIG) Cython comes
with an extension to the GNU debugger that helps debug Python, Cython and C code. In [1]: import cos_module
Note: The autogenerated C code uses the Python-C-Api. In [2]: cos_module?

Type: module

String Form:<module 'cos_module' from 'cos_module.so'> * out_array. */

File: /home/esc/git-working/scipy-lecture-notes/advanced/interfacing_with_c/cython/cos_module.so void cos_doubles(double * in_array, double * out_array, int size){
Docstring: <no docstring> int i;
In [3]: dir(cos_module) out_array[i] = cos(in_array[i]);
Out[3]: }
['__builtins__', }
'__doc__',
'__file__', This is wrapped as cos_doubles_func using the following Cython code:
'__name__',
'__package__', """ Example of wrapping a C function that takes C double arrays as input using
'__test__', the Numpy declarations from Cython """
'cos_func']
# cimport the Cython declarations for numpy
In [4]: cos_module.cos_func(1.0) cimport numpy as np
Out[4]: 0.5403023058681398
# if you want to use the Numpy-C-API from Cython
In [5]: cos_module.cos_func(0.0) # (not strictly necessary for this example, but good practice)
Out[5]: 1.0 np.import_array()
In [6]: cos_module.cos_func(3.14159265359) # cdefine the signature of our c function

Out[6]: -1.0 cdef extern from "cos_doubles.h":
void cos_doubles (double * in_array, double * out_array, int size)
And, testing a little for robustness, we can see that we get good error messages:
# create the wrapper code, with numpy type annotations
In [7]: cos_module.cos_func('foo') def cos_doubles_func(np.ndarray[double, ndim=1, mode="c"] in_array not None,
--------------------------------------------------------------------------- np.ndarray[double, ndim=1, mode="c"] out_array not None):
TypeError Traceback (most recent call last) cos_doubles(<double*> np.PyArray_DATA(in_array),
<ipython-input-7-11bee483665d> in <module>() <double*> np.PyArray_DATA(out_array),
----> 1 cos_module.cos_func('foo') in_array.shape[0])
And (cos_module.c:506)()
/home/esc/git-working/scipy-lecture-notes/advanced/interfacing_with_c/cython/cos_module.so in cos_module.cos_func can be compiled using distutils:
TypeError: a float is required from distutils.core import setup, Extension

import numpy
Additionally, it is worth noting that Cython ships with complete declarations for the C math library, which from Cython.Distutils import build_ext
simplifies the code above to become:
setup(
""" Simpler example of wrapping cos function from math.h using Cython. """ cmdclass={'build_ext': build_ext},
ext_modules=[Extension("cos_doubles",
from libc.math cimport cos sources=["_cos_doubles.pyx", "cos_doubles.c"],
include_dirs=[numpy.get_include()])],
def cos_func(arg): )
return cos(arg)
As with the previous compiled Numpy examples, we need the include_dirs option.
In this case the cimport statement is used to import the cos function.
$ ls
cos_doubles.c cos_doubles.h _cos_doubles.pyx setup.py test_cos_doubles.py
$ python setup.py build_ext -i
Numpy Support
running build_ext
cythoning _cos_doubles.pyx to _cos_doubles.c
Cython has support for Numpy via the numpy.pyx file which allows you to add the Numpy array type to your building 'cos_doubles' extension
Cython code. I.e. like specifying that variable i is of type int, you can specify that variable a is of type creating build
numpy.ndarray with a given dtype. Also, certain optimizations such as bounds checking are supported. creating build/temp.linux-x86_64-2.7
Look at the corresponding section in the Cython documentation. In case you want to pass Numpy arrays as C gcc -pthread -fno-strict-aliasing -g -O2 -DNDEBUG -g -fwrapv -O3 -Wall -Wstrict-prototypes -fPIC -
arrays to your Cython wrapped C functions, there is a section about this in the Cython wiki. In file included from /home/esc/anaconda/lib/python2.7/site-packages/numpy/core/include/numpy/ndar
from /home/esc/anaconda/lib/python2.7/site-packages/numpy/core/include/numpy/ndar
In the following example, we will show how to wrap the familiar cos_doubles function using Cython. from /home/esc/anaconda/lib/python2.7/site-packages/numpy/core/include/numpy/arra
void cos_doubles(double * in_array, double * out_array, int size); from _cos_doubles.c:253:
/home/esc/anaconda/lib/python2.7/site-packages/numpy/core/include/numpy/npy_deprecated_api.h:11:2:
#include <math.h> /home/esc/anaconda/lib/python2.7/site-packages/numpy/core/include/numpy/__ufunc_api.h:236: warning
gcc -pthread -fno-strict-aliasing -g -O2 -DNDEBUG -g -fwrapv -O3 -Wall -Wstrict-prototypes -fPIC -
/* Compute the cosine of each element in in_array, storing the result in gcc -pthread -shared build/temp.linux-x86_64-2.7/_cos_doubles.o build/temp.linux-x86_64-2.7/cos_do
$ ls

build/ _cos_doubles.c cos_doubles.c cos_doubles.h _cos_doubles.pyx cos_doubles.so* setup.py (d) Empty array
test_cos_doubles.py
(e) Arrays with non-double types
And, as before, we convince ourselves that it worked:
4. Use the %timeit IPython magic to measure the execution time of the various solutions
import numpy as np
import pylab
import cos_doubles Python-C-API
x = np.arange(0, 2 * np.pi, 0.1) 1. Modify the Numpy example such that the function takes two input arguments, where the second is the
y = np.empty_like(x)
preallocated output array, making it similar to the other Numpy examples.
cos_doubles.cos_doubles_func(x, y) 2. Modify the example such that the function only takes a single input array and modifies this in place.
pylab.plot(x, y)
pylab.show() 3. Try to fix the example to use the new Numpy iterator protocol. If you manage to obtain a working solution,
please submit a pull-request on github.
4. You may have noticed, that the Numpy-C-API example is the only Numpy example that does not wrap
cos_doubles but instead applies the cos function directly to the elements of the Numpy array. Does
this have any advantages over the other techniques.
5. Can you wrap cos_doubles using only the Numpy-C-API. You may need to ensure that the arrays have
the correct type, are one dimensional and contiguous in memory.
2.8.6 Summary Ctypes
In this section four different techniques for interfacing with native code have been presented. The table below 1. Modify the Numpy example such that cos_doubles_func handles the preallocation for you, thus mak-
roughly summarizes some of the aspects of the techniques. ing it more like the Numpy-C-API example.
x Part of CPython Compiled Autogenerated Numpy Support
Python-C-API True True False True SWIG
Ctypes True False False True
Swig False True True True 1. Look at the code that SWIG autogenerates, how much of it do you understand?
Cython False True True True
2. Modify the Numpy example such that cos_doubles_func handles the preallocation for you, thus mak-
Of all three presented techniques, Cython is the most modern and advanced. In particular, the ability to optimize ing it more like the Numpy-C-API example.
code incrementally by adding types to your Python code is unique.
3. Modify the cos_doubles C function so that it returns an allocated array. Can you wrap this using SWIG
typemaps? If not, why not? Is there a workaround for this specific situation? (Hint: you know the size of
2.8.7 Further Reading and References the output array, so it may be possible to construct a Numpy array from the returned double *.)
Gal Varoquauxs blog post about avoiding data copies provides some insight on how to handle memory Cython
management cleverly. If you ever run into issues with large datasets, this is a reference to come back to for
some inspiration. 1. Look at the code that Cython autogenerates. Take a closer look at some of the comments that Cython inserts.
What do you see?
2.8.8 Exercises 2. Look at the section Working with Numpy from the Cython documentation to learn how to incrementally
optimize a pure python script that uses Numpy.
Since this is a brand new section, the exercises are considered more as pointers as to what to look at next, so pick
3. Modify the Numpy example such that cos_doubles_func handles the preallocation for you, thus mak-
the ones that you find more interesting. If you have good ideas for exercises, please let us know!
ing it more like the Numpy-C-API example.
1. Download the source code for each example and compile and run them on your machine.
2. Make trivial changes to each example and convince yourself that this works. ( E.g. change cos for sin.)
3. Most of the examples, especially the ones involving Numpy may still be fragile and respond badly to input
errors. Look for ways to crash the examples, figure what the problem is and devise a potential solution.
Here are some ideas:
(a) Numerical overflow.
(b) Input and output arrays that have different lengths.
(c) Multidimensional array.

CHAPTER 3 Contents
Data representation and interaction (page 272)
Data as a table (page 272)
The panda data-frame (page 272)
Hypothesis testing: comparing two groups (page 276)
Students t-test: the simplest statistical test (page 277)
Paired tests: repeated measurements on the same indivuals (page 278)
Packages and applications Linear models, multiple factors, and analysis of variance (page 279)
formulas to specify statistical models in Python (page 279)
Multiple Regression: including multiple factors (page 282)
Post-hoc hypothesis testing: analysis of variance (ANOVA) (page 284)
More visualization: seaborn for statistical exploration (page 284)
Pairplot: scatter matrices (page 284)
lmplot: plotting a univariate regression (page 286)
This part of the Scipy lecture notes is dedicated to various scientific packages useful for extended needs. Testing for interactions (page 287)
Full examples (page 288)
Tip: In this document, the Python inputs are represented with the sign >>>.
3.1 Statistics in Python

Disclaimer: Gender questions
Requirements Some of the examples of this tutorial are chosen around gender questions. The reason is that on such questions
controlling the truth of a claim actually matters to many people.
Standard scientific Python environment (numpy, scipy, matplotlib)
Pandas
Statsmodels
Seaborn 3.1.1 Data representation and interaction
To install Python and these dependencies, we recommend that you download Anaconda Python or Enthought
Canopy, or preferably use the package manager if you are under Ubuntu or other linux. Data as a table
See also: The setting that we consider for statistical analysis is that of multiple observations or samples described by
a set of different attributes or features. The data can than be seen as a 2D table, or matrix, with columns
Bayesian statistics in Python giving the different attributes of the data, and rows the observations. For instance, the data contained in
This chapter does not cover tools for Bayesian statistics. Of particular interest for Bayesian modelling is PyMC, examples/brain_size.csv:
which implements a probabilistic programming language in Python. "";"Gender";"FSIQ";"VIQ";"PIQ";"Weight";"Height";"MRI_Count"
"1";"Female";133;132;124;"118";"64.5";816932
"2";"Male";140;150;124;".";"72.5";1001121
"3";"Male";139;123;150;"143";"73.3";1038437
"4";"Male";133;129;128;"172";"68.8";965353
"5";"Female";137;132;134;"147";"65.0";951545
Tip: Why Python for statistics?

R is a language dedicated to statistics. Python is a general-purpose language with statistics modules. R has more The panda data-frame
statistical analysis features than Python, and specialized syntaxes. However, when it comes to building complex
analysis pipelines that mix statistics with e.g. image analysis, text mining, or control of a physical experiment, the
Tip: We will store and manipulate this data in a pandas.DataFrame, from the pandas module. It is the
richness of Python is an invaluable asset. Python equivalent of the spreadsheet table. It is different from a 2D numpy array as it has named columns, can
contain a mixture of different data types by column, and has elaborate selection and pivotal mechanisms.
Creating dataframes: reading data files or converting arrays
271 3.1. Statistics in Python 272

Manipulating data
Separator
It is a CSV file, but the separator is ; data is a pandas.DataFrame, that resembles Rs dataframe:
>>> data.shape # 40 rows and 8 columns
Reading from a CSV file: Using the above CSV file that gives observations of brain size and weight and IQ (40, 8)
(Willerman et al. 1991), the data are a mixture of numerical and categorical values:
>>> data.columns # It has columns
>>> import pandas Index([u'Unnamed: 0', u'Gender', u'FSIQ', u'VIQ', u'PIQ', u'Weight', u'Height', u'MRI_Count'], dty
>>> data = pandas.read_csv('examples/brain_size.csv', sep=';', na_values=".")
>>> data >>> print(data['Gender']) # Columns can be addressed by name
Unnamed: 0 Gender FSIQ VIQ PIQ Weight Height MRI_Count 0 Female
0 1 Female 133 132 124 118 64.5 816932 1 Male
1 2 Male 140 150 124 NaN 72.5 1001121 2 Male
2 3 Male 139 123 150 143 73.3 1038437 3 Male
3 4 Male 133 129 128 172 68.8 965353 4 Female
4 5 Female 137 132 134 147 65.0 951545 ...
...
>>> # Simpler selector
Warning: Missing values >>> data[data['Gender'] == 'Female']['VIQ'].mean()
109.45
The weight of the second individual is missing in the CSV file. If we dont specify the missing value (NA =
not available) marker, we will not be able to do statistical analysis.
Note: For a quick view on a large dataframe, use its describe method:
pandas.DataFrame.describe().
Creating from arrays: A pandas.DataFrame can also be seen as a dictionary of 1D series, eg arrays or
lists. If we have 3 numpy arrays:
groupby: splitting a dataframe on values of categorical variables:
>>> t = np.linspace(-6, 6, 20) >>> groupby_gender = data.groupby('Gender')
>>> sin_t = np.sin(t) >>> for gender, value in groupby_gender['VIQ']:
>>> cos_t = np.cos(t) ... print((gender, value.mean()))
('Female', 109.45)
We can expose them as a pandas.DataFrame: ('Male', 115.25)
>>> pandas.DataFrame({'t': t, 'sin': sin_t, 'cos': cos_t}) groupby_gender is a powerful object that exposes many operations on the resulting group of dataframes:
cos sin t
0 0.960170 0.279415 -6.000000 >>> groupby_gender.mean()
1 0.609977 0.792419 -5.368421 Unnamed: 0 FSIQ VIQ PIQ Weight Height MRI_Count
2 0.024451 0.999701 -4.736842 Gender
3 -0.570509 0.821291 -4.105263 Female 19.65 111.9 109.45 110.45 137.200000 65.765000 862654.6
4 -0.945363 0.326021 -3.473684 Male 21.35 115.0 115.25 111.60 166.444444 71.431579 954855.4
5 -0.955488 -0.295030 -2.842105
6 -0.596979 -0.802257 -2.210526 Tip: Use tab-completion on groupby_gender to find more. Other common grouping functions are median,
7 -0.008151 -0.999967 -1.578947 count (useful for checking to see the amount of missing values in different subsets) or sum. Groupby evaluation
8 0.583822 -0.811882 -0.947368 is lazy, no work is done until an aggregation function is applied.
...
Other inputs: pandas can input data from SQL, excel files, or other formats. See the pandas documentation.
3.1. Statistics in Python 273 3.1. Statistics in Python 274

>>> plotting.scatter_matrix(data[['PIQ', 'VIQ', 'FSIQ']])
Two populations
The IQ metrics are bimodal, as if there are 2 sub-populations.
Exercise
What is the mean value for VIQ for the full population?
How many males/females were included in this study?
Hint use tab completion to find out the methods that can be called, instead of mean in the above
example.
What is the average value of MRI counts expressed in log units, for males and females?
Note: groupby_gender.boxplot is used for the plots above (see this example (page 289)).
Plotting data Exercise

Plot the scatter matrix for males only, and for females only. Do you think that the 2 sub-populations corre-
Pandas comes with some plotting tools (pandas.tools.plotting, using matplotlib behind the scene) to spond to gender?
display statistics of the data in dataframes:
Scatter matrices:
>>> from pandas.tools import plotting 3.1.2 Hypothesis testing: comparing two groups
>>> plotting.scatter_matrix(data[['Weight', 'Height', 'MRI_Count']])
For simple statistical tests, we will use the scipy.stats sub-module of scipy:
>>> from scipy import stats
See also:
Scipy is a vast library. For a quick summary to the whole library, see the scipy (page 103) chapter.

Students t-test: the simplest statistical test Paired tests: repeated measurements on the same indivuals
1-sample t-test: testing the value of a population mean
PIQ, VIQ, and FSIQ give 3 measures of IQ. Let us test if FISQ
and PIQ are significantly different. We can use a 2 sample test:
>>> stats.ttest_ind(data['FSIQ'], data['PIQ'])
(...0.46563759638..., 0.64277250...)
The problem with this approach is that it forgets that there are links between observations: FSIQ and PIQ are
measured on the same individuals. Thus the variance due to inter-subject variability is confounding, and can be
removed, using a paired test, or repeated measures test:
>>> stats.ttest_rel(data['FSIQ'], data['PIQ'])
scipy.stats.ttest_1samp() tests if the population mean of data is likely to be equal to a given value (...1.784201940..., 0.082172638183...)
(technically if observations are drawn from a Gaussian distributions of given population mean). It returns the T
statistic, and the p-value (see the functions help):
>>> stats.ttest_1samp(data['VIQ'], 0)
(...30.088099970..., 1.32891964...e-28)
Tip: With a p-value of 10^-28 we can claim that the population mean for the IQ (VIQ measure) is not 0.
2-sample t-test: testing for difference across populations
This is equivalent to a 1-sample test on the difference:

We have seen above that the mean VIQ in the male and female populations were different. To test if this is
significant, we do a 2-sample t-test with scipy.stats.ttest_ind(): >>> stats.ttest_1samp(data['FSIQ'] - data['PIQ'], 0)
(...1.784201940..., 0.082172638...)
>>> female_viq = data[data['Gender'] == 'Female']['VIQ']
>>> male_viq = data[data['Gender'] == 'Male']['VIQ']
>>> stats.ttest_ind(female_viq, male_viq)
(...-0.77261617232..., 0.4445287677858...)
T-tests assume Gaussian errors. We can use a Wilcoxon signed-rank test, that relaxes this assumption:
>>> stats.wilcoxon(data['FSIQ'], data['PIQ'])
(274.5, 0.106594927...)
Note: The corresponding test in the non paired case is the MannWhitney U test,
scipy.stats.mannwhitneyu().

Exercice
Test the difference between weights in males and females.
Use non parametric statistics to test the difference between VIQ in males and females. Then we specify an OLS model and fit it:
Conclusion: we find that the data does not support the hypothesis that males and females have different
VIQ. >>> from statsmodels.formula.api import ols
>>> model = ols("y ~ x", data).fit()
We can inspect the various statistics derived from the fit:

>>> print(model.summary())
OLS Regression Results
==============================================================================
3.1.3 Linear models, multiple factors, and analysis of variance Dep. Variable: y R-squared: 0.804
Model: OLS Adj. R-squared: 0.794
Method: Least Squares F-statistic: 74.03
formulas to specify statistical models in Python
Date: ... Prob (F-statistic): 8.56e-08
Time: ... Log-Likelihood: -57.988
A simple linear regression No. Observations: 20 AIC: 120.0
Df Residuals: 18 BIC: 122.0
Df Model: 1
==============================================================================
coef std err t P>|t| [95.0% Conf. Int.]
------------------------------------------------------------------------------
Intercept -5.5335 1.036 -5.342 0.000 -7.710 -3.357
x 2.9369 0.341 8.604 0.000 2.220 3.654
==============================================================================
Omnibus: 0.100 Durbin-Watson: 2.956
Prob(Omnibus): 0.951 Jarque-Bera (JB): 0.322
Skew: -0.058 Prob(JB): 0.851
Kurtosis: 2.390 Cond. No. 3.03
=========================================================================...
Terminology:
Given two set of observations, x and y, we want to test the Statsmodel uses a statistical terminology: the y variable in statsmodel is called endogenous while the x
hypothesis that y is a linear function of x. In other terms: variable is called exogenous. This is discussed in more detail here.
= * + + To simplify, y (endogenous) is the value you are trying to predict, while x (exogenous) represents the features
you are using to make the prediction.
where e is observation noise. We will use the statmodels module to:
1. Fit a linear model. We will use the simplest strategy, ordinary least squares (OLS).
2. Test that coef is non zero. Exercise
Retrieve the estimated parameters from the model above. Hint: use tab-completion to find the relevent
attribute.
First, we generate simulated data according to the model:

>>> x = np.linspace(-5, 5, 20)
>>> np.random.seed(1) Categorical variables: comparing groups or multiple categories
>>> # normal distributed noise
>>> y = -5 + 3*x + 4 * np.random.normal(size=x.shape) Let us go back the data on brain size:
>>> # Create a data frame containing all the relevant variables
>>> data = pandas.DataFrame({'x': x, 'y': y}) >>> data = pandas.read_csv('examples/brain_size.csv', sep=';', na_values=".")
We can write a comparison between IQ of male and female using a linear model:
formulas for statistics in Python >>> model = ols("VIQ ~ Gender + 1", data).fit()
See the statsmodels documentation >>> print(model.summary())

OLS Regression Results Link to t-tests between different FSIQ and PIQ
==============================================================================
Dep. Variable: VIQ R-squared: 0.015
To compare different type of IQ, we need to create a long-form table, listing IQs, where the type of IQ is
Model: OLS Adj. R-squared: -0.010 indicated by a categorical variable:
Method: Least Squares F-statistic: 0.5969 >>> data_fisq = pandas.DataFrame({'iq': data['FSIQ'], 'type': 'fsiq'})
Date: ... Prob (F-statistic): 0.445 >>> data_piq = pandas.DataFrame({'iq': data['PIQ'], 'type': 'piq'})
Time: ... Log-Likelihood: -182.42 >>> data_long = pandas.concat((data_fisq, data_piq))
No. Observations: 40 AIC: 368.8 >>> print(data_long)
Df Residuals: 38 BIC: 372.2 iq type
Df Model: 1 0 133 fsiq
=======================================================================... 1 140 fsiq
coef std err t P>|t| [95.0% Conf. Int.] 2 139 fsiq
-----------------------------------------------------------------------... ...
Intercept 109.4500 5.308 20.619 0.000 98.704 120.196 31 137 piq
Gender[T.Male] 5.8000 7.507 0.773 0.445 -9.397 20.997 32 110 piq
=======================================================================... 33 86 piq
Omnibus: 26.188 Durbin-Watson: 1.709 ...
Skew: 0.010 Prob(JB): 0.157 >>> model = ols("iq ~ type", data_long).fit()
Kurtosis: 1.510 Cond. No. 2.62 >>> print(model.summary())
=======================================================================... OLS Regression Results
...
=======================================================================...
Tips on specifying model coef std err t P>|t| [95.0% Conf. Int.]
-----------------------------------------------------------------------...
Forcing categorical: the Gender is automatical detected as a categorical variable, and thus each of its
Intercept 113.4500 3.683 30.807 0.000 106.119 120.781
different values are treated as different entities. type[T.piq] -2.4250 5.208 -0.466 0.643 -12.793 7.943
An integer column can be forced to be treated as categorical using: ...
>>> model = ols('VIQ ~ C(Gender)', data).fit() We can see that we retrieve the same values for t-test and corresponding p-values for the effect of the type
Intercept: We can remove the intercept using - 1 in the formula, or force the use of an intercept using + of iq than the previous t-test:
1. >>> stats.ttest_ind(data['FSIQ'], data['PIQ'])
(...0.46563759638..., 0.64277250...)
Tip: By default, statsmodel treats a categorical variable with K possible values as K-1 dummy
boolean variables (the last level being absorbed into the intercept term). This is almost always a
good default choice - however, it is possible to specify different encodings for categorical variables
(http://statsmodels.sourceforge.net/devel/contrasts.html). Multiple Regression: including multiple factors
Consider a linear model explaining a variable z (the dependent variable) with 2 variables x and y:
= 1 + 2 + +
Such a model can be seen in 3D as fitting a plane to a cloud of (x, y, z) points.

Post-hoc hypothesis testing: analysis of variance (ANOVA)
In the above iris example, we wish to test if the petal length is different between versicolor and virginica, after
Example: the iris data (examples/iris.csv) removing the effect of sepal width. This can be formulated as testing the difference between the coefficient
associated to versicolor and virginica in the linear model estimated above (it is an Analysis of Variance, ANOVA).
Tip: Sepal and petal size tend to be related: bigger flowers are bigger! But is there in addition a systematic effect For this, we write a vector of contrast on the parameters estimated: we want to test "name[T.versicolor]
of species? - name[T.virginica]", with an F-test:
>>> print(model.f_test([0, 1, -1, 0]))
<F test: F=array([[ 3.24533535]]), p=[[ 0.07369059]], df_denom=146, df_num=1>
Is this difference significant?
Exercice
Going back to the brain size + IQ data, test if the VIQ of male and female are different after removing the
effect of brain size, height and weight.
3.1.4 More visualization: seaborn for statistical exploration
Seaborn combines simple statistical fits with plotting on pandas dataframes.

Let us consider a data giving wages and many other personal information on 500 individuals (Berndt, ER. The
>>> data = pandas.read_csv('examples/iris.csv') Practice of Econometrics. 1991. NY: Addison-Wesley).
>>> model = ols('sepal_width ~ name + petal_length', data).fit()
>>> print(model.summary()) Tip: The full code loading and plotting of the wages data is found in corresponding example (page 303).
============================================================================== >>> print data
Dep. Variable: sepal_width R-squared: 0.478 EDUCATION SOUTH SEX EXPERIENCE UNION WAGE AGE RACE \
Model: OLS Adj. R-squared: 0.468 0 8 0 1 21 0 0.707570 35 2
Method: Least Squares F-statistic: 44.63 1 9 0 1 42 0 0.694605 57 3
Date: ... Prob (F-statistic): 1.58e-20 2 12 0 0 1 0 0.824126 19 3
Time: ... Log-Likelihood: -38.185 3 12 0 0 4 0 0.602060 22 3
No. Observations: 150 AIC: 84.37 ...
Df Model: 3
===========================================================================...
Pairplot: scatter matrices
---------------------------------------------------------------------------...
Intercept 2.9813 0.099 29.989 0.000 2.785 3.178 We can easily have an intuition on the interactions between continuous variables using seaborn.pairplot()
name[T.versicolor] -1.4821 0.181 -8.190 0.000 -1.840 -1.124 to display a scatter matrix:
name[T.virginica] -1.6635 0.256 -6.502 0.000 -2.169 -1.158
>>> import seaborn
petal_length 0.2983 0.061 4.920 0.000 0.178 0.418
>>> seaborn.pairplot(data, vars=['WAGE', 'AGE', 'EDUCATION'],
==============================================================================
... kind='reg')
Skew: -0.082 Prob(JB): 0.236
==============================================================================

Look and feel and matplotlib settings

Seaborn changes the default of matplotlib figures to achieve a more modern, excel-like look. It does that
upon import. You can reset the default using:
>>> from matplotlib import pyplot as plt
>>> plt.rcdefaults()
Tip: To switch back to seaborn settings, or understand better styling in seaborn, see the relevent section of
the seaborn documentation.
lmplot: plotting a univariate regression
A regression capturing the relation between one variable and another, eg wage and eduction, can be plotted using
seaborn.lmplot():
>>> seaborn.lmplot(y='WAGE', x='EDUCATION', data=data)
Categorical variables can be plotted as the

hue:
>>> seaborn.pairplot(data, vars=['WAGE', 'AGE', 'EDUCATION'],
... kind='reg', hue='SEX')
Robust regression
Tip: Given that, in the above plot, there seems to be a couple of data points that are outside of the main
cloud to the right, they might be outliers, not representative of the population, but driving the regression.
To compute a regression that is less sentive to outliers, one must use a robust model. This is done in seaborn
using robust=True in the plotting functions, or in statsmodels by replacing the use of the OLS by a
Robust Linear Model, statsmodels.formula.api.rlm().

3.1.5 Testing for interactions 3.1.6 Full examples
Examples
Code examples
Code examples for the statistics in Python tutorial.
Do wages increase more with education for

males than females?
Tip: The plot above is made of two different fits. We need to formulate a single model that tests for a variance of
slope across the to population. This is done via an interaction.
>>> result = sm.ols(formula='wage ~ education + gender + education * gender',

... data=data).fit()
>>> print(result.summary())
...
------------------------------------------------------------------------------
Intercept 0.2998 0.072 4.173 0.000 0.159 0.441
gender[T.male] 0.2750 0.093 2.972 0.003 0.093 0.457
education 0.0415 0.005 7.647 0.000 0.031 0.052 Figure 3.1: Boxplots and paired differences (page 288)
education:gender[T.male] -0.0134 0.007 -1.919 0.056 -0.027 0.000
==============================================================================
...
Boxplots and paired differences Plot boxplots for FSIQ, PIQ, and the paired difference between the two: while
Can we conclude that education benefits males more than females? the spread (error bars) for FSIQ and PIQ are very large, there is a systematic (common) effect due to the subjects.
This effect is cancelled out in the difference and the spread of the difference (paired by subject) is much smaller
than the spread of the individual measures.
Take home messages

Hypothesis testing and p-value give you the significance of an effect / difference
Formulas (with categorical variables) enable you to express rich links in your data
Visualizing your data and simple model fits matters!
Conditionning (adding factors that can explain all or part of the variation) is important modeling
aspect that changes the interpretation.

Figure 3.2: Plotting simple quantities of a pandas dataframe (page 289)
Plotting simple quantities of a pandas dataframe This example loads from a CSV file data with mixed nu-
merical and categorical entries, and plots a few quantities, separately for females and males, thanks to the pandas
integrated plotting tool (that uses matplotlib behind the scene).
See http://pandas.pydata.org/pandas-docs/stable/visualization.html

Python source code: plot_paired_boxplots.py
import pandas
data = pandas.read_csv('brain_size.csv', sep=';', na_values='.')
# Box plot of FSIQ and PIQ (different measures od IQ)

plt.figure(figsize=(4, 3))
data.boxplot(column=['FSIQ', 'PIQ'])
# Boxplot of the difference

plt.boxplot(data['FSIQ'] - data['PIQ'])
plt.xticks((1, ), ('FSIQ - PIQ', ))
plt.show()
Total running time of the example: 0.29 seconds ( 0 minutes 0.29 seconds)


Python source code: plot_pandas.py
import pandas
data = pandas.read_csv('brain_size.csv', sep=';', na_values='.')
# Box plots of different columns for each gender

groupby_gender = data.groupby('Gender')
groupby_gender.boxplot(column=['FSIQ', 'VIQ', 'PIQ'])
from pandas.tools import plotting
# Scatter matrices for different columns

plotting.scatter_matrix(data[['Weight', 'Height', 'MRI_Count']])
plotting.scatter_matrix(data[['PIQ', 'VIQ', 'FSIQ']])

plt.show()
Analysis of Iris petal and sepal sizes Ilustrate an analysis on a real dataset:
Visualizing the data to formulate intuitions
Fitting of a linear model
Hypothesis test of the effect of a categorical variable in the presence of a continuous confound


==============================================================================
Dep. Variable: sepal_width R-squared: 0.478
Date: Mon, 21 Sep 2015 Prob (F-statistic): 1.58e-20
Time: 19:08:35 Log-Likelihood: -38.185
No. Observations: 150 AIC: 84.37
Df Model: 3
======================================================================================
--------------------------------------------------------------------------------------
Intercept 2.9813 0.099 29.989 0.000 2.785 3.178
name[T.versicolor] -1.4821 0.181 -8.190 0.000 -1.840 -1.124
name[T.virginica] -1.6635 0.256 -6.502 0.000 -2.169 -1.158
petal_length 0.2983 0.061 4.920 0.000 0.178 0.418
==============================================================================
Skew: -0.082 Prob(JB): 0.236
==============================================================================
Testing the difference between effect of versicolor and virginica
<F test: F=array([[ 3.24533535]]), p=[[ 0.07369059]], df_denom=146, df_num=1>
Figure 3.3: Analysis of Iris petal and sepal sizes (page 292) Python source code: plot_iris_analysis.py
import pandas
from pandas.tools import plotting
from statsmodels.formula.api import ols
# Load the data

data = pandas.read_csv('iris.csv')
##############################################################################
# Plot a scatter matrix
# Express the names as categories

categories = pandas.Categorical(data['name'])
# The parameter 'c' is passed to plt.scatter and will control the color
plotting.scatter_matrix(data, c=categories.labels, marker='o')
fig = plt.gcf()
fig.suptitle("blue: setosa, green: versicolor, red: virginica", size=13)
##############################################################################
# Statistical analysis
# Let us try to explain the sepal length as a function of the petal

# width and the category of iris
model = ols('sepal_width ~ name + petal_length', data).fit()

print(model.summary())
# Now formulate a "contrast", to test if the offset for versicolor and

# virginica are identical
Script output:
print('Testing the difference between effect of versicolor and virginica')

print(model.f_test([0, 1, -1, 0])) Script output:

plt.show() OLS Regression Results
==============================================================================
Total running time of the example: 0.83 seconds ( 0 minutes 0.83 seconds) Dep. Variable: y R-squared: 0.804
Df Model: 1
==============================================================================
------------------------------------------------------------------------------
Intercept -5.5335 1.036 -5.342 0.000 -7.710 -3.357
x 2.9369 0.341 8.604 0.000 2.220 3.654
==============================================================================
Skew: -0.058 Prob(JB): 0.851
==============================================================================
ANOVA results
df sum_sq mean_sq F PR(>F)
x 1 1588.873443 1588.873443 74.029383 8.560649e-08
Residual 18 386.329330 21.462741 NaN NaN
Python source code: plot_regression.py

# Original author: Thomas Haslwanter
Figure 3.4: Simple Regression (page 295) import numpy as np

import pandas
Simple Regression Fit a simple linear regression using statsmodels, compute corresponding p-values.
# For statistics. Requires statsmodels 5.0 or more
# Analysis of Variance (ANOVA) on linear models
from statsmodels.stats.anova import anova_lm
##############################################################################
# Generate and show the data
x = np.linspace(-5, 5, 20)
# To get reproducable values, provide a seed value

np.random.seed(1)
y = -5 + 3*x + 4 * np.random.normal(size=x.shape)
# Plot the data

plt.plot(x, y, 'o')
##############################################################################
# Multilinear regression model, calculating fit, P-values, confidence
# intervals etc.
# Convert the data into a Pandas DataFrame to use the formulas framework
# in statsmodels
data = pandas.DataFrame({'x': x, 'y': y})
# Fit the model

model = ols("y ~ x", data).fit()
# Print the summary

# Peform analysis of variance on fitted linear model

anova_results = anova_lm(model)
print('\nANOVA results')
print(anova_results)
##############################################################################
# Plot the fitted model
# Retrieve the parameter estimates

offset, coef = model._results.params
plt.plot(x, x*coef + offset)
plt.xlabel('x')
plt.ylabel('y')
plt.show()
Script output:
==============================================================================
Dep. Variable: wage R-squared: 0.193
Time: 19:08:36 Log-Likelihood: 86.654
No. Observations: 534 AIC: -167.3
Df Residuals: 531 BIC: -154.5
Df Model: 2
==================================================================================
----------------------------------------------------------------------------------
Intercept 0.4053 0.046 8.732 0.000 0.314 0.496
gender[T.male] 0.1008 0.018 5.625 0.000 0.066 0.136
education 0.0334 0.003 9.768 0.000 0.027 0.040
Figure 3.5: Test for an education/gender interaction in wages (page 297) ==============================================================================
Skew: -0.147 Prob(JB): 0.0873
Test for an education/gender interaction in wages Wages depend mostly on education. Here we investigate
how this dependence is related to gender: not only does gender create an offset in wages, it also seems that wages
==============================================================================
increase more with education for males than females. OLS Regression Results
Does our data support this last hypothesis? We will test this using statsmodels formulas ==============================================================================
(http://statsmodels.sourceforge.net/stable/example_formulas.html). Dep. Variable: wage R-squared: 0.198

Time: 19:08:36 Log-Likelihood: 88.503 # joined model for male and female, not separate models for male and
No. Observations: 534 AIC: -169.0 # female. The reason is that a single model enables statistical testing
Df Residuals: 530 BIC: -151.9 result = sm.ols(formula='wage ~ education + gender', data=data).fit()
Df Model: 3 print(result.summary())
============================================================================================
coef std err t P>|t| [95.0% Conf. Int.] # The plots above highlight that there is not only a different offset in
-------------------------------------------------------------------------------------------- # wage but also a different slope
Intercept 0.2998 0.072 4.173 0.000 0.159 0.441
gender[T.male] 0.2750 0.093 2.972 0.003 0.093 0.457 # We need to model this using an interaction
education 0.0415 0.005 7.647 0.000 0.031 0.052 result = sm.ols(formula='wage ~ education + gender + education * gender',
education:gender[T.male] -0.0134 0.007 -1.919 0.056 -0.027 0.000 data=data).fit()
============================================================================== print(result.summary())
Skew: -0.156 Prob(JB): 0.0821 # Looking at the p-value of the interaction of gender and education, the
Kurtosis: 3.356 Cond. No. 194. # data does not support the hypothesis that education benefits males
============================================================================== # more than female (p-value > 0.05).
Python source code: plot_wage_education_gender.py

############################################################################## plt.show()
# Load and massage the data
import pandas
import urllib
import os
if not os.path.exists('wages.txt'):
# Download the file if it is not present
urllib.urlretrieve('http://lib.stat.cmu.edu/datasets/CPS_85_Wages',
'wages.txt')
# EDUCATION: Number of years of education

# SEX: 1=Female, 0=Male
# WAGE: Wage (dollars per hour)
data = pandas.read_csv('wages.txt', skiprows=27, skipfooter=6, sep=None,
header=None, names=['education', 'gender', 'wage'],
usecols=[0, 2, 5],
)
# Convert genders to strings (this is particulary useful so that the

# statsmodels formulas detects that gender is a categorical variable)
import numpy as np
data['gender'] = np.choose(data.gender, ['male', 'female'])
# Log-transform the wages, because they typically are increased with

# multiplicative factors
data['wage'] = np.log10(data['wage'])
##############################################################################
# simple plotting Figure 3.6: Multiple Regression (page 300)
import seaborn
# Plot 2 linear fits for male and female.

seaborn.lmplot(y='wage', x='education', hue='gender', data=data) Multiple Regression Calculate using statsmodels just the best fit, or all the corresponding statistical parame-
ters.
Also shows how to make 3d plots.
##############################################################################
# statistical analysis
import statsmodels.formula.api as sm
# Note that this model is not the plot displayed above: it is one

Residual 438 27576.201607 62.959364 NaN NaN
Python source code: plot_regression_3d.py

# Original author: Thomas Haslwanter
import numpy as np
import pandas
# For 3d plots. This import is necessary to have 3D plotting below

from mpl_toolkits.mplot3d import Axes3D
# For statistics. Requires statsmodels 5.0 or more

# Analysis of Variance (ANOVA) on linear models
from statsmodels.stats.anova import anova_lm
##############################################################################
# Generate and show the data
x = np.linspace(-5, 5, 21)
# We generate a 2D grid
X, Y = np.meshgrid(x, x)
# To get reproducable values, provide a seed value

np.random.seed(1)
# Z is the elevation of this 2D grid

Z = -5 + 3*X - 0.5*Y + 8 * np.random.normal(size=X.shape)
# Plot the data

fig = plt.figure()
Script output: ax = fig.gca(projection='3d')
OLS Regression Results surf = ax.plot_surface(X, Y, Z, cmap=plt.cm.coolwarm,
============================================================================== rstride=1, cstride=1)
Dep. Variable: z R-squared: 0.594 ax.view_init(20, -120)
Model: OLS Adj. R-squared: 0.592 ax.set_xlabel('X')
Method: Least Squares F-statistic: 320.4 ax.set_ylabel('Y')
Date: Mon, 21 Sep 2015 Prob (F-statistic): 1.89e-86 ax.set_zlabel('Z')
No. Observations: 441 AIC: 3081. ##############################################################################
Df Residuals: 438 BIC: 3094. # Multilinear regression model, calculating fit, P-values, confidence
Df Model: 2 # intervals etc.
==============================================================================
coef std err t P>|t| [95.0% Conf. Int.] # Convert the data into a Pandas DataFrame to use the formulas framework
------------------------------------------------------------------------------ # in statsmodels
Intercept -4.5052 0.378 -11.924 0.000 -5.248 -3.763
x 3.1173 0.125 24.979 0.000 2.872 3.363 # First we need to flatten the data: it's 2D layout is not relevent.
y -0.5109 0.125 -4.094 0.000 -0.756 -0.266 X = X.flatten()
============================================================================== Y = Y.flatten()
Omnibus: 0.260 Durbin-Watson: 2.057 Z = Z.flatten()
Skew: -0.052 Prob(JB): 0.903 data = pandas.DataFrame({'x': X, 'y': Y, 'z': Z})
============================================================================== # Fit the model
model = ols("z ~ x + y", data).fit()
Retrieving manually the parameter estimates:
[-4.50523303 3.11734237 -0.51091248] # Print the summary
ANOVA results
df sum_sq mean_sq F PR(>F) print("\nRetrieving manually the parameter estimates:")
x 1 39284.301219 39284.301219 623.962799 2.888238e-86 print(model._results.params)
y 1 1055.220089 1055.220089 16.760336 5.050899e-05 # should be array([-4.99754526, 3.00250049, -0.50514907])

# Peform analysis of variance on fitted linear model

anova_results = anova_lm(model)
print('\nANOVA results')
print(anova_results)
plt.show()
Figure 3.7: Visualizing factors influencing wages (page 303)

Visualizing factors influencing wages This example uses seaborn to quickly plot various factors relating wages,
experience and eduction.
Seaborn (http://stanford.edu/~mwaskom/software/seaborn/) is a library that combines visualization and statistical
fits to show trends in data.
Note that importing seaborn changes the matplotlib style to have an excel-like feeling. This changes affect other
matplotlib figures. To restore defaults once this example is run, we would need to call plt.rcdefaults().



Script output:
loading seaborn

Python source code: plot_wage_data.py
# Standard library imports
import urllib
import os
##############################################################################
# Load the data
import pandas
if not os.path.exists('wages.txt'):
# Download the file if it is not present
urllib.urlretrieve('http://lib.stat.cmu.edu/datasets/CPS_85_Wages',
'wages.txt')
# Give names to the columns

names = [
'EDUCATION: Number of years of education',
'SOUTH: 1=Person lives in South, 0=Person lives elsewhere',
'SEX: 1=Female, 0=Male',
'EXPERIENCE: Number of years of work experience',
'UNION: 1=Union member, 0=Not union member',
'WAGE: Wage (dollars per hour)',
'AGE: years',

'RACE: 1=Other, 2=Hispanic, 3=White',

'OCCUPATION: 1=Management, 2=Sales, 3=Clerical, 4=Service, 5=Professional, 6=Other',
'SECTOR: 0=Other, 1=Manufacturing, 2=Construction',
'MARR: 0=Unmarried, 1=Married',
]
short_names = [n.split(':')[0] for n in names]
data = pandas.read_csv('wages.txt', skiprows=27, skipfooter=6, sep=None,

header=None)
data.columns = short_names
# Log-transform the wages, because they typically are increased with

# multiplicative factors
import numpy as np
data['WAGE'] = np.log10(data['WAGE'])
##############################################################################
# Plot scatter matrices highlighting different aspects
import seaborn
seaborn.pairplot(data, vars=['WAGE', 'AGE', 'EDUCATION'],
kind='reg')

kind='reg', hue='SEX')
plt.suptitle('Effect of gender: 1=Female, 0=Male') Figure 3.8: Air fares before and after 9/11 (page 309)
kind='reg', hue='RACE')
plt.suptitle('Effect of race: 1=Other, 2=Hispanic, 3=White')

kind='reg', hue='UNION')
plt.suptitle('Effect of union: 1=Union member, 0=Not union member')
##############################################################################
# Plot a simple regression
seaborn.lmplot(y='WAGE', x='EDUCATION', data=data)
plt.show()
Air fares before and after 9/11 This is a business-intelligence (BI) like application.
What is interesting here is that we may want to study fares as a function of the year, paired accordingly to the
trips, or forgetting the year, only as a function of the trip endpoints.
Using statsmodels linear models, we find that both with an OLS (ordinary least square) and a robust fit, the
intercept and the slope are significantly non-zero: the air fares have decreased between 2000 and 2001, and their
dependence on distance travelled has also decreased


Script output:
==============================================================================
Dep. Variable: fare R-squared: 0.275
Method: Least Squares F-statistic: 1585.
Date: Mon, 21 Sep 2015 Prob (F-statistic): 0.00
Time: 19:08:53 Log-Likelihood: -45532.
No. Observations: 8352 AIC: 9.107e+04
Df Residuals: 8349 BIC: 9.109e+04
Df Model: 2
=================================================================================
---------------------------------------------------------------------------------
Intercept 211.2428 2.466 85.669 0.000 206.409 216.076

dist 0.0484 0.001 48.149 0.000 0.046 0.050 import os

nb_passengers -32.8925 1.127 -29.191 0.000 -35.101 -30.684
============================================================================== ##############################################################################
Omnibus: 604.051 Durbin-Watson: 1.446 # Load the data
Prob(Omnibus): 0.000 Jarque-Bera (JB): 740.733 import pandas
Skew: 0.710 Prob(JB): 1.42e-161
Kurtosis: 3.338 Cond. No. 5.23e+03 if not os.path.exists('airfares.txt'):
============================================================================== # Download the file if it is not present
urllib.urlretrieve(
Warnings: 'http://www.stat.ufl.edu/~winner/data/airq4.dat',
[1] The condition number is large, 5.23e+03. This might indicate that there are 'airfares.txt')
strong multicollinearity or other numerical problems.
Robust linear Model Regression Results # As a seperator, ' +' is a regular expression that means 'one of more
============================================================================== # space'
Dep. Variable: fare No. Observations: 8352 data = pandas.read_csv('airfares.txt', sep=' +', header=0,
Model: RLM Df Residuals: 8349 names=['city1', 'city2', 'pop1', 'pop2',
Method: IRLS Df Model: 2 'dist', 'fare_2000', 'nb_passengers_2000',
Norm: HuberT 'fare_2001', 'nb_passengers_2001'])
Scale Est.: mad
Cov Type: H1 # we log-transform the number of passengers
Date: Mon, 21 Sep 2015 import numpy as np
Time: 19:08:53 data['nb_passengers_2000'] = np.log10(data['nb_passengers_2000'])
No. Iterations: 12 data['nb_passengers_2001'] = np.log10(data['nb_passengers_2001'])
=================================================================================
coef std err z P>|z| [95.0% Conf. Int.] ##############################################################################
--------------------------------------------------------------------------------- # Make a dataframe whith the year as an attribute, instead of separate columns
Intercept 215.0848 2.448 87.856 0.000 210.287 219.883
dist 0.0460 0.001 46.166 0.000 0.044 0.048 # This involves a small danse in which we separate the dataframes in 2,
nb_passengers -35.2686 1.119 -31.526 0.000 -37.461 -33.076 # one for year 2000, and one for 2001, before concatenating again.
=================================================================================
# Make an index of each flight
If the model instance has been used for another fit with different fit data_flat = data.reset_index()
parameters, then the fit options might not be the correct ones anymore .
OLS Regression Results data_2000 = data_flat[['city1', 'city2', 'pop1', 'pop2',
============================================================================== 'dist', 'fare_2000', 'nb_passengers_2000']]
Dep. Variable: fare_2001 R-squared: 0.159 # Rename the columns
Model: OLS Adj. R-squared: 0.159 data_2000.columns = ['city1', 'city2', 'pop1', 'pop2', 'dist', 'fare',
Method: Least Squares F-statistic: 791.7 'nb_passengers']
Date: Mon, 21 Sep 2015 Prob (F-statistic): 1.20e-159 # Add a column with the year
Time: 19:08:53 Log-Likelihood: -22640. data_2000['year'] = 2000
No. Observations: 4176 AIC: 4.528e+04
Df Residuals: 4174 BIC: 4.530e+04 data_2001 = data_flat[['city1', 'city2', 'pop1', 'pop2',
Df Model: 1 'dist', 'fare_2001', 'nb_passengers_2001']]
============================================================================== # Rename the columns
coef std err t P>|t| [95.0% Conf. Int.] data_2001.columns = ['city1', 'city2', 'pop1', 'pop2', 'dist', 'fare',
------------------------------------------------------------------------------ 'nb_passengers']
Intercept 148.0279 1.673 88.480 0.000 144.748 151.308 # Add a column with the year
dist 0.0388 0.001 28.136 0.000 0.036 0.041 data_2001['year'] = 2001
==============================================================================
Omnibus: 136.558 Durbin-Watson: 1.544 data_flat = pandas.concat([data_2000, data_2001])
Skew: 0.462 Prob(JB): 3.23e-33
Kurtosis: 2.920 Cond. No. 2.40e+03 ##############################################################################
============================================================================== # Plot scatter matrices highlighting different aspects
Warnings: import seaborn

[1] The condition number is large, 2.4e+03. This might indicate that there are seaborn.pairplot(data_flat, vars=['fare', 'dist', 'nb_passengers'],
strong multicollinearity or other numerical problems. kind='reg', markers='.')
Python source code: plot_airfare.py # A second plot, to show the effect of the year (ie the 9/11 effect)
seaborn.pairplot(data_flat, vars=['fare', 'dist', 'nb_passengers'],
# Standard library imports kind='reg', hue='year', markers='.')
import urllib

##############################################################################
# Plot the difference in fare
seaborn.boxplot(data.fare_2001 - data.fare_2000)
plt.title('Fare: 2001 - 2000')
plt.subplots_adjust()
seaborn.boxplot(data.nb_passengers_2001 - data.nb_passengers_2000)
plt.title('NB passengers: 2001 - 2000')
plt.subplots_adjust()
##############################################################################
# Statistical testing: dependence of fare on distance and number of
# passengers
import statsmodels.formula.api as sm
result = sm.ols(formula='fare ~ 1 + dist + nb_passengers', data=data_flat).fit()

print(result.summary())
# Using a robust fit

result = sm.rlm(formula='fare ~ 1 + dist + nb_passengers', data=data_flat).fit()
print(result.summary()) Figure 3.9: Relating Gender and IQ (page 315)
##############################################################################
# Statistical testing: regression of fare on distance: 2001/2000 difference
result = sm.ols(formula='fare_2001 - fare_2000 ~ 1 + dist', data=data).fit()

print(result.summary())
# Plot the corresponding regression

data['fare_difference'] = data['fare_2001'] - data['fare_2000']
seaborn.lmplot(x='dist', y='fare_difference', data=data)
plt.show()
Solutions to the exercises of the course
Relating Gender and IQ Going back to the brain size + IQ data, test if the VIQ of male and female are different
after removing the effect of brain size, height and weight.
Notice that here Gender is a categorical value. As it is a non-float data type, statsmodels is able to automatically
infer this.
Script output:

OLS Regression Results c=(data['Gender'] == 'Female'), marker='o',

============================================================================== alpha=1, cmap='winter')
Dep. Variable: VIQ R-squared: 0.246
Model: OLS Adj. R-squared: 0.181 fig = plt.gcf()
Method: Least Squares F-statistic: 3.809 fig.suptitle("blue: male, green: female", size=13)
Date: Mon, 21 Sep 2015 Prob (F-statistic): 0.0184
Time: 07:37:45 Log-Likelihood: -172.34 plt.show()
Df Residuals: 35 BIC: 359.3 Total running time of the example: 0.62 seconds ( 0 minutes 0.62 seconds)
Df Model: 3
==================================================================================
---------------------------------------------------------------------------------- 3.2 Sympy : Symbolic Mathematics in Python
Intercept 166.6258 88.824 1.876 0.069 -13.696 346.948
Gender[T.Male] 8.8524 10.710 0.827 0.414 -12.890 30.595 Author: Fabian Pedregosa
MRI_Count 0.0002 6.46e-05 2.615 0.013 3.78e-05 0.000
Height -3.0837 1.276 -2.417 0.021 -5.674 -0.494
============================================================================== Objectives
Prob(Omnibus): 0.025 Jarque-Bera (JB): 2.252 1. Evaluate expressions with arbitrary precision.
Skew: 0.005 Prob(JB): 0.324 2. Perform algebraic manipulations on symbolic expressions.
Kurtosis: 1.823 Cond. No. 2.40e+07 3. Perform basic calculus tasks (limits, differentiation and integration) with symbolic expressions.
============================================================================== 4. Solve polynomial and transcendental equations.
5. Solve some differential equations.
Warnings:
[1] The condition number is large, 2.4e+07. This might indicate that there are
strong multicollinearity or other numerical problems. What is SymPy? SymPy is a Python library for symbolic mathematics. It aims to be an alternative to systems
<F test: F=array([[ 0.68319608]]), p=[[ 0.41408784]], df_denom=35, df_num=1> such as Mathematica or Maple while keeping the code as simple as possible and easily extensible. SymPy is
written entirely in Python and does not require any external libraries.
Python source code: plot_brain_size.py
Sympy documentation and packages for installation can be found on http://www.sympy.org/
import pandas
Chapters contents
data = pandas.read_csv('../brain_size.csv', sep=';', na_values='.') First Steps with SymPy (page 318)
Using SymPy as a calculator (page 318)
model = ols('VIQ ~ Gender + MRI_Count + Height', data).fit()
Exercises (page 319)
Symbols (page 319)
# Here, we don't need to define a contrast, as we are testing a single Algebraic manipulations (page 320)
# coefficient of our model, and not a combination of coefficients. Expand (page 320)
# However, defining a contrast, which would then be a 'unit contrast', Simplify (page 320)
# will give us the same results Calculus (page 320)
print(model.f_test([0, 1, 0, 0])) Limits (page 320)
Differentiation (page 321)
############################################################################### Series expansion (page 321)
# Here we plot a scatter matrix to get intuitions on our results. Integration (page 322)
# This goes beyond what was asked in the exercise
Exercises (page 322)
# This plotting is useful to get an intuitions on the relationships between
Equation solving (page 322)
# our different variables Exercises (page 323)
Linear Algebra (page 323)
from pandas.tools import plotting Matrices (page 323)
import matplotlib.pyplot as plt Differential Equations (page 324)
# Fill in the missing values for Height for plotting

data['Height'].fillna(method='pad', inplace=True)
3.2.1 First Steps with SymPy
# The parameter 'c' is passed to plt.scatter and will control the color
# The same holds for parameters 'marker', 'alpha' and 'cmap', that
Using SymPy as a calculator
# control respectively the type of marker used, their transparency and
# the colormap
plotting.scatter_matrix(data[['VIQ', 'MRI_Count', 'Height']], SymPy defines three numerical types: Real, Rational and Integer.
3.1. Statistics in Python 317 3.2. Sympy : Symbolic Mathematics in Python 318
The Rational class represents a rational number as a pair of two Integers: the numerator and the denominator, so
Rational(1,2) represents 1/2, Rational(5,2) 5/2 and so on: Printing
>>> from sympy import * Here we use the following setting for printing
>>> a = Rational(1,2)
>>> sympy.init_printing(use_unicode=False, wrap_line=True)
>>> a
1/2
>>> a*2 3.2.2 Algebraic manipulations

1
SymPy is capable of performing powerful algebraic manipulations. Well take a look into some of the most
SymPy uses mpmath in the background, which makes it possible to perform computations using arbitrary- frequently used: expand and simplify.
precision arithmetic. That way, some special constants, like e, pi, oo (Infinity), are treated as symbols and can be
evaluated with arbitrary precision:
Expand
>>> pi**2
2
pi Use this to expand an algebraic expression. It will try to denest powers and multiplications:
>>> expand((x + y)**3)
>>> pi.evalf() 3 2 2 3
3.14159265358979 x + 3*x *y + 3*x*y + y
>>> 3*x*y**2 + 3*y*x**2 + x**3 + y**3
>>> (pi + exp(1)).evalf() 3 2 2 3
5.85987448204884 x + 3*x *y + 3*x*y + y
as you see, evalf evaluates the expression to a floating-point number.
Further options can be given in form on keywords:
There is also a class representing mathematical infinity, called oo: >>> expand(x + y, complex=True)
>>> oo > 99999 re(x) + re(y) + I*im(x) + I*im(y)
True >>> I*im(x) + I*im(y) + re(x) + re(y)
>>> oo + 1 re(x) + re(y) + I*im(x) + I*im(y)
oo
>>> expand(cos(x + y), trig=True)
-sin(x)*sin(y) + cos(x)*cos(y)
>>> cos(x)*cos(y) - sin(x)*sin(y)
Exercises
-sin(x)*sin(y) + cos(x)*cos(y)

1. Calculate 2 with 100 decimals.
2. Calculate 1/2 + 1/3 in rational arithmetic. Simplify
Use simplify if you would like to transform an expression into a simpler form:
Symbols
>>> simplify((x + x*y) / x)
y + 1
In contrast to other Computer Algebra Systems, in SymPy you have to declare symbolic variables explicitly:
>>> from sympy import * Simplification is a somewhat vague term, and more precises alternatives to simplify exists: powsimp (simplifica-
>>> x = Symbol('x') tion of exponents), trigsimp (for trigonometric expressions) , logcombine, radsimp, together.
>>> y = Symbol('y')
Then you can manipulate them: Exercises

>>> x + y + x - y 1. Calculate the expanded form of ( + )6 .
2*x 2. Simplify the trigonometric expression sin()/ cos()
>>> (x + y)**2
2
(x + y) 3.2.3 Calculus
Symbols can now be manipulated using some of python operators: +, -, *, ** (arithmetic), &, |, ~ , >>, <<
Limits
(boolean).
Limits are easy to use in SymPy, they follow the syntax limit(function, variable, point), so to compute the limit of
() as 0, you would issue limit(f, x, 0):
3.2. Sympy : Symbolic Mathematics in Python 319 3.2. Sympy : Symbolic Mathematics in Python 320
>>> limit(sin(x)/x, x, 0)
1 Exercises
you can also calculate the limit at infinity: 1. Calculate lim0 sin()/
2. Calculate the derivative of () for .
>>> limit(x, x, oo)
oo
>>> limit(1/x, x, oo) Integration

0
SymPy has support for indefinite and definite integration of transcendental elementary and special functions via
>>> limit(x**x, x, 0) integrate() facility, which uses powerful extended Risch-Norman algorithm and some heuristics and pattern
1
matching. You can integrate elementary functions:
>>> integrate(6*x**5, x)
Differentiation 6
x
You can differentiate any SymPy expression using diff(func, var). Examples: >>> integrate(sin(x), x)
-cos(x)
>>> diff(sin(x), x) >>> integrate(log(x), x)
cos(x) x*log(x) - x
>>> diff(sin(2*x), x) >>> integrate(2*x + sinh(x), x)
2*cos(2*x) 2
x + cosh(x)
>>> diff(tan(x), x)
2 Also special functions are handled easily:
tan (x) + 1
>>> integrate(exp(-x**2)*erf(x), x)
You can check, that it is correct by: ____ 2
\/ pi *erf (x)
>>> limit((tan(x+y) - tan(x))/y, y, 0) --------------
2 4
tan (x) + 1
It is possible to compute definite integral:
Higher derivatives can be calculated using the diff(func, var, n) method:
>>> integrate(x**3, (x, -1, 1))
>>> diff(sin(2*x), x, 1) 0
2*cos(2*x) >>> integrate(sin(x), (x, 0, pi/2))
1
>>> diff(sin(2*x), x, 2) >>> integrate(cos(x), (x, -pi/2, pi/2))
-4*sin(2*x) 2
>>> diff(sin(2*x), x, 3) Also improper integrals are supported as well:

-8*cos(2*x)
>>> integrate(exp(-x), (x, 0, oo))
1
>>> integrate(exp(-x**2), (x, -oo, oo))
Series expansion ____
\/ pi
SymPy also knows how to compute the Taylor series of an expression at a point. Use series(expr, var):
>>> series(cos(x), x)
2 4 Exercises
x x / 6\
1 - -- + -- + O\x / 3.2.4 Equation solving
2 24
>>> series(1/cos(x), x) SymPy is able to solve algebraic equations, in one and several variables:
2 4
x 5*x / 6\ In [7]: solve(x**4 - 1, x)
1 + -- + ---- + O\x / Out[7]: [I, 1, -1, -I]
2 24
As you can see it takes as first argument an expression that is supposed to be equaled to 0. It is able to solve a large
part of polynomial equations, and is also capable of solving multiple equations with respect to multiple variables
giving a tuple as second argument:
3.2. Sympy : Symbolic Mathematics in Python 321 3.2. Sympy : Symbolic Mathematics in Python 322
In [8]: solve([x + 5*y - 2, -3*x + 6*y - 15], [x, y]) Differential Equations
Out[8]: {y: 1, x: -3}
SymPy is capable of solving (some) Ordinary Differential. To solve differential equations, use dsolve. First, create
It also has (limited) support for trascendental equations: an undefined function by passing cls=Function to the symbols function:
In [9]: solve(exp(x) + 1, x) >>> f, g = symbols('f g', cls=Function)
Out[9]: [pi*I]
f and g are now undefined functions. We can call f(x), and it will represent an unknown function:
Another alternative in the case of polynomial equations is factor. factor returns the polynomial factorized
>>> f(x)
into irreducible terms, and is capable of computing the factorization over various domains:
f(x)
In [10]: f = x**4 - 3*x**2 + 1
In [11]: factor(f) >>> f(x).diff(x, x) + f(x)
Out[11]: (1 + x - x**2)*(1 - x - x**2) 2
d
In [12]: factor(f, modulus=5) f(x) + ---(f(x))
Out[12]: (2 + x)**2*(2 - x)**2 2
dx
SymPy is also able to solve boolean equations, that is, to decide if a certain boolean expression is satisfiable or
not. For this, we use the function satisfiable: >>> dsolve(f(x).diff(x, x) + f(x), f(x))
f(x) = C1*sin(x) + C2*cos(x)
In [13]: satisfiable(x & y)
Out[13]: {x: True, y: True} Keyword arguments can be given to this function in order to help if find the best possible resolution system. For
example, if you know that it is a separable equations, you can use keyword hint=separable to force dsolve to
This tells us that (x & y) is True whenever x and y are both True. If an expression cannot be true, i.e. no values of resolve it as a separable equation:
its arguments can make the expression True, it will return False:
>>> dsolve(sin(x)*cos(f(x)) + cos(x)*sin(f(x))*f(x).diff(x), f(x), hint='separable')
In [14]: satisfiable(x & ~x) / _____________\ / _____________\
Out[14]: False | / C1 | | / C1 |
[f(x) = - asin| / ------- + 1 | + pi, f(x) = asin| / ------- + 1 | + pi,
| / 2 | | / 2 |
Exercises \\/ cos (x) / \\/ cos (x) /
1. Solve the system of equations + = 2, 2 + = 0 / _____________\ / _____________\

| / C1 | | / C1 |
2. Are there boolean values x, y that make (~x | y) & (~y | x) true? f(x) = -asin| / ------- + 1 |, f(x) = asin| / ------- + 1 |]
| / 2 | | / 2 |
\\/ cos (x) / \\/ cos (x) /
3.2.5 Linear Algebra
Matrices Exercises
1. Solve the Bernoulli differential equation
Matrices are created as instances from the Matrix class:
>>> from sympy import Matrix
()
>>> Matrix([[1,0], [0,1]]) + () ()2 = 0
[1 0]
[ ] 2. Solve the same equation using hint=Bernoulli. What do you observe ?
[0 1]
unlike a NumPy array, you can also put Symbols in it:

>>> x = Symbol('x') 3.3 Scikit-image: image processing
>>> y = Symbol('y')
>>> A = Matrix([[1,x], [y,1]])
>>> A Author: Emmanuelle Gouillart
[1 x]
[ ] scikit-image is a Python package dedicated to image processing, and using natively NumPy arrays as image
[y 1] objects. This chapter describes how to use scikit-image on various image processing tasks, and insists on the
link with other scientific Python modules such as NumPy and SciPy.
>>> A**2
See also:
[x*y + 1 2*x ]
[ ] For basic image manipulation, such as image cropping or simple filtering, a large number of simple operations
[ 2*y x*y + 1] can be realized with NumPy and SciPy only. See Image manipulation and processing using Numpy and Scipy
3.2. Sympy : Symbolic Mathematics in Python 323 3.3. Scikit-image: image processing 324
(page 222).
Note that you should be familiar with the content of the previous chapter before reading the current one, as basic
operations such as masking and labeling are a prerequisite.
Chapters contents
Introduction and concepts (page 325)
scikit-image and the SciPy ecosystem (page 326)
Whats to be found in scikit-image (page 326)
Input/output, data types and colorspaces (page 327)
Data types (page 328)
Colorspaces (page 328)
Image preprocessing / enhancement (page 329)
Local filters (page 329)
Non-local filters (page 330)
Mathematical morphology (page 330)
Image segmentation (page 332)
Binary segmentation: foreground + background (page 332)
Marker based methods (page 333)
Measuring regions properties (page 335)
Data visualization and interaction (page 335)
Feature extraction for computer vision (page 337)
scikit-image and the SciPy ecosystem
3.3.1 Introduction and concepts Recent versions of scikit-image is packaged in most Scientific Python distributions, such as Anaconda or
Enthought Canopy. It is also packaged for Ubuntu/Debian.
Images are NumPys arrays np.ndarray
>>> import skimage
image np.ndarray >>> from skimage import data # most functions are in subpackages
pixels array values: a[2, 3]
Most scikit-image functions take NumPy ndarrays as arguments
channels array dimensions
>>> camera = data.camera()
image encoding dtype (np.uint8, np.uint16, np.float) >>> camera.dtype
dtype('uint8')
filters functions (numpy, skimage, scipy) >>> camera.shape
(512, 512)
>>> from skimage import restoration
>>> check = np.zeros((9, 9))
>>> filtered_camera = restoration.denoise_bilateral(camera)
>>> check[::2, 1::2] = 1
>>> type(filtered_camera)
>>> check[1::2, ::2] = 1
<type 'numpy.ndarray'>
>>> plt.imshow(check, cmap='gray', interpolation='nearest')
Other Python packages are available for image processing and work with NumPy arrays:
scipy.ndimage : for nd-arrays. Basic filtering, mathematical morphology, regions properties
Mahotas
Also, powerful image processing libraries have Python bindings:
OpenCV (computer vision)
ITK (3D images and registration)
and many others
(but they are less Pythonic and NumPy friendly, to a variable extent).
Whats to be found in scikit-image
Website: http://scikit-image.org/
3.3. Scikit-image: image processing 325 3.3. Scikit-image: image processing 326
Gallery of examples (as in matplotlib or scikit-learn): http://scikit-image.org/docs/stable/auto_examples/ Data types

Different kinds of functions, from boilerplate utility functions to high-level recent algorithms.
Filters: functions transforming images into other images.
NumPy machinery
Common filtering algorithms
Data reduction functions: computation of image histogram, position of local maxima, of corners, etc.
Other actions: I/O, visualization, etc.
Image ndarrays can be represented either by integers

3.3.2 Input/output, data types and colorspaces (signed or unsigned) or floats.
Careful with overflows with integer data types
I/O: skimage.io
>>> camera = data.camera()
>>> from skimage import io
>>> camera.dtype
dtype('uint8')
Reading from files: skimage.io.imread() >>> camera_multiply = 3 * camera
>>> import os
>>> filename = os.path.join(skimage.data_dir, 'camera.png') Different integer sizes are possible: 8-, 16- or 32-bytes, signed or unsigned.
>>> camera = io.imread(filename)
Warning: An important (if questionable) skimage convention: float images are supposed to lie in [-1, 1]
(in order to have comparable contrast for all float images)
>>> from skimage import img_as_float

>>> camera_float = img_as_float(camera)
>>> camera.max(), camera_float.max()
(255, 1.0)
Some image processing routines need to work with float arrays, and may hence output an array with a different
type and the data range from the input array
>>> try:
... from skimage import filters
... except ImportError:
... from skimage import filter as filters
>>> camera_sobel = filters.sobel(camera)
>>> camera_sobel.max()
0.591502...
Warning: In the example above, we use the filters submodule of scikit-image, that has been renamed
Works with all data formats supported by the Python from filter to filters between versions 0.10 and 0.11, in order to avoid a collision with Pythons built-in
Imaging Library (or any other I/O plugin provided to imread with the plugin keyword argument). name filter.
Also works with URL image paths:
Utility functions are provided in skimage to convert both the dtype and the data range, following skimages
>>> logo = io.imread('http://scikit-image.org/_static/img/logo.png')
conventions: util.img_as_float, util.img_as_ubyte, etc.
Saving to files: See the user guide for more details.
>>> io.imsave('local_logo.png', logo)
Colorspaces
(imsave also uses an external plugin such as PIL)
Color images are of shape (N, M, 3) or (N, M, 4) (when an alpha channel encodes transparency)
>>> lena = data.lena()
>>> lena.shape
(512, 512, 3)
Routines converting between different colorspaces (RGB, HSV, LAB etc.) are available in skimage.color : Non-local filters
color.rgb2hsv, color.lab2rgb, etc. Check the docstring for the expected dtype (and data range) of input
images. Non-local filters use a large region of the image (or all the image) to transform the value of one pixel:
>>> from skimage import exposure
3D images >>> camera = data.camera()
>>> camera_equalized = exposure.equalize_hist(camera)
Most functions of skimage can take 3D images as input arguments. Check the docstring to know if a
function can be used on 3D images (for example MRI or CT images).
Exercise
Open a color image on your disk as a NumPy array.
Find a skimage function computing the histogram of an image and plot the histogram of each
color channel
Convert the image to grayscale and plot its histogram.
Enhances contrast in large almost uniform regions.

3.3.3 Image preprocessing / enhancement
Mathematical morphology
Goals: denoising, feature (edges) extraction, ...
See https://en.wikipedia.org/wiki/Mathematical_morphology
Local filters Probe an image with a simple shape (a structuring element), and modify this image according to how the shape
locally fits or misses the image.
Local filters replace the value of pixels by a function of the values of neighboring pixels. The function can be
linear or non-linear. Default structuring element: 4-connectivity of a pixel
Neighbourhood: square (choose size), disk, or more complicated structuring element. >>> from skimage import morphology
>>> morphology.diamond(1)
array([[0, 1, 0],
[1, 1, 1],
[0, 1, 0]], dtype=uint8)
Example : horizontal Sobel filter

>>> text = data.text()
>>> hsobel_text = filters.hsobel(text)
Uses the following linear kernel for computing horizontal gradients:

1 2 1 Erosion = minimum filter. Replace the value of a pixel by the minimal value covered by the structuring element.:
0 0 0
-1 -2 -1 >>> a = np.zeros((7,7), dtype=np.int)
>>> a[1:6, 2:5] = 1
>>> a
array([[0, 0, 0, 0, 0, 0, 0],
[0, 0, 1, 1, 1, 0, 0],
[0, 0, 1, 1, 1, 0, 0],
[0, 0, 1, 1, 1, 0, 0],
[0, 0, 1, 1, 1, 0, 0],
[0, 0, 1, 1, 1, 0, 0],
[0, 0, 0, 0, 0, 0, 0]])
>>> morphology.binary_erosion(a, morphology.diamond(1)).astype(np.uint8)
array([[0, 0, 0, 0, 0, 0, 0], scipy.ndimage implementation works on arbitrary-dimensional arrays.

[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 1, 0, 0, 0],
[0, 0, 0, 1, 0, 0, 0],
[0, 0, 0, 1, 0, 0, 0], Example of filters comparison: image denoising
[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0]], dtype=uint8) >>> from skimage.morphology import disk
>>> #Erosion removes objects smaller than the structure >>> coins = data.coins()
>>> morphology.binary_erosion(a, morphology.diamond(2)).astype(np.uint8) >>> coins_zoom = coins[10:80, 300:370]
array([[0, 0, 0, 0, 0, 0, 0], >>> median_coins = filters.rank.median(coins_zoom, disk(1))
[0, 0, 0, 0, 0, 0, 0], >>> from skimage import restoration
[0, 0, 0, 0, 0, 0, 0], >>> tv_coins = restoration.denoise_tv_chambolle(coins_zoom, weight=0.1)
[0, 0, 0, 0, 0, 0, 0], >>> gaussian_coins = filters.gaussian_filter(coins, sigma=2)
[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0]], dtype=uint8)
Dilation: maximum filter:

>>> a = np.zeros((5, 5))
>>> a[2, 2] = 1
>>> a
array([[ 0., 0., 0., 0., 0.],
[ 0., 0., 0., 0., 0.],
[ 0., 0., 1., 0., 0.],
[ 0., 0., 0., 0., 0.],
[ 0., 0., 0., 0., 0.]])
>>> morphology.binary_dilation(a, morphology.diamond(1)).astype(np.uint8)
3.3.4 Image segmentation
array([[0, 0, 0, 0, 0],
[0, 0, 1, 0, 0], Image segmentation is the attribution of different labels to different regions of the image, for example in order to
[0, 1, 1, 1, 0], extract the pixels of an object of interest.
[0, 0, 1, 0, 0],
[0, 0, 0, 0, 0]], dtype=uint8)
Binary segmentation: foreground + background
Opening: erosion + dilation:
Histogram-based method: Otsu thresholding
>>> a[1:4, 1:4] = 1; a[4, 4] = 1
>>> a Tip: The Otsu method is a simple heuristic to find a threshold to separate the foreground from the background.
array([[0, 0, 0, 0, 0],
[0, 1, 1, 1, 0],
[0, 1, 1, 1, 0],
[0, 1, 1, 1, 0], Earlier scikit-image versions
[0, 0, 0, 0, 1]])
skimage.filters is called skimage.filter in earlier versions of scikit-image
>>> morphology.binary_opening(a, morphology.diamond(1)).astype(np.uint8)
array([[0, 0, 0, 0, 0],
[0, 0, 1, 0, 0], from skimage import data
[0, 1, 1, 1, 0], from skimage import filters
[0, 0, 1, 0, 0], camera = data.camera()
[0, 0, 0, 0, 0]], dtype=uint8) val = filters.threshold_otsu(camera)
mask = camera < val
Opening removes small objects and smoothes corners.
Grayscale mathematical morphology

Mathematical morphology operations are also available for (non-binary) grayscale images (int or float type).
Erosion and dilation correspond to minimum (resp. maximum) filters.
Higher-level mathematical morphology are available: tophat, skeletonization, etc.

See also:
Basic mathematical morphology is also implemented in scipy.ndimage.morphology. The
Watershed segmentation
The Watershed (skimage.morphology.watershed()) is a region-growing approach that fills basins in

the image
>>> from skimage.morphology import watershed
>>> from skimage.feature import peak_local_max
>>>
>>> # Generate an initial image with two overlapping circles
>>> x, y = np.indices((80, 80))
>>> x1, y1, x2, y2 = 28, 28, 44, 52
>>> r1, r2 = 16, 20
>>> mask_circle1 = (x - x1) ** 2 + (y - y1) ** 2 < r1 ** 2
>>> mask_circle2 = (x - x2) ** 2 + (y - y2) ** 2 < r2 ** 2
>>> image = np.logical_or(mask_circle1, mask_circle2)
Labeling connected components of a discrete image
>>> # Now we want to separate the two objects in image
>>> # Generate the markers as local maxima of the distance
>>> # to the background
Tip: Once you have separated foreground objects, it is use to separate them from each other. For this, we can
assign a different integer labels to each one. >>> from scipy import ndimage
>>> distance = ndimage.distance_transform_edt(image)
>>> local_maxi = peak_local_max(distance, indices=False, footprint=np.ones((3, 3)), labels=image)
Synthetic data: >>> markers = morphology.label(local_maxi)
>>> labels_ws = watershed(-distance, markers, mask=image)
>>> n = 20
>>> l = 256
>>> im = np.zeros((l, l))
>>> points = l * np.random.random((2, n ** 2)) Random walker segmentation
>>> im = filters.gaussian_filter(im, sigma=l / (4. * n)) The random walker algorithm (skimage.segmentation.random_walker()) is similar to the Watershed,
>>> blobs = im > im.mean() but with a more probabilistic approach. It is based on the idea of the diffusion of labels in the image:
Label all connected components: >>> from skimage import segmentation
>>> # Transform markers image so that 0-valued pixels are to
>>> from skimage import measure >>> # be labelled, and -1-valued pixels represent background
>>> all_labels = measure.label(blobs) >>> markers[~image] = -1
>>> labels_rw = segmentation.random_walker(image, markers)
Label only foreground connected components:
>>> blobs_labels = measure.label(blobs, background=0)
Postprocessing label images

skimage provides several utility functions that can be used on label images
(ie images where different discrete values identify different regions). Functions
names are often self-explaining: skimage.segmentation.clear_border(),
See also: skimage.segmentation.relabel_from_one(), skimage.morphology.remove_small_objects(),
scipy.ndimage.find_objects() is useful to return slices on object in an image. etc.
Marker based methods
If you have markers inside a set of regions, you can use these to segment the regions.
Exercise >>> coins_edges = segmentation.mark_boundaries(coins, clean_border.astype(np.int))

Load the coins image from the data submodule.
Separate the coins from the background by testing several segmentation methods: Otsu thresholding,
adaptive thresholding, and watershed or random walker segmentation.
If necessary, use a postprocessing function to improve the coins / background segmentation.
3.3.5 Measuring regions properties
>>> from skimage import measure
Example: compute the size and perimeter of the two segmented regions:
>>> properties = measure.regionprops(labels_rw)
>>> [prop.area for prop in properties]
[770.0, 1168.0] The (ex-
>>> [prop.perimeter for prop in properties]
perimental) scikit-image viewer
[100.91..., 126.81...]
skimage.viewer = matplotlib-based canvas for displaying images + experimental Qt-based GUI-toolkit
See also:
>>> from skimage import viewer
for some properties, functions are available as well in scipy.ndimage.measurements with a different API >>> new_viewer = viewer.ImageViewer(coins)
(a list is returned). >>> new_viewer.show()
Useful for displaying pixel values.

Exercise (continued)
For more interaction, plugins can be added to the viewer:
Use the binary image of the coins and background from the previous exercise.
Compute an image of labels for the different coins. >>> new_viewer = viewer.ImageViewer(coins)
Compute the size and eccentricity of all coins. >>> from skimage.viewer.plugins import lineprofile
>>> new_viewer += lineprofile.LineProfile()
>>> new_viewer.show()
3.3.6 Data visualization and interaction
Meaningful visualizations are useful when testing a given processing pipeline.

Some image processing operations:
>>> coins = data.coins()
>>> mask = coins > filters.threshold_otsu(coins)
>>> clean_border = segmentation.clear_border(mask)
Visualize binary result:

>>> plt.figure()
>>> plt.imshow(clean_border, cmap='gray')
Visualize contour
>>> plt.figure()
>>> plt.imshow(coins, cmap='gray')
>>> plt.contour(clean_border, [0.5])
<matplotlib.contour.QuadContourSet ...>
Use skimage dedicated utility function:
>>> from skimage import feature
Example: detecting corners using Harris detector

from skimage.feature import corner_harris, corner_subpix, corner_peaks
from skimage.transform import warp, AffineTransform
tform = AffineTransform(scale=(1.3, 1.1), rotation=1, shear=0.7,

translation=(210, 50))
image = warp(data.checkerboard(), tform.inverse, output_shape=(350, 350))
coords = corner_peaks(corner_harris(image), min_distance=5)

coords_subpix = corner_subpix(image, coords, window_size=13)
(this ex-
ample is taken from http://scikit-image.org/docs/stable/auto_examples/plot_corner.html)
Points of interest such as corners can then be used to match objects in different images, as described in http://scikit-
image.org/docs/stable/auto_examples/plot_matching.html
3.4 Traits: building interactive dialogs
Author: Didrik Pinte

3.3.7 Feature extraction for computer vision The Traits project allows you to simply add validation, initialization, delegation, notification and a graphical user
interface to Python object attributes.
Geometric or textural descriptor can be extracted from images in order to
Tip: In this tutorial we will explore the Traits toolset and learn how to dramatically reduce the amount of
classify parts of the image (e.g. sky vs. buildings) boilerplate code you write, do rapid GUI application development, and understand the ideas which underly other
match parts of different images (e.g. for object detection) parts of the Enthought Tool Suite.
and many other applications of Computer Vision Traits and the Enthought Tool Suite are open source projects licensed under a BSD-style license.
3.3. Scikit-image: image processing 337 3.4. Traits: building interactive dialogs 338
Intended Audience
Intermediate to advanced Python programmers
Requirements
Python 2.6 or 2.7 (www.python.org)
Either wxPython (http://www.wxpython.org/), PyQt (https://riverbankcomputing.com/software/pyqt/intro)
or PySide (https://pyside.github.io/docs/pyside/)
Numpy and Scipy (http://www.scipy.org)
Enthought Tool Suite 3.x or higher (http://code.enthought.com/projects)
All required software can be obtained by installing the EPD Free (https://store.enthought.com/)
Tutorial content
Introduction (page 339)
Example (page 340)
What are Traits (page 341)
Initialisation (page 342)
Validation (page 342)
Documentation (page 343)
Visualization: opening a dialog (page 343)
Deferral (page 345)
Notification (page 350)
Some more advanced traits (page 353) The main packages of the Enthought Tool Suite are:
Traits - component based approach to build our applications.
Kiva - 2D primitives supporting path based rendering, affine transforms, alpha blending and more.
3.4.1 Introduction
Enable - object based 2D drawing canvas.
Tip: The Enthought Tool Suite enable the construction of sophisticated application frameworks for data analysis, Chaco - plotting toolkit for building complex interactive 2D plots.
2D plotting and 3D visualization. These powerful, reusable components are released under liberal BSD-style
licenses. Mayavi - 3D visualization of scientific data based on VTK.
Envisage - application plugin framework for building scriptable and extensible applications
In this tutorial, we will focus on Traits.
3.4.2 Example
Throughout this tutorial, we will use an example based on a water resource management simple case. We will try
to model a dam and reservoir system. The reservoir and the dams do have a set of parameters :
Name
Minimal and maximal capacity of the reservoir [hm3]
Height and length of the dam [m]
Catchment area [km2]
Hydraulic head [m]
Power of the turbines [MW]
Minimal and maximal release [m3/s]
Efficiency of the turbines
3.4. Traits: building interactive dialogs 339 3.4. Traits: building interactive dialogs 340
The reservoir has a known behaviour. One part is related to the energy production based on the water released. A reservoir = Reservoir(name='Lac de Vouglans', max_storage=605)
simple formula for approximating electric power production at a hydroelectric plant is = , where:
is Power in watts,
Initialisation
is the density of water (~1000 kg/m3),
is height in meters, All the traits do have a default value that initialise the variables. For example, the basic python types do have the
following trait equivalents:
is flow rate in cubic meters per second,
Trait Python Type Built-in Default Value
is acceleration due to gravity of 9.8 m/s2, Bool Boolean False
is a coefficient of efficiency ranging from 0 to 1. Complex Complex number 0+0j
Float Floating point number 0.0
Tip: Annual electric energy production depends on the available water supply. In some installations the water Int Plain integer 0
flow rate can vary by a factor of 10:1 over the course of a year. Long Long integer 0L
Str String
The second part of the behaviour is the state of the storage that depends on controlled and uncontrolled parameters Unicode Unicode u
:
A number of other predefined trait type do exist : Array, Enum, Range, Event, Dict, List, Color, Set, Expression,
+1 = + Code, Callable, Type, Tuple, etc.
Custom default values can be defined in the code:
Warning: The data used in this tutorial are not real and might even not have sense in the reality.
from traits.api import HasTraits, Str, Float
class Reservoir(HasTraits):
3.4.3 What are Traits
name = Str
A trait is a type definition that can be used for normal Python object attributes, giving the attributes some additional max_storage = Float(100)
characteristics:
reservoir = Reservoir(name='Lac de Vouglans')
Standardization:
Initialization
Complex initialisation
Validation
When a complex initialisation is required for a trait, a _XXX_default magic method can be implemented. It
Deferral will be lazily called when trying to access the XXX trait. For example:
Notification def _name_default(self):
""" Complex initialisation of the reservoir name. """
Visualization
Documentation return 'Undefined'
A class can freely mix trait-based attributes with normal Python attributes, or can opt to allow the use of only a
fixed or open set of trait attributes within the class. Trait attributes defined by a class are automatically inherited
by any subclass derived from the class. Validation
The common way of creating a traits class is by extending from the HasTraits base class and defining class traits
: Every trait does validation when the user tries to set its content:
reservoir = Reservoir(name='Lac de Vouglans', max_storage=605)
reservoir.max_storage = '230'
---------------------------------------------------------------------------
TraitError Traceback (most recent call last)
name = Str
/Users/dpinte/projects/scipy-lecture-notes/advanced/traits/<ipython-input-7-979bdff9974a> in <modu
max_storage = Float
----> 1 reservoir.max_storage = '230'
Warning: For Traits 3.x users /Users/dpinte/projects/ets/traits/traits/trait_handlers.pyc in error(self, object, name, value)
If using Traits 3.x, you need to adapt the namespace of the traits packages: 166 """
traits.api should be enthought.traits.api 167 raise TraitError( object, name, self.full_info( object, name, value ),
traitsui.api should be enthought.traits.ui.api --> 168 value )
169
170 def arg_error ( self, method, arg_num, object, name, value ):
Using a traits class like that is as simple as any other Python class. Note that the trait value are passed using
keyword arguments:
reservoir1 = Reservoir()
TraitError: The 'max_storage' trait of a Reservoir instance must be a float, but a value of '23' <type 'str'> was specified.
reservoir1.edit_traits()
Documentation
By essence, all the traits do provide documentation about the model itself. The declarative approach to the creation
of classes makes it self-descriptive:
name = Str
max_storage = Float(100)
The desc metadata of the traits can be used to provide a more descriptive information about the trait :
name = Str TraitsUI simplifies the way user interfaces are created. Every trait on a HasTraits class has a default editor that
max_storage = Float(100, desc='Maximal storage [hm3]')
will manage the way the trait is rendered to the screen (e.g. the Range trait is displayed as a slider, etc.).
Lets now define the complete reservoir class: In the very same vein as the Traits declarative way of creating classes, TraitsUI provides a declarative interface to
build user interfaces code:
from traits.api import HasTraits, Str, Float, Range
from traits.api import HasTraits, Str, Float, Range
class Reservoir(HasTraits): from traitsui.api import View
name = Str
max_storage = Float(1e6, desc='Maximal storage [hm3]') class Reservoir(HasTraits):
max_release = Float(10, desc='Maximal release [m3/s]') name = Str
head = Float(10, desc='Hydraulic head [m]') max_storage = Float(1e6, desc='Maximal storage [hm3]')
efficiency = Range(0, 1.) max_release = Float(10, desc='Maximal release [m3/s]')
head = Float(10, desc='Hydraulic head [m]')
def energy_production(self, release): efficiency = Range(0, 1.)
''' Returns the energy production [Wh] for the given release [m3/s]
''' traits_view = View(
power = 1000 * 9.81 * self.head * release * self.efficiency 'name', 'max_storage', 'max_release', 'head', 'efficiency',
return power * 3600 title = 'Reservoir',
resizable = True,
)
if __name__ == '__main__':
reservoir = Reservoir( def energy_production(self, release):
name = 'Project A', ''' Returns the energy production [Wh] for the given release [m3/s]
max_storage = 30, '''
max_release = 100.0, power = 1000 * 9.81 * self.head * release * self.efficiency
head = 60, return power * 3600
efficiency = 0.8
)
if __name__ == '__main__':
release = 80 reservoir = Reservoir(
print 'Releasing {} m3/s produces {} kWh'.format( name = 'Project A',
release, reservoir.energy_production(release) max_storage = 30,
) max_release = 100.0,
head = 60,
efficiency = 0.8
Visualization: opening a dialog )
The Traits library is also aware of user interfaces and can pop up a default view for the Reservoir class: reservoir.configure_traits()
A special trait allows to manage events and trigger function calls using the magic _xxxx_fired method:
from traits.api import HasTraits, Instance, DelegatesTo, Float, Range, Event
from reservoir import Reservoir
class ReservoirState(HasTraits):
"""Keeps track of the reservoir state given the initial storage.
For the simplicity of the example, the release is considered in

hm3/timestep and not in m3/s.
"""
reservoir = Instance(Reservoir, ())
min_storage = Float
max_storage = DelegatesTo('reservoir')
min_release = Float
Deferral
max_release = DelegatesTo('reservoir')
Being able to defer the definition of a trait and its value to another object is a powerful feature of Traits. # state attributes
from traits.api import HasTraits, Instance, DelegatesTo, Float, Range storage = Range(low='min_storage', high='max_storage')
from reservoir import Reservoir # control attributes

inflows = Float(desc='Inflows [hm3]')
class ReservoirState(HasTraits): release = Range(low='min_release', high='max_release')
"""Keeps track of the reservoir state given the initial storage. spillage = Float(desc='Spillage [hm3]')
"""
reservoir = Instance(Reservoir, ()) update_storage = Event(desc='Updates the storage to the next time step')
min_storage = Float
max_storage = DelegatesTo('reservoir') def _update_storage_fired(self):
min_release = Float # update storage state
max_release = DelegatesTo('reservoir') new_storage = self.storage - self.release + self.inflows
self.storage = min(new_storage, self.max_storage)
# state attributes overflow = new_storage - self.max_storage
storage = Range(low='min_storage', high='max_storage') self.spillage = max(overflow, 0)
# control attributes def print_state(self):

inflows = Float(desc='Inflows [hm3]') print 'Storage\tRelease\tInflows\tSpillage'
release = Range(low='min_release', high='max_release') str_format = '\t'.join(['{:7.2f}'for i in range(4)])
spillage = Float(desc='Spillage [hm3]') print str_format.format(self.storage, self.release, self.inflows,
self.spillage)
def print_state(self): print '-' * 79
print 'Storage\tRelease\tInflows\tSpillage'
str_format = '\t'.join(['{:7.2f}'for i in range(4)])
print str_format.format(self.storage, self.release, self.inflows, if __name__ == '__main__':
self.spillage) projectA = Reservoir(
print '-' * 79 name = 'Project A',
max_storage = 30,
max_release = 5.0,
if __name__ == '__main__': hydraulic_head = 60,
projectA = Reservoir( efficiency = 0.8
name = 'Project A', )
max_storage = 30,
max_release = 100.0, state = ReservoirState(reservoir=projectA, storage=15)
hydraulic_head = 60, state.release = 5
efficiency = 0.8 state.inflows = 0
)
# release the maximum amount of water during 3 time steps
state = ReservoirState(reservoir=projectA, storage=10) state.update_storage = True
state.release = 90 state.print_state()
state.inflows = 0 state.update_storage = True
state.print_state() state.print_state()
state.update_storage = True
print 'How do we update the current storage ?' state.print_state()
Dependency between objects can be made automatic using the trait Property. The depends_on attribute ex- )
presses the dependency between the property and other traits. When the other traits gets changed, the property is
invalidated. Again, Traits uses magic method names for the property : state = ReservoirState(reservoir=projectA, storage=25)
state.release = 4
_get_XXX for the getter of the XXX Property trait
state.inflows = 0
_set_XXX for the setter of the XXX Property trait
state.print_state()
from traits.api import HasTraits, Instance, DelegatesTo, Float, Range
from traits.api import Property
Note: Caching property
from reservoir import Reservoir Heavy computation or long running computation might be a problem when accessing a property where the inputs
have not changed. The @cached_property decorator can be used to cache the value and only recompute them once
class ReservoirState(HasTraits): invalidated.
For the simplicity of the example, the release is considered in Lets extend the TraitsUI introduction with the ReservoirState example:
from traits.api import HasTraits, Instance, DelegatesTo, Float, Range, Property
"""
from traitsui.api import View, Item, Group, VGroup
min_release = Float
# state attributes
storage = Property(depends_on='inflows, release')
# control attributes
"""
release = Range(low='min_release', high='max_release')
name = DelegatesTo('reservoir')
spillage = Property(
desc='Spillage [hm3]', depends_on=['storage', 'inflows', 'release']
)
min_release = Float
### Private traits. ######################################################
# state attributes
_storage = Float
storage = Property(depends_on='inflows, release')
### Traits property implementation. ######################################
def _get_storage(self):
new_storage = self._storage - self.release + self.inflows
release = Range(low='min_release', high='max_release')
return min(new_storage, self.max_storage)
spillage = Property(
desc='Spillage [hm3]', depends_on=['storage', 'inflows', 'release']
def _set_storage(self, storage_value):
)
self._storage = storage_value
### Traits view ##########################################################
def _get_spillage(self):
traits_view = View(
Group(
overflow = new_storage - self.max_storage
VGroup(Item('name'), Item('storage'), Item('spillage'),
return max(overflow, 0)
label = 'State', style = 'readonly'
),
def print_state(self):
VGroup(Item('inflows'), Item('release'), label='Control'),
)
)
print str_format.format(self.storage, self.release, self.inflows,
self.spillage)
### Private traits. ######################################################
print '-' * 79
_storage = Float
if __name__ == '__main__':
projectA = Reservoir(
name = 'Project A',
max_storage = 30,
max_release = 5,
hydraulic_head = 60,
efficiency = 0.8
self._storage = storage_value head = Float(10, desc='Hydraulic head [m]')

efficiency = Range(0, 1.)
new_storage = self._storage - self.release + self.inflows turbine = Instance(Turbine)
overflow = new_storage - self.max_storage installed_capacity = PrototypedFrom('turbine', 'power')
def print_state(self): if __name__ == '__main__':

print 'Storage\tRelease\tInflows\tSpillage' turbine = Turbine(turbine_type='type1', power=5.0)
print str_format.format(self.storage, self.release, self.inflows, reservoir = Reservoir(
self.spillage) name = 'Project A',
print '-' * 79 max_storage = 30,
max_release = 100.0,
if __name__ == '__main__': head = 60,
projectA = Reservoir( efficiency = 0.8,
name = 'Project A', turbine = turbine,
max_storage = 30, )
max_release = 5,
hydraulic_head = 60, print 'installed capacity is initialised with turbine.power'
efficiency = 0.8 print reservoir.installed_capacity
)
print '-' * 15
state = ReservoirState(reservoir=projectA, storage=25) print 'updating the turbine power updates the installed capacity'
state.release = 4 turbine.power = 10
state.inflows = 0 print reservoir.installed_capacity
state.print_state() print '-' * 15

state.configure_traits() print 'setting the installed capacity breaks the link between turbine.power'
print 'and the installed_capacity trait'
reservoir.installed_capacity = 8
print turbine.power, reservoir.installed_capacity
Notification
Traits implements a Listener pattern. For each trait a list of static and dynamic listeners can be fed with callbacks.
When the trait does change, all the listeners are called.
Static listeners are defined using the _XXX_changed magic methods:
"""
Some use cases need the delegation mechanism to be broken by the user when setting the value of the trait. The reservoir = Instance(Reservoir, ())
min_storage = Float
PrototypeFrom trait implements this behaviour.
from traits.api import HasTraits, Str, Float, Range, PrototypedFrom, Instance min_release = Float
class Turbine(HasTraits):
turbine_type = Str # state attributes
power = Float(1.0, desc='Maximal power delivered by the turbine [Mw]') storage = Range(low='min_storage', high='max_storage')
class Reservoir(HasTraits): inflows = Float(desc='Inflows [hm3]')
name = Str release = Range(low='min_release', high='max_release')
max_storage = Float(1e6, desc='Maximal storage [hm3]') spillage = Float(desc='Spillage [hm3]')
max_release = Float(10, desc='Maximal release [m3/s]')
state = ReservoirState(reservoir=projectA, storage=10)
#register the dynamic listener
state.on_trait_change(wake_up_watchman_if_spillage, name='spillage')
self.spillage)
print '-' * 79
state.release = 90
state.inflows = 0
### Traits listeners #####################################################
state.print_state()
def _release_changed(self, new):
"""When the release is higher than zero, warn all the inhabitants of
print 'Forcing spillage'
the valley.
state.inflows = 100
"""
state.release = 0
if new > 0:
print 'Why do we have two executions of the callback ?'
print 'Warning, we are releasing {} hm3 of water'.format(new)
The dynamic trait notification signatures are not the same as the static ones :
if __name__ == '__main__': def wake_up_watchman(): pass
name = 'Project A', def wake_up_watchman(new): pass
max_storage = 30,
def wake_up_watchman(name, new): pass
hydraulic_head = 60, def wake_up_watchman(object, name, new): pass
efficiency = 0.8
) def wake_up_watchman(object, name, old, new): pass
Removing a dynamic listener can be done by:
state.release = 90 calling the remove_trait_listener method on the trait with the listener method as argument,
state.inflows = 0
state.print_state() calling the on_trait_change method with listener method and the keyword remove=True,
deleting the instance that holds the listener.
The static trait notification signatures can be:
Listeners can also be added to classes using the on_trait_change decorator:
def _release_changed(self): pass
def _release_changed(self, new): pass from traits.api import Property, on_trait_change
def _release_changed(self, old, new): pass
def _release_changed(self, name, old, new pass
Listening to all the changes
To listen to all the changes on a HasTraits class, the magic _any_trait_changed method can be implemented.
"""
In many situations, you do not know in advance what type of listeners need to be activated. Traits offers the ability reservoir = Instance(Reservoir, ())
to register listeners on the fly with the dynamic listeners max_storage = DelegatesTo('reservoir')
min_release = Float
from reservoir import Reservoir max_release = DelegatesTo('reservoir')
from reservoir_state_property import ReservoirState
# state attributes
def wake_up_watchman_if_spillage(new_value): storage = Property(depends_on='inflows, release')
if new_value > 0:
print 'Wake up watchman! Spilling {} hm3'.format(new_value) # control attributes
if __name__ == '__main__': release = Range(low='min_release', high='max_release')
projectA = Reservoir( spillage = Property(
name = 'Project A', desc='Spillage [hm3]', depends_on=['storage', 'inflows', 'release']
max_storage = 30, )
efficiency = 0.8 ### Private traits. ######################################################
) _storage = Float
power = 1000 * 9.81 * self.head * release * self.efficiency

return power * 3600
traits_view = View(
Item('name'),
Item('max_storage'),
Item('max_release'),
self._storage = storage_value
Item('head'),
Item('efficiency'),
Item('irrigated_areas'),
resizable = True
overflow = new_storage - self.max_storage
)
if __name__ == '__main__':
@on_trait_change('storage')
upper_block = IrrigationArea(name='Section C', surface=2000, crop='Wheat')
reservoir = Reservoir(
name='Project A',
max_storage=30,
self.spillage)
max_release=100.0,
print '-' * 79
head=60,
efficiency=0.8,
if __name__ == '__main__':
irrigated_areas=[upper_block]
)
name = 'Project A',
max_storage = 30,
release = 80
max_release = 5,
print 'Releasing {} m3/s produces {} kWh'.format(
release, reservoir.energy_production(release)
efficiency = 0.8
)
)

Trait listeners can be used to listen to changes in the content of the list to e.g. keep track of the total crop surface
state.release = 4 on linked to a given reservoir.
state.inflows = 0 from traits.api import HasTraits, Str, Float, Range, Enum, List, Property
from traitsui.api import View, Item
The patterns supported by the on_trait_change method and decorator are powerful. The reader should look at the
docstring of HasTraits.on_trait_change for the details. class IrrigationArea(HasTraits):
name = Str
surface = Float(desc='Surface [ha]')
Some more advanced traits crop = Enum('Alfalfa', 'Wheat', 'Cotton')
The following example demonstrate the usage of the Enum and List traits :
from traits.api import HasTraits, Str, Float, Range, Enum, List
name = Str
from traitsui.api import View, Item
max_storage = Float(1e6, desc='Maximal storage [hm3]')
class IrrigationArea(HasTraits):
name = Str
surface = Float(desc='Surface [ha]')
crop = Enum('Alfalfa', 'Wheat', 'Cotton')
irrigated_areas = List(IrrigationArea)
total_crop_surface = Property(depends_on='irrigated_areas.surface')
name = Str
max_storage = Float(1e6, desc='Maximal storage [hm3]')
def _get_total_crop_surface(self):
return sum([iarea.surface for iarea in self.irrigated_areas])
def energy_production(self, release):
irrigated_areas = List(IrrigationArea)
'''
def energy_production(self, release):
power = 1000 * 9.81 * self.head * release * self.efficiency
return power * 3600
'''
traits_view = View( ### Traits properties ####################################################

Item('name'), def _get_stock(self):
Item('max_storage'), """
Item('max_release'), fixme: should handle cases where we go over the max storage
Item('head'), """
Item('efficiency'), return self.initial_stock + (self.inflows - self.releases).cumsum()
Item('irrigated_areas'),
Item('total_crop_surface'), def _get_month(self):
resizable = True return np.arange(self.stock.size)
)
if __name__ == '__main__':
if __name__ == '__main__': reservoir = Reservoir(
upper_block = IrrigationArea(name='Section C', surface=2000, crop='Wheat') name = 'Project A',
max_storage = 30,
reservoir = Reservoir( max_release = 100.0,
name='Project A', head = 60,
max_storage=30, efficiency = 0.8
max_release=100.0, )
head=60,
efficiency=0.8, initial_stock = 10.
irrigated_areas=[upper_block], inflows_ts = np.array([6., 6, 4, 4, 1, 2, 0, 0, 3, 1, 5, 3])
) releases_ts = np.array([4., 5, 3, 5, 3, 5, 5, 3, 2, 1, 3, 3])
release = 80 view = ReservoirEvolution(

print 'Releasing {} m3/s produces {} kWh'.format( reservoir = reservoir,
release, reservoir.energy_production(release) inflows = inflows_ts,
) releases = releases_ts
)
The next example shows how the Array trait can be used to feed a specialised TraitsUI Item, the ChacoPlotItem: view.configure_traits()
import numpy as np
from traits.api import HasTraits, Array, Instance, Float, Property

from traits.api import DelegatesTo
from traitsui.api import View, Item, Group
from chaco.chaco_plot_editor import ChacoPlotItem
class ReservoirEvolution(HasTraits):
reservoir = Instance(Reservoir)
name = DelegatesTo('reservoir')
inflows = Array(dtype=np.float64, shape=(None))

releass = Array(dtype=np.float64, shape=(None))
initial_stock = Float
stock = Property(depends_on='inflows, releases, initial_stock')
month = Property(depends_on='stock') See also:

References
### Traits view ##########################################################
traits_view = View( ETS repositories: http://github.com/enthought
Item('name'),
Group( Traits manual: http://docs.enthought.com/traits/traits_user_manual/index.html
ChacoPlotItem('month', 'stock', show_label=False), Traits UI manual: http://docs.enthought.com/traitsui/traitsui_user_manual/index.html
),
width = 500, Mailing list : enthought-dev@enthought.com
resizable = True
)
3.5 3D plotting with Mayavi 3D plotting functions
Points
Tip: Mayavi is an interactive 3D plotting package. matplotlib (page 81) can also do simple 3D plotting, but
Mayavi relies on a more powerful engine ( VTK ) and is more suited to displaying large or complex data.
Chapters contents
Mlab: the scripting interface (page 357)
3D plotting functions (page 358)
* Points (page 358)
* Lines (page 358)
* Elevation surface (page 359)
* Arbitrary regular mesh (page 359)
* Volumetric data (page 360)
Figures and decorations (page 360) Hint: Points in 3D, represented with markers (or glyphs) and optionaly different sizes.
* Figure management (page 360)
x, y, z, value = np.random.random((4, 40))
* Changing plot properties (page 360)
mlab.points3d(x, y, z, value)
* Decorations (page 362)
Interactive work (page 363)
The pipeline dialog (page 363)
The script recording button (page 364) Lines
Slicing and dicing data: sources, modules and filters (page 364)
An example: inspecting magnetic fields (page 364)
Different views on data: sources and modules (page 364)
* Different sources: scatters and fields (page 365)
* Transforming data: filters (page 366)
* mlab.pipeline: the scripting layer (page 366)
Animating the data (page 366)
Making interactive dialogs (page 367)
A simple dialog (page 367)
Making it interactive (page 368)
Putting it together (page 369)
3.5.1 Mlab: the scripting interface Hint: A line connecting points in 3D, with optional thickness and varying color.
The mayavi.mlab module provides simple plotting functions to apply to numpy arrays, similar to matplotlib or mlab.clf() # Clear the figure
matlabs plotting interface. Try using them in IPython, by starting IPython with the switch --gui=wx. t = np.linspace(0, 20, 200)
mlab.plot3d(np.sin(t), np.cos(t), 0.1*t, t)
3.5. 3D plotting with Mayavi 357 3.5. 3D plotting with Mayavi 358
Elevation surface Volumetric data
Hint: A surface given by its elevation, coded as a 2D array Hint: If your data is dense in 3D, it is more difficult to display. One option is to take iso-contours of the data.
mlab.clf() mlab.clf()
x, y = np.mgrid[-10:10:100j, -10:10:100j] x, y, z = np.mgrid[-5:5:64j, -5:5:64j, -5:5:64j]
r = np.sqrt(x**2 + y**2) values = x*x*0.5 + y*y + z*z*2.0
z = np.sin(r)/r mlab.contour3d(values)
mlab.surf(z, warp_scale='auto')
Arbitrary regular mesh
This function works with a regular orthogonal grid: the value array is a 3D array that gives the shape of the
grid.
Hint: A surface mesh given by x, y, z positions of its node points

Figures and decorations
mlab.clf()
phi, theta = np.mgrid[0:np.pi:11j, 0:2*np.pi:11j] Figure management
x = np.sin(phi) * np.cos(theta)
y = np.sin(phi) * np.sin(theta)
z = np.cos(phi) Tip: Here is a list of functions useful to control the current figure
mlab.mesh(x, y, z)
mlab.mesh(x, y, z, representation='wireframe', color=(0, 0, 0)) Get the current figure: mlab.gcf()
Clear the current figure: mlab.clf()
Set the current figure: mlab.figure(1, bgcolor=(1, 1, 1), fgcolor=(0.5, 0.5,
Note: A surface is defined by points connected to form triangles or polygones. In mayavi.mlab.surf()
and mayavi.mlab.mesh(), the connectivity is implicity given by the layout of the arrays. See also 0.5)
mayavi.mlab.triangular_mesh(). Save figure to image mlab.savefig(foo.png, size=(300, 300))
file:
Change the view: mlab.view(azimuth=45, elevation=54, distance=1.)
Our data is often more than points and values: it needs some connectivity information
Changing plot properties
Tip: In general, many properties of the various objects on the figure can be changed. If these visualization are
created via mlab functions, the easiest way to change them is to use the keyword arguments of these functions,
as described in the docstrings.
Example docstring: mlab.mesh

Plots a surface using grid-spaced data supplied as 2D arrays.
Function signatures:
mesh(x, y, z, ...)
x, y, z are 2D arrays, all of the same shape, giving the positions of the vertices of the surface. The connectivity
between these points is implied by the connectivity on the arrays.
For simple structures (such as orthogonal grids) prefer the surf function, as it will create more efficient data
structures.
Keyword arguments:
color the color of the vtk object. Overides the colormap, if any, when specified.
This is specified as a triplet of float ranging from 0 to 1, eg (1, 1, 1) for white.
colormap type of colormap to use.
extent [xmin, xmax, ymin, ymax, zmin, zmax] Default is the x, y, z arrays extents.
Use this to change the extent of the object created.
figure Figure to populate.
line_width The with of the lines, if any used. Must be a float. Default: 2.0 Example:
mask boolean mask array to suppress some data points.
mask_points If supplied, only one out of mask_points data point is displayed. In [1]: import numpy as np
This option is usefull to reduce the number of points displayed on large datasets
In [2]: r, theta = np.mgrid[0:10, -np.pi:np.pi:10j]
Must be an integer or None.
mode the mode of the glyphs. Must be 2darrow or 2dcircle or 2dcross or In [3]: x = r * np.cos(theta)
2ddash or 2ddiamond or 2dhooked_arrow or 2dsquare or 2dthick_arrow
or 2dthick_cross or 2dtriangle or 2dvertex or arrow or cone or cube or In [4]: y = r * np.sin(theta)
cylinder or point or sphere. Default: sphere
name the name of the vtk object created. In [5]: z = np.sin(r)/r
representation the representation type used for the surface. Must be surface or
wireframe or points or mesh or fancymesh. Default: surface In [6]: from mayavi import mlab
resolution The resolution of the glyph created. For spheres, for instance, this is the
number of divisions along theta and phi. Must be an integer. Default: 8 In [7]: mlab.mesh(x, y, z, colormap='gist_earth', extent=[0, 1, 0, 1, 0, 1])
scalars optional scalar data. Out[7]: <mayavi.modules.surface.Surface object at 0xde6f08c>
scale_factor scale factor of the glyphs used to represent the vertices, in fancy_mesh
In [8]: mlab.mesh(x, y, z, extent=[0, 1, 0, 1, 0, 1],
mode. Must be a float. Default: 0.05 ...: representation='wireframe', line_width=1, color=(0.5, 0.5, 0.5))
scale_mode the scaling mode for the glyphs (vector, scalar, or none). Out[8]: <mayavi.modules.surface.Surface object at 0xdd6a71c>
transparent make the opacity of the actor depend on the scalar.
tube_radius radius of the tubes used to represent the lines, in mesh mode. If None,
simple lines are used. Decorations
tube_sides number of sides of the tubes used to represent the lines. Must be an
integer. Default: 6
vmax vmax is used to scale the colormap If None, the max of the data will be used Tip: Different items can be added to the figure to carry extra information, such as a colorbar or a title.
vmin vmin is used to scale the colormap If None, the min of the data will be used
In [9]: mlab.colorbar(Out[7], orientation='vertical')
Out[9]: <tvtk_classes.scalar_bar_actor.ScalarBarActor object at 0xd897f8c>
In [10]: mlab.title('polar mesh')

Out[10]: <enthought.mayavi.modules.text.Text object at 0xd8ed38c>
In [11]: mlab.outline(Out[7])
Out[11]: <enthought.mayavi.modules.outline.Outline object at 0xdd21b6c>
In [12]: mlab.axes(Out[7])
Out[12]: <enthought.mayavi.modules.axes.Axes object at 0xd2e4bcc>
The script recording button
To find out what code can be used to program these changes, click on the red button as you modify those properties,
and it will generate the corresponding lines of code.
3.5.3 Slicing and dicing data: sources, modules and filters
An example: inspecting magnetic fields
Suppose we are simulating the magnetic field generated by Helmholtz coils. The
examples/compute_field.py script does this computation and gives you a B array, that is (3 x n),
where the first axis is the direction of the field (Bx, By, Bz), and the second axis the index number of the point.
Arrays X, Y and Z give the positions of these data points.
Excercise
Visualize this field. Your goal is to make sure that the simulation code is correct.
Warning: extent: If we specified extents for a plotting object, mlab.outline and mlab.axes
dont get them by default.
3.5.2 Interactive work
Tip: The quickest way to create beautiful visualization with Mayavi is probably to interactively tweak the various
settings.
The pipeline dialog
Click on the Mayavi button in the scene, and you can control properties of objects with dialogs.
Suggestions
If you compute the norm of the vector field, you can apply an isosurface to it.
using mayavi.mlab.quiver3d() you can plot vectors. You can also use the masking options
(in the GUI) to make the plot a bit less dense.
Different views on data: sources and modules
Set the background of the figure in the Mayavi Scene node Tip: As we see above, it may be desirable to look at the same data in different ways.
Set the colormap in the Colors and legends node
Mayavi visualization are created by loading the data in a data source and then displayed on the screen using
Right click on the node to add modules or filters modules.
This can be seen by looking at the pipeline view. By right-clicking on the nodes of the pipeline, you can add Transforming data: filters
new modules.
If you create a vector field, you may want to visualize the iso-contours of its magnitude. But the isosurface module
Quiz can only be applied to scalar data, and not vector data. We can use a filter, ExtractVectorNorm to add this
scalar value to the vector field.
Why is it not possible to add a VectorCutPlane to the vectors created by
mayavi.mlab.quiver3d()? Filters apply a transformation to data, and can be added between sources and modules
Excercice
Different sources: scatters and fields Using the GUI, add the ExtractVectorNorm filter to display iso-contours of the field magnitude.
Tip: Data comes in different descriptions.

A 3D block of regularly-spaced value is structured: it is easy to know how one measurement is related mlab.pipeline: the scripting layer
to another neighboring and how to continuously interpolate between these. We can call such data a field,
borrowing from terminology used in physics, as it is continuously defined in space. The mlab scripting layer builds pipelines for you. You can reproduce these pipelines programmatically with
the mlab.pipeline interface: each step has a corresponding mlab.pipeline function (simply convert
A set of data points measured at random positions in a random order gives rise to much more difficult and
the name of the step to lower-case underscore-separated: ExtractVectorNorm gives extract_vector_norm). This
ill-posed interpolation problems: the data structure itself does not tell us what are the neighbors of a data
function takes as an argument the node that it applies to, as well as optional parameters, and returns the new node.
point. We call such data a scatter.
For example, iso-contours of the magnitude are coded as:
mlab.pipeline.iso_surface(mlab.pipeline.extract_vector_norm(field),
contours=[0.1*Bmax, 0.4*Bmax],
opacity=0.5)
Unstructured and unconnected data: a scatter Structured and connected data: a field
mlab.points3d, mlab.quiver3d mlab.contour3d
Data sources corresponding to scatters can be created with mayavi.mlab.pipeline.scalar_scatter()
or mayavi.mlab.pipeline.vector_scatter(); field data sources can be created with
mlab.pipeline.scalar_field() or mlab.pipeline.vector_field().
Exercice:
1. Create a contour (for instance of the magnetic field norm) by using one of those functions and adding Excercice
the right module by clicking on the GUI dialog.
Using the mlab.pipeline interface, generate a complete visualization, with iso-contours of the field magni-
2. Create the right source to apply a vector_cut_plane and reproduce the picture of the magnetic field
tude, and a vector cut plane.
shown previously.
(click on the figure for a solution)
Note that one of the difficulties is providing the data in the right form (number of arrays, shape) to the
functions. This is often the case with real-life data.
See also: 3.5.4 Animating the data

Sources are described in details in the Mayavi manual.
Tip: To make movies, or interactive application, you may want to change the data represented on a given
visualization.
If you have built a visualization, using the mlab plotting functions, or the mlab.pipeline function, we can
update the data by assigning new values to the mlab_source attributes
x , y , z = np.ogrid[-5:5:100j ,-5:5:100j, -5:5:100j] for embedding it in the dialog. The view of this dialog is defined by the view attribute of the object. In the init of
scalars = np.sin(x * y * z) / (x * y * z) this object, we populate the 3D scene with a curve.
Finally, the configure_traits method creates the dialog and starts the event loop.
iso = mlab.contour3d(scalars, transparent=True, contours=[0.5])
for i in range(1, 20):
scalars = np.sin(i * x * y * z) /(x * y * z) See also:
iso.mlab_source.scalars = scalars
There are a few things to be aware of when doing dialogs with Mayavi. Please read the Mayavi documentation
See also:
More details in the Mayavi documentation Making it interactive
We can combine the Traits events handler (page 350) with the mlab_source to modify the visualization with
Event loops the dialog.
For the interaction with the user (for instance changing the view with the mouse), Mayavi needs some time We will enable the user to vary the n_turns parameter in the definition of the curve. For this, we need:
to process these events. The for loop above prevents this. The Mayavi documentation details a workaround
to define an n_turns attribute on our visualization object, so that it can appear in the dialog. We use a
Range type.
to wire modification of this attribute to a recomputation of the curve. For this, we use the
3.5.5 Making interactive dialogs
on_traits_change decorator.
It is very simple to make interactive dialogs with Mayavi using the Traits library (see the dedicated chapter Traits:
building interactive dialogs (page 338)).
A simple dialog
from traits.api import HasTraits, Instance

from traitsui.api import View, Item, HGroup
from mayavi.core.ui.api import SceneEditor, MlabSceneModel
def curve(n_turns):
"The function creating the x, y, z coordinates needed to plot"
phi = np.linspace(0, 2*np.pi, 2000)
return [np.cos(phi) * (1 + 0.5*np.cos(n_turns*phi)),
np.sin(phi) * (1 + 0.5*np.cos(n_turns*phi)),
0.5*np.sin(n_turns*phi)]
class Visualization(HasTraits):
"The class that contains the dialog"
scene = Instance(MlabSceneModel, ())
def __init__(self):
HasTraits.__init__(self)
x, y, z = curve(n_turns=2) from traits.api import Range, on_trait_change
# Populating our plot
self.plot = self.scene.mlab.plot3d(x, y, z) class Visualization(HasTraits):
n_turns = Range(0, 30, 11)
# Describe the dialog scene = Instance(MlabSceneModel, ())
view = View(Item('scene', height=300, show_label=False,
editor=SceneEditor()), def __init__(self):
HGroup('n_turns'), resizable=True) HasTraits.__init__(self)
x, y, z = curve(self.n_turns)
# Fire up the dialog self.plot = self.scene.mlab.plot3d(x, y, z)
Visualization().configure_traits()
@on_trait_change('n_turns')
def update_plot(self):
Tip: Let us read a bit the code above (examples/mlab_dialog.py). x, y, z = curve(self.n_turns)
First, the curve function is used to compute the coordinate of the curve we want to plot. self.plot.mlab_source.set(x=x, y=y, z=z)
Second, the dialog is defined by an object inheriting from HasTraits, as it is done with Traits (page 338). The view = View(Item('scene', height=300, show_label=False,
important point here is that a Mayavi scene is added as a specific Traits attribute (Instance). This is important
editor=SceneEditor()), Chapters contents

HGroup('n_turns'), resizable=True)
Loading an example dataset (page 370)
# Fire up the dialog Learning and Predicting (page 371)
Visualization().configure_traits() Classification (page 371)
k-Nearest neighbors classifier (page 371)
Support vector machines (SVMs) for classification (page 372)
Tip: Full code of the example: examples/mlab_dialog.py. Clustering: grouping observations together (page 373)
K-means clustering (page 374)
Dimension Reduction with Principal Component Analysis (page 375)
3.5.6 Putting it together Putting it all together: face recognition (page 376)
Linear model: from regression to sparsity (page 377)
Sparse models (page 377)
Exercise Model selection: choosing estimators and their parameters (page 378)
Grid-search and cross-validated estimators (page 378)
Using the code from the magnetic field simulation, create a dialog that enable to move the 2 coils: change
their parameters.
Hint: to define a dialog entry for a vector of dimension 3
direction = Array(float, value=(0, 0, 1), cols=3, shape=(3,))
3.6.1 Loading an example dataset
You can look at the example_coil_application.py to see a full-blown application for coil design in 270 lines of
code.
3.6 scikit-learn: machine learning in Python
Authors: Fabian Pedregosa, Gael Varoquaux
First we will load some data to play with. The data we will use is a very simple flower database known as the Iris
dataset.
Prerequisites
We have 150 observations of the iris flower specifying some measurements: sepal length, sepal width, petal length
Numpy, Scipy and petal width together with its subtype: Iris setosa, Iris versicolor, Iris virginica.
IPython
To load the dataset into a Python object:
matplotlib
scikit-learn (http://scikit-learn.org) >>> from sklearn import datasets
>>> iris = datasets.load_iris()
This data is stored in the .data member, which is a (n_samples, n_features) array.
>>> iris.data.shape
(150, 4)
The class of each observation is stored in the .target attribute of the dataset. This is an integer 1D array of
length n_samples:
>>> iris.target.shape
(150,)
>>> np.unique(iris.target)
array([0, 1, 2])
3.6. scikit-learn: machine learning in Python 369 3.6. scikit-learn: machine learning in Python 370
An example of reshaping data: the digits dataset
The digits dataset consists of 1797 images, where each one is an 8x8 pixel image representing a hand-written
digit
>>> digits = datasets.load_digits()
>>> digits.images.shape
(1797, 8, 8)
>>> pl.imshow(digits.images[0], cmap=pl.cm.gray_r) The k-nearest neighbors classifier internally uses an algorithm based on ball trees to represent the samples it is
<matplotlib.image.AxesImage object at ...> trained on.
To use this dataset with the scikit, we transform each 8x8 image into a vector of length 64 KNN (k-nearest neighbors) classification example:
>>> data = digits.images.reshape((digits.images.shape[0], -1)) >>> # Create and fit a nearest-neighbor classifier
>>> from sklearn import neighbors
>>> knn = neighbors.KNeighborsClassifier()
>>> knn.fit(iris.data, iris.target)
Learning and Predicting KNeighborsClassifier(...)
>>> knn.predict([[0.1, 0.2, 0.3, 0.4]])
array([0])
Now that weve got some data, we would like to learn from it and predict on new one. In scikit-learn, we
learn from existing data by creating an estimator and calling its fit(X, Y) method.
>>> from sklearn import svm Training set and testing set
>>> clf = svm.LinearSVC()
>>> clf.fit(iris.data, iris.target) # learn from the data When experimenting with learning algorithms, it is important not to test the prediction of an estimator on
LinearSVC(...) the data used to fit the estimator. Indeed, with the kNN estimator, we would always get perfect prediction on
the training set.
Once we have learned from the data, we can use our model to predict the most likely outcome on unseen data:
>>> perm = np.random.permutation(iris.target.size)
>>> clf.predict([[ 5.0, 3.6, 1.3, 0.25]]) >>> iris.data = iris.data[perm]
array([0]) >>> iris.target = iris.target[perm]
>>> knn.fit(iris.data[:100], iris.target[:100])
KNeighborsClassifier(...)
Note: We can access the parameters of the model via its attributes ending with an underscore: >>> knn.score(iris.data[100:], iris.target[100:])
>>> clf.coef_ 0.95999...
array([[ 0...]]) Bonus question: why did we use a random permutation?
3.6.2 Classification Support vector machines (SVMs) for classification
k-Nearest neighbors classifier Linear Support Vector Machines
The simplest possible classifier is the nearest neighbor: given a new observation, take the label of the training SVMs try to construct a hyperplane maximizing the margin between the two classes. It selects a subset of the
samples closest to it in n-dimensional space, where n is the number of features in each sample. input, called the support vectors, which are the observations closest to the separating hyperplane.
K-means clustering
The simplest clustering algorithm is k-means. This divides a set into k clusters, assigning each observation to a
cluster so as to minimize the distance of that observation (in n-dimensional space) to the clusters mean; the means
are then recomputed. This operation is run iteratively until the clusters converge, for a maximum for max_iter
rounds.
(An alternative implementation of k-means is available in SciPys cluster package. The scikit-learn
implementation differs from that by offering an object API and several additional features, including smart initial-
ization.)
>>> from sklearn import cluster, datasets
>>> iris = datasets.load_iris()
>>> k_means = cluster.KMeans(n_clusters=3)
>>> k_means.fit(iris.data)
KMeans(...)
>>> print(k_means.labels_[::10])
>>> from sklearn import svm
[1 1 1 1 1 0 0 0 0 0 2 2 2 2 2]
>>> svc = svm.SVC(kernel='linear')
>>> print(iris.target[::10])
>>> svc.fit(iris.data, iris.target)
[0 0 0 0 0 1 1 1 1 1 2 2 2 2 2]
SVC(...)
There are several support vector machine implementations in scikit-learn. The most commonly used ones
are svm.SVC, svm.NuSVC and svm.LinearSVC; SVC stands for Support Vector Classifier (there also exist
SVMs for regression, which are called SVR in scikit-learn).
Excercise
Train an svm.SVC on the digits dataset. Leave out the last 10% and test prediction performance on these
observations. Ground truth K-means (3 clusters) K-means (8 clusters)
Application to Image Compression
Using kernels Clustering can be seen as a way of choosing a small number of observations from the information. For
instance, this can be used to posterize an image (conversion of a continuous gradation of tone to several
regions of fewer tones):
Classes are not always separable by a hyperplane, so it would be desirable to have a decision function that is not
linear but that may be for instance polynomial or exponential: >>> from scipy import misc
>>> lena = misc.lena().astype(np.float32)
Linear kernel Polynomial kernel RBF kernel (Radial Basis Func- >>> X = lena.reshape((-1, 1)) # We need an (n_sample, n_feature) array
tion) >>> k_means = cluster.KMeans(n_clusters=5)
>>> k_means.fit(X)
KMeans(...)
>>> values = k_means.cluster_centers_.squeeze()
>>> labels = k_means.labels_
>>> lena_compressed = np.choose(labels, values)
>>> lena_compressed.shape = lena.shape
>>> svc = svm.SVC(kernel='linear')

>>> svc = svm.SVC(kernel='poly',
>>> svc = svm.SVC(kernel='rbf')
... >>> # gamma: inverse of size of
degree=3)
>>> # degree: polynomial degree
>>> # radial kernel
Exercise
Which of the kernels noted above has a better prediction performance on the digits dataset?
Raw image K-means quantization
3.6.3 Clustering: grouping observations together
Given the iris dataset, if we knew that there were 3 types of iris, but did not have access to their labels, we could
try unsupervised learning: we could cluster the observations into several groups by some criterion.
3.6.4 Dimension Reduction with Principal Component Analysis 3.6.5 Putting it all together: face recognition
An example showcasing face recognition using Principal Component Analysis for dimension reduction and Sup-
port Vector Machines for classification.
The cloud of points spanned by the observations above is very flat in one direction, so that one feature can almost
be exactly computed using the 2 other. PCA finds the directions in which the data is not flat and it can reduce the
dimensionality of the data by projecting on a subspace.
Warning: Depending on your version of scikit-learn PCA will be in module decomposition or pca.
>>> from sklearn import decomposition

>>> pca = decomposition.PCA(n_components=2)
>>> pca.fit(iris.data)
PCA(copy=True, n_components=2, whiten=False)
>>> X = pca.transform(iris.data) """
Stripped-down version of the face recognition example by Olivier Grisel
Now we can visualize the (transformed) iris dataset:
http://scikit-learn.org/dev/auto_examples/applications/face_recognition.html
>>> pl.scatter(X[:, 0], X[:, 1], c=iris.target) ## original shape of images: 50, 37
<matplotlib.collections...Collection object at ...> """
import numpy as np
import pylab as pl
from sklearn import cross_val, datasets, decomposition, svm
# ..
# .. load data ..
lfw_people = datasets.fetch_lfw_people(min_faces_per_person=70, resize=0.4)
perm = np.random.permutation(lfw_people.target.size)
lfw_people.data = lfw_people.data[perm]
lfw_people.target = lfw_people.target[perm]
faces = np.reshape(lfw_people.data, (lfw_people.target.shape[0], -1))
train, test = iter(cross_val.StratifiedKFold(lfw_people.target, k=4)).next()
X_train, X_test = faces[train], faces[test]
y_train, y_test = lfw_people.target[train], lfw_people.target[test]
# ..
# .. dimension reduction ..
PCA is not just useful for visualization of high dimensional datasets. It can also be used as a preprocessing step pca = decomposition.RandomizedPCA(n_components=150, whiten=True)
to help speed up supervised methods that are not efficient with high dimensions. pca.fit(X_train)
X_train_pca = pca.transform(X_train)
X_test_pca = pca.transform(X_test)
Different algorithms for a same problem

# ..
# .. classification .. Different algorithms can be used to solve the same mathematical problem. For instance the Lasso object in
clf = svm.SVC(C=5., gamma=0.001) the sklearn solves the lasso regression using a coordinate descent method, that is efficient on large datasets.
clf.fit(X_train_pca, y_train) However, the sklearn also provides the LassoLARS object, using the LARS which is very efficient for prob-
lems in which the weight vector estimated is very sparse, that is problems with very few observations.
# ..
# .. predict on new images ..
for i in range(10):
print(lfw_people.target_names[clf.predict(X_test_pca[i])[0]])
_ = pl.imshow(X_test[i].reshape(50, 37), cmap=pl.cm.gray)
3.6.7 Model selection: choosing estimators and their parameters
_ = raw_input()
Grid-search and cross-validated estimators
Grid-search
3.6.6 Linear model: from regression to sparsity
The scikit-learn provides an object that, given data, computes the score during the fit of an estimator on a parameter
Diabetes dataset grid and chooses the parameters to maximize the cross-validation score. This object takes an estimator during the
construction and exposes an estimator API:
The diabetes dataset consists of 10 physiological variables (age, sex, weight, blood pressure) measure on
442 patients, and an indication of disease progression after one year: >>> from sklearn import svm, grid_search
>>> gammas = np.logspace(-6, -1, 10)
>>> diabetes = datasets.load_diabetes() >>> svc = svm.SVC()
>>> diabetes_X_train = diabetes.data[:-20] >>> clf = grid_search.GridSearchCV(estimator=svc, param_grid=dict(gamma=gammas),
>>> diabetes_X_test = diabetes.data[-20:] ... n_jobs=-1)
>>> diabetes_y_train = diabetes.target[:-20] >>> clf.fit(digits.data[:1000], digits.target[:1000])
>>> diabetes_y_test = diabetes.target[-20:] GridSearchCV(cv=None,...)
The task at hand is to predict disease prediction from physiological variables. >>> clf.best_score_
0.9...
>>> clf.best_estimator_.gamma
0.00059948425031894088
Sparse models
By default the GridSearchCV uses a 3-fold cross-validation. However, if it detects that a classifier is passed, rather
than a regressor, it uses a stratified 3-fold.
To improve the conditioning of the problem (uninformative variables, mitigate the curse of dimensionality, as a
feature selection preprocessing, etc.), it would be interesting to select only the informative features and set non-
informative ones to 0. This penalization approach, called Lasso, can set some coefficients to zero. Such methods Cross-validated estimators
are called sparse method, and sparsity can be seen as an application of Occams razor: prefer simpler models to
complex ones.
Cross-validation to set a parameter can be done more efficiently on an algorithm-by-algorithm basis. This is
>>> from sklearn import linear_model why, for certain estimators, the scikit-learn exposes CV estimators, that set their parameter automatically by
>>> regr = linear_model.Lasso(alpha=.3) cross-validation:
>>> regr.fit(diabetes_X_train, diabetes_y_train)
Lasso(...) >>> from sklearn import linear_model, datasets
>>> regr.coef_ # very sparse coefficients >>> lasso = linear_model.LassoCV()
array([ 0. , -0. , 497.34075682, 199.17441034, >>> diabetes = datasets.load_diabetes()
-0. , -0. , -118.89291545, 0. , >>> X_diabetes = diabetes.data
430.9379595 , 0. ]) >>> y_diabetes = diabetes.target
>>> regr.score(diabetes_X_test, diabetes_y_test) >>> lasso.fit(X_diabetes, y_diabetes)
0.5510835453... LassoCV(alphas=None, ...)
>>> # The estimator chose automatically its lambda:
being the score very similar to linear regression (Least Squares): >>> lasso.alpha_
0.012...
>>> lin = linear_model.LinearRegression()
>>> lin.fit(diabetes_X_train, diabetes_y_train) These estimators are called similarly to their counterparts, with CV appended to their name.
LinearRegression(...)
>>> lin.score(diabetes_X_test, diabetes_y_test)
0.5850753022... Exercise
On the diabetes dataset, find the optimal regularization parameter alpha.
Index
D
diff, 321, 323
differentiation, 321
dsolve, 323
E
equations
algebraic, 322
differential, 323
I
integration, 322
M
Matrix, 323
P
Python Enhancement Proposals
PEP 255, 142
PEP 3118, 178
PEP 3129, 152
PEP 318, 145, 152
PEP 342, 142
PEP 343, 152
PEP 380, 144
PEP 380#id13, 144
PEP 8, 147
S
solve, 322
379

Python Scientific

Uploaded by

Python Scientific

Uploaded by

Contents

Python Scientific lecture notes

3 Packages and applications 271

September 26, 2015 (2013.2-beta-328-g3cd80a5)

One document to learn numerics, science,

CHAPTER 1 Existing solutions

Which solutions do scientists use to work?

3 1.1. Scientific computing with tools and workflow 4

Development tools (automatic testing, documentation generation)

Command line interaction

Getting help by using the ? operator after an object:

Prints the values to a stream, or to sys.stdout by default.

Elaboration of the algorithm in an editor

1.2 The Python language >>> print("Hello, world!")

Tip: Python supports the following numerical, scalar types:

1.2. The Python language 9 1.2. The Python language 10

1.5 Warning: Integer division

>>> 3 > 4 >>> 3 / 2

Note: The behaviour of the division operator has changed in Python 3.

>>> l = ['red', 'blue', 'green', 'black', 'white']

Indexing: accessing individual objects contained in the list:

Counting from the end with negative indices:

1.2. The Python language 11 1.2. The Python language 12

>>> l[-1] >>> L = ['red', 'blue', 'green', 'black', 'white']

Add and remove elements:

1.2. The Python language 13 1.2. The Python language 14

Discovering methods: >>> a[2:10:2] # Syntax: a[start:stop:step]

1.2. The Python language 15 1.2. The Python language 16

Sets: unordered, unique items:

Things to note: >>> if a == 1:

1.2. The Python language 17 1.2. The Python language 18

>>> for i in range(4): >>> 1 == 1.

>>> while abs(z) < 100: >>> for i in 'powerful':

1.2. The Python language 19 1.2. The Python language 20

Mandatory parameters (positional arguments)

Exercise In [83]: double_it()

Function definition In [85]: double_it()

1.2. The Python language 21 1.2. The Python language 22

In [103]: slicer(rhyme) Global variables

1.2. The Python language 23 1.2. The Python language 24

This doesnt work: String Form: <function funcname at 0xeaa0f0>

In [118]: setx(10) Extended summary which can contain multiple paragraphs.

In [123]: x assigned to a variable

Special forms of parameters: In [39]: va('three', x=1, y=2)

In [36]: variable_args('one', 'two', x=1, y=2, z=3)

Exercise: Fibonacci sequence

1.2. The Python language 25 1.2. The Python language 26

Scripts In [1]: import os

In [1]: %run test.py In [4]: from os import listdir

In [2]: message Warning:

1.2. The Python language 27 1.2. The Python language 28

Creating modules '__file__',

In [2]: demo.print_a() In [11]: print_a()

In [3]: demo.print_b() Warning: Module caching

Introspection In [10]: reload(demo)

In [7]: dir(demo) # print_b() runs on import

1.2. The Python language 29 1.2. The Python language 30

Scripts or modules? How to organize your code Packages

1.2. The Python language 31 1.2. The Python language 32

Base Class: <type 'function'>

Use meaningful object names To read from a file

1.2. The Python language 33 1.2. The Python language 34

Binary mode: b In [44]: fp = open('junk.txt', 'w')

In [46]: 'junk.txt' in os.listdir(os.curdir)

Note: Reference document for this section: In [47]: os.remove('junk.txt')

os.path: path manipulations

Directory and file manipulation In [71]: fp.close()

Current directory: In [72]: a = os.path.abspath('junk.txt')

In [17]: os.getcwd() In [73]: a

List a directory: In [74]: os.path.split(a)

In [37]: 'junkdir' in os.listdir(os.curdir) In [92]: os.path.join(os.path.expanduser('~'), 'local', 'bin')

In [38]: 'foodir' in os.listdir(os.curdir)

1.2. The Python language 35 1.2. The Python language 36

In [21]: print com

glob: Pattern matching on files

1.2. The Python language 37 1.2. The Python language 38

Creating modules 'file',