PythonScientific Simple PDF

Python Scientific lecture notes
Release 2013.1
EuroScipy tutorial team

Editors: Valentin Haenel, Emmanuelle Gouillart, Gal Varoquaux
http://scipy-lectures.github.com
February 10, 2013 (2013.1-2-g9a33667)

Contents
I Getting started with Python for science 2

1 Scientific computing with tools and workflow 3
1.1 Why Python? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Scientific Python building blocks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3 The interactive workflow: IPython and a text editor . . . . . . . . . . . . . . . . . . . . . . . . . 6
2 The Python language 9

2.1 First steps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.2 Basic types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.3 Control Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.4 Defining functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.5 Reusing code: scripts and modules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.6 Input and Output . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.7 Standard Library . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
2.8 Exception handling in Python . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.9 Object-oriented programming (OOP) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3 NumPy: creating and manipulating numerical data 42

3.1 The numpy array object . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
3.2 Numerical operations on arrays . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.3 More elaborate arrays . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
3.4 Advanced operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
4 Matplotlib: plotting 81
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
4.2 Simple plot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
4.3 Figures, Subplots, Axes and Ticks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
4.4 Other Types of Plots: examples and exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
4.5 Beyond this tutorial . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
4.6 Quick references . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
5 Scipy : high-level scientific computing 104

5.1 File input/output: scipy.io . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
5.2 Special functions: scipy.special . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
5.3 Linear algebra operations: scipy.linalg . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
5.4 Fast Fourier transforms: scipy.fftpack . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
5.5 Optimization and fit: scipy.optimize . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
5.6 Statistics and random numbers: scipy.stats . . . . . . . . . . . . . . . . . . . . . . . . . . 115
i
5.7 Interpolation: scipy.interpolate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
5.8 Numerical integration: scipy.integrate . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
5.9 Signal processing: scipy.signal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
5.10 Image processing: scipy.ndimage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
5.11 Summary exercises on scientific computing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
6 Getting help and finding documentation 139
II Advanced topics 143

7 Advanced Python Constructs 144
7.1 Iterators, generator expressions and generators . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
7.2 Decorators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
7.3 Context managers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
8 Advanced Numpy 160

8.1 Life of ndarray . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
8.2 Universal functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
8.3 Interoperability features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
8.4 Array siblings: chararray, maskedarray, matrix . . . . . . . . . . . . . . . . . . . . . . 186
8.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
8.6 Contributing to Numpy/Scipy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
9 Debugging code 193

9.1 Avoiding bugs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
9.2 Debugging workflow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
9.3 Using the Python debugger . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
9.4 Debugging segmentation faults using gdb . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
10 Optimizing code 204

10.1 Optimization workflow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
10.2 Profiling Python code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205
10.3 Making code go faster . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
10.4 Writing faster numerical code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
11 Sparse Matrices in SciPy 212

11.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212
11.2 Storage Schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
11.3 Linear System Solvers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 226
11.4 Other Interesting Packages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 230
12 Image manipulation and processing using Numpy and Scipy 232

12.1 Opening and writing to image files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
12.2 Displaying images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234
12.3 Basic manipulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 236
12.4 Image filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 238
12.5 Feature extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243
12.6 Measuring objects properties: ndimage.measurements . . . . . . . . . . . . . . . . . . . . 246
13 Mathematical optimization: finding minima of functions 252

13.1 Knowing your problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253
13.2 A review of the different optimizers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255
13.3 Practical guide to optimization with scipy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262
13.4 Special case: non-linear least-squares . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 264
13.5 Optimization with constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 266
14 Traits 268
14.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269
14.2 Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269
ii
14.3 What are Traits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 270
14.4 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 285
15 3D plotting with Mayavi 287

15.1 Mlab: the scripting interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 287
15.2 Interactive work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 293
16 Sympy : Symbolic Mathematics in Python 294

16.1 First Steps with SymPy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 295
16.2 Algebraic manipulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 296
16.3 Calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 297
16.4 Equation solving . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 298
16.5 Linear Algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 299
17 scikit-learn: machine learning in Python 301

17.1 Loading an example dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 302
17.2 Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 303
17.3 Clustering: grouping observations together . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 306
17.4 Dimension Reduction with Principal Component Analysis . . . . . . . . . . . . . . . . . . . . . 307
17.5 Putting it all together: face recognition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 308
17.6 Linear model: from regression to sparsity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 310
17.7 Model selection: choosing estimators and their parameters . . . . . . . . . . . . . . . . . . . . . 311
18 Interfacing with C 312

18.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 312
18.2 Python-C-Api . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 313
18.3 Ctypes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 317
18.4 SWIG . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 320
18.5 Cython . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 324
18.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 328
18.7 Further Reading and References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 328
18.8 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 329
Index 331
iii
Python Scientific lecture notes, Release 2013.1
Contents 1
Part I
Getting started with Python for science
2
CHAPTER 1
Scientific computing with tools and

workflow
authors Fernando Perez, Emmanuelle Gouillart, Gal Varoquaux, Valentin Haenel
1.1 Why Python?
1.1.1 The scientists needs
Get data (simulation, experiment control)

Manipulate and process data.
Visualize results... to understand what we are doing!
Communicate results: produce figures for reports or publications, write presentations.
1.1.2 Specifications
Rich collection of already existing bricks corresponding to classical numerical methods or basic actions: we
dont want to re-program the plotting of a curve, a Fourier transform or a fitting algorithm. Dont reinvent
the wheel!
Easy to learn: computer science is neither our job nor our education. We want to be able to draw a curve,
smooth a signal, do a Fourier transform in a few minutes.
Easy communication with collaborators, students, customers, to make the code live within a lab or a com-
pany: the code should be as readable as a book. Thus, the language should contain as few syntax symbols or
unneeded routines as possible that would divert the reader from the mathematical or scientific understanding
of the code.
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.
A single environment/language for everything, if possible, to avoid learning a new software for each new
problem.
1.1.3 Existing solutions
Which solutions do scientists use to work?
3
Compiled languages: C, C++, Fortran, etc.

Advantages:
Very fast. Very optimized compilers. For heavy computations, its difficult to outperform these lan-
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-
puter scientists.
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.
Pleasant development environment: comprehensive and well organized help, integrated editor, etc.
Commercial support is available.
Drawbacks:
Base language is quite poor and can become restrictive for advanced users.
Not free.
Other scripting languages: Scilab, Octave, Igor, R, IDL, etc.
Advantages:
Open-source, free, or at least cheaper than Matlab.
Some features can be very advanced (statistics in R, figures in Igor, etc.)
Drawbacks:
Fewer available algorithms than in Matlab, and the language is not more advanced.
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.
What about Python?
Advantages:
Very rich scientific computing libraries (a bit less than Matlab, though)
Well thought out language, allowing to write very readable and well structured code: we code what
we think.
Many libraries for other tasks than scientific computing (web server management, serial port access,
etc.)
Free and open-source software, widely spread, with a vibrant community.
Drawbacks:
less pleasant development environment than, for example, Matlab. (More geek-oriented).
Not all the algorithms that can be found in more specialized software or toolboxes.
1.1. Why Python? 4

1.2 Scientific Python building blocks
Unlike Matlab, Scilab or R, Python does not come with a pre-bundled set of modules for scientific computing.
Below are the basic building blocks that can be combined to obtain a scientific computing environment:
Python, a generic and modern computing language
Python language: data types (string, int), flow control, data collections (lists, dictionaries), pat-
terns, etc.
Modules of the standard library.
A large number of specialized modules or applications written in Python: web protocols, web frame-
work, etc. ... and scientific computing.
Development tools (automatic testing, documentation generation)
IPython, an advanced Python shell http://ipython.scipy.org/moin/

Numpy : provides powerful numerical arrays objects, and routines to manipulate them.
http://www.numpy.org/
Scipy : high-level data processing routines. Optimization, regression, interpolation, etc
http://www.scipy.org/
Matplotlib : 2-D visualization, publication-ready plots http://matplotlib.sourceforge.net/
Mayavi : 3-D visualization http://code.enthought.com/projects/mayavi/
1.2. Scientific Python building blocks 5

1.3 The interactive workflow: IPython and a text editor
Interactive work to test and understand algorithms: In this section, we describe an interactive workflow with
IPython that is handy to explore and understand algorithms.
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
used to write programs, in web servers, or embedded devices.
Note: Reference document for this section:

IPython user manual: http://ipython.org/ipython-doc/dev/index.html
1.3.1 Command line interaction
Start ipython:
In [1]: print(Hello world)
Hello world
Getting help by using the ? operator after an object:
In [2]: print?
Type: builtin_function_or_method
Base Class: <type builtin_function_or_method>
String Form: <built-in function print>
Namespace: Python builtin
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.
1.3.2 Elaboration of the algorithm in an editor
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
1.3. The interactive workflow: IPython and a text editor 6

From a script to functions

While it is tempting to work only with scripts, that is a file full of instructions following each other, do plan
to progressively evolve the script to a set of functions:
A script is not reusable, functions are.
Thinking in terms of functions helps breaking the problem in small blocks.
1.3.3 IPython Tips and Tricks
The IPython user manual contains a wealth of information about using IPython, but to get you started we want to
give you a quick introduction to three useful features: history, magic functions, aliases and tab completion.
Like a UNIX shell, IPython supports command history. Type up and down to navigate previously typed commands:
In [1]: x = 10
In [2]: <UP>
In [2]: x = 10
IPython supports so called magic functions by prefixing a command with the % character. For example, the run
and whos functions from the previous section are magic functions. Note that, the setting automagic, which is
enabled by default, allows you to omit the preceding % sign. Thus, you can just type the magic function and it will
work.
Other useful magic functions are:
%cd to change the current directory.
In [2]: cd /tmp
/tmp
%timeit allows you to time the execution of short snippets using the timeit module from the standard
library:
In [3]: timeit x = 10
10000000 loops, best of 3: 39 ns per loop
%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]):
In [5]: cpaste
Pasting code; enter -- alone on the line to stop or use Ctrl-D.
:In [3]: timeit x = 10
:--
10000000 loops, best of 3: 85.9 ns per loop
In [6]: cpaste
Pasting code; enter -- alone on the line to stop or use Ctrl-D.
:>>> timeit x = 10
:--
10000000 loops, best of 3: 86 ns per loop
%debug allows you to enter post-mortem debugging. That is to say, if the code you try to execute, raises
an exception, using %debug will enter the debugger at the point where the exception was thrown.
In [7]: x === 10
File "<ipython-input-6-12fd421b5f28>", line 1
x === 10
^
SyntaxError: invalid syntax

In [8]: debug
> /home/esc/anaconda/lib/python2.7/site-packages/IPython/core/compilerop.py(87)ast_parse()
86 and are passed to the built-in compile function."""
---> 87 return compile(source, filename, symbol, self.flags | PyCF_ONLY_AST, 1)
88
ipdb>locals()
{source: ux === 10\n, symbol: exec, self:
<IPython.core.compilerop.CachingCompiler instance at 0x2ad8ef0>,
filename: <ipython-input-6-12fd421b5f28>}
Note: The built-in IPython cheat-sheet is accessible via the %quickref magic function.
Note: A list of all available magic functions is shown when typing %magic.
Furthermore IPython ships with various aliases which emulate common UNIX command line tools such as ls to
list files, cp to copy files and rm to remove files. A list of aliases is shown when typing alias:
In [1]: alias
Total number of aliases: 16
Out[1]:
[(cat, cat),
(clear, clear),
(cp, cp -i),
(ldir, ls -F -o --color %l | grep /$),
(less, less),
(lf, ls -F -o --color %l | grep ^-),
(lk, ls -F -o --color %l | grep ^l),
(ll, ls -F -o --color),
(ls, ls -F --color),
(lx, ls -F -o --color %l | grep ^-..x),
(man, man),
(mkdir, mkdir),
(more, more),
(mv, mv -i),
(rm, rm -i),
(rmdir, rmdir)]
Lastly, we would like to mention the tab completion feature, whose description we cite directly from the IPython
manual:
Tab completion, especially for attributes, is a convenient way to explore the structure of any object youre dealing
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.
In [1]: x = 10
In [2]: x.<TAB>
x.bit_length x.conjugate x.denominator x.imag x.numerator
x.real
In [3]: x.real.
x.real.bit_length x.real.denominator x.real.numerator
x.real.conjugate x.real.imag x.real.real
In [4]: x.real.

CHAPTER 2
The Python language
authors Chris Burns, Christophe Combelles, Emmanuelle Gouillart, Gal Varoquaux
Python for scientific computing

We introduce here the Python language. Only the bare minimum necessary for getting started with Numpy
and Scipy is addressed here. To learn more about the language, consider going through the excellent tutorial
http://docs.python.org/tutorial. Dedicated books are also available, such as http://diveintopython.org/.
Python is a programming language, as are C, Fortran, BASIC, PHP, etc. Some specific features of Python are as
follows:
an interpreted (as opposed to compiled) language. Contrary to e.g. C or Fortran, one does not compile
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.
a free software released under an open-source license: Python can be used and distributed free of charge,
even for building commercial software.
multi-platform: Python is available for all major operating systems, Windows, Linux/Unix, MacOS X,
most likely your mobile phone OS, etc.
a very readable language with clear non-verbose syntax
a language for which a large variety of high-quality packages are available for various applications, from
web frameworks to scientific computing.
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
of a program).
See http://www.python.org/about/ for more information about distinguishing features of Python.
9
2.1 First steps
Start the Ipython shell (an enhanced interactive Python shell):

by typing ipython from a Linux/Mac terminal, or from the Windows cmd shell,
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.
If you dont have Ipython installed on your computer, other Python shells are available, such as the plain Python
shell started by typing python in a terminal, or the Idle interpreter. However, we advise to use the Ipython shell
because of its enhanced features, especially for interactive scientific computing.
Once you have started the interpreter, type
>>> print "Hello, world!"
Hello, world!
The message Hello, world! is then displayed. You just executed your first Python instruction, congratulations!
To get yourself started, type the following stack of instructions
>>> a = 3
>>> b = 2*a
>>> type(b)
<type int>
>>> print b
6
>>> a*b
18
>>> b = hello
>>> type(b)
<type str>
>>> b + b
hellohello
>>> 2*b
hellohello
Two variables a and b have been defined above. Note that one does not declare the type of an variable before
assigning its value. In C, conversely, one should write:
int a = 3;
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
integer, but it became equal to a string when it was assigned the value hello. Operations on integers (b=2*a)
are coded natively in Python, and so are some operations on strings such as additions and multiplications, which
amount respectively to concatenation and repetition.
2.2 Basic types
2.2.1 Numerical types
Python supports the following numerical, scalar types:

Integer
>>> 1 + 1
2
>>> a = 4
>>> type(a)
<type int>
2.1. First steps 10

Floats
>>> c = 2.1
>>> type(c)
<type float>
Complex
>>> a = 1.5 + 0.5j
>>> a.real
1.5
>>> a.imag
0.5
>>> type(1. + 0j )
<type complex>
Booleans
>>> 3 > 4
False
>>> test = (3 > 4)
>>> test
False
>>> type(test)
<type bool>
A Python shell can therefore replace your pocket calculator, with the basic arithmetic operations +, -, *, /, %
(modulo) natively implemented:
>>> 7 * 3.
21.0
>>> 2**10
1024
>>> 8 % 3
2
Type conversion (casting):

>>> float(1)
1.0
2.2. Basic types 11

Warning: Integer division
>>> 3 / 2
1
Trick: use floats:
>>> 3 / 2.
1.5
>>> a = 3
>>> b = 2
>>> a / b
1
>>> a / float(b)
1.5
If you explicitly want integer division use //
>>> 3.0 // 2
1.0
Note: The behaviour of the division operator has changed in Python 3. Please look at the python3porting
website for details.
2.2.2 Containers
Python provides many efficient types of containers, in which collections of objects can be stored.
Lists
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]
white
>>> L[-2]
black
Warning: Indexing starts at 0 (as in C), not at 1 (as in Fortran or Matlab)!
Slicing: obtaining sublists of regularly-spaced elements:

>>> L
[red, blue, green, black, white]
>>> L[2:4]
[green, black]
2.2. Basic types 12

Warning: Note that L[start:stop] contains the elements with indices i such as start<= i < stop
(i ranging from start to stop-1). Therefore, L[start:stop] has (stop-start) elements.
Slicing syntax: L[start:stop:stride]

All slicing parameters are optional:
>>> L
>>> L[3:]
[black, white]
>>> L[:3]
[red, blue, green]
>>> L[::2]
[red, green, white]
Lists are mutable objects and can be modified:

>>> L[0] = yellow
>>> L
[yellow, blue, green, black, white]
>>> L[2:4] = [gray, purple]
>>> L
[yellow, blue, gray, purple, white]
Note: The elements of a list may have different types:

>>> L = [3, -200, hello]
>>> L
[3, -200, hello]
>>> L[1], L[2]
(-200, hello)
For collections of numerical data that all have the same type, it is often more efficient to use the array type
provided by the numpy module. A NumPy array is a chunk of memory containing fixed-sized items. With
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.
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:
>>> L = [red, blue, green, black, white]
>>> L.append(pink)
>>> L
[red, blue, green, black, white, pink]
>>> L.pop() # removes and returns the last item
pink
>>> L
>>> L.extend([pink, purple]) # extend L, in-place
>>> L
[red, blue, green, black, white, pink, purple]
>>> L = L[:-2]
>>> L
Reverse:
>>> r = L[::-1]
>>> r
[white, black, green, blue, red]
2.2. Basic types 13

>>> r2 = list(L)
>>> r2
>>> r2.reverse() # in-place
>>> r2
Concatenate and repeat lists:

>>> r + L
[white, black, green, blue, red, red, blue, green, black, white]
>>> r * 2
[white, black, green, blue, red, white, black, green, blue, red]
Sort:
>>> sorted(r) # new object
[black, blue, green, red, white]
>>> r
>>> r.sort() # in-place
>>> r
[black, blue, green, red, white]
Note: Methods and Object-Oriented Programming

The notation r.method() (r.sort(), r.append(3), 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 notation ..
No further knowledge of OOP than understanding the notation . is necessary for going through this tutorial.
Note: Discovering methods:

Reminder: in Ipython: tab-completion (press tab)
In [28]: r.<TAB>
r.__add__ r.__iadd__ r.__setattr__
r.__class__ r.__imul__ r.__setitem__
r.__contains__ r.__init__ r.__setslice__
r.__delattr__ r.__iter__ r.__sizeof__
r.__delitem__ r.__le__ r.__str__
r.__delslice__ r.__len__ r.__subclasshook__
r.__doc__ r.__lt__ r.append
r.__eq__ r.__mul__ r.count
r.__format__ r.__ne__ r.extend
r.__ge__ r.__new__ r.index
r.__getattribute__ r.__reduce__ r.insert
r.__getitem__ r.__reduce_ex__ r.pop
r.__getslice__ r.__repr__ r.remove
r.__gt__ r.__reversed__ r.reverse
r.__hash__ r.__rmul__ r.sort
Strings
Different string syntaxes (simple, double or triple quotes):

s = Hello, how are you?
s = "Hi, whats up"
s = Hello, # tripling the quotes allows the
how are you # the string to span more than one line
2.2. Basic types 14

s = """Hi,
whats up?"""
In [1]: Hi, whats up?

------------------------------------------------------------
File "<ipython console>", line 1
Hi, whats up?
^
SyntaxError: invalid syntax
The newline character is \n, and the tab character is \t.

Strings are collections like lists. Hence they can be indexed and sliced, using the same syntax and rules.
Indexing:
>>> a = "hello"
>>> a[0]
h
>>> a[1]
e
>>> a[-1]
o
(Remember that negative indices correspond to counting from the right end.)
Slicing:
>>> a = "hello, world!"
>>> a[3:6] # 3rd to 6th (excluded) elements: elements 3, 4, 5
lo,
>>> a[2:10:2] # Syntax: a[start:stop:step]
lo o
>>> a[::3] # every three characters, from beginning to end
hl r!
Accents and special characters can also be handled in Unicode strings (see
http://docs.python.org/tutorial/introduction.html#unicode-strings).
A string is an immutable object and it is not possible to modify its contents. One may however create new strings
from the original one.
In [53]: a = "hello, world!"
In [54]: a[2] = z
---------------------------------------------------------------------------
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
TypeError: str object does not support item assignment
In [55]: a.replace(l, z, 1)
Out[55]: hezlo, world!
In [56]: a.replace(l, z)
Out[56]: hezzo, worzd!
Strings have many useful methods, such as a.replace as seen above. Remember the a. object-oriented
notation and use tab completion or help(str) to search for new methods.
Note: Python offers advanced possibilities for manipulating strings, looking for patterns or format-
ting. The interested reader is referred to http://docs.python.org/library/stdtypes.html#string-methods and
http://docs.python.org/library/string.html#new-string-formatting
String substitution:
2.2. Basic types 15

>>> An integer: %i; a float: %f; another string: %s % (1, 0.1, string)
An integer: 1; a float: 0.100000; another string: string
>>> i = 102
>>> filename = processing_of_dataset_%d.txt % i
>>> filename
processing_of_dataset_102.txt
Dictionaries
A dictionary is basically an efficient table that maps keys to values. It is an unordered container:
>>> tel = {emmanuelle: 5752, sebastian: 5578}
>>> tel[francis] = 5915
>>> tel
{sebastian: 5578, francis: 5915, emmanuelle: 5752}
>>> tel[sebastian]
5578
>>> tel.keys()
[sebastian, francis, emmanuelle]
>>> tel.values()
[5578, 5915, 5752]
>>> francis in tel
True
It can be used to conveniently store and retrieve values associated with a name (a string for a date, a name, etc.).
See http://docs.python.org/tutorial/datastructures.html#dictionaries for more information.
A dictionary can have keys (resp. values) with different types:
>>> d = {a:1, b:2, 3:hello}
>>> d
{a: 1, 3: hello, b: 2}
More container types
Tuples
Tuples are basically immutable lists. The elements of a tuple are written between parentheses, or just separated
by commas:
>>> t = 12345, 54321, hello!
>>> t[0]
12345
>>> t
(12345, 54321, hello!)
>>> u = (0, 2)
Sets: unordered, unique items:

>>> s = set((a, b, c, a))
>>> s
set([a, c, b])
>>> s.difference((a, b))
set([c])
2.2.3 Assignment operator
Python library reference says:
2.2. Basic types 16

Assignment statements are used to (re)bind names to values and to modify attributes or items of
mutable objects.
In short, it works as follows (simple assignment):
1. an expression on the right hand side is evaluated, the corresponding object is created/obtained
2. a name on the left hand side is assigned, or bound, to the r.h.s. object
Things to note:
a single object can have several names bound to it:
In [1]: a = [1, 2, 3]
In [2]: b = a
In [3]: a
Out[3]: [1, 2, 3]
In [4]: b
Out[4]: [1, 2, 3]
In [5]: a is b
Out[5]: True
In [6]: b[1] = hi!
In [7]: a
Out[7]: [1, hi!, 3]
to change a list in place, use indexing/slices:

In [1]: a = [1, 2, 3]
In [3]: a
Out[3]: [1, 2, 3]
In [4]: a = [a, b, c] # Creates another object.
In [5]: a
Out[5]: [a, b, c]
In [6]: id(a)
Out[6]: 138641676
In [7]: a[:] = [1, 2, 3] # Modifies object in place.
In [8]: a
Out[8]: [1, 2, 3]
In [9]: id(a)
Out[9]: 138641676 # Same as in Out[6], yours will differ...
the key concept here is mutable vs. immutable

mutable objects can be changed in place
immutable objects cannot be modified once created
A very good and detailed explanation of the above issues can be found in David M. Beazleys article Types and
Objects in Python.
2.3 Control Flow
Controls the order in which the code is executed.
2.3.1 if/elif/else
>>> if 2**2 == 4:
... print Obvious!
...
Obvious!
Blocks are delimited by indentation
2.3. Control Flow 17

Type the following lines in your Python interpreter, and be careful to respect the indentation depth. The Ipython
shell automatically increases the indentation depth after a column : sign; to decrease the indentation depth, go
four spaces to the left with the Backspace key. Press the Enter key twice to leave the logical block.
In [1]: a = 10
In [2]: if a == 1:
...: print(1)
...: elif a == 2:
...: print(2)
...: else:
...: print(A lot)
...:
A lot
Indentation is compulsory in scripts as well. As an exercise, re-type the previous lines with the same indentation
in a script condition.py, and execute the script with run condition.py in Ipython.
2.3.2 for/range
Iterating with an index:

>>> for i in range(4):
... print(i)
0
1
2
3
But most often, it is more readable to iterate over values:

>>> for word in (cool, powerful, readable):
... print(Python is %s % word)
Python is cool
Python is powerful
Python is readable
2.3.3 while/break/continue
Typical C-style while loop (Mandelbrot problem):

>>> z = 1 + 1j
>>> while abs(z) < 100:
... z = z**2 + 1
>>> z
(-134+352j)
More advanced features

break out of enclosing for/while loop:
>>> z = 1 + 1j
>>> while abs(z) < 100:

... if z.imag == 0:
... break
... z = z**2 + 1
continue the next iteration of a loop.:

>>> a = [1, 0, 2, 4]
>>> for element in a:

... if element == 0:
... continue
... print 1. / element
1.0
0.5
0.25
2.3.4 Conditional Expressions
if <OBJECT>
Evaluates to False:
any number equal to zero (0, 0.0, 0+0j)
an empty container (list, tuple, set, dictionary, ...)
False, None
Evaluates to True:
everything else
a == b Tests equality, with logics:
>>> 1 == 1.
True
a is b Tests identity: both sides are the same object:

>>> 1 is 1.
False
>>> a = 1
>>> b = 1
>>> a is b
True
a in b For any collection b: b contains a

>>> b = [1, 2, 3]
>>> 2 in b
True
>>> 5 in b
False
If b is a dictionary, this tests that a is a key of b.
2.3.5 Advanced iteration
Iterate over any sequence
You can iterate over any sequence (string, list, keys in a dictionary, lines in a file, ...):
>>> vowels = aeiouy
>>> for i in powerful:

... if i in vowels:
... print(i),
o e u

>>> message = "Hello how are you?"

>>> message.split() # returns a list
[Hello, how, are, you?]
>>> for word in message.split():
... print word
...
Hello
how
are
you?
Few languages (in particular, languages for scientific computing) allow to loop over anything but integers/indices.
With Python it is possible to loop exactly over the objects of interest without bothering with indices you often
dont care about.
Warning: Not safe to modify the sequence you are iterating over.
Keeping track of enumeration number
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:
>>> words = (cool, powerful, readable)
>>> for i in range(0, len(words)):
... print i, words[i]
0 cool
1 powerful
2 readable
But, Python provides enumerate keyword for this:

>>> for index, item in enumerate(words):
... print index, item
0 cool
1 powerful
2 readable
Looping over a dictionary
Use iteritems:
>>> d = {a: 1, b:1.2, c:1j}
>>> for key, val in d.iteritems():

... print(Key: %s has value: %s % (key, val))
Key: a has value: 1
Key: c has value: 1j
Key: b has value: 1.2
2.3.6 List Comprehensions
>>> [i**2 for i in range(4)]

[0, 1, 4, 9]

Exercise
Compute the decimals of Pi using the Wallis formula:

Y 4i2
=2
i=1
4i21
2.4 Defining functions
2.4.1 Function definition
In [56]: def test():

....: print(in test function)
....:
....:
In [57]: test()
in test function
Warning: Function blocks must be indented as other control-flow blocks.
2.4.2 Return statement
Functions can optionally return values.

In [6]: def disk_area(radius):
...: return 3.14 * radius * radius
...:
In [8]: disk_area(1.5)
Out[8]: 7.0649999999999995
Note: By default, functions return None.
Note: Note the syntax to define a function:

the def keyword;
is followed by the functions name, then
the arguments of the function are given between parentheses followed by a colon.
the function body;
and return object for optionally returning values.
2.4.3 Parameters
Mandatory parameters (positional arguments)
2.4. Defining functions 21

In [81]: def double_it(x):

....: return x * 2
....:
In [82]: double_it(3)
Out[82]: 6
In [83]: double_it()
---------------------------------------------------------------------------
TypeError: double_it() takes exactly 1 argument (0 given)
Optional parameters (keyword or named arguments)

In [84]: def double_it(x=2):
....: return x * 2
....:
Out[85]: 4
In [86]: double_it(3)
Out[86]: 6
Keyword arguments allow you to specify default values.
Warning: Default values are evaluated when the function is defined, not when it is called. This can be
problematic when using mutable types (e.g. dictionary or list) and modifying them in the function body, since
the modifications will be persistent across invocations of the function.
In [124]: bigx = 10
In [125]: def double_it(x=bigx):

.....: return x * 2
.....:
In [126]: bigx = 1e9 # Now really big
Out[128]: 20
More involved example implementing pythons slicing:

In [98]: def slicer(seq, start=None, stop=None, step=None):
....: """Implement basic python slicing."""
....: return seq[start:stop:step]
....:
In [101]: rhyme = one fish, two fish, red fish, blue fish.split()
In [102]: rhyme
Out[102]: [one, fish,, two, fish,, red, fish,, blue, fish]
In [103]: slicer(rhyme)
Out[103]: [one, fish,, two, fish,, red, fish,, blue, fish]
In [104]: slicer(rhyme, step=2)

Out[104]: [one, two, red, blue]
In [105]: slicer(rhyme, 1, step=2)

Out[105]: [fish,, fish,, fish,, fish]

In [106]: slicer(rhyme, start=1, stop=4, step=2)

Out[106]: [fish,, fish,]
The order of the keyword arguments does not matter:

In [107]: slicer(rhyme, step=2, start=1, stop=4)
Out[107]: [fish,, fish,]
but it is good practice to use the same ordering as the functions definition.
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.
2.4.4 Passing by value
Can you modify the value of a variable inside a function? Most languages (C, Java, ...) distinguish passing
by value and passing by reference. In Python, such a distinction is somewhat artificial, and it is a bit subtle
whether your variables are going to be modified or not. Fortunately, there exist clear rules.
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.
If the value is immutable, the function does not modify the callers variable. If the value is mutable, the function
may modify the callers variable in-place:
>>> def try_to_modify(x, y, z):
... x = 23
... y.append(42)
... z = [99] # new reference
... print(x)
... print(y)
... print(z)
...
>>> a = 77 # immutable variable
>>> b = [99] # mutable variable
>>> c = [28]
>>> try_to_modify(a, b, c)
23
[99, 42]
[99]
>>> print(a)
77
>>> print(b)
[99, 42]
>>> print(c)
[28]
Functions have a local variable table called a local namespace.

The variable x only exists within the function try_to_modify.
2.4.5 Global variables
Variables declared outside the function can be referenced within the function:
In [114]: x = 5
In [115]: def addx(y):

.....: return x + y
.....:

In [116]: addx(10)
Out[116]: 15
But these global variables cannot be modified within the function, unless declared global in the function.
This doesnt work:
In [117]: def setx(y):
.....: x = y
.....: print(x is %d % x)
.....:
.....:
In [118]: setx(10)
x is 10
In [120]: x
Out[120]: 5
This works:
In [121]: def setx(y):
.....: global x
.....: x = y
.....: print(x is %d % x)
.....:
.....:
In [122]: setx(10)
x is 10
In [123]: x
Out[123]: 10
2.4.6 Variable number of parameters
Special forms of parameters:

*args: any number of positional arguments packed into a tuple
**kwargs: any number of keyword arguments packed into a dictionary
In [35]: def variable_args(*args, **kwargs):
....: print args is, args
....: print kwargs is, kwargs
....:
In [36]: variable_args(one, two, x=1, y=2, z=3)

args is (one, two)
kwargs is {y: 2, x: 1, z: 3}
2.4.7 Docstrings
Documentation about what the function does and its parameters. General convention:
In [67]: def funcname(params):
....: """Concise one-line sentence describing the function.
....:
....: Extended summary which can contain multiple paragraphs.
....: """
....: # function body
....: pass

....:
In [68]: funcname?
Type: function
Base Class: type function>
String Form: <function funcname at 0xeaa0f0>
Namespace: Interactive
File: <ipython console>
Definition: funcname(params)
Docstring:
Concise one-line sentence describing the function.
Extended summary which can contain multiple paragraphs.
Note: Docstring guidelines

For the sake of standardization, the Docstring Conventions webpage documents the semantics and conventions
associated with Python docstrings.
Also, the Numpy and Scipy modules have defined a precise standard for documenting scientific func-
tions, that you may want to follow for your own functions, with a Parameters section, an
Examples section, etc. See http://projects.scipy.org/numpy/wiki/CodingStyleGuidelines#docstring-standard
and http://projects.scipy.org/numpy/browser/trunk/doc/example.py#L37
2.4.8 Functions are objects
Functions are first-class objects, which means they can be:

assigned to a variable
an item in a list (or any collection)
passed as an argument to another function.
In [38]: va = variable_args
In [39]: va(three, x=1, y=2)

args is (three,)
kwargs is {y: 2, x: 1}
2.4.9 Methods
Methods are functions attached to objects. Youve seen these in our examples on lists, dictionaries, strings, etc...
2.4.10 Exercises
Exercise: Fibonacci sequence

Write a function that displays the n first terms of the Fibonacci sequence, defined by:
u_0 = 1; u_1 = 1
u_(n+2) = u_(n+1) + u_n

Exercise: Quicksort
Implement the quicksort algorithm, as defined by wikipedia:
function quicksort(array) var list less, greater if length(array) < 2
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
return concatenate(quicksort(less), pivot, quicksort(greater))
2.5 Reusing code: scripts and modules
For now, we have typed all instructions in the interpreter. For longer sets of instructions we need to change tack
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)).
2.5.1 Scripts
Let us first write a script, that is a file with a sequence of instructions that are executed each time the script is
called.
Instructions may be e.g. copied-and-pasted from the interpreter (but take care to respect indentation rules!). The
extension for Python files is .py. Write or copy-and-paste the following lines in a file called test.py
message = "Hello how are you?"
for word in message.split():
print word
Let us now execute the script interactively, that is inside the Ipython interpreter. This is maybe the most common
use of scripts in scientific computing.
Note: in Ipython, the syntax to execute a script is %run script.py. For example,
In [1]: %run test.py

Hello
how
are
you?
In [2]: message
Out[2]: Hello how are you?
The script has been executed. Moreover the variables defined in the script (such as message) are now available
inside the interpeters namespace.
Other interpreters also offer the possibility to execute scripts (e.g., execfile in the plain Python interpreter,
etc.).
It is also possible In order to execute this script as a standalone program, by executing the script inside a shell
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
Hello
how
are
you?
2.5. Reusing code: scripts and modules 26

Standalone scripts may also take command-line arguments

In file.py:
import sys
print sys.argv
$ python file.py test arguments

[file.py, test, arguments]
Note: Dont implement option parsing yourself. Use modules such as optparse or argparse .
2.5.2 Importing objects from modules
In [1]: import os
In [2]: os
Out[2]: <module os from /usr/lib/python2.6/os.pyc>
In [3]: os.listdir(.)
Out[3]:
[conf.py,
basic_types.rst,
control_flow.rst,
functions.rst,
python_language.rst,
reusing.rst,
file_io.rst,
exceptions.rst,
workflow.rst,
index.rst]
And also:
In [4]: from os import listdir
Importing shorthands:
In [5]: import numpy as np
Warning:
from os import *
This is called the star import and please, Use it with caution
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
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.
Makes the code impossible to statically check for undefined symbols.
Modules are thus a good way to organize code in a hierarchical way. Actually, all the scientific computing tools
we are going to use are modules:
>>> import numpy as np # data arrays
>>> np.linspace(0, 10, 6)
array([ 0., 2., 4., 6., 8., 10.])
>>> import scipy # scientific computing

In Python(x,y), Ipython(x,y) executes the following imports at startup:

>>> import numpy
>>> import numpy as np
>>> from pylab import *
>>> import scipy
and it is not necessary to re-import these modules.
2.5.3 Creating modules
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 modules.
Let us create a module demo contained in the file demo.py:
"A demo module."
def print_b():
"Prints b."
print b
def print_a():
"Prints a."
print a
c = 2
d = 2
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 function
print_a, we are rather going to import it as a module. The syntax is as follows.
In [1]: import demo
In [2]: demo.print_a()
a
In [3]: demo.print_b()
b
Importing the module gives access to its objects, using the module.object syntax. Dont forget to put the
modules name before the objects name, otherwise Python wont recognize the instruction.
Introspection
In [4]: demo?
Type: module
Base Class: <type module>
String Form: <module demo from demo.py>
File: /home/varoquau/Projects/Python_talks/scipy_2009_tutorial/source/demo.py
Docstring:
A demo module.
In [5]: who
demo
In [6]: whos

------------------------------
demo module <module demo from demo.py>
In [7]: dir(demo)
Out[7]:
[__builtins__,
__doc__,
__file__,
__name__,
__package__,
c,
d,
print_a,
print_b]
In [8]: demo.
demo.__builtins__ demo.__init__ demo.__str__
demo.__class__ demo.__name__ demo.__subclasshook__
demo.__delattr__ demo.__new__ demo.c
demo.__dict__ demo.__package__ demo.d
demo.__doc__ demo.__reduce__ demo.print_a
demo.__file__ demo.__reduce_ex__ demo.print_b
demo.__format__ demo.__repr__ demo.py
demo.__getattribute__ demo.__setattr__ demo.pyc
demo.__hash__ demo.__sizeof__
Importing objects from modules into the main namespace

In [9]: from demo import print_a, print_b
In [10]: whos
--------------------------------
demo module <module demo from demo.py>
print_a function <function print_a at 0xb7421534>
print_b function <function print_b at 0xb74214c4>
In [11]: print_a()
a
Warning: Module caching

Modules are cached: if you modify demo.py and re-import it in the old session, you will get the
old one.
Solution:
In [10]: reload(demo)
2.5.4 __main__ and module loading
File demo2.py:
import sys
def print_a():
"Prints a."
print a
print sys.argv

if __name__ == __main__:
print_a()
Importing it:
In [11]: import demo2
b
In [12]: import demo2
Running it:
In [13]: %run demo2
b
a
2.5.5 Scripts or modules? How to organize your code
Note: Rule of thumb

Sets of instructions that are called several times should be written inside functions for better code reusability.
Functions (or other bits of code) that are called from several scripts should be written inside a module,
so that only the module is imported in the different scripts (do not copy-and-paste your functions in the
different scripts!).
Note: How to import a module from a remote directory?

Many solutions exist, depending mainly on your operating system. When the import mymodule statement is
executed, the module mymodule is searched in a given list of directories. This list includes a list of installation-
dependent default path (e.g., /usr/lib/python) as well as the list of directories specified by the environment
variable PYTHONPATH.
The list of directories searched by Python is given by the sys.path variable
In [1]: import sys
In [2]: sys.path
Out[2]:
[,
/usr/bin,
/usr/local/include/enthought.traits-1.1.0,
/usr/lib/python2.6,
/usr/lib/python2.6/plat-linux2,
/usr/lib/python2.6/lib-tk,
/usr/lib/python2.6/lib-old,
/usr/lib/python2.6/lib-dynload,
/usr/lib/python2.6/dist-packages,
/usr/lib/pymodules/python2.6,
/usr/lib/pymodules/python2.6/gtk-2.0,
/usr/lib/python2.6/dist-packages/wx-2.8-gtk2-unicode,
/usr/local/lib/python2.6/dist-packages,
/usr/lib/python2.6/dist-packages,
/usr/lib/pymodules/python2.6/IPython/Extensions,
u/home/gouillar/.ipython]
Modules must be located in the search path, therefore you can:

write your own modules within directories already defined in the search path (e.g.
/usr/local/lib/python2.6/dist-packages). You may use symbolic links (on Linux)
to keep the code somewhere else.

modify the environment variable PYTHONPATH to include the directories containing the user-defined mod-
ules. On Linux/Unix, add the following line to a file read by the shell at startup (e.g. /etc/profile, .profile)
export PYTHONPATH=$PYTHONPATH:/home/emma/user_defined_modules
On Windows, http://support.microsoft.com/kb/310519 explains how to handle environment variables.

or modify the sys.path variable itself within a Python script.
import sys
new_path = /home/emma/user_defined_modules
if new_path not in sys.path:
sys.path.append(new_path)
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 directory.
See http://docs.python.org/tutorial/modules.html for more information about modules.
2.5.6 Packages
A directory that contains many modules is called a package. A package is a module with submodules (which can
have submodules themselves, etc.). A special file called __init__.py (which may be empty) tells Python that
the directory is a Python package, from which modules can be imported.
$ ls
cluster/ io/ README.txt@ stsci/
__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@
__init__.py@ maxentropy/ signal/ version.py@
__init__.pyc misc/ sparse/ version.pyc
INSTALL.txt@ ndimage/ spatial/ weave/
integrate/ odr/ special/
interpolate/ optimize/ stats/
$ cd ndimage
$ ls
doccer.py@ fourier.pyc interpolation.py@ morphology.pyc setup.pyc
doccer.pyc info.py@ interpolation.pyc _nd_image.so
setupscons.py@
filters.py@ info.pyc measurements.py@ _ni_support.py@
setupscons.pyc
filters.pyc __init__.py@ measurements.pyc _ni_support.pyc tests/
fourier.py@ __init__.pyc morphology.py@ setup.py@
From Ipython:
In [1]: import scipy
In [2]: scipy.__file__
Out[2]: /usr/lib/python2.6/dist-packages/scipy/__init__.pyc
In [3]: import scipy.version
In [4]: scipy.version.version
Out[4]: 0.7.0
In [5]: import scipy.ndimage.morphology
In [6]: from scipy.ndimage import morphology

In [17]: morphology.binary_dilation?
Type: function
Base Class: <type function>
String Form: <function binary_dilation at 0x9bedd84>
File: /usr/lib/python2.6/dist-packages/scipy/ndimage/morphology.py
Definition: morphology.binary_dilation(input, structure=None,
iterations=1, mask=None, output=None, border_value=0, origin=0,
brute_force=False)
Docstring:
Multi-dimensional binary dilation with the given structure.
An output array can optionally be provided. The origin parameter

controls the placement of the filter. If no structuring element is
provided an element is generated with a squared connectivity equal
to one. The dilation operation is repeated iterations times. If
iterations is less than 1, the dilation is repeated until the
result does not change anymore. If a mask is given, only those
elements with a true value at the corresponding mask element are
modified at each iteration.
2.5.7 Good practices
Note: Good practices

Indentation: no choice!
Indenting is compulsory in Python! Every command block following a colon bears an additional 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 new block is entered, etc.)
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
------------------------------------------------------------
IndentationError: unexpected indent (test.py, line 2)
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.
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 editor to map the Tab key to a
4-space indentation. In Python(x,y), the editor Scite is already configured this way.
Style guidelines
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
>>> long_line = "Here is a very very long line \
... that we break in two parts."
Spaces
Write well-spaced code: put whitespaces after commas, around arithmetic operators, etc.:
>>> a = 1 # yes
>>> a=1 # too cramped
A certain number of rules for writing beautiful code (and more importantly using the same conventions as
anybody else!) are given in the Style Guide for Python Code.

Use meaningful object names
2.6 Input and Output
To be exhaustive, here are some information about input and output in Python. Since we will use the Numpy
methods to read and write files, you may skip this chapter at first reading.
We write or read strings to/from files (other types must be converted to strings). To write in a file:
>>> f = open(workfile, w) # opens the workfile file
>>> type(f)
<type file>
>>> f.write(This is a test \nand another test)
>>> f.close()
To read from a file

In [1]: f = open(workfile, r)
In [2]: s = f.read()
In [3]: print(s)
This is a test
and another test
In [4]: f.close()
For more details: http://docs.python.org/tutorial/inputoutput.html
2.6.1 Iterating over a file
In [6]: f = open(workfile, r)
In [7]: for line in f:

...: print line
...:
...:
This is a test
and another test
In [8]: f.close()
File modes
Read-only: r
Write-only: w
Note: Create a new file or overwrite existing file.
Append a file: a
Read and Write: r+
Binary mode: b
Note: Use for binary files, especially on Windows.
2.6. Input and Output 33

2.7 Standard Library
Note: Reference document for this section:

The Python Standard Library documentation: http://docs.python.org/library/index.html
Python Essential Reference, David Beazley, Addison-Wesley Professional
2.7.1 os module: operating system functionality
A portable way of using operating system dependent functionality.
Directory and file manipulation
Current directory:
In [17]: os.getcwd()
Out[17]: /Users/cburns/src/scipy2009/scipy_2009_tutorial/source
List a directory:
In [31]: os.listdir(os.curdir)
Out[31]:
[.index.rst.swo,
.python_language.rst.swp,
.view_array.py.swp,
_static,
_templates,
basic_types.rst,
conf.py,
control_flow.rst,
debugging.rst,
...
Make a directory:
In [32]: os.mkdir(junkdir)
In [33]: junkdir in os.listdir(os.curdir)

Out[33]: True
Rename the directory:

In [36]: os.rename(junkdir, foodir)
In [37]: junkdir in os.listdir(os.curdir)

Out[37]: False
In [38]: foodir in os.listdir(os.curdir)

Out[38]: True
In [41]: os.rmdir(foodir)
In [42]: foodir in os.listdir(os.curdir)

Out[42]: False
Delete a file:
In [44]: fp = open(junk.txt, w)
2.7. Standard Library 34

In [45]: fp.close()
In [46]: junk.txt in os.listdir(os.curdir)

Out[46]: True
In [47]: os.remove(junk.txt)
In [48]: junk.txt in os.listdir(os.curdir)

Out[48]: False
os.path: path manipulations
os.path provides common operations on pathnames.

In [70]: fp = open(junk.txt, w)
In [71]: fp.close()
In [72]: a = os.path.abspath(junk.txt)
In [73]: a
Out[73]: /Users/cburns/src/scipy2009/scipy_2009_tutorial/source/junk.txt
In [74]: os.path.split(a)
Out[74]: (/Users/cburns/src/scipy2009/scipy_2009_tutorial/source,
junk.txt)
In [78]: os.path.dirname(a)
Out[78]: /Users/cburns/src/scipy2009/scipy_2009_tutorial/source
In [79]: os.path.basename(a)
Out[79]: junk.txt
In [80]: os.path.splitext(os.path.basename(a))
Out[80]: (junk, .txt)
In [84]: os.path.exists(junk.txt)
Out[84]: True
In [86]: os.path.isfile(junk.txt)
Out[86]: True
In [87]: os.path.isdir(junk.txt)
Out[87]: False
In [88]: os.path.expanduser(~/local)
Out[88]: /Users/cburns/local
In [92]: os.path.join(os.path.expanduser(~), local, bin)

Out[92]: /Users/cburns/local/bin
Running an external command
In [8]: os.system(ls)
basic_types.rst demo.py functions.rst python_language.rst standard_library.rst
control_flow.rst exceptions.rst io.rst python-logo.png
demo2.py first_steps.rst oop.rst reusing_code.rst
Note: Alternative to os.system

A noteworthy alternative to os.system is the sh module. Which provides much more convenient ways to obtain
the output, error stream and exit code of the external command.
In [20]: import sh
In [20]: com = sh.ls()
In [21]: print com

basic_types.rst exceptions.rst oop.rst standard_library.rst
control_flow.rst first_steps.rst python_language.rst
demo2.py functions.rst python-logo.png
demo.py io.rst reusing_code.rst
In [22]: print com.exit_code

0
In [23]: type(com)
Out[23]: sh.RunningCommand
Walking a directory
os.path.walk generates a list of filenames in a directory tree.

In [10]: for dirpath, dirnames, filenames in os.walk(os.curdir):
....: for fp in filenames:
....: print os.path.abspath(fp)
....:
....:
/Users/cburns/src/scipy2009/scipy_2009_tutorial/source/.index.rst.swo
/Users/cburns/src/scipy2009/scipy_2009_tutorial/source/.view_array.py.swp
/Users/cburns/src/scipy2009/scipy_2009_tutorial/source/basic_types.rst
/Users/cburns/src/scipy2009/scipy_2009_tutorial/source/conf.py
/Users/cburns/src/scipy2009/scipy_2009_tutorial/source/control_flow.rst
...
Environment variables:
In [9]: import os
In [11]: os.environ.keys()
Out[11]:
[_,
FSLDIR,
TERM_PROGRAM_VERSION,
FSLREMOTECALL,
USER,
HOME,
PATH,
PS1,
SHELL,
EDITOR,
WORKON_HOME,
PYTHONPATH,
...
In [12]: os.environ[PYTHONPATH]
Out[12]: .:/Users/cburns/src/utils:/Users/cburns/src/nitools:
/Users/cburns/local/lib/python2.5/site-packages/:
/usr/local/lib/python2.5/site-packages/:
/Library/Frameworks/Python.framework/Versions/2.5/lib/python2.5
In [16]: os.getenv(PYTHONPATH)

Out[16]: .:/Users/cburns/src/utils:/Users/cburns/src/nitools:
/Users/cburns/local/lib/python2.5/site-packages/:
/usr/local/lib/python2.5/site-packages/:
/Library/Frameworks/Python.framework/Versions/2.5/lib/python2.5
2.7.2 shutil: high-level file operations
The shutil provides useful file operations:

shutil.rmtree: Recursively delete a directory tree.
shutil.move: Recursively move a file or directory to another location.
shutil.copy: Copy files or directories.
2.7.3 glob: Pattern matching on files
The glob module provides convenient file pattern matching.

Find all files ending in .txt:
In [18]: import glob
In [19]: glob.glob(*.txt)
Out[19]: [holy_grail.txt, junk.txt, newfile.txt]
2.7.4 sys module: system-specific information
System-specific information related to the Python interpreter.

Which version of python are you running and where is it installed:
In [117]: sys.platform
Out[117]: darwin
In [118]: sys.version
Out[118]: 2.5.2 (r252:60911, Feb 22 2008, 07:57:53) \n
[GCC 4.0.1 (Apple Computer, Inc. build 5363)]
In [119]: sys.prefix
Out[119]: /Library/Frameworks/Python.framework/Versions/2.5
List of command line arguments passed to a Python script:

In [100]: sys.argv
Out[100]: [/Users/cburns/local/bin/ipython]
sys.path is a list of strings that specifies the search path for modules. Initialized from PYTHONPATH:
In [121]: sys.path
Out[121]:
[,
/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,
/Users/cburns/local/lib/python2.5/site-packages/urwid-0.9.7.1-py2.5.egg,
/Users/cburns/local/lib/python2.5/site-packages/yolk-0.4.1-py2.5.egg,
/Users/cburns/local/lib/python2.5/site-packages/virtualenv-1.2-py2.5.egg,
...

2.7.5 pickle: easy persistence
Useful to store arbitrary objects to a file. Not safe or fast!

In [1]: import pickle
In [2]: l = [1, None, Stan]
In [3]: pickle.dump(l, file(test.pkl, w))
In [4]: pickle.load(file(test.pkl))
Out[4]: [1, None, Stan]
Exercise
Write a program to search your PYTHONPATH for the module site.py.
path_site
2.8 Exception handling in Python
It is highly unlikely that you havent yet raised Exceptions if you have typed all the previous commands of the
tutorial. For example, you may have raised an exception if you entered a command with a typo.
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.
2.8.1 Exceptions
Exceptions are raised by errors in Python:

In [1]: 1/0
---------------------------------------------------------------------------
ZeroDivisionError: integer division or modulo by zero
In [2]: 1 + e
---------------------------------------------------------------------------
TypeError: unsupported operand type(s) for +: int and str
In [3]: d = {1:1, 2:2}
In [4]: d[3]
---------------------------------------------------------------------------
KeyError: 3
In [5]: l = [1, 2, 3]
In [6]: l[4]
---------------------------------------------------------------------------
IndexError: list index out of range
In [7]: l.foobar
---------------------------------------------------------------------------
AttributeError: list object has no attribute foobar
As you can see, there are different types of exceptions for different errors.
2.8. Exception handling in Python 38

2.8.2 Catching exceptions
try/except
In [8]: while True:

....: try:
....: x = int(raw_input(Please enter a number: ))
....: break
....: except ValueError:
....: print(That was no valid number. Try again...)
....:
....:
Please enter a number: a
That was no valid number. Try again...
Please enter a number: 1
In [9]: x
Out[9]: 1
try/finally
In [10]: try:
....: x = int(raw_input(Please enter a number: ))
....: finally:
....: print(Thank you for your input)
....:
....:
Please enter a number: a
Thank you for your input
---------------------------------------------------------------------------
ValueError: invalid literal for int() with base 10: a
Important for resource management (e.g. closing a file)
Easier to ask for forgiveness than for permission
In [11]: def print_sorted(collection):

....: try:
....: collection.sort()
....: except AttributeError:
....: pass
....: print(collection)
....:
....:
In [12]: print_sorted([1, 3, 2])

[1, 2, 3]
In [13]: print_sorted(set((1, 3, 2)))

set([1, 2, 3])
In [14]: print_sorted(132)
132
2.8.3 Raising exceptions
Capturing and reraising an exception:
2.8. Exception handling in Python 39

In [15]: def filter_name(name):

....: try:
....: name = name.encode(ascii)
....: except UnicodeError, e:
....: if name == Gal:
....: print(OK, Gal)
....: else:
....: raise e
....: return name
....:
In [16]: filter_name(Gal)
OK, Gal
Out[16]: Ga\xc3\xabl
In [17]: filter_name(Stfan)
---------------------------------------------------------------------------
UnicodeDecodeError: ascii codec cant decode byte 0xc3 in position 2: ordinal not in range(
Exceptions to pass messages between parts of the code:

In [17]: def achilles_arrow(x):
....: if abs(x - 1) < 1e-3:
....: raise StopIteration
....: x = 1 - (1-x)/2.
....: return x
....:
In [18]: x = 0
In [19]: while True:

....: try:
....: x = achilles_arrow(x)
....: except StopIteration:
....: break
....:
....:
In [20]: x
Out[20]: 0.9990234375
Use exceptions to notify certain conditions are met (e.g. StopIteration) or not (e.g. custom error raising)
2.9 Object-oriented programming (OOP)
Python supports object-oriented programming (OOP). The goals of OOP are:

to organize the code, and
to re-use code in similar contexts.
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):
... def __init__(self, name):
... self.name = name
... def set_age(self, age):
... self.age = age
... def set_major(self, major):
... self.major = major
...
2.9. Object-oriented programming (OOP) 40

>>> anna = Student(anna)

>>> anna.set_age(21)
>>> anna.set_major(physics)
In the previous example, the Student class has __init__, set_age and set_major methods. Its at-
tributes are name, age and major. We can call these methods and attributes with the following notation:
classinstance.method or classinstance.attribute. The __init__ constructor is a special
method we call with: MyClass(init parameters if any).
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:
>>> class MasterStudent(Student):
... internship = mandatory, from March to June
...
>>> james = MasterStudent(james)
>>> james.internship
mandatory, from March to June
>>> james.set_age(23)
>>> james.age
23
The MasterStudent class inherited from the Student attributes and methods.
Thanks to classes and object-oriented programming, we can organize code with different classes corresponding to
different objects we encounter (an Experiment class, an Image class, a Flow class, etc.), with their own methods
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.
2.9. Object-oriented programming (OOP) 41

CHAPTER 3
NumPy: creating and manipulating

numerical data
authors Emmanuelle Gouillart, Didrik Pinte, Gal Varoquaux, and Pauli Virtanen
This chapter gives an overview of Numpy, the core tool for performant numerical computing with Python.
3.1 The numpy array object
Section contents
What are Numpy and numpy arrays (page 42)
Reference documentation (page 43)
Creating arrays (page 44)
Basic data types (page 46)
Basic visualization (page 46)
Indexing and slicing (page 49)
Copies and views (page 50)
Adding Axes (page 51)
Fancy indexing (page 52)
3.1.1 What are Numpy and numpy arrays
Python has built-in:

containers: lists (costless insertion and append), dictionaries (fast lookup)
high-level number objects: integers, floating point
Numpy is:
extension package to Python for multi-dimensional arrays
closer to hardware (efficiency)
designed for scientific computation (convenience)
42

>>> a = np.array([0, 1, 2, 3])
>>> a
array([0, 1, 2, 3])
For example:
An array containing:
values of an experiment/simulation at discrete time steps
signal recorded by a measurement device, e.g. sound wave
pixels of an image, grey-level or colour
3-D data measured at different X-Y-Z positions, e.g. MRI scan
...
Why it is useful: Memory-efficient container that provides fast numerical operations.

In [1]: L = range(1000)
In [2]: %timeit [i**2 for i in L]

1000 loops, best of 3: 403 us per loop
In [3]: a = np.arange(1000)
In [4]: %timeit a**2

100000 loops, best of 3: 12.7 us per loop
3.1.2 Reference documentation
On the web: http://docs.scipy.org/

Interactive help:
>>> help(np.array)
Help on built-in function array in module numpy.core.multiarray:
array(...)
array(object, dtype=None, copy=True, order=None, subok=False, ...
...
In [5]: np.array?
String Form:<built-in function array>
Docstring:
array(object, dtype=None, copy=True, order=None, subok=False, ndmin=0, ...
...
Looking for something:

>>> np.lookfor(create array)
Search results for create array
---------------------------------
numpy.array
Create an array.
numpy.memmap
Create a memory-map to an array stored in a *binary* file on disk.
...
In [6]: np.con* ?
np.concatenate
np.conj
np.conjugate
np.convolve
3.1. The numpy array object 43

3.1.3 Creating arrays
1-D:
>>> a = np.array([0, 1, 2, 3])
>>> a
array([0, 1, 2, 3])
>>> a.ndim
1
>>> a.shape
(4,)
>>> len(a)
4
2-D, 3-D, ...:

>>> b = np.array([[0, 1, 2], [3, 4, 5]]) # 2 x 3 array
>>> b
array([[0, 1, 2],
[3, 4, 5]])
>>> b.ndim
2
>>> b.shape
(2, 3)
>>> len(b) # returns the size of the first dimension
2
>>> c = np.array([[[1], [2]], [[3], [4]]])

>>> c
array([[[1],
[2]],
[[3],
[4]]])
>>> c.shape
(2, 2, 1)
In practice, we rarely enter items one by one...

Evenly spaced:
>>> a = np.arange(10) # 0 .. n-1 (!)
>>> a
array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9])
>>> b = np.arange(1, 9, 2) # start, end (exlusive), step
>>> b
array([1, 3, 5, 7])
or by number of points:
>>> c = np.linspace(0, 1, 6) # start, end, num-points
>>> c
array([ 0. , 0.2, 0.4, 0.6, 0.8, 1. ])
>>> d = np.linspace(0, 1, 5, endpoint=False)
>>> d
array([ 0. , 0.2, 0.4, 0.6, 0.8])
Common arrays:

>>> a = np.ones((3, 3)) # reminder: (3, 3) is a tuple

>>> a
array([[ 1., 1., 1.],
[ 1., 1., 1.],
[ 1., 1., 1.]])
>>> b = np.zeros((2, 2))
>>> b
array([[ 0., 0.],
[ 0., 0.]])
>>> c = np.eye(3)
>>> c
array([[ 1., 0., 0.],
[ 0., 1., 0.],
[ 0., 0., 1.]])
>>> d = np.diag(np.array([1, 2, 3, 4]))
>>> d
array([[1, 0, 0, 0],
[0, 2, 0, 0],
[0, 0, 3, 0],
[0, 0, 0, 4]])
np.random: random numbers (Mersenne Twister PRNG):

>>> a = np.random.rand(4) # uniform in [0, 1]
>>> a
array([ 0.95799151, 0.14222247, 0.08777354, 0.51887998])
>>> b = np.random.randn(4) # Gaussian

>>> b
array([ 0.37544699, -0.11425369, -0.47616538, 1.79664113])
>>> np.random.seed(1234) # Setting the random seed
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:
[[4, 3, 4, 3, 4, 3],
[2, 1, 2, 1, 2, 1],
[4, 3, 4, 3, 4, 3],
[2, 1, 2, 1, 2, 1]]

3.1.4 Basic data types
You may have noticed that, in some instances, array elements are displayed with a trailing dot (e.g. 2. vs 2). This
is due to a difference in the data-type used:
>>> a = np.array([1, 2, 3])
>>> a.dtype
dtype(int64)
>>> b = np.array([1., 2., 3.])

>>> b.dtype
dtype(float64)
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.
You can explicitly specify which data-type you want:

>>> c = np.array([1, 2, 3], dtype=float)
>>> c.dtype
dtype(float64)
The default data type is floating point:

>>> a = np.ones((3, 3))
>>> a.dtype
dtype(float64)
There are also other types:

Complex
>>> d = np.array([1+2j, 3+4j, 5+6*1j])
>>> d.dtype
dtype(complex128)
Bool
>>> e = np.array([True, False, False, True])
>>> e.dtype
dtype(bool)
Strings
>>> f = np.array([Bonjour, Hello, Hallo,])
>>> f.dtype # <--- strings containing max. 7 letters
dtype(S7)
Much more int32/int64...
3.1.5 Basic visualization
Now that we have our first data arrays, we are going to visualize them.
Start by launching IPython in pylab mode
$ ipython --pylab
Matplotlib is a 2D plotting package. We can import its functions as below:

>>> import matplotlib.pyplot as plt # the tidy way
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
>>> plt.show() # <-- shows the plot (not needed with Ipython)
9
8
7
6
5
4
3
2
1
00.0 0.5 1.0 1.5 2.0 2.5 3.0
2D arrays (such as images):
>>> image = np.random.rand(30, 30)
>>> plt.imshow(image, cmap=plt.cm.gray)
<matplotlib.image.AxesImage object at ...>
>>> plt.colorbar()
<matplotlib.colorbar.Colorbar instance at ...>
>>> plt.show()

1.0
0
0.9
5 0.8
0.7
10
0.6
15 0.5
0.4
20 0.3
25 0.2
0.1
0 5 10 15 20 25
See Also:
More in the matplotlib chapter
3D plotting
For 3D visualization, we can use another package: Mayavi. A quick example: start by relaunching iPython
with these options: ipython pylab=wx (or ipython -pylab -wthread in IPython < 0.10).
In [58]: from mayavi import mlab

In [61]: mlab.surf(image)
Out[61]: <enthought.mayavi.modules.surface.Surface object at ...>
In [62]: mlab.axes()
Out[62]: <enthought.mayavi.modules.axes.Axes object at ...>
The mayavi/mlab window that opens is interactive: by clicking on the left mouse button you can rotate the

image, zoom with the mouse wheel, etc.

For more information on Mayavi : http://github.enthought.com/mayavi/mayavi
See Also:
More in the Mayavi chapter (page 287)
3.1.6 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.
For multidimensional arrays, indexes are tuples of integers:

>>> a = np.diag(np.arange(3))
>>> a
array([[0, 0, 0],
[0, 1, 0],
[0, 0, 2]])
>>> a[1, 1]
1
>>> a[2, 1] = 10 # third line, second column
>>> a
array([[ 0, 0, 0],
[ 0, 1, 0],
[ 0, 10, 2]])
>>> a[1]
array([0, 1, 0])
Note that:
In 2D, the first dimension corresponds to rows, the second to columns.
for multidimensional a, a[0] is interpreted by taking all elements in the unspecified dimensions.
Slicing Arrays, like other Python sequences can also be sliced:
>>> a
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!

>>> a[:4]
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:
>>> a[1:3]
array([1, 2])
>>> a[::2]
array([0, 2, 4, 6, 8])
>>> a[3:]
array([3, 4, 5, 6, 7, 8, 9])

A small illustrated summary of Numpy indexing and slicing...
3.1.7 Copies and views
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.
When modifying the view, the original array is modified as well:
>>> a
array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9])
>>> b = a[::2]; b
array([0, 2, 4, 6, 8])
>>> b[0] = 12
>>> b
array([12, 2, 4, 6, 8])
>>> a # (!)
array([12, 1, 2, 3, 4, 5, 6, 7, 8, 9])
>>> b = a[::2].copy() # force a copy
>>> b[0] = 12
>>> a
array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9])
This behavior can be surprising at first sight... but it allows to save both memory and time.

Warning: The transpose is a view

As a result, a matrix cannot be made symmetric in-place:
>>> a = np.ones((100, 100))

>>> a += a.T
>>> a
array([[ 2., 2., 2., ..., 2., 2., 2.],
[ 2., 2., 2., ..., 2., 2., 2.],
[ 2., 2., 2., ..., 2., 2., 2.],
...,
[ 3., 3., 3., ..., 2., 2., 2.],
[ 3., 3., 3., ..., 2., 2., 2.],
[ 3., 3., 3., ..., 2., 2., 2.]])
Worked example: Prime number sieve
Compute prime numbers in 099, with a sieve

Construct a shape (100,) boolean array is_prime, filled with True in the beginning:
>>> is_prime = np.ones((100,), dtype=bool)
Cross out 0 and 1 which are not primes:
>>> is_prime[:2] = 0
For each integer j starting from 2, cross out its higher multiples:
>>> N_max = int(np.sqrt(len(is_prime)))
>>> for j in range(2, N_max):
... is_prime[2*j::j] = False
Skim through help(np.nonzero), and print the prime numbers
Follow-up:
Move the above code into a script file named prime_sieve.py
Run it to check it works
Convert the simple sieve to the sieve of Eratosthenes:
1. Skip j which are already known to not be primes
2. The first number to cross out is j 2
3.1.8 Adding Axes
Indexing with the np.newaxis object allows us to add an axis to an array:

>>> z = np.array([1, 2, 3])

>>> z
array([1, 2, 3])
>>> z[:, np.newaxis]

array([[1],
[2],
[3]])
>>> z[np.newaxis, :]
array([[1, 2, 3]])
3.1.9 Fancy indexing
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)
>>> extract_from_a = a[mask] # or, a[a%3==0]
>>> 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
array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9])
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([2, 3, 2, 4, 2])
New values can be assigned with this kind of indexing:

>>> a[[9, 7]] = -10
>>> a
array([ 0, 1, 2, 3, 4, 5, 6, -10, 8, -10])
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:

>>> idx = np.array([[3, 4], [9, 7]])
>>> a[idx]
array([[3, 4],
[9, 7]])
>>> b = np.arange(10)
The image below illustrates various fancy indexing applications
We can even use fancy indexing and broadcasting (page 59) at the same time:
>>> a = np.arange(12).reshape(3,4)
>>> a
array([[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11]])
>>> i = np.array([[0, 1], [1, 2]])
>>> a[i, 2] # same as a[i, 2*np.ones((2, 2), dtype=int)]
array([[ 2, 6],
[ 6, 10]])
3.2 Numerical operations on arrays
Section contents
Elementwise operations (page 53)
Basic reductions (page 55)
Broadcasting (page 59)
Array shape manipulation (page 62)
Sorting data (page 65)
Some exercises (page 66)
Summary (page 72)
3.2.1 Elementwise operations
With scalars:
3.2. Numerical operations on arrays 53

>>> a = np.array([1, 2, 3, 4])

>>> a + 1
array([2, 3, 4, 5])
>>> 2**a
array([ 2, 4, 8, 16])
All arithmetic operates elementwise:

>>> b = np.ones(4) + 1
>>> a - b
array([-1., 0., 1., 2.])
>>> a * b
array([ 2., 4., 6., 8.])
>>> j = np.arange(5)
>>> 2**(j + 1) - j
array([ 2, 3, 6, 13, 28])
Warning: Array multiplication is not matrix multiplication:
>>> c = np.ones((3, 3))

>>> c * c # NOT matrix multiplication!
array([[ 1., 1., 1.],
[ 1., 1., 1.],
[ 1., 1., 1.]])
Note: Matrix multiplication:

>>> c.dot(c)
array([[ 3., 3., 3.],
[ 3., 3., 3.],
[ 3., 3., 3.]])
Comparisons:
>>> a = np.array([1, 2, 3, 4])
>>> b = np.array([4, 2, 2, 4])
>>> a == b
array([False, True, False, True], dtype=bool)
>>> a > b
array([False, False, True, False], dtype=bool)
Logical operations:
>>> a = np.array([1, 1, 0, 0], dtype=bool)
>>> b = np.array([1, 0, 1, 0], dtype=bool)
>>> np.logical_or(a, b)
array([ True, True, True, False], dtype=bool)
>>> np.logical_and(a, b)
array([ True, False, False, False], dtype=bool)
Shape mismatches
>>> a + np.array([1, 2])
ValueError: operands could not be broadcast together with shapes (4) (2)
Broadcasting? Well return to that later (page 59).

Transposition:

>>> a = np.triu(np.ones((3, 3)), 1) # see help(np.triu)

>>> a
array([[ 0., 1., 1.],
[ 0., 0., 1.],
[ 0., 0., 0.]])
>>> a.T
array([[ 0., 0., 0.],
[ 1., 0., 0.],
[ 1., 1., 0.]])
Note: Linear algebra

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 106)
Exercise
Generate arrays [2**0, 2**1, 2**2, 2**3, 2**4] and a_j = 2^(3*j) - j
3.2.2 Basic reductions
Computing sums:
>>> x = np.array([1, 2, 3, 4])
>>> np.sum(x)
10
>>> x.sum()
10
Sum by rows and by columns:

>>> x = np.array([[1, 1], [2, 2]])
>>> x
array([[1, 1],
[2, 2]])
>>> x.sum(axis=0) # columns (first dimension)
array([3, 3])
>>> x[:, 0].sum(), x[:, 1].sum()
(3, 3)
>>> x.sum(axis=1) # rows (second dimension)
array([2, 4])

>>> x[0, :].sum(), x[1, :].sum()

(2, 4)
Same idea in higher dimensions:

>>> x = np.random.rand(2, 2, 2)
>>> x.sum(axis=2)[0, 1]
1.14764...
>>> x[0, 1, :].sum()
1.14764...
Other reductions works the same way (and take axis=)

Statistics:
>>> x = np.array([1, 2, 3, 1])
>>> y = np.array([[1, 2, 3], [5, 6, 1]])
>>> x.mean()
1.75
>>> np.median(x)
1.5
>>> np.median(y, axis=-1) # last axis
array([ 2., 5.])
>>> x.std() # full population standard dev.

0.82915619758884995
Extrema:
>>> x = np.array([1, 3, 2])
>>> x.min()
1
>>> x.max()
3
>>> x.argmin() # index of minimum

0
>>> x.argmax() # index of maximum
1
Logical operations:
>>> np.all([True, True, False])
False
>>> np.any([True, True, False])
True
Note: Can be used for array comparisons:

>>> a = np.zeros((100, 100))
>>> np.any(a != 0)
False
>>> np.all(a == a)
True
>>> a = np.array([1, 2, 3, 2])

>>> b = np.array([2, 2, 3, 2])
>>> c = np.array([6, 4, 4, 5])
>>> ((a <= b) & (b <= c)).all()
True
... and many more (best to learn as you go).

Example: data statistics

Data in populations.txt describes the populations of hares and lynxes (and carrots) in northern Canada
during 20 years.
We can first plot the data:
>>> data = np.loadtxt(data/populations.txt)
>>> year, hares, lynxes, carrots = data.T # trick: columns to variables
>>> from matplotlib import pyplot as plt

>>> 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 ...>
80000
70000
60000
Hare
50000 Lynx
40000
Carrot
30000
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])

Example: diffusion simulation using a random walk algorithm
What is the typical distance from the origin of a random walker after t left or right jumps?
>>> 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:
>>> mean_sq_distance = np.mean(sq_distance, axis=0)
Plot the results:
>>> plt.figure(figsize=(4, 3))
<matplotlib.figure.Figure object at ...>
>>> plt.plot(t, np.sqrt(mean_sq_distance), g., t, np.sqrt(t), y-)
[<matplotlib.lines.Line2D object at ...>, <matplotlib.lines.Line2D object at ...>]
>>> plt.xlabel(r"$t$")
<matplotlib.text.Text object at ...>
>>> plt.ylabel(r"$\sqrt{\langle (\delta x)^2 \rangle}$")
16
14
12
10
(x)2
8
6
q
4
2
00 50 100 150 200
t
The RMS distance grows as the square root of the time!
3.2.3 Broadcasting
Basic operations on numpy arrays (addition, etc.) are elementwise

This works on arrays of the same size.
Nevertheless, Its also possible to do operations on arrays of different
sizes if Numpy can transform these arrays so that they all have
the same size: this conversion is called broadcasting.
The image below gives an example of broadcasting:
Lets verify:
>>> a = np.tile(np.arange(0, 40, 10), (3, 1)).T
>>> a
array([[ 0, 0, 0],
[10, 10, 10],
[20, 20, 20],
[30, 30, 30]])
>>> b = np.array([0, 1, 2])
>>> a + b
array([[ 0, 1, 2],
[10, 11, 12],
[20, 21, 22],
[30, 31, 32]])
An useful trick:
>>> a = np.arange(0, 40, 10)
>>> a.shape

(4,)
>>> a = a[:, np.newaxis] # adds a new axis -> 2D array
>>> a.shape
(4, 1)
>>> a
array([[ 0],
[10],
[20],
[30]])
>>> a + b
array([[ 0, 1, 2],
[10, 11, 12],
[20, 21, 22],
[30, 31, 32]])
We have already used broadcasting without knowing it!:

>>> a = np.ones((4, 5))
>>> a[0] = 2 # we assign an array of dimension 0 to an array of dimension 1
>>> a
array([[ 2., 2., 2., 2., 2.],
[ 1., 1., 1., 1., 1.],
[ 1., 1., 1., 1., 1.],
[ 1., 1., 1., 1., 1.]])
Broadcasting seems a bit magical, but it is actually quite natural to use it when we want to solve a problem whose
output data is an array with more dimensions than input data.
Example
Lets construct an array of distances (in miles) between cities of Route 66: Chicago, Springfield, Saint-Louis,
Tulsa, Oklahoma City, Amarillo, Santa Fe, Albuquerque, Flagstaff and Los Angeles.
>>> mileposts = np.array([0, 198, 303, 736, 871, 1175, 1475, 1544,
... 1913, 2448])
>>> distance_array = np.abs(mileposts - mileposts[:, np.newaxis])
>>> distance_array
array([[ 0, 198, 303, 736, 871, 1175, 1475, 1544, 1913, 2448],
[ 198, 0, 105, 538, 673, 977, 1277, 1346, 1715, 2250],
[ 303, 105, 0, 433, 568, 872, 1172, 1241, 1610, 2145],
[ 736, 538, 433, 0, 135, 439, 739, 808, 1177, 1712],
[ 871, 673, 568, 135, 0, 304, 604, 673, 1042, 1577],
[1175, 977, 872, 439, 304, 0, 300, 369, 738, 1273],
[1475, 1277, 1172, 739, 604, 300, 0, 69, 438, 973],
[1544, 1346, 1241, 808, 673, 369, 69, 0, 369, 904],
[1913, 1715, 1610, 1177, 1042, 738, 438, 369, 0, 535],
[2448, 2250, 2145, 1712, 1577, 1273, 973, 904, 535, 0]])

A lot of grid-based or network-based problems can also use broadcasting. For instance, if we want to compute the
distance from the origin of points on a 10x10 grid, we can do
>>> x, y = np.arange(5), np.arange(5)
>>> distance = np.sqrt(x ** 2 + y[:, np.newaxis] ** 2)
>>> distance
array([[ 0. , 1. , 2. , 3. , 4. ],
[ 1. , 1.41421356, 2.23606798, 3.16227766, 4.12310563],
[ 2. , 2.23606798, 2.82842712, 3.60555128, 4.47213595],
[ 3. , 3.16227766, 3.60555128, 4.24264069, 5. ],
[ 4. , 4.12310563, 4.47213595, 5. , 5.65685425]])
Or in color:
>>> plt.pcolor(distance)
<matplotlib.collections.PolyCollection object at ...>
>>> plt.colorbar()
<matplotlib.colorbar.Colorbar instance at ...>
>>> plt.axis(equal)
(0.0, 200.0, 0.0, 16.0)
5 5.4
4.8
4
4.2
3.6
3
3.0
2 2.4
1.8
1 1.2
0.6
0 0.0
0 1 2 3 4 5
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)

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 broadcasting:
>>> x, y = np.mgrid[0:4, 0:4]
>>> x
array([[0, 0, 0, 0],
[1, 1, 1, 1],
[2, 2, 2, 2],
[3, 3, 3, 3]])
>>> y
array([[0, 1, 2, 3],
[0, 1, 2, 3],
[0, 1, 2, 3],
[0, 1, 2, 3]])
However, in practice, this is rarely needed!
3.2.4 Array shape manipulation
Flattening
>>> a = np.array([[1, 2, 3], [4, 5, 6]])

>>> a.ravel()
array([1, 2, 3, 4, 5, 6])
>>> a.T
array([[1, 4],
[2, 5],
[3, 6]])
>>> a.T.ravel()
array([1, 4, 2, 5, 3, 6])
Higher dimensions: last dimensions ravel out first.
Reshaping
The inverse operation to flattening:

>>> a.shape
(2, 3)
>>> b = a.ravel()
>>> b.reshape((2, 3))
array([[1, 2, 3],
[4, 5, 6]])
Creating an array with a different shape, from another array:

>>> b = a.reshape((6, 6))
>>> b
array([[ 0, 1, 2, 3, 4, 5],
[ 6, 7, 8, 9, 10, 11],
[12, 13, 14, 15, 16, 17],
[18, 19, 20, 21, 22, 23],
[24, 25, 26, 27, 28, 29],
[30, 31, 32, 33, 34, 35]])
Or,
>>> b = a.reshape((6, -1)) # unspecified (-1) value is inferred

Views and copies
ndarray.reshape may return a view (cf help(np.reshape))), not a copy:

>>> b[0, 0] = 99
>>> a
array([99, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16,
17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33,
34, 35])
Beware: reshape may also return a copy!:

>>> a = np.zeros((3, 2))
>>> b = a.T.reshape(3*2)
>>> b[0] = 9
>>> a
array([[ 0., 0.],
[ 0., 0.],
[ 0., 0.]])
To understand, see the section on the memory layout of an array below.
Dimension shuffling
>>> a = np.arange(4*3*2).reshape(4, 3, 2)
>>> a.shape
(4, 3, 2)
>>> a[0, 2, 1]
5
>>> b = a.transpose(1, 2, 0)
>>> b.shape
(3, 2, 4)
>>> b[2, 1, 0]
5
Also creates a view:

>>> b[2, 1, 0] = -1
>>> a[0, 2, 1]
-1
Resizing
Size of an array can be changed with ndarray.resize:

>>> a.resize((8,))
>>> a
array([0, 1, 2, 3, 0, 0, 0, 0])
However, it must not be referred to somewhere else:

>>> b = a
>>> a.resize((4,))
ValueError: cannot resize an array that has been referenced or is
referencing another array in this way. Use the resize function
Some examples of real-world use cases

Case 2.a: Calling (legacy) Fortran code

Shape-preserving functions with elementwise non-Python routines. For instance, Fortran
! 2_a_fortran_module.f90
subroutine some_function(n, a, b)
integer :: n
double precision, dimension(n), intent(in) :: a
double precision, dimension(n), intent(out) :: b
b = a + 1
end subroutine some_function
We can use f2py to wrap this fortran code in Python: f2py -c -m fortran_module
2_a_fortran_module.f90
import numpy as np
import fortran_module
def some_function(input):
"""
Call a Fortran routine, and preserve input shape
"""
input = np.asarray(input)
# fortran_module.some_function() takes 1-D arrays!
output = fortran_module.some_function(input.ravel())
return output.reshape(input.shape)
print some_function(np.array([1, 2, 3]))

print some_function(np.array([[1, 2], [3, 4]]))
# ->
# [ 2. 3. 4.]
# [[ 2. 3.]
# [ 4. 5.]]

Case 2.b: Block matrices and vectors (and tensors)

Vector space: quantum level spin

1 2
= 1 , 1 = 2 =
2 1 2
In short: for block matrices and vectors, it can be useful to preserve the block structure.
In Numpy:
>>> psi = np.zeros((2, 2)) # dimensions: level, spin
>>> psi[0, 1] # <-- psi_{1,downarrow}
0.0
Linear operators on such block vectors have similar block structure:

H = h11 V , 11 = 1,
h
0
, ...
V h 22 0 1,
>>> H = np.zeros((2, 2, 2, 2)) # dimensions: level1, level2, spin1, spin2

>>> h_11 = H[0,0,:,:]
>>> V = H[0,1]
Doing the matrix product: get rid of the block structure, do the 4x4 matrix product, then put it back

H
>>> def mdot(operator, psi):

... return operator.transpose(0, 2, 1, 3).reshape(4, 4).dot(
... psi.reshape(4)).reshape(2, 2)
I.e., reorder dimensions first to level1, spin1, level2, spin2 and then reshape => correct matrix
product.
3.2.5 Sorting data
Sorting along an axis:

>>> a = np.array([[4, 3, 5], [1, 2, 1]])
>>> b = np.sort(a, axis=1)
>>> b
array([[3, 4, 5],
[1, 1, 2]])
Note: Sorts each row separately!
In-place sort:
>>> a.sort(axis=1)
>>> a
array([[3, 4, 5],
[1, 1, 2]])
Sorting with fancy indexing:

>>> a = np.array([4, 3, 1, 2])
>>> j = np.argsort(a)
>>> j
array([2, 3, 1, 0])

>>> a[j]
array([1, 2, 3, 4])
Finding minima and maxima:

>>> a = np.array([4, 3, 1, 2])
>>> j_max = np.argmax(a)
>>> j_min = np.argmin(a)
>>> j_max, j_min
(0, 2)
3.2.6 Some exercises

Worked example: Framing Lena

Lets do some manipulations on numpy arrays by starting with the famous image of Lena
(http://www.cs.cmu.edu/~chuck/lennapg/). scipy provides a 2D array of this image with the
scipy.lena function:
>>> from scipy import misc
>>> lena = misc.lena()
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, change some parts of the image.
Lets use the imshow function of pylab to display the image.

In [3]: import pylab as plt
In [4]: lena = misc.lena()
In [5]: plt.imshow(lena)
Lena is then displayed in false colors. A colormap must be specified for her to be displayed in grey.
In [6]: plt.imshow(lena, cmap=plt.cm.gray)
Create an array of the image with a narrower centering : for example, remove 30 pixels from all the
borders of the image. To check the result, display this new array with imshow.
In [9]: crop_lena = lena[30:-30,30:-30]
We will now frame Lenas face with a black locket. For this, we need to create a mask corre-
sponding to the pixels we want to be black. The mask is defined by this condition (y-256)**2
+ (x-256)**2
In [15]: y, x = np.ogrid[0:512,0:512] # x and y indices of pixels
In [16]: y.shape, x.shape
Out[16]: ((512, 1), (1, 512))
In [17]: centerx, centery = (256, 256) # center of the image
In [18]: mask = ((y - centery)**2 + (x - centerx)**2) > 230**2 # circle
then we assign the value 0 to the pixels of the image corresponding to the mask. The syntax is
extremely simple and intuitive:
In [19]: lena[mask] = 0
In [20]: plt.imshow(lena)
Out[20]: <matplotlib.image.AxesImage object at 0xa36534c>
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.
Change the circle to an ellipsoid.

Exercise: Array manipulations

1. Form the 2-D array (without typing it in explicitly):
[[1, 6, 11],
[2, 7, 12],
[3, 8, 13],
[4, 9, 14],
[5, 10, 15]]
and generate a new array containing its 2nd and 4th rows.
2. Divide each column of the array:
>>> a = np.arange(25).reshape(5, 5)
elementwise with the array b = np.array([1., 5, 10, 15, 20]). (Hint: np.newaxis).
3. Harder one: Generate a 10 x 3 array of random numbers (in range [0,1]). For each row, pick the
number closest to 0.5.
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.)

Exercise: Data statistics

The data in populations.txt:: describes the populations of hares and lynxes (and carrots) in northern
Canada during 20 years:
>>> year, hares, lynxes, carrots = data.T # trick: columns to variables
>>> plt.axes([0.2, 0.1, 0.5, 0.8])

<matplotlib.axes.Axes object at ...>
>>> plt.plot(year, hares, year, lynxes, year, carrots)
>>> plt.legend((Hare, Lynx, Carrot), loc=(1.05, 0.5))
<matplotlib.legend.Legend object at ...>
80000
70000
60000
Hare
50000 Lynx
40000
Carrot
30000
20000
10000
0
1900 1905 1910 1915 1920
Computes and print, based on the data in populations.txt...
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.
3. Which species has the largest population for each year. (Hint: argsort & fancy indexing of
np.array([H, L, C]))
4. Which years any of the populations is above 50000. (Hint: comparisons and np.any)
5. The top 2 years for each species when they had the lowest populations. (Hint: argsort, fancy
indexing)
6. Compare (plot) the change in hare population (see help(np.gradient)) and the number of
lynxes. Check correlation (see help(np.corrcoef)).
... all without for-loops.

Exercise: Crude integral approximations

Write a function f(a, b, c) that returns ab c. Form a 24x12x6 array containing its values in parameter
ranges [0,1] x [0,1] x [0,1].
Approximate the 3-d integral
Z 1 Z 1 Z 1
(ab c)da db dc
0 0 0
over this volume with the mean. The exact result is: ln 2 12 0.1931 . . . what is your relative error?
(Hints: use elementwise operations and broadcasting. You can make np.ogrid give a number of points in
given range with np.ogrid[0:1:20j].)
Reminder Python functions:
def f(a, b, c):
return some_result

Exercise: Mandelbrot set
1.5
1.0
0.5
0.0
0.5
1.0
1.52.0 1.5 1.0 0.5 0.0 0.5 1.0

Write a script that computes the Mandelbrot fractal. The Mandelbrot iteration:
N_max = 50
some_threshold = 50
c = x + 1j*y
for j in xrange(N_max):
z = z**2 + c
Point (x, y) belongs to the Mandelbrot set if |c| < some_threshold.
Do this computation by:
1. Construct a grid of c = x + 1j*y values in range [-2, 1] x [-1.5, 1.5]
2. Do the iteration
3. Form the 2-d boolean mask indicating which points are in the set
4. Save the result to an image with:
>>> import matplotlib.pyplot as plt
>>> plt.imshow(mask.T, extent=[-2, 1, -1.5, 1.5])
>>> plt.gray()
>>> plt.savefig(mandelbrot.png)

Exercise: Markov chain
Markov chain transition matrix P, and probability distribution on the states p:

1. 0 <= P[i,j] <= 1: probability to go from state i to state j
2. Transition rule: pnew = P T pold
3. all(sum(P, axis=1) == 1), p.sum() == 1: normalization
Write a script that works with 5 states, and:
Constructs a random matrix, and normalizes each row so that it is a transition matrix.
Starts from a random (normalized) probability distribution p and takes 50 steps => p_50
Computes the stationary distribution: the eigenvector of P.T with eigenvalue 1 (numeri-
cally: closest to 1) => p_stationary
Remember to normalize the eigenvector I didnt...
Checks if p_50 and p_stationary are equal to tolerance 1e-5
Toolbox: np.random.rand, .dot(), np.linalg.eig, reductions, abs(), argmin,
comparisons, all, np.linalg.norm, etc.
3.2.7 Summary
What do you need to know to get started?

Know how to create arrays : array, arange, ones, zeros.
Know the shape of the array with array.shape, then use slicing to obtain different views of the array:
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:
>>> a[a < 0] = 0
Know miscellaneous operations on arrays, such as finding the mean or max (array.max(),
array.mean()). No need to retain everything, but have the reflex to search in the documentation (online
docs, help(), lookfor())!!
For advanced use: master the indexing with arrays of integers, as well as broadcasting. Know more Numpy
functions to handle various array operations.
3.3 More elaborate arrays
3.3. More elaborate arrays 72

Section contents
More data types (page 73)
Structured data types (page 74)
maskedarray: dealing with (propagation of) missing data (page 75)
3.3.1 More data types
Casting
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])
>>> a.dtype
dtype(int64)
>>> a[0] = 1.9 # <-- float is truncated to integer
>>> a
array([1, 2, 3])
Forced casts:
>>> a = np.array([1.7, 1.2, 1.6])
>>> b = a.astype(int) # <-- truncates to integer
>>> b
array([1, 1, 1])
Rounding:
>>> a = np.array([1.2, 1.5, 1.6, 2.5, 3.5, 4.5])
>>> b = np.around(a)
>>> b # still floating-point
array([ 1., 2., 2., 2., 4., 4.])
>>> c = np.around(a).astype(int)
>>> c
array([1, 2, 2, 2, 4, 4])
Different data type sizes
Integers (signed):
int8 8 bits
int16 16 bits
int32 32 bits (same as int on 32-bit platform)
int64 64 bits (same as int on 64-bit platform)
>>> np.array([1], dtype=int).dtype
dtype(int64)
>>> np.iinfo(np.int32).max, 2**31 - 1
(2147483647, 2147483647)
>>> np.iinfo(np.int64).max, 2**63 - 1
(9223372036854775807, 9223372036854775807L)
Unsigned integers:

uint8 8 bits
uint16 16 bits
uint32 32 bits
uint64 64 bits
>>> np.iinfo(np.uint32).max, 2**32 - 1
(4294967295, 4294967295)
>>> np.iinfo(np.uint64).max, 2**64 - 1
(18446744073709551615L, 18446744073709551615L)
Floating-point numbers:
float16 16 bits
float32 32 bits
float64 64 bits (same as float)
float96 96 bits, platform-dependent (same as np.longdouble)
float128 128 bits, platform-dependent (same as np.longdouble)
>>> np.finfo(np.float32).eps
1.1920929e-07
>>> np.finfo(np.float64).eps
2.2204460492503131e-16
>>> np.float32(1e-8) + np.float32(1) == 1

True
>>> np.float64(1e-8) + np.float64(1) == 1
False
Complex floating-point numbers:

complex64 two 32-bit floats
complex128 two 64-bit floats
complex192 two 96-bit floats, platform-dependent
complex256 two 128-bit floats, platform-dependent
Smaller data types

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
Half the memory bandwidth required (may be a bit faster in some operations)
In [1]: a = np.zeros((1e6,), dtype=np.float64)
In [2]: b = np.zeros((1e6,), dtype=np.float32)
In [3]: %timeit a*a

1000 loops, best of 3: 1.78 ms per loop
In [4]: %timeit b*b

But: bigger rounding errors sometimes in surprising places (i.e., dont use them unless you really
need them)
3.3.2 Structured data types
sensor_code (4-character string)

position (float)
value (float)

>>> samples = np.zeros((6,), dtype=[(sensor_code, S4),

... (position, float), (value, float)])
>>> samples.ndim
1
>>> samples.shape
(6,)
>>> samples.dtype.names
(sensor_code, position, value)
>>> 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)])
Field access works by indexing with field names:

>>> samples[sensor_code]
array([ALFA, BETA, TAU, ALFA, ALFA, TAU],
dtype=|S4)
>>> samples[value]
array([ 0.37, 0.11, 0.13, 0.37, 0.11, 0.13])
>>> samples[0]
(ALFA, 1.0, 0.37)
>>> samples[0][sensor_code] = TAU

>>> samples[0]
(TAU, 1.0, 0.37)
Multiple fields at once:

>>> samples[[position, value]]
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)])
Fancy indexing works, as usual:

>>> samples[samples[sensor_code] == ALFA]
array([(ALFA, 1.5, 0.37), (ALFA, 3.0, 0.11)],
dtype=[(sensor_code, S4), (position, <f8), (value, <f8)])
Note: There are a bunch of other syntaxes for constructing structured arrays, see here and here.
3.3.3 maskedarray: dealing with (propagation of) missing data
For floats one could use NaNs, but masks work for all types:
>>> x = np.ma.array([1, 2, 3, 4], mask=[0, 1, 0, 1])
>>> x
masked_array(data = [1 -- 3 --],
mask = [False True False True],
fill_value = 999999)
>>> y = np.ma.array([1, 2, 3, 4], mask=[0, 1, 1, 1])

>>> x + y
masked_array(data = [2 -- -- --],
mask = [False True True True],

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 186)
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:
Good practices
Explicit variable names (no need of a comment to explain what is in the variable)
Style: spaces after commas, around =, etc.
A certain number of rules for writing beautiful code (and, more importantly, using the same conven-
tions as everybody else!) are given in the Style Guide for Python Code and the Docstring Conventions
page (to manage help strings).
Except some rare cases, variable names and comments in English.
3.4 Advanced operations
Section contents
Polynomials (page 76)
Loading data files (page 78)
3.4.1 Polynomials
Numpy also contains polynomials in different bases:

For example, 3x2 + 2x 1:
>>> p = np.poly1d([3, 2, -1])
>>> p(0)
-1
>>> p.roots
array([-1. , 0.33333333])
>>> p.order
2
>>> x = np.linspace(0, 1, 20)

>>> y = np.cos(x) + 0.3*np.random.rand(20)
>>> p = np.poly1d(np.polyfit(x, y, 3))
>>> t = np.linspace(0, 1, 200)

>>> plt.plot(x, y, o, t, p(t), -)
[<matplotlib.lines.Line2D object at ...>, <matplotlib.lines.Line2D object at ...>]
3.4. Advanced operations 76

1.3
1.2
1.1
1.0
0.9
0.8
0.7
0.6
0.50.0 0.2 0.4 0.6 0.8 1.0
See http://docs.scipy.org/doc/numpy/reference/routines.polynomials.poly1d.html for more.
More polynomials (with more bases)
Numpy also has a more sophisticated polynomial interface, which supports e.g. the Chebyshev basis.
3x2 + 2x 1:
>>> p = np.polynomial.Polynomial([-1, 2, 3]) # coefs in different order!
>>> p(0)
-1.0
>>> p.roots()
array([-1. , 0.33333333])
>>> p.degree() # In general polynomials do not always expose order
2
Example using polynomials in Chebyshev basis, for polynomials in range [-1, 1]:
>>> x = np.linspace(-1, 1, 2000)
>>> y = np.cos(x) + 0.3*np.random.rand(2000)
>>> p = np.polynomial.Chebyshev.fit(x, y, 90)
>>> t = np.linspace(-1, 1, 200)

>>> plt.plot(x, y, r.)
>>> plt.plot(t, p(t), k-, lw=3)

1.3
1.2
1.1
1.0
0.9
0.8
0.7
0.6
0.51.0 0.5 0.0 0.5 1.0
The Chebyshev polynomials have some advantages in interpolation.
3.4.2 Loading data files
Text files
Example: populations.txt:
# year hare lynx carrot
1900 30e3 4e3 48300
1901 47.2e3 6.1e3 48200
1902 70.2e3 9.8e3 41500
1903 77.4e3 35.2e3 38200

>>> data
array([[ 1900., 30000., 4000., 48300.],
[ 1901., 47200., 6100., 48200.],
[ 1902., 70200., 9800., 41500.],
...
>>> np.savetxt(pop2.txt, data)

>>> data2 = np.loadtxt(pop2.txt)
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
/home/user/stuff/2011-numpy-tutorial/ex
In [3]: ls
populations.txt species.txt
Images
Using Matplotlib:
>>> img = plt.imread(data/elephant.png)
>>> img.shape, img.dtype
((200, 300, 3), dtype(float32))
>>> plt.imshow(img)
>>> plt.savefig(plot.png)
>>> plt.imsave(red_elephant, img[:,:,0], cmap=plt.cm.gray)
This saved only one channel (of RGB):

>>> plt.imshow(plt.imread(red_elephant.png))
Other libraries:
>>> from scipy.misc import imsave
>>> imsave(tiny_elephant.png, img[::6,::6])
>>> plt.imshow(plt.imread(tiny_elephant.png), interpolation=nearest)

0
5
10
15
20
25
30
0 10 20 30 40
Numpys own format
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.mmread
... 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
If you are interested in the Numpy internals, there is a good discussion in Advanced Numpy (page 160).

CHAPTER 4
Matplotlib: plotting
authors Nicolas Rougier, Mike Mller, Gal Varoquaux
Thanks
Many thanks to Bill Wing and Christoph Deil for review and corrections.
81
Chapters contents
Introduction (page 82)
IPython and the pylab mode (page 83)
pylab (page 83)
Simple plot (page 83)
Using defaults (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)
Regular Plots (page 93)
Scatter Plots (page 93)
Bar Plots (page 94)
Contour Plots (page 94)
Imshow (page 95)
Pie Charts (page 95)
Quiver Plots (page 96)
Grids (page 96)
Multi Plots (page 96)
Polar Axis (page 97)
3D Plots (page 97)
Text (page 98)
Beyond this tutorial (page 98)
Tutorials (page 98)
Matplotlib documentation (page 99)
Code documentation (page 99)
Galleries (page 100)
Mailing lists (page 100)
Quick references (page 100)
Line properties (page 101)
Line styles (page 102)
Markers (page 102)
Colormaps (page 102)
4.1 Introduction
Matplotlib is probably the single most used Python package for 2D-graphics. It provides both a very quick 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.
4.1. Introduction 82
4.1.1 IPython and the pylab mode
IPython is an enhanced interactive Python shell that has lots of interesting features including named inputs and
outputs, access to shell commands, improved debugging and many more. When we start it with the command
line argument -pylab (--pylab since IPython version 0.12), it allows interactive matplotlib sessions that have
Matlab/Mathematica-like functionality.
4.1.2 pylab
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 arguments.
Important commands are explained with interactive examples.
4.2 Simple plot
In this section, we want to draw the cosine and sine functions on the same plot. Starting from the default 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 numpy as np
X = np.linspace(-np.pi, np.pi, 256, endpoint=True)

C, S = np.cos(X), np.sin(X)
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).
To run the example, you can type them in an IPython interactive session:
$ ipython --pylab
This brings us to the IPython prompt:

IPython 0.13 -- An enhanced Interactive Python.
? -> Introduction to IPythons features.
%magic -> Information about IPythons magic % functions.
help -> Pythons 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).
or you can download each of the examples and run it using regular python:
$ python exercice_1.py
You can get source for each step by clicking on the corresponding figure.
4.2. Simple plot 83

4.2.1 Using defaults
Hint: Documentation
plot tutorial
plot() command
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, axis and grid
properties, text and font properties and so on.
import pylab as pl
import numpy as np

pl.plot(X, C)
pl.plot(X, S)
pl.show()
4.2.2 Instantiating defaults
Hint: Documentation
Customizing matplotlib
In the script below, weve instantiated (and commented) all the figure settings that influence the appearance of
the plot. 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
import numpy as np
4.2. Simple plot 84

# Create a figure of size 8x6 points, 80 dots per inch

pl.figure(figsize=(8, 6), dpi=80)
# Create a new subplot from a grid of 1x1

pl.subplot(1, 1, 1)

# 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="-")
# Set x limits
pl.xlim(-4.0, 4.0)
# Set x ticks
pl.xticks(np.linspace(-4, 4, 9, endpoint=True))
# Set y limits
pl.ylim(-1.0, 1.0)
# Set y ticks
pl.yticks(np.linspace(-1, 1, 5, endpoint=True))
# Save figure using 72 dots per inch

# savefig("exercice_2.png", dpi=72)
# Show result on screen

pl.show()
4.2.3 Changing colors and line widths
Hint: Documentation
Controlling line properties
Line API
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.figure(figsize=(10, 6), dpi=80)
pl.plot(X, C, color="blue", linewidth=2.5, linestyle="-")
pl.plot(X, S, color="red", linewidth=2.5, linestyle="-")
...
4.2. Simple plot 85

4.2.4 Setting limits
Hint: Documentation
xlim() command
ylim() command
Current limits of the figure are a bit too tight and we want to make some space in order to clearly see all data
points.
...
pl.xlim(X.min() * 1.1, X.max() * 1.1)
pl.ylim(C.min() * 1.1, C.max() * 1.1)
...
4.2.5 Setting ticks
Hint: Documentation
xticks() command
yticks() command
Tick container
Tick locating and formatting
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.
...
pl.xticks([-np.pi, -np.pi/2, 0, np.pi/2, np.pi])
pl.yticks([-1, 0, +1])
...
4.2. Simple plot 86

4.2.6 Setting tick labels
Hint: Documentation
Working with text
xticks() command
yticks() command
set_xticklabels()
set_yticklabels()
Ticks are now properly placed but their label is not very explicit. We could guess that 3.142 is but it would
be better to make it explicit. When we set tick values, we can also provide a corresponding label in the second
argument list. Note that well use latex to allow for nice rendering of the label.
...
pl.xticks([-np.pi, -np.pi/2, 0, np.pi/2, np.pi],
[r$-\pi$, r$-\pi/2$, r$0$, r$+\pi/2$, r$+\pi$])
pl.yticks([-1, 0, +1],
[r$-1$, r$0$, r$+1$])
...
4.2.7 Moving spines
Hint: Documentation
Spines
Axis container
Transformations tutorial
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 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.
4.2. Simple plot 87

...
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))
...
4.2.8 Adding a legend
Hint: Documentation
Legend guide
legend() command
Legend API
Lets add a legend in the upper left corner. This only requires adding the keyword argument label (that will 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")
pl.legend(loc=upper left)
...
4.2.9 Annotate some points
Hint: Documentation
Annotating axis
annotate() command
4.2. Simple plot 88

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. Then,
well use the annotate command to display some text with an arrow.
...
t = 2 * np.pi / 3
pl.plot([t, t], [0, np.cos(t)], color=blue, linewidth=2.5, linestyle="--")
pl.scatter([t, ], [np.cos(t), ], 50, color=blue)
pl.annotate(r$sin(\frac{2\pi}{3})=\frac{\sqrt{3}}{2}$,
xy=(t, np.sin(t)), xycoords=data,
xytext=(+10, +30), textcoords=offset points, fontsize=16,
arrowprops=dict(arrowstyle="->", connectionstyle="arc3,rad=.2"))
pl.plot([t, t],[0, np.sin(t)], color=red, linewidth=2.5, linestyle="--")

pl.scatter([t, ],[np.sin(t), ], 50, color=red)
pl.annotate(r$cos(\frac{2\pi}{3})=-\frac{1}{2}$,
xy=(t, np.cos(t)), xycoords=data,
xytext=(-90, -50), textcoords=offset points, fontsize=16,
arrowprops=dict(arrowstyle="->", connectionstyle="arc3,rad=.2"))
...
4.2.10 Devil is in the details
Hint: Documentation
Artists
BBox
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)
label.set_bbox(dict(facecolor=white, edgecolor=None, alpha=0.65))
...
4.3 Figures, Subplots, Axes and Ticks
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. A figure in matplotlib means the whole window in the user
interface. Within this figure there can be subplots. 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 current
4.3. Figures, Subplots, Axes and Ticks 89

axes and gca in turn calls gcf() to get the current figure. If there is none it calls figure() to make one, strictly
speaking, to make a subplot(111). Lets look at the details.
4.3.1 Figures
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 that
determine what the figure looks like:
Argument Default Description
num 1 number of figure
figsize figure.figsize figure size in in inches (width, height)
dpi figure.dpi resolution in dots per inch
facecolor figure.facecolor color of the drawing background
edgecolor figure.edgecolor color of edge around the drawing background
frameon True draw figure frame or not
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.
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).
As with other objects, you can set figure properties also setp or with the set_something methods.
4.3.2 Subplots
With subplot you can arrange plots in a regular grid. You need to specify the number of rows and
columns and the number of the plot. Note that the gridspec command is a more powerful alternative.
4.3.3 Axes
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.
4.3. Figures, Subplots, Axes and Ticks 90

4.3.4 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
Tick locators control the positions of the ticks. They are set as follows:
ax = pl.gca()
ax.xaxis.set_major_locator(eval(locator))
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.
4.4 Other Types of Plots: examples and exercises
(page 93)
(page 93) (page 94)
4.4. Other Types of Plots: examples and exercises 91

(page 94) (page 95)
(page 96) (page 95)
(page 96) (page 96)
(page 97) (page 97)

(page 98)

4.4.1 Regular Plots
Hint: You need to use the fill_between command.
Starting from the code below, try to reproduce the graphic on the right taking care of filled areas:
n = 256
X = np.linspace(-np.pi, np.pi, n, endpoint=True)
Y = np.sin(2 * X)
pl.plot(X, Y + 1, color=blue, alpha=1.00)

pl.plot(X, Y - 1, color=blue, alpha=1.00)
Click on the figure for solution.
4.4.2 Scatter 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)
Y = np.random.normal(0,1,n)
pl.scatter(X,Y)
Click on figure for solution.

4.4.3 Bar Plots
Hint: You need to take care of text alignment.
Starting from the code below, try to reproduce the graphic on the right by adding labels for red bars.
n = 12
X = np.arange(n)
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.bar(X, +Y1, facecolor=#9999ff, edgecolor=white)

pl.bar(X, -Y2, facecolor=#ff9999, edgecolor=white)
for x, y in zip(X, Y1):

pl.text(x + 0.4, y + 0.05, %.2f % y, ha=center, va=bottom)
pl.ylim(-1.25, +1.25)
4.4.4 Contour Plots
Hint: You need to use the clabel command.
Starting from the code below, try to reproduce the graphic on the right taking care of the colormap (see Colormaps
(page 102) below).
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)

C = pl.contour(X, Y, f(X, Y), 8, colors=black, linewidth=.5)
4.4.5 Imshow
Hint: You need to take care of the origin of the image in the imshow command and use a colorbar
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):
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.
4.4.6 Pie Charts
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.

4.4.7 Quiver Plots
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]
pl.quiver(X, Y)
4.4.8 Grids
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([])
4.4.9 Multi Plots

Hint: You can use several subplots with different partition.
Starting from the code below, try to reproduce the graphic on the right.
pl.subplot(2, 2, 1)
pl.subplot(2, 2, 3)
pl.subplot(2, 2, 4)
4.4.10 Polar Axis
Hint: You only need to modify the axes line
pl.axes([0, 0, 1, 1])
N = 20
theta = np.arange(0., 2 * np.pi, 2 * np.pi / N)
radii = 10 * np.random.rand(N)
width = np.pi / 4 * np.random.rand(N)
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)
4.4.11 3D Plots
Hint: You need to use contourf

from mpl_toolkits.mplot3d import Axes3D
fig = pl.figure()
ax = Axes3D(fig)
X = np.arange(-4, 4, 0.25)
Y = np.arange(-4, 4, 0.25)
X, Y = np.meshgrid(X, Y)
R = np.sqrt(X**2 + Y**2)
Z = np.sin(R)
ax.plot_surface(X, Y, Z, rstride=1, cstride=1, cmap=hot)

See Also:
3D plotting with Mayavi (page 287)
4.4.12 Text
Hint: Have a look at the matplotlib logo.
Try to do the same from scratch !

4.5 Beyond this tutorial
Matplotlib benefits from extensive documentation as well as a large community of users and developpers. Here
are some links of interest:
4.5.1 Tutorials
Pyplot tutorial
Introduction
Controlling line properties
Working with multiple figures and axes
Working with text
Image tutorial
Startup commands
Importing image data into Numpy arrays
Plotting numpy arrays as images
4.5. Beyond this tutorial 98

Text tutorial
Text introduction
Basic text commands
Text properties and layout
Writing mathematical expressions
Text rendering With LaTeX
Annotating text
Artist tutorial
Introduction
Customizing your objects
Object containers
Figure container
Axes container
Axis containers
Tick containers
Path tutorial
Introduction
Bzier example
Compound paths
Transforms tutorial
Introduction
Data coordinates
Axes coordinates
Blended transformations
Using offset transforms to create a shadow effect
The transformation pipeline
4.5.2 Matplotlib documentation
User guide
FAQ
Installation
Usage
How-To
Troubleshooting
Environment Variables
Screenshots
4.5.3 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
argument, allowing for multiple *x*, *y* pairs with an
optional format string. For example, each of the following is
legal::
plot(x, y) # plot x and y using default line style and color
4.5. Beyond this tutorial 99

plot(x, y, bo) # plot x and y using blue circle markers

plot(y) # plot y using x as index array 0..N-1
plot(y, r+) # ditto, but with red plusses
If *x* and/or *y* is 2-dimensional, then the corresponding columns

will be plotted.
...
4.5.4 Galleries
The matplotlib gallery is also incredibly useful when you search how to render a given graphic. Each example
comes with its source.
A smaller gallery is also available here.
4.5.5 Mailing lists
Finally, there is a user mailing list where you can ask for help and a developers mailing list that is more technical.
4.6 Quick references
Here is a set of tables that show main properties and styles.
4.6. Quick references 100

4.6.1 Line properties
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)

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

4.6.2 Line styles
4.6.3 Markers
4.6.4 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.


CHAPTER 5
Scipy : high-level scientific computing
authors Adrien Chauve, Andre Espaze, Emmanuelle Gouillart, Gal Varoquaux, Ralf Gommers
Scipy
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.
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;
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-
mented in Scipy. As non-professional programmers, scientists often tend to re-invent the wheel, which
leads to buggy, non-optimal, difficult-to-share and unmaintainable code. By contrast, Scipys routines are
optimized and tested, and should therefore be used when possible.
Chapters contents
File input/output: scipy.io (page 105)
Special functions: scipy.special (page 105)
Linear algebra operations: scipy.linalg (page 106)
Fast Fourier transforms: scipy.fftpack (page 107)
Optimization and fit: scipy.optimize (page 111)
Statistics and random numbers: scipy.stats (page 115)
Interpolation: scipy.interpolate (page 117)
Numerical integration: scipy.integrate (page 118)
Signal processing: scipy.signal (page 120)
Image processing: scipy.ndimage (page 121)
Summary exercises on scientific computing (page 126)
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
general idea of how to use scipy for scientific computing.
scipy is composed of task-specific sub-modules:
104
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
They all depend on numpy, but are mostly independent of each other. The standard way of importing Numpy and
these Scipy modules is:
>>> from scipy import stats # same for other sub-modules
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
your code.
5.1 File input/output: scipy.io
Loading and saving matlab files:

>>> from scipy import io as spio
>>> a = np.ones((3, 3))
>>> spio.savemat(file.mat, {a: a}) # savemat expects a dictionary
>>> data = spio.loadmat(file.mat, struct_as_record=True)
>>> data[a]
array([[ 1., 1., 1.],
[ 1., 1., 1.],
[ 1., 1., 1.]])
Reading images:
>>> misc.imread(fname.png)
>>> # Matplotlib also has a similar function
>>> plt.imread(fname.png)
See also:
Load text files: numpy.loadtxt()/numpy.savetxt()
Clever loading of text/csv files: numpy.genfromtxt()/numpy.recfromcsv()
Fast and efficient, but numpy-specific, binary format: numpy.save()/numpy.load()
5.2 Special functions: scipy.special
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:
5.1. File input/output: scipy.io 105

Bessel function, such as scipy.special.jn() (nth integer order Bessel function)

Elliptic function (scipy.special.ellipj() for the Jacobian elliptic function, ...)
Gamma function: scipy.special.gamma(), also note scipy.special.gammaln() which will
give the log of Gamma to a higher numerical precision.
Erf, the area under a Gaussian curve: scipy.special.erf()
5.3 Linear algebra operations: scipy.linalg
The scipy.linalg module provides standard linear algebra operations, relying on an underlying efficient
implementation (BLAS, LAPACK).
The scipy.linalg.det() function computes the determinant of a square matrix:
>>> from scipy import linalg
>>> arr = np.array([[1, 2],
... [3, 4]])
>>> linalg.det(arr)
-2.0
>>> arr = np.array([[3, 2],
... [6, 4]])
>>> linalg.det(arr)
0.0
>>> linalg.det(np.ones((3, 4)))
...
ValueError: expected square matrix
The scipy.linalg.inv() function computes the inverse of a square matrix:

>>> arr = np.array([[1, 2],
... [3, 4]])
>>> iarr = linalg.inv(arr)
>>> iarr
array([[-2. , 1. ],
[ 1.5, -0.5]])
>>> np.allclose(np.dot(arr, iarr), np.eye(2))
True
Finally computing the inverse of a singular matrix (its determinant is zero) will raise LinAlgError:
>>> arr = np.array([[3, 2],
... [6, 4]])
>>> linalg.inv(arr)
...
LinAlgError: singular matrix
More advanced operations are available, for example singular-value decomposition (SVD):
>>> arr = np.arange(9).reshape((3, 3)) + np.diag([1, 0, 1])
>>> uarr, spec, vharr = linalg.svd(arr)
The resulting array spectrum is:

>>> spec
array([ 14.88982544, 0.45294236, 0.29654967])
The original matrix can be re-composed by matrix multiplication of the outputs of svd with np.dot:
>>> sarr = np.diag(spec)
>>> svd_mat = uarr.dot(sarr).dot(vharr)
5.3. Linear algebra operations: scipy.linalg 106

>>> np.allclose(svd_mat, arr)

True
SVD is commonly used in statistics and signal processing. Many other standard decompositions (QR, LU,
Cholesky, Schur), as well as solvers for linear systems, are available in scipy.linalg.
5.4 Fast Fourier transforms: scipy.fftpack
The scipy.fftpack module allows to compute fast Fourier transforms. As an illustration, a (noisy) input
signal may look like:
>>> time_step = 0.02
>>> period = 5.
>>> time_vec = np.arange(0, 20, time_step)
>>> sig = np.sin(2 * np.pi / period * time_vec) + \
... 0.5 * np.random.randn(time_vec.size)
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
scipy.fftpack.fft() will compute the fast Fourier transform:
>>> from scipy import fftpack
>>> sample_freq = fftpack.fftfreq(sig.size, d=time_step)
>>> sig_fft = fftpack.fft(sig)
Because the resulting power is symmetric, only the positive part of the spectrum needs to be used for finding the
frequency:
>>> pidxs = np.where(sample_freq > 0)
>>> freqs = sample_freq[pidxs]
>>> power = np.abs(sig_fft)[pidxs]
600
Peak frequency
500
400
plower
300
200
0.050.100.150.200.250.300.350.400.45
100
00 5 10 15 20 25
Frequency [Hz]
The signal frequency can be found by:
>>> freq = freqs[power.argmax()]
>>> 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:
5.4. Fast Fourier transforms: scipy.fftpack 107

>>> sig_fft[np.abs(sample_freq) > freq] = 0
The resulting filtered signal can be computed by the scipy.fftpack.ifft() function:

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

>>> import pylab as plt
>>> plt.figure()
>>> plt.plot(time_vec, sig)
>>> plt.plot(time_vec, main_sig, linewidth=3)
>>> plt.xlabel(Time [s])
>>> plt.ylabel(Amplitude)
1
Amplitude
30 5 10 15 20
Time [s]
numpy.fft
Numpy also has an implementation of FFT (numpy.fft). However, in general the scipy one should be
preferred, as it uses more efficient underlying implementations.

Worked example: Crude periodicity finding
80
hare
70 lynx
60 carrot
Population number ( 103 )
50
40
30
20
10
0
1900 1905 1910 1915 1920
Year
300
250
200
Power ( 103 )
150
100
50
00 5 10 15 20
Period

Worked example: Gaussian image blur

Convolution:
Z
f1 (t) = dt0 K(t t0 )f0 (t0 )
f1 () = K()
f0 ()
50
100
150
0 50 100 150 200 250

Exercise: Denoise moon landing image
1. Examine the provided image moonlanding.png, which is heavily contaminated with periodic noise. In
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?
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.
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()
5.5. Optimization and fit: scipy.optimize 111

120
100
80
60
40
20
0
2010 5 0 5 10
This function has a global minimum around -1.3 and a local minimum around 3.8.
The general and efficient way to find a minimum for this function is to conduct a gradient descent starting from a
given initial point. The BFGS algorithm is a good way of doing this:
>>> optimize.fmin_bfgs(f, 0)
Optimization terminated successfully.
Current function value: -7.945823
Iterations: 5
Function evaluations: 24
Gradient evaluations: 8
array([-1.30644003])
A possible issue with this approach is that, if the function has local minima the algorithm may find these local
minima instead of the global minimum depending on the initial point:
>>> optimize.fmin_bfgs(f, 3, disp=0)
array([ 3.83746663])
If we dont know the neighborhood of the global minimum to choose the initial point, we need to resort to costlier
global optimization. To find the global minimum, the simplest algorithm is the brute force algorithm, in which the
function is evaluated on each point of a given grid:
>>> grid = (-10, 10, 0.1)
>>> xmin_global = optimize.brute(f, (grid,))
>>> xmin_global
array([-1.30641113])
For larger grid sizes, scipy.optimize.brute() becomes quite slow. scipy.optimize.anneal()

provides an alternative, using simulated annealing. More efficient algorithms for different classes of global opti-
mization problems exist, but this is out of the scope of scipy. Some useful packages for global optimization are
OpenOpt, IPOPT, PyGMO and PyEvolve.
To find the local minimum, lets constraint the variable to the interval (0, 10) using

scipy.optimize.fminbound():
>>> xmin_local = optimize.fminbound(f, 0, 10)
>>> xmin_local
3.8374671...
Note: Finding minima of function is discussed in more details in the advanced chapter: Mathematical optimiza-
tion: finding minima of functions (page 252).
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
scipy.optimize.fsolve():
>>> root = optimize.fsolve(f, 1) # our initial guess is 1
>>> root
array([ 0.])
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:
>>> root2 = optimize.fsolve(f, -2.5)
>>> root2
array([-2.47948183])
Curve fitting
Suppose we have data sampled from f with some noise:
>>> xdata = np.linspace(-10, 10, num=20)
>>> ydata = f(xdata) + np.random.randn(xdata.size)
Now if we know the functional form of the function from which the samples were drawn (x^2 + sin(x) in
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:
>>> def f2(x, a, b):
... return a*x**2 + b*np.sin(x)
Then we can use scipy.optimize.curve_fit() to find a and b:

>>> guess = [2, 2]
>>> params, params_covariance = optimize.curve_fit(f2, xdata, ydata, guess)
>>> params
array([ 0.99925147, 9.76065551])
Now we have found the minima and roots of f and used curve fitting on it, we put all those resuls together in a
single plot:

120
f(x)
100 Curve fit result
Minima
80 Roots
60
f(x)
40
20
0
2010 5 0 5 10
x
Note: In Scipy >= 0.11 unified interfaces to all minimization and root finding algorithms
are available: scipy.optimize.minimize(), scipy.optimize.minimize_scalar() and
scipy.optimize.root(). They allow comparing various algorithms easily through the method keyword.
You can find algorithms with the same functionalities for multi-dimensional problems in scipy.optimize.
Exercise: Curve fitting of temperature data

The temperature extremes in Alaska for each month, starting in January, are given by (in degrees
Celcius):
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
1. Plot these temperature extremes.
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().
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?

Exercise: 2-D minimization
Six-hump Camelback function
5
4
3
f(x, y)
2
1
0
1
1.0
0.5
2.01.51.0 0.0y
0.50.0 0.5
x 0.51.01.5 1.0
2.0
The six-hump camelback function
x4 2
f (x, y) = (4 2.1x2 + )x + xy + (4y 2 4)y 2
3
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.
Use numpy.meshgrid() and pylab.imshow() to find visually the regions.
Use scipy.optimize.fmin_bfgs() or another multi-dimensional minimizer.
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)?
See the summary exercise on Non linear least squares curve fitting: application to point extraction in topograph-
ical lidar data (page 129) for another, more advanced example.
5.6 Statistics and random numbers: scipy.stats
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.
5.6.1 Histogram and probability density function
Given observations of a random process, their histogram is an estimator of the random processs PDF (probability
density function):
5.6. Statistics and random numbers: scipy.stats 115

>>> a = np.random.normal(size=1000)
>>> bins = np.arange(-4, 5)
>>> bins
array([-4, -3, -2, -1, 0, 1, 2, 3, 4])
>>> histogram = np.histogram(a, bins=bins, normed=True)[0]
>>> bins = 0.5*(bins[1:] + bins[:-1])
>>> bins
array([-3.5, -2.5, -1.5, -0.5, 0.5, 1.5, 2.5, 3.5])
>>> from scipy import stats
>>> b = stats.norm.pdf(bins) # norm is a distribution
In [1]: pl.plot(bins, histogram)

In [2]: pl.plot(bins, b)
0.40
0.35
0.30
0.25
0.20
0.15
0.10
0.05
0.00 6 4 2 0 2 4 6
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.
Here we fit a normal process to the observed data:
>>> loc, std = stats.norm.fit(a)
>>> loc
-0.045256707490...
>>> std
0.9870331586690...
Exercise: Probability distributions

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)?
Extra: the distributions have a number of useful methods. Explore them by reading the docstring or by using
IPython tab completion. Can you find the shape parameter of 1 back by using the fit method on your
random variates?
5.6.2 Percentiles
The median is the value with half of the observations below, and half above:
>>> np.median(a)
-0.058028034...
5.6. Statistics and random numbers: scipy.stats 116

It is also called the percentile 50, because 50% of the observation are below it:
>>> stats.scoreatpercentile(a, 50)
-0.0580280347...
Similarly, we can calculate the percentile 90:

>>> stats.scoreatpercentile(a, 90)
1.231593551...
The percentile is an estimator of the CDF: cumulative distribution function.
5.6.3 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
different:
>>> a = np.random.normal(0, 1, size=100)
>>> b = np.random.normal(1, 1, size=10)
>>> stats.ttest_ind(a, b)
(-3.75832707..., 0.00027786...)
The resulting output is composed of:

The T statistic value: it is a number the sign of which is proportional to the difference between the two
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
certainly identical. The closer it is to zero, the more likely it is that the processes have different means.
5.7 Interpolation: scipy.interpolate
The scipy.interpolate is useful for fitting a function from experimental data and thus evaluating points
where no measure exists. The module is based on the FITPACK Fortran subroutines from the netlib project.
By imagining experimental data close to a sine function:
>>> 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
The scipy.interpolate.interp1d class can build a linear interpolation function:

>>> from scipy.interpolate import interp1d
>>> linear_interp = interp1d(measured_time, measures)
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)
A cubic interpolation can also be selected by providing the kind optional keyword argument:
>>> cubic_interp = interp1d(measured_time, measures, kind=cubic)
>>> cubic_results = cubic_interp(computed_time)
The results are now gathered on the following Matplotlib figure:
5.7. Interpolation: scipy.interpolate 117

1.0
measures
linear interp
0.5 cubic interp
0.0
0.5
1.0
1.50.0 0.2 0.4 0.6 0.8 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-
mary exercise on Maximum wind speed prediction at the Sprog station (page 126) for a more advance spline
interpolation example.
5.8 Numerical integration: scipy.integrate
The most generic integration routine is scipy.integrate.quad():

>>> from scipy.integrate import quad
>>> res, err = quad(np.sin, 0, np.pi/2)
>>> np.allclose(res, 1)
True
>>> np.allclose(err, 1 - res)
True
Others integration schemes are available with fixed_quad, quadrature, romberg.

scipy.integrate also features routines for integrating Ordinary Differential Equations (ODE). In particular,
scipy.integrate.odeint() is a general-purpose integrator using LSODA (Livermore Solver for Ordinary
Differential equations with Automatic method switching for stiff and non-stiff problems), see the ODEPACK
Fortran library for more details.
odeint solves first-order ODE systems of the form:
dy/dt = rhs(y1, y2, .., t0,...)
As an introduction, let us solve the ODE dy/dt = -2y between t = 0..4, with the initial condition y(t=0)
= 1. First the function computing the derivative of the position needs to be defined:
>>> def calc_derivative(ypos, time, counter_arr):
... counter_arr += 1
5.8. Numerical integration: scipy.integrate 118

... return -2 * ypos

...
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:
>>> counter = np.zeros((1,), dtype=np.uint16)
The trajectory will now be computed:

>>> from scipy.integrate import odeint
>>> time_vec = np.linspace(0, 4, 40)
>>> 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):
>>> counter
array([129], dtype=uint16)
and the cumulative number of iterations for each of the 10 first time steps can be obtained by:
>>> 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
be plotted:
1.0
0.8
y position [m]
0.6
0.4
0.2
0.00.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0

Time [s]
Another example with scipy.integrate.odeint() will be a damped spring-mass oscillator (2nd order
oscillator). The position of a mass attached to a spring obeys the 2nd order ODE y + 2 eps wo y +
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
damping coefficient. For this example, we choose the parameters as:
>>> mass = 0.5 # kg
>>> kspring = 4 # N/m
>>> cviscous = 0.4 # N s/m
so the system will be underdamped, because:

>>> eps = cviscous / (2 * mass * np.sqrt(kspring/mass))
>>> eps < 1
True
5.8. Numerical integration: scipy.integrate 119

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
/ m and om = wo^2 = k/m:
>>> nu_coef = cviscous / mass
>>> om_coef = kspring / mass
Thus the function will calculate the velocity and acceleration by:
>>> def calc_deri(yvec, time, nuc, omc):
... return (yvec[1], -nuc * yvec[1] - omc * yvec[0])
...
>>> time_vec = np.linspace(0, 10, 100)
>>> yarr = odeint(calc_deri, (1, 0), time_vec, args=(nu_coef, om_coef))
The final position and velocity are shown on the following Matplotlib figure:
1.5
y
1.0 y'
0.5
0.0
0.5
1.0
1.5
2.0
2.50 2 4 6 8 10
There is no Partial Differential Equations (PDE) solver in Scipy. Some Python packages for solving PDEs are
available, such as fipy or SfePy.
5.9 Signal processing: scipy.signal

>>> from scipy import signal
scipy.signal.detrend(): remove linear trend from signal:

t = np.linspace(0, 5, 100)
x = t + np.random.normal(size=100)
pl.plot(t, x, linewidth=3)
pl.plot(t, signal.detrend(x), linewidth=3)
5.9. Signal processing: scipy.signal 120

40 1 2 3 4 5
scipy.signal.resample(): resample a signal to n points using FFT.

x = np.sin(t)
pl.plot(t, x, linewidth=3)
pl.plot(t[::2], signal.resample(x, 50), ko)
1.5
1.0
0.5
0.0
0.5
1.0
1.50 1 2 3 4 5
Notice how on the side of the window the resampling is less accurate and has a rippling effect.
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.
5.10 Image processing: scipy.ndimage
The submodule dedicated to image processing in scipy is scipy.ndimage.
5.10. Image processing: scipy.ndimage 121

>>> from scipy import ndimage
Image processing routines may be sorted according to the category of processing they perform.
5.10.1 Geometrical transformations on images
Changing orientation, resolution, ..

>>> shifted_lena = ndimage.shift(lena, (50, 50))
>>> shifted_lena2 = ndimage.shift(lena, (50, 50), mode=nearest)
>>> rotated_lena = ndimage.rotate(lena, 30)
>>> cropped_lena = lena[50:-50, 50:-50]
>>> zoomed_lena = ndimage.zoom(lena, 2)
>>> zoomed_lena.shape
(1024, 1024)
In [35]: subplot(151)
Out[35]: <matplotlib.axes.AxesSubplot object at 0x925f46c>
In [36]: pl.imshow(shifted_lena, cmap=cm.gray)

Out[36]: <matplotlib.image.AxesImage object at 0x9593f6c>
In [37]: axis(off)
Out[37]: (-0.5, 511.5, 511.5, -0.5)
In [39]: # etc.
5.10.2 Image filtering

>>> noisy_lena = np.copy(lena).astype(np.float)
>>> noisy_lena += lena.std()*0.5*np.random.standard_normal(lena.shape)
>>> blurred_lena = ndimage.gaussian_filter(noisy_lena, sigma=3)
>>> median_lena = ndimage.median_filter(blurred_lena, size=5)
>>> from scipy import signal
>>> wiener_lena = signal.wiener(blurred_lena, (5,5))
Many other filters in scipy.ndimage.filters and scipy.signal can be applied to images.

Exercise
Compare histograms for the different filtered images.
5.10.3 Mathematical morphology
Mathematical morphology is a mathematical theory that stems from set theory. It characterizes and transforms
geometrical structures. Binary (black and white) images, in particular, can be transformed using this theory:
the sets to be transformed are the sets of neighboring non-zero-valued pixels. The theory was also extended to
gray-valued images.
Elementary mathematical-morphology operations use a structuring element in order to modify other geometrical
structures.
Let us first generate a 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)
array([[0, 1, 0],
[1, 1, 1],
[0, 1, 0]])
Erosion
>>> 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]])
>>> ndimage.binary_erosion(a).astype(a.dtype)
array([[0, 0, 0, 0, 0, 0, 0],
[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],
[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0]])
>>> #Erosion removes objects smaller than the structure

>>> ndimage.binary_erosion(a, structure=np.ones((5,5))).astype(a.dtype)

array([[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],
[0, 0, 0, 0, 0, 0, 0]])
Dilation
>>> 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.]])
>>> ndimage.binary_dilation(a).astype(a.dtype)
array([[ 0., 0., 0., 0., 0.],
[ 0., 0., 1., 0., 0.],
[ 0., 1., 1., 1., 0.],
[ 0., 0., 1., 0., 0.],
[ 0., 0., 0., 0., 0.]])
Opening
>>> a[1:4, 1:4] = 1; a[4, 4] = 1
>>> a
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],
[0, 1, 1, 1, 0],
[0, 0, 0, 0, 0]])
>>> # Opening can also smooth corners
>>> ndimage.binary_opening(a).astype(np.int)
array([[0, 0, 0, 0, 0],
[0, 0, 1, 0, 0],
[0, 1, 1, 1, 0],
[0, 0, 1, 0, 0],
[0, 0, 0, 0, 0]])
Closing: ndimage.binary_closing
Exercise
Check that opening amounts to eroding, then dilating.
An opening operation removes small structures, while a closing operation fills small holes. Such operations can
therefore be used to clean an image.
>>> a = np.zeros((50, 50))
>>> a[10:-10, 10:-10] = 1
>>> a += 0.25*np.random.standard_normal(a.shape)

>>> mask = a>=0.5

>>> opened_mask = ndimage.binary_opening(mask)
>>> closed_mask = ndimage.binary_closing(opened_mask)
Exercise
Check that the area of the reconstructed square is smaller than the area of the initial square. (The
opposite would occur if the closing step was performed before the opening).
For gray-valued images, eroding (resp. dilating) amounts to replacing a pixel by the minimal (resp. maximal)
value among pixels covered by the structuring element centered on the pixel of interest.
>>> a[1:6, 1:6] = 3
>>> a[4,4] = 2; a[2,3] = 1
>>> a
array([[0, 0, 0, 0, 0, 0, 0],
[0, 3, 3, 3, 3, 3, 0],
[0, 3, 3, 1, 3, 3, 0],
[0, 3, 3, 3, 3, 3, 0],
[0, 3, 3, 3, 2, 3, 0],
[0, 3, 3, 3, 3, 3, 0],
[0, 0, 0, 0, 0, 0, 0]])
>>> ndimage.grey_erosion(a, size=(3,3))
array([[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0],
[0, 0, 1, 1, 1, 0, 0],
[0, 0, 1, 1, 1, 0, 0],
[0, 0, 3, 2, 2, 0, 0],
[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0]])
5.10.4 Measurements on images
Let us first generate a nice synthetic binary image.

>>> x, y = np.indices((100, 100))
>>> sig = np.sin(2*np.pi*x/50.)*np.sin(2*np.pi*y/50.)*(1+x*y/50.**2)**2
>>> mask = sig > 1
Now we look for various information about the objects in the image:
>>> labels, nb = ndimage.label(mask)
>>> nb
8
>>> areas = ndimage.sum(mask, labels, xrange(1, labels.max()+1))
>>> areas
array([ 190., 45., 424., 278., 459., 190., 549., 424.])
>>> maxima = ndimage.maximum(sig, labels, xrange(1, labels.max()+1))
>>> maxima
array([ 1.80238238, 1.13527605, 5.51954079, 2.49611818,

6.71673619, 1.80238238, 16.76547217, 5.51954079])

>>> ndimage.find_objects(labels==4)
[(slice(30L, 48L, None), slice(30L, 48L, None))]
>>> sl = ndimage.find_objects(labels==4)
>>> pl.imshow(sig[sl[0]])
See the summary exercise on Image processing application: counting bubbles and unmolten grains (page 134) for
a more advanced example.
5.11 Summary exercises on scientific computing
The summary exercises use mainly Numpy, Scipy and Matplotlib. They provide some real-life examples of
scientific computing with Python. Now that the basics of working with Numpy and Scipy have been introduced,
the interested user is invited to try these exercises.
5.11.1 Maximum wind speed prediction at the Sprog station
The exercise goal is to predict the maximum wind speed occurring every 50 years even if no measure exists for
such a period. The available data are only measured over 21 years at the Sprog meteorological station located
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
approach.
Statistical approach
The annual maxima are supposed to fit a normal probability density function. However such function is not going
to be estimated because it gives a probability from a wind speed maxima. Finding the maximum wind speed
occurring every 50 years requires the opposite approach, the result needs to be found from a defined probability.
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.
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
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
scipy.interpolate module will be very useful for fitting the quantile function. Finally the 50 years maxima is going
to be evaluated from the cumulative probability of the 2% quantile.
5.11. Summary exercises on scientific computing 126

Computing the cumulative probabilities
The annual wind speeds maxima have already been computed and saved in the numpy format in the file
examples/max-speeds.npy, thus they will be loaded by using numpy:
>>> max_speeds = np.load(intro/summary-exercises/examples/max-speeds.npy)
>>> years_nb = max_speeds.shape[0]
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
BivariateSpline class family is provided. All those classes for 1D and 2D splines use the FITPACK For-
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
parameters are also provided for simpler use (see interp1d, interp2d, barycentric_interpolate
and so on).
For the Sprog maxima wind speeds, the UnivariateSpline will be used because a spline of degree 3 seems
to correctly fit the data:
>>> from scipy.interpolate import UnivariateSpline
>>> quantile_func = UnivariateSpline(cprob, sorted_max_speeds)
The quantile function is now going to be evaluated from the full range of probabilities:
>>> nprob = np.linspace(0, 1, 1e2)
>>> fitted_max_speeds = quantile_func(nprob)
2%
In the current model, the maximum wind speed occurring every 50 years is defined as the upper 2% quantile. As
a result, the cumulative probability value will be:
>>> fifty_prob = 1. - 0.02
So the storm wind speed occurring every 50 years can be guessed by:
>>> fifty_wind = quantile_func(fifty_prob)
>>> fifty_wind
32.97989825...
The results are now gathered on a Matplotlib figure:
Exercise with the Gumbell distribution
The interested readers are now invited to make an exercise by using the wind speeds measured over 21 years.
The measurement period is around 90 minutes (the original period was around 10 minutes but the file size
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 5.1: 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.
The second step will be to use the Gumbell distribution on cumulative probabilities p_i defined as -log(
-log(p_i) ) for fitting a linear quantile function (remember that you can define the degree of the
UnivariateSpline). Plotting the annual maxima versus the Gumbell distribution should give you
the following figure.
The last step will be to find 34.23 m/s for the maximum wind speed occurring every 50 years.
5.11.2 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 131).
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).
1 Mallet, C. and Bretar, F. Full-Waveform Topographic Lidar: State-of-the-Art. ISPRS Journal of Photogrammetry and Remote Sensing
64(1), pp.1-16, January 2009 http://dx.doi.org/10.1016/j.isprsjprs.2008.09.007


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
footprint of the laser beam is around 1m on the Earth surface, the beam can hit multiple targets during the two-way
propagation (for example the ground and the top of a tree or building). The sum of the contributions of each target
hit by the laser beam then produces a complex signal with multiple peaks, each one containing information about
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()
As you can notice, this waveform is a 80-bin-length signal with a single peak.
2 The data used for this tutorial are part of the demonstration data available for the FullAnalyze software and were kindly provided by the
GIS DRAIX.

Fitting a waveform with a simple Gaussian model
The signal is very simple and can be modeled as a single Gaussian function and an offset corresponding to the
background noise. To fit the signal with the function, we must:
define the model
propose an initial solution
call scipy.optimize.leastsq
Model
A Gaussian function defined by

( 2 )
t
B + A 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 B (noise)
coeffs[1] is A (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. Basically,
the function to minimize is the residuals (the difference between the data and the model):
>>> def residuals(coeffs, y, t):
... return y - model(t, coeffs)
So lets get our solution by calling scipy.optimize.leastsq with the following arguments:
the function to minimize
an initial solution
the additional arguments to pass to the function
>>> from scipy.optimize import leastsq
>>> x, flag = leastsq(residuals, x0, args=(waveform_1, t))
>>> print x
[ 2.70363341 27.82020742 15.47924562 3.05636228]
And visualize the solution:

>>> plt.plot(t, waveform_1, t, model(t, x))

>>> plt.legend([waveform, model])
>>> 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.
Going further
Try with a more complex waveform (for instance data/waveform_2.npy) that contains three signif-
icant peaks. You must adapt the model which is now a sum of Gaussian functions instead of only one
Gaussian peak.
In some cases, writing an explicit function to compute the Jacobian is faster than letting leastsq esti-
mate it numerically. Create a function to compute the Jacobian of the residuals and use it as an input for
leastsq.
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 model enables to overcome such limitations. An example of a priori knowledge we can add is the sign
of our variables (which are all positive).
With the following initial solution:
>>> x0 = np.array([3, 50, 20, 1], dtype=float)
compare the result of scipy.optimize.leastsq and what you can get with
scipy.optimize.fmin_slsqp when adding boundary constraints.

5.11.3 Image processing application: counting bubbles and unmolten grains
Statement of the problem
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).
This Scanning Element Microscopy image shows a glass sample (light gray matrix) with some bubbles (on black)
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.
2. Crop the image to remove the lower panel with measure information.
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


5.11.4 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).
>>> dat = pl.imread(data/MV_HFV_012.jpg)
2. Crop the image to remove the lower panel with measure information.
>>> dat = dat[60:]
3. Slightly filter the image with a median filter in order to refine its histogram. Check how the histogram
changes.
>>> filtdat = ndimage.median_filter(dat, size=(7,7))
>>> hi_dat = np.histogram(dat, bins=np.arange(256))
>>> hi_filtdat = np.histogram(filtdat, bins=np.arange(256))

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.
>>> void = filtdat <= 50
>>> sand = np.logical_and(filtdat > 50, filtdat <= 114)
>>> glass = filtdat > 114
5. Display an image in which the three phases are colored with three different colors.
>>> phases = void.astype(np.int) + 2*glass.astype(np.int) + 3*sand.astype(np.int)

6. Use mathematical morphology to clean the different phases.

>>> sand_op = ndimage.binary_opening(sand, iterations=2)
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.
>>> sand_labels, sand_nb = ndimage.label(sand_op)
>>> sand_areas = np.array(ndimage.sum(sand_op, sand_labels, np.arange(sand_labels.max()+1)))
>>> mask = sand_areas > 100
>>> 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)

CHAPTER 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:
In Ipython, help function opens the docstring of the function. Only type the beginning of the func-
tions name and use tab completion to display the matching functions.
In [204]: help np.v
np.vander np.vdot np.version np.void0 np.vstack
np.var np.vectorize np.void np.vsplit
In [204]: help np.vander
In Ipython it is not possible to open a separated window for help and documentation; however one can always
open a second Ipython shell just to display help and docstrings...
Numpys and Scipys documentations can be browsed online on http://docs.scipy.org/doc.
The search button is quite useful inside the reference documentation of the two packages
(http://docs.scipy.org/doc/numpy/reference/ and http://docs.scipy.org/doc/scipy/reference/).
Tutorials on various topics as well as the complete API with all docstrings are found on this website.
139
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
you may want to read directly the documentation on the wiki instead of the official documentation website.
Note that anyone can create an account on the wiki and write better documentation; this is an easy way to
contribute to an open-source project and improve the tools you are using!
Scipys cookbook http://www.scipy.org/Cookbook gives recipes on many common problems frequently en-
countered, such as fitting data points, solving ODE, etc.
Matplotlibs website http://matplotlib.sourceforge.net/ features a very nice gallery with a large number of
plots, each of them shows both the source code and the resulting plot. This is very useful for learning by
example. More standard documentation is also available.
Mayavis website http://code.enthought.com/projects/mayavi/docs/development/html/mayavi/ also has a

very nice gallery of examples http://code.enthought.com/projects/mayavi/docs/development/html/mayavi/auto/examples.html
in which one can browse for different visualization solutions.
140
Finally, two more technical possibilities are useful as well:

In Ipython, the magical function %psearch search for objects matching patterns. This is useful if, for
example, one does not know the exact name of a function.
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
Returns the discrete, linear convolution of two one-dimensional
sequences.
numpy.bartlett
Return the Bartlett window.
numpy.correlate
Discrete, linear correlation of two 1-dimensional sequences.
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.
If everything listed above fails (and Google doesnt have the answer)... dont despair! Write to the mailing-
list suited to your problem: you should have a quick answer if you describe your problem well. Experts on
scientific python often give very enlightening explanations on the mailing-list.
Numpy discussion (numpy-discussion@scipy.org): all about numpy arrays, manipulating them, in-
dexation questions, etc.
141
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.
142
Part II
Advanced topics
143
CHAPTER 7
Advanced Python Constructs
author Zbigniew Jedrzejewski-Szmek

This chapter is about some features of the Python language which can be considered advanced in the sense that
not every language has them, and also in the sense that they are more useful in more complicated programs or
libraries, but not in the sense of being particularly specialized, or particularly complicated.
It is important to underline that this chapter is purely about the language itself about features supported through
special syntax complemented by functionality of the Python stdlib, which could not be implemented through
clever external modules.
The process of developing the Python programming language, its syntax, is unique because it is very transparent,
proposed changes are evaluated from various angles and discussed on public mailing lists, and the final decision
takes into account the balance between the importance of envisioned use cases, the burden of carrying more
language features, consistency with the rest of the syntax, and whether the proposed variant is the easiest to read,
write, and understand. This process is formalised in Python Enhancement Proposals PEPs. As a result, features
described in this chapter were added after it was shown that they indeed solve real problems and that their use is
as simple as possible.
Chapters contents
Iterators, generator expressions and generators (page 145)
Iterators (page 145)
Generator expressions (page 146)
Generators (page 146)
Bidirectional communication (page 147)
Chaining generators (page 149)
Decorators (page 149)
Replacing or tweaking the original object (page 150)
Decorators implemented as classes and as functions (page 150)
Copying the docstring and other attributes of the original function (page 152)
Examples in the standard library (page 153)
Deprecation of functions (page 155)
A while-loop removing decorator (page 156)
A plugin registration system (page 156)
More examples and reading (page 157)
Context managers (page 157)
Catching exceptions (page 158)
Using generators to define context managers (page 159)
144
7.1 Iterators, generator expressions and generators
7.1.1 Iterators
Simplicity
Duplication of effort is wasteful, and replacing the various home-grown approaches with a standard feature
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.
An iterator object allows to loop just once. It holds the state (position) of a single iteration, or from the other side,
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.
>>> nums = [1,2,3] # note that ... varies: these are different objects
>>> iter(nums)
<listiterator object at ...>
>>> nums.__iter__()
<listiterator object at ...>
>>> nums.__reversed__()
<listreverseiterator object at ...>
>>> it = iter(nums)
>>> next(it) # next(obj) simply calls obj.next()
1
>>> it.next()
2
>>> next(it)
3
>>> next(it)
StopIteration
When used in a loop, StopIteration is swallowed and causes the loop to finish. But with explicit invocation,
we can see that once the iterator is exhausted, accessing it raises an exception.
Using the for..in loop also uses the __iter__ method. This allows us to transparently start the iteration over a
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
to achieve this, iterators in addition to next are also required to have a method called __iter__ which returns
the iterator (self).
Support for iteration is pervasive in Python: all sequences and unordered containers in the standard library allow
this. The concept is also stretched to other things: e.g. file objects support iteration over lines.
>>> f = open(/etc/fstab)
>>> f is f.__iter__()
True
The file is an iterator itself and its __iter__ method doesnt create a separate object: only a single thread of
sequential access is allowed.
7.1. Iterators, generator expressions and generators 145

7.1.2 Generator expressions
A second way in which iterator objects are created is through generator expressions, the basis for list compre-
hensions. To increase clarity, a generator expression must always be enclosed in parentheses or an expression. If
round parentheses are used, then a generator iterator is created. If rectangular parentheses are used, the process is
short-circuited and we get a list.
>>> (i for i in nums)
<generator object <genexpr> at 0x...>
>>> [i for i in nums]
[1, 2, 3]
>>> list(i for i in nums)
[1, 2, 3]
In Python 2.7 and 3.x the list comprehension syntax was extended to dictionary and set comprehensions. A
set is created when the generator expression is enclosed in curly braces. A dict is created when the generator
expression contains pairs of the form key:value:
>>> {i for i in range(3)}
set([0, 1, 2])
>>> {i:i**2 for i in range(3)}
{0: 0, 1: 1, 2: 4}
If you are stuck at some previous Python version, the syntax is only a bit worse:
>>> set(i for i in abc)
set([a, c, b])
>>> dict((i, ord(i)) for i in abc)
{a: 97, c: 99, b: 98}
Generator expression are fairly simple, not much to say here. Only one gotcha should be mentioned: in old
Pythons the index variable (i) would leak, and in versions >= 3 this is fixed.
7.1.3 Generators
Generators
A generator is a function that produces a sequence of results instead of a single value.
David Beazley A Curious Course on Coroutines and Concurrency
A third way to create iterator objects is to call a generator function. A generator is a function containing the
keyword yield. It must be noted that the mere presence of this keyword completely changes the nature of the
function: this yield statement doesnt have to be invoked, or even reachable, but causes the function to be
marked as a generator. When a normal function is called, the instructions contained in the body start to be
executed. When a generator is called, the execution stops before the first instruction in the body. An invocation
of a generator function creates a generator object, adhering to the iterator protocol. As with normal function
invocations, concurrent and recursive invocations are allowed.
When next is called, the function is executed until the first yield. Each encountered yield statement gives a
value becomes the return value of next. After executing the yield statement, the execution of this function is
suspended.
>>> def f():
... yield 1
... yield 2
>>> f()
<generator object f at 0x...>
>>> gen = f()
>>> gen.next()
1

>>> gen.next()
2
>>> gen.next()
StopIteration
Lets go over the life of the single invocation of the generator function.
>>> def f():
... print("-- start --")
... yield 3
... print("-- middle --")
... yield 4
... print("-- finished --")
>>> gen = f()
>>> next(gen)
-- start --
3
>>> next(gen)
-- middle --
4
>>> next(gen)
-- finished --
...
StopIteration
Contrary to a normal function, where executing f() would immediately cause the first print to be executed,
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
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
is stored in the generator object. From the point of view of the generator function, is looks almost as if it was
running in a separate thread, but this is just an illusion: execution is strictly single-threaded, but the interpreter
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
create an iterator object. Everything that can be done with yield statements, could also be done with next
methods. Nevertheless, using a function and having the interpreter perform its magic to create an iterator has
advantages. A function can be much shorter than the definition of a class with the required next and __iter__
methods. What is more important, it is easier for the author of the generator to understand the state which is
kept in local variables, as opposed to instance attributes, which have to be used to pass data between consecutive
invocations of next on an iterator object.
A broader question is why are iterators useful? When an iterator is used to power a loop, the loop becomes very
simple. The code to initialise the state, to decide if the loop is finished, and to find the next value is extracted into
a separate place. This highlights the body of the loop the interesting part. In addition, it is possible to reuse the
iterator code in other places.
7.1.4 Bidirectional communication
Each yield statement causes a value to be passed to the caller. This is the reason for the introduction of generators
by PEP 255 (implemented in Python 2.2). But communication in the reverse direction is also useful. One obvious
way would be some external state, either a global variable or a shared mutable object. Direct communication is
possible thanks to PEP 342 (implemented in 2.5). It is achieved by turning the previously boring yield statement
into an expression. When the generator resumes execution after a yield statement, the caller can call a method
on the generator object to either pass a value into the generator, which then is returned by the yield statement,
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.
The second of the new methods is throw(type, value=None, traceback=None) which is equivalent
to:
raise type, value, traceback
at the point of the yield statement.

Unlike raise (which immediately raises an exception from the current execution point), throw() first resumes
the generator, and only then raises the exception. The word throw was picked because it is suggestive of putting
the exception in another location, and is associated with exceptions in other languages.
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
the caller.
For completeness sake, its worth mentioning that generator iterators also have a close() method, which can
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 define a generator which just prints what is passed in through send and throw.
>>> import itertools
>>> def g():
... print --start--
... for i in itertools.count():
... print --yielding %i-- % i
... try:
... ans = yield i
... except GeneratorExit:
... print --closing--
... raise
... except Exception as e:
... print --yield raised %r-- % e
... else:
... print --yield returned %s-- % ans
>>> it = g()
>>> next(it)
--start--
--yielding 0--
0
>>> it.send(11)
--yield returned 11--
--yielding 1--
1
>>> it.throw(IndexError)
--yield raised IndexError()--
--yielding 2--
2
>>> it.close()
--closing--
Note: next or __next__?

In Python 2.x, the iterator method to retrieve the next value is called next. It is invoked implicitly through the
global function next, which means that it should be called __next__. Just like the global function iter
calls __iter__. This inconsistency is corrected in Python 3.x, where it.next becomes it.__next__.
For other generator methods send and throw the situation is more complicated, because they are not
called implicitly by the interpreter. Nevertheless, theres a proposed syntax extension to allow continue to take

an argument which will be passed to send of the loops iterator. If this extension 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.
7.1.5 Chaining generators
Note: This is a preview of PEP 380 (not yet implemented, but accepted for Python 3.3).
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
loop such as
subgen = some_other_generator()
for v in subgen:
yield v
However, if the subgenerator is to interact properly with the caller in the case of calls to send(), throw()
and close(), things become considerably more difficult. The yield statement has to be guarded by a
try..except..finally structure similar to the one defined in the previous section to debug the generator function.
Such code is provided in PEP 380, here it suffices to say that new syntax to properly yield from a subgenerator is
being introduced in Python 3.3:
yield from some_other_generator()
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.
7.2 Decorators
Summary
This amazing feature appeared in the language almost apologetically and with concern that it might not be
that useful.
Bruce Eckel An Introduction to Python Decorators
Since a function or a class are objects, they can be passed around. Since they are mutable objects, they can be
modified. The act of altering a function or class object after it has been constructed but before is is bound to its
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
be a simple expression, usually this is just the name of a function or class. This part is evaluated first, and
after the function defined below is ready, the decorator is called with the newly defined function object as
the single argument. The value returned by the decorator is attached to the original name of the function.
7.2. Decorators 149

Decorators can be applied to functions and to classes. For classes the semantics are identical the original class
definition is used as an argument to call the decorator and whatever is returned is assigned under the original name.
Before the decorator syntax was implemented (PEP 318), it was possible to achieve the same effect by assigning
the function or class object to a temporary variable and then invoking the decorator explicitly and then assigning
the return value to the name of the function. This sounds like more typing, and it is, and also the name of the
decorated function doubling as a temporary variable must be used at least three times, which is prone to errors.
Nevertheless, the example above is equivalent to:
def function(): #
pass
function = decorator(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
decorator is used as an argument for the second decorator, ..., and whatever is returned by the last decorator is
attached under the name of the original function.
The decorator syntax was chosen for its readability. Since the decorator is specified before the header of the
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
function. Because the expression is prefixed with @ is stands out and is hard to miss (in your face, according to
the PEP :) ). When more than one decorator is applied, each one is placed on a separate line in an easy to read
way.
7.2.1 Replacing or tweaking the original object
Decorators can either return the same function or class object or they can return a completely different object.
In the first case, the decorator can exploit the fact that function and class objects are mutable and add attributes,
e.g. add a docstring to a class. A decorator might do something useful even without modifying the object, for
example register the decorated class in a global registry. In the second case, virtually anything is possible: when
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,
the new function usually calls the original function, after doing some preparatory work. Likewise, when a class is
decorated by replacing if with a new class, the new class is usually derived from the original class. When the
purpose of the decorator is to do something every time, like to log every call to a decorated function, only the
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.
7.2.2 Decorators implemented as classes and as functions
The only requirement on decorators is that they can be called with a single argument. This means that decorators
can be implemented as normal functions, or as classes with a __call__ method, or in theory, even as lambda
functions.
Lets compare the function and class approaches. The decorator expression (the part after @) can be either just a
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:
>>> def simple_decorator(function):
... print "doing decoration"
... return function
>>> @simple_decorator
... def function():
... print "inside function"
doing decoration
>>> function()
inside function
7.2. Decorators 150

>>> def decorator_with_arguments(arg):

... print "defining the decorator"
... def _decorator(function):
... # in this inner function, arg is available too
... print "doing decoration,", arg
... return function
... return _decorator
>>> @decorator_with_arguments("abc")
... def function():
defining the decorator
doing decoration, abc
>>> function()
inside function
The two trivial decorators above fall into the category of decorators which return the original function. If they
were to return a new function, an extra level of nestedness would be required. In the worst case, three levels of
nested functions.
>>> def replacing_decorator_with_args(arg):
... # in this inner function, arg is available too
... def _wrapper(*args, **kwargs):
... print "inside wrapper,", args, kwargs
... return function(*args, **kwargs)
... return _wrapper
>>> @replacing_decorator_with_args("abc")
... def function(*args, **kwargs):
... print "inside function,", args, kwargs
... return 14
>>> function(11, 12)
inside wrapper, (11, 12) {}
inside function, (11, 12) {}
14
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.
Compared to decorators defined as functions, complex decorators defined as classes are simpler. When an object
is created, the __init__ method is only allowed to return None, and the type of the created object cannot be
changed. This means that when a decorator is defined as a class, it doesnt make much sense to use the argument-
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
in the decorator expression and the decorator __init__ method is used for decorator construction.
>>> class decorator_class(object):
... def __init__(self, arg):
... # this method is called in the decorator expression
... print "in decorator init,", arg
... self.arg = arg
... def __call__(self, function):
... # this method is called to do the job
... print "in decorator call,", self.arg
... return function
>>> deco_instance = decorator_class(foo)
in decorator init, foo
>>> @deco_instance
7.2. Decorators 151


... print "in function,", args, kwargs
in decorator call, foo
>>> function()
in function, () {}
Contrary to normal rules (PEP 8) decorators written as classes behave more like functions and therefore their
name often starts with a lowercase letter.
In reality, it doesnt make much sense to create a new class just to have a decorator which returns the original
function. Objects are supposed to hold state, and such decorators are more useful when the decorator returns a
new object.
>>> class replacing_decorator_class(object):
... def __init__(self, arg):
... # this method is called in the decorator expression
... print "in decorator init,", arg
... self.arg = arg
... def __call__(self, function):
... # this method is called to do the job
... print "in decorator call,", self.arg
... self.function = function
... return self._wrapper
... def _wrapper(self, *args, **kwargs):
... print "in the wrapper,", args, kwargs
... return self.function(*args, **kwargs)
>>> deco_instance = replacing_decorator_class(foo)
in decorator init, foo
>>> @deco_instance
... print "in function,", args, kwargs
in decorator call, foo
>>> function(11, 12)
in the wrapper, (11, 12) {}
in function, (11, 12) {}
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.
7.2.3 Copying the docstring and other attributes of the original function
When a new function is returned by the decorator to replace the original function, an unfortunate consequence
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),
__module__ and __name__ (the full name of the function), and __annotations__ (extra information
about arguments and the return value of the function available in Python 3). This can be done automatically by
using functools.update_wrapper.
functools.update_wrapper(wrapper, wrapped)
Update a wrapper function to look like the wrapped function.
>>> import functools

>>> def better_replacing_decorator_with_args(arg):
... def _wrapper(*args, **kwargs):
... print "inside wrapper,", args, kwargs
... return function(*args, **kwargs)
7.2. Decorators 152

... return functools.update_wrapper(_wrapper, function)

>>> @better_replacing_decorator_with_args("abc")
... def function():
... "extensive documentation"
... return 14
>>> function
<function function at 0x...>
>>> print function.__doc__
extensive documentation
One important thing is missing from the list of attributes which can be copied to the replacement func-
tion: the argument list. The default values for arguments can be modified through the __defaults__,
__kwdefaults__ attributes, but unfortunately the argument list itself cannot be set as an attribute. This means
that help(function) will display a useless argument list which will be confusing for the user of the function.
An effective but ugly way around this problem is to create the wrapper dynamically, using eval. This can be
automated by using the external decorator module. It provides support for the decorator decorator, which
takes a wrapper and turns it into a decorator which preserves the function signature.
To sum things up, decorators should always use functools.update_wrapper or some other means of
copying function attributes.
7.2.4 Examples in the standard library
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
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
the first parameter, often called cls.
Class methods are still accessible through the class namespace, so they dont pollute the modules names-
pace. Class methods can be used to provide alternative constructors:
class Array(object):
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
class, despite an implementation which doesnt require this.
property is the pythonic answer to the problem of getters and setters. A method decorated with
property becomes a getter which is automatically called on attribute access.
>>> class A(object):
... @property
... def a(self):
... "an important attribute"
... return "a value"
>>> A.a
7.2. Decorators 153

<property object at 0x...>

>>> A().a
a value
In this example, A.a is an read-only attribute. It is also documented: help(A) includes the docstring for
attribute a taken from the getter method. Defining a as a property allows it to be a calculated on the fly, and
has the side effect of making it read-only, because no setter is defined.
To have a setter and a getter, two methods are required, obviously. Since Python 2.6 the following syntax is
preferred:
class Rectangle(object):
def __init__(self, edge):
self.edge = edge
@property
def area(self):
"""Computed area.
Setting this updates the edge length to the proper value.

"""
return self.edge**2
@area.setter
def area(self, area):
self.edge = area ** 0.5
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.
Their job is to set the getter, setter and deleter of the property object (stored as attributes fget, fset, and
fdel). The getter can be set like in the example above, when creating the object. When defining the setter,
we already have the property object under area, and we add the setter to it by using the setter method.
All this happens when we are creating the class.
Afterwards, when an instance of the class has been created, the property object is special. When the inter-
preter executes attribute access, assignment, or deletion, the job is delegated to the methods of the property
object.
To make everything crystal clear, lets define a debug example:
>>> class D(object):
... @property
... def a(self):
... print "getting", 1
... return 1
... @a.setter
... def a(self, value):
... print "setting", value
... @a.deleter
... def a(self):
... print "deleting"
>>> D.a
<property object at 0x...>
>>> D.a.fget
<function a at 0x...>
>>> D.a.fset
>>> D.a.fdel
>>> d = D() # ... varies, this is not the same a function
>>> d.a
getting 1
1
>>> d.a = 2
7.2. Decorators 154

setting 2
>>> del d.a
deleting
>>> d.a
getting 1
1
Properties are a bit of a stretch for the decorator syntax. One of the premises of the decorator syntax that
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.
Some newer examples include:
functools.lru_cache memoizes an arbitrary function maintaining a limited cache of argu-
ments:answer pairs (Python 3.2)
functools.total_ordering is a class decorator which fills in missing ordering methods (__lt__,
__gt__, __le__, ...) based on a single available one (Python 2.7).
7.2.5 Deprecation of functions
Lets say we want to print a deprecation warning on stderr on the first invocation of a function we dont like
anymore. If we dont want to modify the function, we can use a decorator:
class deprecated(object):
"""Print a deprecation warning once on first use of the function.
>>> @deprecated() # doctest: +SKIP

... def f():
... pass
>>> f() # doctest: +SKIP
f is deprecated
"""
def __call__(self, func):
self.func = func
self.count = 0
return self._wrapper
def _wrapper(self, *args, **kwargs):
self.count += 1
if self.count == 1:
print self.func.__name__, is deprecated
return self.func(*args, **kwargs)
It can also be implemented as a function:

def deprecated(func):
"""Print a deprecation warning once on first use of the function.
>>> @deprecated # doctest: +SKIP

... def f():
... pass
>>> f() # doctest: +SKIP
f is deprecated
"""
count = [0]
def wrapper(*args, **kwargs):
count[0] += 1
if count[0] == 1:
print func.__name__, is deprecated
return func(*args, **kwargs)
return wrapper
7.2. Decorators 155

7.2.6 A while-loop removing decorator
Lets say we have function which returns a lists of things, and this list created by running a loop. If we dont know
how many objects will be needed, the standard way to do this is something like:
def find_answers():
answers = []
while True:
ans = look_for_next_answer()
if ans is None:
break
answers.append(ans)
return answers
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
would have to explicitly call list(find_answers()).
We can define a decorator which constructs the list for us:
def vectorized(generator_func):
def wrapper(*args, **kwargs):
return list(generator_func(*args, **kwargs))
return functools.update_wrapper(wrapper, generator_func)
Our function then becomes:

@vectorized
def find_answers():
while True:
ans = look_for_next_answer()
if ans is None:
break
yield ans
7.2.7 A plugin registration system
This is a class decorator which doesnt modify the class, but just puts it in a global registry. It falls into the category
of decorators returning the original object:
class WordProcessor(object):
PLUGINS = []
def process(self, text):
for plugin in self.PLUGINS:
text = plugin().cleanup(text)
return text
@classmethod
def plugin(cls, plugin):
cls.PLUGINS.append(plugin)
@WordProcessor.plugin
class CleanMdashesExtension(object):
def cleanup(self, text):
return text.replace(—, u\N{em dash})
Here we use a decorator to decentralise the registration of plugins. We call our decorator with a noun, instead of
a verb, because we use it to declare that our class is a plugin for WordProcessor. Method plugin simply
appends the class to the list of plugins.
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).
7.2. Decorators 156

If the Unicode character was inserted directly, it would be impossible to distinguish it from an en-dash in the
source of a program.
7.2.8 More examples and reading
PEP 318 (function and method decorator syntax)

PEP 3129 (class decorator syntax)
http://wiki.python.org/moin/PythonDecoratorLibrary
http://docs.python.org/dev/library/functools.html
http://pypi.python.org/pypi/decorator
Bruce Eckel
Decorators I: Introduction to Python Decorators
Python Decorators II: Decorator Arguments
Python Decorators III: A Decorator-Based Build System
7.3 Context managers
A context manager is an object with __enter__ and __exit__ methods which can be used in the with state-
ment:
with manager as var:
do_something(var)
is in the simplest case equivalent to

var = manager.__enter__()
try:
do_something(var)
finally:
manager.__exit__()
In other words, the context manager protocol defined in PEP 343 permits the extraction of the boring part of a
try..except..finally structure into a separate class leaving only the interesting do_something block.
1. The __enter__ method is called first. It can return a value which will be assigned to var. The as-part
is optional: if it isnt present, the value returned by __enter__ is simply ignored.
2. The block of code underneath with is executed. Just like with try clauses, it can either execute success-
fully to the end, or it can break, continue or return, or it can throw an exception. Either way, after the
block is finished, the __exit__ method is called. If an exception was thrown, the information about the
exception is passed to __exit__, which is described below in the next subsection. In the normal case,
exceptions can be ignored, just like in a finally clause, and will be rethrown after __exit__ is finished.
Lets say we want to make sure that a file is closed immediately after we are done writing to it:
>>> class closing(object):
... def __init__(self, obj):
... self.obj = obj
... def __enter__(self):
... return self.obj
... def __exit__(self, *args):
... self.obj.close()
>>> with closing(open(/tmp/file, w)) as f:
... f.write(the contents\n)
7.3. Context managers 157

Here we have made sure that the f.close() is called when the with block is exited. Since closing files is such
a common operation, the support for this is already present in the file class. It has an __exit__ method which
calls close and can be used as a context manager itself:
>>> with open(/tmp/file, a) as f:
... f.write(more contents\n)
The common use for try..finally is releasing resources. Various different cases are implemented similarly:
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
it is most convenient to have the support built in. With each release, Python provides support in more places:
all file-like objects:
file automatically closed
fileinput, tempfile (py >= 3.2)
bz2.BZ2File, gzip.GzipFile, tarfile.TarFile, zipfile.ZipFile
ftplib, nntplib close connection (py >= 3.2 or 3.3)
locks
multiprocessing.RLock lock and unlock
multiprocessing.Semaphore
memoryview automatically release (py >= 3.2 and 2.7)
decimal.localcontext modify precision of computations temporarily
_winreg.PyHKEY open and close hive key
warnings.catch_warnings kill warnings temporarily
contextlib.closing the same as the example above, call close
parallel programming
concurrent.futures.ThreadPoolExecutor invoke in parallel then kill thread pool (py
>= 3.2)
concurrent.futures.ProcessPoolExecutor invoke in parallel then kill process pool
(py >= 3.2)
nogil solve the GIL problem temporarily (cython only :( )
7.3.1 Catching exceptions
When an exception is thrown in the with-block, it is passed as arguments to __exit__. Three arguments are
used, the same as returned by sys.exc_info(): type, value, traceback. When no exception is thrown, None
is used for all three arguments. The context manager can swallow the exception by returning a true value from
__exit__. Exceptions can be easily ignored, because if __exit__ doesnt use return and just falls of the
end, None is returned, a false value, and therefore the exception is rethrown after __exit__ is finished.
The ability to catch exceptions opens interesting possibilities. A classic example comes from unit-tests we
want to make sure that some code throws the right kind of exception:
class assert_raises(object):
# based on pytest and unittest.TestCase
def __init__(self, type):
self.type = type
def __enter__(self):
pass
def __exit__(self, type, value, traceback):
if type is None:
raise AssertionError(exception expected)

if issubclass(type, self.type):
return True # swallow the expected exception
raise AssertionError(wrong exception type)
with assert_raises(KeyError):
{}[foo]
7.3.2 Using generators to define context managers
When discussing generators (page 146), it was said that we prefer generators to iterators implemented as classes
because they are shorter, sweeter, and the state is stored as local, not instance, variables. On the other hand, as
described in Bidirectional communication (page 147), 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
context managers as special generator functions. In fact, the generator protocol was designed to support this use
case.
@contextlib.contextmanager
def some_generator(<arguments>):
<setup>
try:
yield <value>
finally:
<cleanup>
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
exactly once. The part before the yield is executed from __enter__, the block of code protected by the con-
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
function then throws it at the point of the yield statement. Through the use of generators, the context manager
is shorter and simpler.
Lets rewrite the closing example as a generator:
def closing(obj):
try:
yield obj
finally:
obj.close()
Lets rewrite the assert_raises example as a generator:

def assert_raises(type):
try:
yield
except type:
return
except Exception as value:
raise AssertionError(wrong exception type)
else:
raise AssertionError(exception expected)
Here we use a decorator to turn generator functions into context managers!

CHAPTER 8
Advanced Numpy
author Pauli Virtanen

Numpy is at the base of Pythons scientific stack of tools. Its purpose is simple: implementing efficient operations
on many items in a block of memory. Understanding how it works in detail helps in making efficient use of its
flexibility, taking useful shortcuts, and in building new work based on it.
This tutorial aims to cover:
Anatomy of Numpy arrays, and its consequences. Tips and tricks.
Universal functions: what, why, and what to do if you want a new one.
Integration with other tools: Numpy offers several ways to wrap any data in an ndarray, without unnecessary
copies.
Recently added features, and whats in them for me: PEP 3118 buffers, generalized ufuncs, ...
Prerequisites
Numpy (>= 1.2; preferably newer...)
Cython (>= 0.12, for the Ufunc example)
PIL (used in a couple of examples)
In this section, numpy will be imported as follows:

160
Chapter contents
Life of ndarray (page 161)
Its... (page 161)
Block of memory (page 162)
Data types (page 163)
Indexing scheme: strides (page 168)
Findings in dissection (page 174)
Universal functions (page 174)
What they are? (page 174)
Exercise: building an ufunc from scratch (page 176)
Solution: building an ufunc from scratch (page 179)
Generalized ufuncs (page 182)
Interoperability features (page 183)
Sharing multidimensional, typed data (page 183)
The old buffer protocol (page 183)
The old buffer protocol (page 184)
Array interface protocol (page 185)
Array siblings: chararray, maskedarray, matrix (page 186)
chararray: vectorized string operations (page 186)
masked_array missing data (page 186)
recarray: purely convenience (page 189)
matrix: convenience? (page 189)
Summary (page 189)
Contributing to Numpy/Scipy (page 189)
Why (page 189)
Reporting bugs (page 189)
Contributing to documentation (page 190)
Contributing features (page 191)
How to help, in general (page 192)
8.1 Life of ndarray
8.1.1 Its...
ndarray =
block of memory + indexing scheme + data type descriptor
raw data
how to locate an element
how to interpret an element
8.1. Life of ndarray 161

typedef struct PyArrayObject {

PyObject_HEAD
/* Block of memory */
char *data;
/* Data type descriptor */

PyArray_Descr *descr;
/* Indexing scheme */
int nd;
npy_intp *dimensions;
npy_intp *strides;
/* Other stuff */
PyObject *base;
int flags;
PyObject *weakreflist;
} PyArrayObject;
8.1.2 Block of memory
>>> x = np.array([1, 2, 3, 4], dtype=np.int32)

>>> x.data
<read-write buffer for ..., size 16, offset 0 at ...>
>>> str(x.data)
\x01\x00\x00\x00\x02\x00\x00\x00\x03\x00\x00\x00\x04\x00\x00\x00
Memory address of the data:

>>> x.__array_interface__[data][0]
64803824
The whole __array_interface__:

>>> x.__array_interface__
{data: (35828928, False),
descr: [(, <i4)],
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]

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

>>> x = 1234
>>> y = np.frombuffer(x, dtype=np.int8)
>>> y.data
<read-only buffer for ..., size 4, offset 0 at ...>
>>> y.base is x
True
>>> y.flags
C_CONTIGUOUS : True
F_CONTIGUOUS : True
OWNDATA : False
WRITEABLE : False
ALIGNED : True
UPDATEIFCOPY : False
The owndata and writeable flags indicate status of the memory block.
See also : array interface
8.1.3 Data types
The descriptor
dtype describes a single item in the array:

type scalar type of the data, one of:
int8, int16, float64, et al. (fixed size)
str, unicode, void (flexible size)
itemsize size of the data block
byteorder byte order: big-endian > / little-endian < / not applicable |
fields sub-dtypes, if its a structured data type
shape shape of the array, if its a sub-array
>>> np.dtype(int).type
<type numpy.int64>
>>> np.dtype(int).itemsize
8
>>> np.dtype(int).byteorder
=
Example: reading .wav files
The .wav file header:

chunk_id "RIFF"
chunk_size 4-byte unsigned little-endian integer
format "WAVE"
fmt_id "fmt "
fmt_size 4-byte unsigned little-endian integer
audio_fmt 2-byte unsigned little-endian integer
num_channels 2-byte unsigned little-endian integer
sample_rate 4-byte unsigned little-endian integer
byte_rate 4-byte unsigned little-endian integer
block_align 2-byte unsigned little-endian integer
bits_per_sample 2-byte unsigned little-endian integer
data_id "data"
data_size 4-byte unsigned little-endian integer
44-byte block of raw data (in the beginning of the file)
... followed by data_size bytes of actual sound data.
The .wav file header as a Numpy structured data type:
>>> wav_header_dtype = np.dtype([
... ("chunk_id", (str, 4)), # flexible-sized scalar type, item size 4
... ("chunk_size", "<u4"), # little-endian unsigned 32-bit integer
... ("format", "S4"), # 4-byte string
... ("fmt_id", "S4"),
... ("fmt_size", "<u4"),
... ("audio_fmt", "<u2"), #
... ("num_channels", "<u2"), # .. more of the same ...
... ("sample_rate", "<u4"), #
... ("byte_rate", "<u4"),
... ("block_align", "<u2"),
... ("bits_per_sample", "<u2"),
... ("data_id", ("S1", (2, 2))), # sub-array, just for fun!
... ("data_size", "u4"),
... #
... # the sound data itself cannot be represented here:
... # it does not have a fixed size
... ])
See Also:
wavreader.py
>>> wav_header_dtype[format]
dtype(|S4)
>>> wav_header_dtype.fields
<dictproxy object at ...>
>>> wav_header_dtype.fields[format]
(dtype(|S4), 8)
The first element is the sub-dtype in the structured data, corresponding to the name format
The second one is its offset (in bytes) from the beginning of the item

Exercise
Mini-exercise, make a sparse dtype by using offsets, and only some of the fields:
>>> wav_header_dtype = np.dtype(dict(
... names=[format, sample_rate, data_id],
... offsets=[offset_1, offset_2, offset_3], # counted from start of structure in bytes
... formats=list of dtypes for each of the fields,
... ))
and use that to read the sample rate, and data_id (as sub-array).
>>> f = open(data/test.wav, r)
>>> wav_header = np.fromfile(f, dtype=wav_header_dtype, count=1)
>>> f.close()
>>> print(wav_header)
[ (RIFF, 17402L, WAVE, fmt , 16L, 1, 1, 16000L, 32000L, 2, 16, [[d, a], [t, a]], 173
>>> wav_header[sample_rate]
array([16000], dtype=uint32)
Lets try accessing the sub-array:

>>> wav_header[data_id]
array([[[d, a],
[t, a]]],
dtype=|S1)
>>> wav_header.shape
(1,)
>>> wav_header[data_id].shape
(1, 2, 2)
When accessing sub-arrays, the dimensions get added to the end!
Note: There are existing modules such as wavfile, audiolab, etc. for loading sound data...
Casting and re-interpretation/views
casting
on assignment
on array construction
on arithmetic
etc.
and manually: .astype(dtype)
data re-interpretation
manually: .view(dtype)
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
array([ 1., 2., 3., 4.])
>>> y = x.astype(np.int8)
>>> y
array([1, 2, 3, 4], dtype=int8)
>>> y + 1
>>> y + 256
>>> y + 256.0
array([ 257., 258., 259., 260.])
>>> y + np.array([256], dtype=np.int32)
array([257, 258, 259, 260], dtype=int32)
Casting on setitem: dtype of the array is not changed on item assignment:

>>> y[:] = y + 1.5
>>> y
Note: Exact rules: see documentation: http://docs.scipy.org/doc/numpy/reference/ufuncs.html#casting-rules
Re-interpretation / viewing
Data block in memory (4 bytes)

0x01 || 0x02 || 0x03 || 0x04
4 of uint8, OR,
4 of int8, OR,
2 of int16, OR,
1 of int32, OR,
1 of float32, OR,
...
How to switch from one to another?
1. Switch the dtype:
>>> x = np.array([1, 2, 3, 4], dtype=np.uint8)
>>> 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
Note:
.view() makes views, does not copy (or alter) the memory block
only changes the dtype (and adjusts array shape):
>>> x[1] = 5
>>> y
>>> y.base is x
True
Mini-exercise: data re-interpretation
See Also:
view-colors.py
You have RGBA data in an array:
>>> x = np.zeros((10, 10, 4), dtype=np.int8)
>>> x[:, :, 0] = 1
>>> x[:, :, 1] = 2
>>> x[:, :, 2] = 3
>>> x[:, :, 3] = 4
where the last three dimensions are the R, B, and G, and alpha channels.
How to make a (10, 10) structured array with field names r, g, b, a without copying data?
>>> y = ...
>>> assert (y[r] == 1).all()

>>> assert (y[g] == 2).all()
>>> assert (y[b] == 3).all()
>>> assert (y[a] == 4).all()
Solution
>>> y = x.view([(r, i1),
... (g, i1),
... (b, i1),
... (a, i1)]
... )[:, :, 0]

Warning: Another array taking exactly 4 bytes of memory:
>>> y = np.array([[1, 3], [2, 4]], dtype=np.uint8).transpose()

>>> x = y.copy()
>>> x
array([[1, 2],
[3, 4]], dtype=uint8)
>>> y
array([[1, 2],
[3, 4]], dtype=uint8)
>>> x.view(np.int16)
array([[ 513],
[1027]], dtype=int16)
>>> 0x0201, 0x0403
(513, 1027)
>>> y.view(np.int16)
array([[ 769, 1026]], dtype=int16)
What happened?
... we need to look into what x[0,1] actually means
>>> 0x0301, 0x0402

(769, 1026)
8.1.4 Indexing scheme: strides
Main point
The question
>>> x = np.array([[1, 2, 3],
... [4, 5, 6],
... [7, 8, 9]], dtype=np.int8)
>>> str(x.data)
\x01\x02\x03\x04\x05\x06\x07\x08\t
At which byte in x.data does the item x[1,2] begin?

The answer (in Numpy)
strides: the number of bytes to jump to find the next element
1 stride per dimension
>>> x.strides
(3, 1)
>>> byte_offset = 3*1 + 1*2 # to find x[1,2]
>>> x.data[byte_offset]
\x06
>>> x[1, 2]
6
simple, flexible
C and Fortran order
>>> x = np.array([[1, 2, 3],

... [4, 5, 6],
... [7, 8, 9]], dtype=np.int16, order=C)
>>> x.strides

(6, 2)
>>> str(x.data)
\x01\x00\x02\x00\x03\x00\x04\x00\x05\x00\x06\x00\x07\x00\x08\x00\t\x00
Need to jump 6 bytes to find the next row

Need to jump 2 bytes to find the next column
>>> y = np.array(x, order=F)
>>> y.strides
(2, 6)
>>> str(y.data)
\x01\x00\x04\x00\x07\x00\x02\x00\x05\x00\x08\x00\x03\x00\x06\x00\t\x00
Need to jump 2 bytes to find the next row

Need to jump 6 bytes to find the next column
Similarly to higher dimensions:
C: last dimensions vary fastest (= smaller strides)
F: first dimensions vary fastest
shape = (d1 , d2 , ..., dn )

strides = (s1 , s2 , ..., sn )
sC
j = dj+1 dj+2 ...dn itemsize
sF
j = d1 d2 ...dj1 itemsize
Note: Now we can understand the behavior of .view():

>>> y = np.array([[1, 3], [2, 4]], dtype=np.uint8).transpose()
>>> x = y.copy()
Transposition does not affect the memory layout of the data, only strides
>>> x.strides
(2, 1)
>>> y.strides
(1, 2)
>>> str(x.data)
\x01\x02\x03\x04
>>> str(y.data)
\x01\x03\x02\x04
the results are different when interpreted as 2 of int16

.copy() creates new arrays in the C order (by default)
Slicing with integers
Everything can be represented by changing only shape, strides, and possibly adjusting the data
pointer!
Never makes copies of the data
>>> x = np.array([1, 2, 3, 4, 5, 6], dtype=np.int32)
>>> y = x[::-1]
>>> y
array([6, 5, 4, 3, 2, 1], dtype=int32)

>>> y.strides
(-4,)
>>> y = x[2:]
>>> y.__array_interface__[data][0] - x.__array_interface__[data][0]
8
>>> x = np.zeros((10, 10, 10), dtype=np.float)

>>> x.strides
(800, 80, 8)
>>> x[::2,::3,::4].strides
(1600, 240, 32)
Similarly, transposes never make copies (it just swaps strides)

>>> x = np.zeros((10, 10, 10), dtype=np.float)
>>> x.strides
(800, 80, 8)
>>> x.T.strides
(8, 80, 800)
But: not all reshaping operations can be represented by playing with strides.
>>> a = np.arange(6, dtype=np.int8).reshape(3, 2)
>>> b = a.T
>>> b.strides
(1, 2)
So far, so good. However:

>>> str(a.data)
\x00\x01\x02\x03\x04\x05
>>> b
array([[0, 2, 4],
[1, 3, 5]], dtype=int8)
>>> c = b.reshape(3*2)
>>> c
array([0, 2, 4, 1, 3, 5], dtype=int8)
Here, there is no way to represent the array c given one stride and the block of memory for a. Therefore, the
reshape operation needs to make a copy here.
Example: fake dimensions with strides
Stride manipulation
>>> from numpy.lib.stride_tricks import as_strided

>>> help(as_strided)
as_strided(x, shape=None, strides=None)
Make an ndarray from the given array with the given shape and strides
Warning: as_strided does not check that you stay inside the memory block bounds...

>>> as_strided(x, strides=(2*2, ), shape=(2, ))
array([1, 3], dtype=int16)
>>> x[::2]
See Also:
stride-fakedims.py

Exercise
array([1, 2, 3, 4], dtype=np.int8)
-> array([[1, 2, 3, 4],

[1, 2, 3, 4],
[1, 2, 3, 4]], dtype=np.int8)
using only as_strided.:

Hint: byte_offset = stride[0]*index[0] + stride[1]*index[1] + ...
Spoiler
Stride can also be 0:
>>> y = as_strided(x, strides=(0, 1), shape=(3, 4))
>>> y
array([[1, 2, 3, 4],
[1, 2, 3, 4],
[1, 2, 3, 4]], dtype=int8)
>>> y.base.base is x
True
Broadcasting
Doing something useful with it: outer product of [1, 2, 3, 4] and [5, 6, 7]
>>> x2 = as_strided(x, strides=(0, 1*2), shape=(3, 4))
>>> x2
array([[1, 2, 3, 4],
[1, 2, 3, 4],
[1, 2, 3, 4]], dtype=int16)
>>> y = np.array([5, 6, 7], dtype=np.int16)

>>> y2 = as_strided(y, strides=(1*2, 0), shape=(3, 4))
>>> y2
array([[5, 5, 5, 5],
[6, 6, 6, 6],
[7, 7, 7, 7]], dtype=int16)
>>> x2 * y2
array([[ 5, 10, 15, 20],
[ 6, 12, 18, 24],
[ 7, 14, 21, 28]], dtype=int16)
... seems somehow familiar ...

>>> y = np.array([5, 6, 7], dtype=np.int16)
>>> x[np.newaxis,:] * y[:,np.newaxis]
array([[ 5, 10, 15, 20],
[ 6, 12, 18, 24],
[ 7, 14, 21, 28]], dtype=int16)
Internally, array broadcasting is indeed implemented using 0-strides.

More tricks: diagonals
See Also:
stride-diagonals.py
Challenge
Pick diagonal entries of the matrix: (assume C memory order):
>>> x = np.array([[1, 2, 3],
... [4, 5, 6],
... [7, 8, 9]], dtype=np.int32)
>>> x_diag = as_strided(x, shape=(3,), strides=(???,))
Pick the first super-diagonal entries [2, 6].

And the sub-diagonals?
(Hint to the last two: slicing first moves the point where striding starts from.)
Solution
Pick diagonals:
>>> x_diag = as_strided(x, shape=(3, ), strides=((3+1)*x.itemsize, ))
>>> x_diag
array([1, 5, 9], dtype=int32)
Slice first, to adjust the data pointer:

>>> as_strided(x[0, 1:], shape=(2, ), strides=((3+1)*x.itemsize, ))
>>> as_strided(x[1:, 0], shape=(2, ), strides=((3+1)*x.itemsize, ))

Note:
>>> y = np.diag(x, k=1)
>>> y
However,
>>> y.flags.owndata
True
It makes a copy?!
See Also:
stride-diagonals.py
Challenge
Compute the tensor trace:
>>> x = np.arange(5*5*5*5).reshape(5,5,5,5)
>>> s = 0
>>> for i in xrange(5):
... for j in xrange(5):
... s += x[j,i,j,i]
by striding, and using sum() on the result.

>>> y = as_strided(x, shape=(5, 5), strides=(TODO, TODO))

>>> s2 = ...
>>> assert s == s2
Solution
>>> y = as_strided(x, shape=(5, 5), strides=((5*5*5 + 5)*x.itemsize,
... (5*5 + 1)*x.itemsize))
>>> s2 = y.sum()
CPU cache effects
Memory layout can affect performance:

In [1]: x = np.zeros((20000,))
In [2]: y = np.zeros((20000*67,))[::67]
In [3]: x.shape, y.shape

((20000,), (20000,))
In [4]: %timeit x.sum()

In [5]: %timeit y.sum()

In [6]: x.strides, y.strides

((8,), (536,))
Smaller strides are faster?
CPU pulls data from main memory to its cache in blocks

If many array items consecutively operated on fit in a single block (small stride):
fewer transfers needed
faster
See Also:
numexpr is designed to mitigate cache effects in array computing.
Example: inplace operations (caveat emptor)
Sometimes,
>>> a -= b
is not the same as

>>> a -= b.copy()
>>> x = np.array([[1, 2], [3, 4]])

>>> x -= x.transpose()
>>> x
array([[ 0, -1],
[ 4, 0]])
>>> y = np.array([[1, 2], [3, 4]])

>>> y -= y.T.copy()
>>> y
array([[ 0, -1],
[ 1, 0]])
x and x.transpose() share data

x -= x.transpose() modifies the data element-by-element...
because x and x.transpose() have different striding, modified data re-appears on the RHS
8.1.5 Findings in dissection
memory block: may be shared, .base, .data

data type descriptor: structured data, sub-arrays, byte order, casting, viewing, .astype(), .view()
strided indexing: strides, C/F-order, slicing w/ integers, as_strided, broadcasting, stride tricks, diag,
CPU cache coherence
8.2 Universal functions
8.2.1 What they are?
Ufunc performs and elementwise operation on all elements of an array.

Examples:
np.add, np.subtract, scipy.special.*, ...
Automatically support: broadcasting, casting, ...

The author of an ufunc only has to supply the elementwise operation, Numpy takes care of the rest.
The elementwise operation needs to be implemented in C (or, e.g., Cython)
8.2. Universal functions 174

Parts of an Ufunc
1. Provided by user
void ufunc_loop(void **args, int *dimensions, int *steps, void *data)
{
/*
* int8 output = elementwise_function(int8 input_1, int8 input_2)
*
* This function must compute the ufunc for many values at once,
* in the way shown below.
*/
char *input_1 = (char*)args[0];
char *output = (char*)args[2];
int i;
for (i = 0; i < dimensions[0]; ++i) {

*output = elementwise_function(*input_1, *input_2);
input_1 += steps[0];
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 */
identity_element,
name,
docstring,
unused)
A ufunc can also support multiple different input-output type combinations.
Making it easier
3. ufunc_loop is of very generic form, and Numpy provides pre-made ones

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)
PyUfunc_dd_ddouble elementwise_func(double input_1, double input_2)
PyUfunc_D_D elementwise_func(npy_cdouble *input, npy_cdouble* output)
PyUfunc_DD_Delementwise_func(npy_cdouble *in1, npy_cdouble *in2,
npy_cdouble* out)
Only elementwise_func needs to be supplied
... except when your elementwise function is not in one of the above forms

8.2.2 Exercise: building an ufunc from scratch
The Mandelbrot fractal is defined by the iteration
z z2 + c
where c = x + iy is a complex number. This iteration is repeated if z stays finite no matter how long the iteration
runs, c belongs to the Mandelbrot set.
Make ufunc called mandel(z0, c) that computes:
z = z0
for k in range(iterations):
z = z*z + c
say, 100 iterations or until z.real**2 + z.imag**2 > 1000. Use it to determine which c are in the
Mandelbrot set.
Our function is a simple one, so make use of the PyUFunc_* helpers.
Write it in Cython
See Also:
mandel.pyx, mandelplot.py
#
# Fix the parts marked by TODO
#
#
# Compile this file by (Cython >= 0.12 required because of the complex vars)
#
# cython mandel.pyx
# python setup.py build_ext -i
#
# and try it out with, in this directory,
#
# >>> import mandel
# >>> mandel.mandel(0, 1 + 2j)
#
#
# The elementwise function

# ------------------------
cdef void mandel_single_point(double complex *z_in,

double complex *c_in,
double complex *z_out) nogil:
#
# The Mandelbrot iteration
#
#
# Some points of note:
#
# - Its *NOT* allowed to call any Python functions here.
#
# The Ufunc loop runs with the Python Global Interpreter Lock released.
# Hence, the nogil.
#
# - And so all local variables must be declared with cdef
#
# - Note also that this function receives *pointers* to the data
#

cdef double complex z = z_in[0]

cdef double complex c = c_in[0]
cdef int k # the integer we use in the for loop
#
# TODO: write the Mandelbrot iteration for one point here,
# as you would write it in Python.
#
# Say, use 100 as the maximum number of iterations, and 1000
# as the cutoff for z.real**2 + z.imag**2.
#
TODO: mandelbrot iteration should go here
# Return the answer for this point

z_out[0] = z
# Boilerplate Cython definitions

#
# The litany below is particularly long, but you dont really need to
# read this part; it just pulls in stuff from the Numpy C headers.
# ----------------------------------------------------------
cdef extern from "numpy/arrayobject.h":

void import_array()
ctypedef int npy_intp
cdef enum NPY_TYPES:
NPY_DOUBLE
NPY_CDOUBLE
NPY_LONG
cdef extern from "numpy/ufuncobject.h":

void import_ufunc()
ctypedef void (*PyUFuncGenericFunction)(char**, npy_intp*, npy_intp*, void*)
object PyUFunc_FromFuncAndData(PyUFuncGenericFunction* func, void** data,
char* types, int ntypes, int nin, int nout,
int identity, char* name, char* doc, int c)
# List of pre-defined loop functions
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)
void PyUFunc_f_f(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_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)
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)
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)
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)
void PyUFunc_FF_F(char** args, npy_intp* dimensions, npy_intp* steps, void* func)
void PyUFunc_GG_G(char** args, npy_intp* dimensions, npy_intp* steps, void* func)
# Required module initialization

# ------------------------------

import_array()
import_ufunc()
# The actual ufunc declaration

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

cdef char input_output_types[3]
cdef void *elementwise_funcs[1]
#
# Reminder: some pre-made Ufunc loops:
#
# ================ =======================================================
# PyUfunc_f_f float elementwise_func(float input_1)
# 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)
# PyUfunc_D_D elementwise_func(complex_double *input, complex_double* complex_double)
# PyUfunc_DD_D elementwise_func(complex_double *in1, complex_double *in2, complex_double* o
# ================ =======================================================
#
# The full list is above.
#
#
# Type codes:
#
# NPY_BOOL, NPY_BYTE, NPY_UBYTE, NPY_SHORT, NPY_USHORT, NPY_INT, NPY_UINT,
# NPY_LONG, NPY_ULONG, NPY_LONGLONG, NPY_ULONGLONG, NPY_FLOAT, NPY_DOUBLE,
# NPY_LONGDOUBLE, NPY_CFLOAT, NPY_CDOUBLE, NPY_CLONGDOUBLE, NPY_DATETIME,
# NPY_TIMEDELTA, NPY_OBJECT, NPY_STRING, NPY_UNICODE, NPY_VOID
#
loop_func[0] = ... TODO: suitable PyUFunc_* ...

input_output_types[0] = ... TODO ...
... TODO: fill in rest of input_output_types ...
# This thing is passed as the data parameter for the generic

# PyUFunc_* loop, to let it know which function it should call.
elementwise_funcs[0] = <void*>mandel_single_point
# Construct the ufunc:
mandel = PyUFunc_FromFuncAndData(
loop_func,
elementwise_funcs,
input_output_types,
1, # number of supported input types
TODO, # number of input args
TODO, # number of output args
0, # identity element, never mind this
"mandel", # function name
"mandel(z, c) -> computes z*z + c", # docstring
0 # unused
)
Reminder: some pre-made Ufunc loops:

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)
PyUfunc_dd_ddouble elementwise_func(double input_1, double input_2)
PyUfunc_D_D elementwise_func(complex_double *input, complex_double*
output)
PyUfunc_DD_Delementwise_func(complex_double *in1, complex_double *in2,
complex_double* out)
Type codes:
NPY_BOOL, NPY_BYTE, NPY_UBYTE, NPY_SHORT, NPY_USHORT, NPY_INT, NPY_UINT,
NPY_LONG, NPY_ULONG, NPY_LONGLONG, NPY_ULONGLONG, NPY_FLOAT, NPY_DOUBLE,
NPY_LONGDOUBLE, NPY_CFLOAT, NPY_CDOUBLE, NPY_CLONGDOUBLE, NPY_DATETIME,
NPY_TIMEDELTA, NPY_OBJECT, NPY_STRING, NPY_UNICODE, NPY_VOID
8.2.3 Solution: building an ufunc from scratch
# The elementwise function

# ------------------------

#
# The Mandelbrot iteration
#
#
# Some points of note:
#
# - Its *NOT* allowed to call any Python functions here.
#
# The Ufunc loop runs with the Python Global Interpreter Lock released.
# Hence, the nogil.
#
# - And so all local variables must be declared with cdef
#
# - Note also that this function receives *pointers* to the data;
# the "traditional" solution to passing complex variables around
#
cdef double complex z = z_in[0]

cdef double complex c = c_in[0]
cdef int k # the integer we use in the for loop
# Straightforward iteration
for k in range(100):
z = z*z + c
if z.real**2 + z.imag**2 > 1000:
break
# Return the answer for this point

z_out[0] = z
# Boilerplate Cython definitions

#
# You dont really need to read this part, it just pulls in
# stuff from the Numpy C headers.

# ----------------------------------------------------------
cdef extern from "numpy/arrayobject.h":

void import_array()
ctypedef int npy_intp
cdef enum NPY_TYPES:
NPY_CDOUBLE
cdef extern from "numpy/ufuncobject.h":

void import_ufunc()
ctypedef void (*PyUFuncGenericFunction)(char**, npy_intp*, npy_intp*, void*)
object PyUFunc_FromFuncAndData(PyUFuncGenericFunction* func, void** data,
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]
cdef void *elementwise_funcs[1]
loop_func[0] = PyUFunc_DD_D
input_output_types[0] = NPY_CDOUBLE
loop_func,
elementwise_funcs,
input_output_types,
1, # number of supported input types
2, # number of input args
1, # number of output args
"mandel(z, c) -> computes iterated z*z + c", # docstring
0 # unused
)
import numpy as np
import mandel
x = np.linspace(-1.7, 0.6, 1000)
y = np.linspace(-1.4, 1.4, 1000)
c = x[None,:] + 1j*y[:,None]
z = mandel.mandel(c, c)
import matplotlib.pyplot as plt

plt.imshow(abs(z)**2 < 1000, extent=[-1.7, 0.6, -1.4, 1.4])
plt.gray()

plt.show()
Note: Most of the boilerplate could be automated by these Cython modules:

http://wiki.cython.org/MarkLodato/CreatingUfuncs
Several accepted input types
E.g. supporting both single- and double-precision versions

...
cdef void mandel_single_point_singleprec(float complex *z_in,

float complex *c_in,
float complex *z_out) nogil:
...
cdef PyUFuncGenericFunction loop_funcs[2]

cdef char input_output_types[3*2]
cdef void *elementwise_funcs[1*2]
loop_funcs[0] = PyUFunc_DD_D
loop_funcs[1] = PyUFunc_FF_F
input_output_types[3] = NPY_CFLOAT
elementwise_funcs[1] = <void*>mandel_single_point_singleprec
loop_func,
elementwise_funcs,
input_output_types,
2, # number of supported input types <----------------
2, # number of input args
1, # number of output args


"mandel(z, c) -> computes iterated z*z + c", # docstring
0 # unused
)
8.2.4 Generalized ufuncs
ufunc
output = elementwise_function(input)
Both output and input can be a single array element only.
generalized ufunc
output and input can be arrays with a fixed number of dimensions
For example, matrix trace (sum of diag elements):
input shape = (n, n)
output shape = () i.e. scalar
(n, n) -> ()
Matrix product:
input_1 shape = (m, n)
input_2 shape = (n, p)
output shape = (m, p)
(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
Status in Numpy
g-ufuncs are in Numpy already ...

new ones can be created with PyUFunc_FromFuncAndDataAndSignature
... but we dont ship with public g-ufuncs, except for testing, ATM
>>> import numpy.core.umath_tests as ut
>>> ut.matrix_multiply.signature
(m,n),(n,p)->(m,p)
>>> x = np.ones((10, 2, 4))

>>> y = np.ones((10, 4, 5))
>>> ut.matrix_multiply(x, y).shape
(10, 2, 5)
the last two dimensions became core dimensions, and are modified as per the signature
otherwise, the g-ufunc operates elementwise
matrix multiplication this way could be useful for operating on many small matrices at once
Generalized ufunc loop
Matrix multiplication (m,n),(n,p) -> (m,p)

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

{
char *input_1 = (char*)args[0]; /* these are as previously */
char *output = (char*)args[2];
int input_1_stride_m = steps[3]; /* strides for the core dimensions */

int input_1_stride_n = steps[4]; /* are added after the non-core */
int input_2_strides_n = steps[5]; /* steps */
int input_2_strides_p = steps[6];
int output_strides_n = steps[7];
int output_strides_p = steps[8];
int m = dimension[1]; /* core dimensions are added after */

int n = dimension[2]; /* the main dimension; order as in */
int p = dimension[3]; /* signature */
int i;
for (i = 0; i < dimensions[0]; ++i) {

matmul_for_strided_matrices(input_1, input_2, output,
strides for each array...);
output += steps[2];
}
}
8.3 Interoperability features
8.3.1 Sharing multidimensional, typed data
Suppose you
1. Write a library than handles (multidimensional) binary data,
2. Want to make it easy to manipulate the data with Numpy, or whatever other library,
3. ... but would not like to have Numpy as a dependency.
Currently, 3 solutions:
1. the old buffer interface
2. the array interface
3. the new buffer interface (PEP 3118)
8.3.2 The old buffer protocol
Only 1-D buffers

No data type information
C-level interface; PyBufferProcs tp_as_buffer in the type object
But its integrated into Python (e.g. strings support it)
Mini-exercise using PIL (Python Imaging Library):
See Also:
8.3. Interoperability features 183

pilbuffer.py
>>> 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)
>>> img.save(test.png)
Q:
Check what happens if data is now modified, and img saved again.
8.3.3 The old buffer protocol
import numpy as np
import Image
# Lets 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.
#
# It turns out that PIL, which knows next to nothing about Numpy,
# happily shares the same data.
#
x[:,:,1] = 254
img.save(test2.png)

8.3.4 Array interface protocol
Multidimensional buffers
Data type information present
Numpy-specific approach; slowly deprecated (but not going away)
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,
}
>>> import Image

>>> img = Image.open(data/test.png)
>>> img.__array_interface__
{data: ...,
shape: (200, 200, 4),
typestr: |u1}
>>> x = np.asarray(img)
>>> x.shape
(200, 200, 4)
>>> x.dtype
dtype(uint8)
Note: A more C-friendly variant of the array interface is also defined.

8.4 Array siblings: chararray, maskedarray, matrix
8.4.1 chararray: vectorized string operations
>>> x = np.array([a, bbb, ccc]).view(np.chararray)

>>> x.lstrip( )
chararray([a, bbb, ccc],
dtype=|S5)
>>> x.upper()
chararray([A, BBB, CCC],
dtype=|S5)
Note: .view() has a second meaning: it can make an ndarray an instance of a specialized ndarray subclass
8.4.2 masked_array missing data
Masked arrays are arrays that may have missing or invalid entries.
For example, suppose we have an array where the fourth entry is invalid:
>>> x = np.array([1, 2, 3, -99, 5])
One way to describe this is to create a masked array:

>>> mx = np.ma.masked_array(x, mask=[0, 0, 0, 1, 0])
>>> mx
masked_array(data = [1 2 3 -- 5],
mask = [False False False True False],
Masked mean ignores masked data:

>>> mx.mean()
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])
The mask
You can modify the mask by assigning:

>>> mx[1] = np.ma.masked
>>> mx
masked_array(data = [1 -- 3 -- 5],
mask = [False True False True False],
The mask is cleared on assignment:
8.4. Array siblings: chararray, maskedarray, matrix 186

>>> mx[1] = 9
>>> mx
masked_array(data = [1 9 3 -- 5],
mask = [False False False True False],
The mask is also available directly:

>>> mx.mask
array([False, False, False, True, False], dtype=bool)
The masked entries can be filled with a given value to get an usual array back:
>>> x2 = mx.filled(-1)
>>> x2
array([ 1, 9, 3, -1, 5])
The mask can also be cleared:

>>> mx.mask = np.ma.nomask
>>> mx
masked_array(data = [1 9 3 -99 5],
mask = [False False False False False],
Domain-aware functions
The masked array package also contains domain-aware functions:

>>> np.ma.log(np.array([1, 2, -1, -2, 3, -5]))
masked_array(data = [0.0 0.69314718056 -- -- 1.09861228867 --],
mask = [False False True True False True],
fill_value = 1e+20)
Note: Streamlined and more seamless support for dealing with missing data in arrays is making its way into
Numpy 1.7. Stay tuned!

Example: Masked statistics

Canadian rangers were distracted when counting hares and lynxes in 1903-1910 and 1917-1918, and got
the numbers are wrong. (Carrot farmers stayed alert, though.) Compute the mean populations over time,
ignoring the invalid numbers.
>>> populations = np.ma.masked_array(data[:,1:])
>>> year = data[:, 0]
>>> bad_years = (((year >= 1903) & (year <= 1910))

... | ((year >= 1917) & (year <= 1918)))
>>> # & means and and | means or
>>> populations[bad_years, 0] = np.ma.masked
>>> populations[bad_years, 1] = np.ma.masked
>>> populations.mean(axis=0)
masked_array(data = [40472.7272727 18627.2727273 42400.0],
mask = [False False False],
fill_value = 1e+20)
>>> populations.std(axis=0)
masked_array(data = [21087.656489 15625.7998142 3322.50622558],
mask = [False False False],
fill_value = 1e+20)
Note that Matplotlib knows about masked arrays:
>>> plt.plot(year, populations, o-)
80000
70000
60000
50000
40000
30000
20000
10000
0
1900 1905 1910 1915 1920

8.4.3 recarray: purely convenience
>>> arr = np.array([(a, 1), (b, 2)], dtype=[(x, S1), (y, int)])
>>> arr2 = arr.view(np.recarray)
>>> arr2.x
chararray([a, b],
dtype=|S1)
>>> arr2.y
array([1, 2])
8.4.4 matrix: convenience?
always 2-D
* is the matrix product, not the elementwise one
>>> np.matrix([[1, 0], [0, 1]]) * np.matrix([[1, 2], [3, 4]])
matrix([[1, 2],
[3, 4]])
8.5 Summary
Anatomy of the ndarray: data, dtype, strides.

Universal functions: elementwise operations, how to make new ones
Ndarray subclasses
Various buffer interfaces for integration with other tools
Recent additions: PEP 3118, generalized ufuncs
8.6 Contributing to Numpy/Scipy
Get this tutorial: http://www.euroscipy.org/talk/882
8.6.1 Why
Theres a bug?
I dont understand what this is supposed to do?
I have this fancy code. Would you like to have it?
Id like to help! What can I do?
8.6.2 Reporting bugs
Bug tracker (prefer this)

http://projects.scipy.org/numpy
http://projects.scipy.org/scipy
Click the Register link to get an account
Mailing lists ( scipy.org/Mailing_Lists )
If youre unsure
8.5. Summary 189

No replies in a week or so? Just file a bug ticket.
Good bug report
Title: numpy.random.permutations fails for non-integer arguments
Im trying to generate random permutations, using numpy.random.permutations
When calling numpy.random.permutation with non-integer arguments

it fails with a cryptic error message::
>>> np.random.permutation(12)
array([ 6, 11, 4, 10, 2, 8, 1, 7, 9, 3, 0, 5])
>>> np.random.permutation(12.)
File "mtrand.pyx", line 3311, in mtrand.RandomState.permutation
File "mtrand.pyx", line 3254, in mtrand.RandomState.shuffle
TypeError: len() of unsized object
This also happens with long arguments, and so

np.random.permutation(X.shape[0]) where X is an array fails on 64
bit windows (where shape is a tuple of longs).
It would be great if it could cast to integer or at least raise a

proper error for non-integer types.
Im using Numpy 1.4.1, built from the official tarball, on Windows

64 with Visual studio 2008, on Python.org 64-bit Python.
0. What are you trying to do?

1. Small code snippet reproducing the bug (if possible)
What actually happens
What youd expect
2. Platform (Windows / Linux / OSX, 32/64 bits, x86/PPC, ...)
3. Version of Numpy/Scipy
>>> print np.__version__
2...
Check that the following is what you expect

>>> print np.__file__
/...
In case you have old/broken Numpy installations lying around.

If unsure, try to remove existing Numpy installations, and reinstall...
8.6.3 Contributing to documentation
1. Documentation editor
http://docs.scipy.org/numpy
Registration
Register an account
Subscribe to scipy-dev mailing list (subscribers-only)
8.6. Contributing to Numpy/Scipy 190

Problem with mailing lists: you get mail
* But: you can turn mail delivery off

* change your subscription options, at the bottom of
http://mail.scipy.org/mailman/listinfo/scipy-dev
Send a mail @ scipy-dev mailing list; ask for activation:
To: scipy-dev@scipy.org
Hi,
Id like to edit Numpy/Scipy docstrings. My account is XXXXX
Cheers,
N. N.
Check the style guide:

http://docs.scipy.org/numpy/
Dont be intimidated; to fix a small thing, just fix it
Edit
2. Edit sources and send patches (as for bugs)
3. Complain on the mailing list
8.6.4 Contributing features
0. Ask on mailing list, if unsure where it should go

1. Write a patch, add an enhancement ticket on the bug tracket
2. OR, create a Git branch implementing the feature + add enhancement ticket.
Especially for big/invasive additions
http://projects.scipy.org/numpy/wiki/GitMirror
http://www.spheredev.org/wiki/Git_for_the_lazy
# Clone numpy repository
git clone --origin svn http://projects.scipy.org/git/numpy.git numpy
cd numpy
# Create a feature branch

git checkout -b name-of-my-feature-branch svn/trunk
<edit stuff>
git commit -a
Create account on http://github.com (or anywhere)

Create a new repository @ Github
Push your work to github
git remote add github git@github:YOURUSERNAME/YOURREPOSITORYNAME.git
git push github name-of-my-feature-branch

8.6.5 How to help, in general
Bug fixes always welcome!

What irks you most
Browse the tracker
Documentation work
API docs: improvements to docstrings
* Know some Scipy module well?

User guide
* Needs to be done eventually.

* Want to think? Come up with a Table of Contents
http://scipy.org/Developer_Zone/UG_Toc
Ask on communication channels:
numpy-discussion list
scipy-dev list

CHAPTER 9
Debugging code
author Gal Varoquaux

This tutorial explores tool 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.
Prerequisites
Numpy
IPython
nosetests (http://readthedocs.org/docs/nose/en/latest/)
pyflakes (http://pypi.python.org/pypi/pyflakes)
gdb for the C-debugging part.
Chapters contents
Avoiding bugs (page 193)
Coding best practices to avoid getting in trouble (page 193)
pyflakes: fast static analysis (page 194)
* Running pyflakes on the current edited file (page 194)
* A type-as-go spell-checker like integration (page 195)
Debugging workflow (page 196)
Using the Python debugger (page 196)
Invoking the debugger (page 197)
* Postmortem (page 197)
* Step-by-step execution (page 198)
* Other ways of starting a debugger (page 200)
Debugger commands and interaction (page 201)
* Getting help when in the debugger (page 201)
Debugging segmentation faults using gdb (page 201)
9.1 Avoiding bugs
9.1.1 Coding best practices to avoid getting in trouble
193
Brian Kernighan
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?
We all write buggy code. Accept it. Deal with it.

Write your code with testing and debugging in mind.
Keep It Simple, Stupid (KISS).
What is the simplest thing that could possibly work?
Dont Repeat Yourself (DRY).
Every piece of knowledge must have a single, unambiguous, authoritative representation within a
system.
Constants, algorithms, etc...
Try to limit interdependencies of your code. (Loose Coupling)
Give your variables, functions and modules meaningful names (not mathematics names)
9.1.2 pyflakes: fast static analysis
They are several static analysis tools in Python; to name a few:

pylint
pychecker
pyflakes
pep8
flake8
Here we focus on pyflakes, which is the simplest tool.
Fast, simple
Detects syntax errors, missing imports, typos on names.
Another good recommendation is the flake8 tool which is a combination of pyflakes and pep8. Thus, in addition
to the types of errors that pyflakes catches, flake8 detects violations of the recommendation in PEP8 style guide.
Integrating pyflakes (or flake8) in your editor or IDE is highly recommended, it does yield productivity gains.
Running pyflakes on the current edited file
You can bind a key to run pyflakes in the current buffer.

In kate Menu: settings -> configure kate
In plugins enable external tools
In external Tools, add pyflakes:
kdialog --title "pyflakes %filename" --msgbox "$(pyflakes %filename)"
In TextMate
Menu: TextMate -> Preferences -> Advanced -> Shell variables, add a shell variable:
TM_PYCHECKER=/Library/Frameworks/Python.framework/Versions/Current/bin/pyflakes
Then Ctrl-Shift-V is binded to a pyflakes report
9.1. Avoiding bugs 194

In vim In your vimrc (binds F5 to pyflakes):

autocmd FileType python let &mp = echo "*** running % ***" ; pyflakes %
autocmd FileType tex,mp,rst,python imap <Esc>[15~ <C-O>:make!^M
autocmd FileType tex,mp,rst,python map <Esc>[15~ :make!^M
autocmd FileType tex,mp,rst,python set autowrite
In emacs In your emacs (binds F5 to pyflakes):

(defun pyflakes-thisfile () (interactive)
(compile (format "pyflakes %s" (buffer-file-name)))
)
(define-minor-mode pyflakes-mode
"Toggle pyflakes mode.
With no argument, this command toggles the mode.
Non-null prefix argument turns on the mode.
Null prefix argument turns off the mode."
;; The initial value.
nil
;; The indicator for the mode line.
" Pyflakes"
;; The minor mode bindings.
( ([f5] . pyflakes-thisfile) )
)
(add-hook python-mode-hook (lambda () (pyflakes-mode t)))
A type-as-go spell-checker like integration
In vim
Use the pyflakes.vim plugin:
1. download the zip file from http://www.vim.org/scripts/script.php?script_id=2441
2. extract the files in ~/.vim/ftplugin/python
3. make sure your vimrc has filetype plugin indent on
Alternatively: use the syntastic plugin. This can be configured to use flake8 too and also handles
on-the-fly checking for many other languages.
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:
9.1. Avoiding bugs 195

(when (load "flymake" t)

(defun flymake-pyflakes-init ()
(let* ((temp-file (flymake-init-create-temp-buffer-copy
flymake-create-temp-inplace))
(local-file (file-relative-name
temp-file
(file-name-directory buffer-file-name))))
(list "pyflakes" (list local-file))))
(add-to-list flymake-allowed-file-name-masks
("\\.py\\" flymake-pyflakes-init)))
(add-hook find-file-hook flymake-find-file-hook)
9.2 Debugging workflow
I you do have a non trivial bug, this is when debugging strategies kick in. There is no silver bullet. Yet, strategies
help:
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-
fail cycles
1. Make it fail reliably. Find a test case that makes the code fail every time.
2. Divide and Conquer. Once you have a failing test case, isolate the failing code.
Which module.
Which function.
Which line of code.
=> isolate a small reproducible failure: a test case
3. Change one thing at a time and re-run the failing test case.
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.
9.3 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.
9.2. Debugging workflow 196

print
Yes, print statements do work as a debugging tool. However to inspect runtime, it is often more efficient
to use the debugger.
9.3.1 Invoking the debugger
Ways to launch the debugger:

1. Postmortem, launch debugger after module errors.
2. Launch the module with the debugger.
3. Call the debugger inside the module
Postmortem
Situation: Youre working in IPython and you get a traceback.

Here we debug the file index_error.py. When running it, an IndexError is raised. Type %debug and
drop into the debugger.
In [1]: %run index_error.py
---------------------------------------------------------------------------
IndexError Traceback (most recent call last)
/home/varoquau/dev/scipy-lecture-notes/advanced/debugging_optimizing/index_error.py in <module>()
6
7 if __name__ == __main__:
----> 8 index_error()
9
/home/varoquau/dev/scipy-lecture-notes/advanced/debugging_optimizing/index_error.py in index_error
3 def index_error():
4 lst = list(foobar)
----> 5 print lst[len(lst)]
6
7 if __name__ == __main__:
In [2]: %debug
> /home/varoquau/dev/scipy-lecture-notes/advanced/debugging_optimizing/index_error.py(5)index_erro
6
ipdb> list
1 """Small snippet to raise an IndexError."""
2
3 def index_error():
6
7 if __name__ == __main__:
8 index_error()
9
ipdb> len(lst)
6
ipdb> print lst[len(lst)-1]
9.3. Using the Python debugger 197

r
ipdb> quit
In [3]:
Post-mortem debugging without IPython

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:
$ python -m pdb index_error.py
> /home/varoquau/dev/scipy-lecture-notes/advanced/debugging_optimizing/index_error.py(1)<module>(
-> """Small snippet to raise an IndexError."""
(Pdb) continue
File "/usr/lib/python2.6/pdb.py", line 1296, in main
pdb._runscript(mainpyfile)
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
exec cmd in globals, locals
File "<string>", line 1, in <module>
File "index_error.py", line 8, in <module>
index_error()
File "index_error.py", line 5, in index_error
print lst[len(lst)]
Uncaught exception. Entering post mortem debugging
Running cont or step will restart the program
> /home/varoquau/dev/scipy-lecture-notes/advanced/debugging_optimizing/index_error.py(5)index_err
-> print lst[len(lst)]
(Pdb)
Step-by-step execution
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
work well.
Run the script in IPython with the debugger using %run -d wiener_filtering.p :
In [1]: %run -d wiener_filtering.py
*** Blank or comment
Breakpoint 1 at /home/varoquau/dev/scipy-lecture-notes/advanced/debugging_optimizing/wiener_f
NOTE: Enter c at the ipdb> prompt to start your script.
> <string>(1)<module>()
Set a break point at line 34 using b 34:

ipdb> n
> /home/varoquau/dev/scipy-lecture-notes/advanced/debugging_optimizing/wiener_filtering.py(4)
3
1---> 4 import numpy as np
5 import scipy as sp
ipdb> b 34
Breakpoint 2 at /home/varoquau/dev/scipy-lecture-notes/advanced/debugging_optimizing/wiener_f

Continue execution to next breakpoint with c(ont(inue)):

ipdb> c
> /home/varoquau/dev/scipy-lecture-notes/advanced/debugging_optimizing/wiener_filtering.py(34
33 """
2--> 34 noisy_img = noisy_img
35 denoised_img = local_mean(noisy_img, size=size)
Step into code with n(ext) and s(tep): next jumps to the next statement in the current execution
context, while step will go across execution contexts, i.e. enable exploring inside function calls:
ipdb> s
2 34 noisy_img = noisy_img
---> 35 denoised_img = local_mean(noisy_img, size=size)
36 l_var = local_var(noisy_img, size=size)
ipdb> n
35 denoised_img = local_mean(noisy_img, size=size)
---> 36 l_var = local_var(noisy_img, size=size)
37 for i in range(3):
Step a few lines and explore the local variables:

ipdb> n
36 l_var = local_var(noisy_img, size=size)
---> 37 for i in range(3):
38 res = noisy_img - denoised_img
ipdb> print l_var
[[5868 5379 5316 ..., 5071 4799 5149]
[5013 363 437 ..., 346 262 4355]
[5379 410 344 ..., 392 604 3377]
...,
[ 435 362 308 ..., 275 198 1632]
[ 548 392 290 ..., 248 263 1653]
[ 466 789 736 ..., 1835 1725 1940]]
ipdb> print l_var.min()
0
Oh dear, nothing but integers, and 0 variation. Here is our bug, we are doing integer arithmetic.

Raising exception on numerical errors

When we run the wiener_filtering.py file, the following warnings are raised:
In [2]: %run wiener_filtering.py
wiener_filtering.py:40: RuntimeWarning: divide by zero encountered in divide
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
our problem more quickly:
In [3]: np.seterr(all=raise)
Out[3]: {divide: print, invalid: print, over: print, under: ignore}
In [4]: %run wiener_filtering.py
---------------------------------------------------------------------------
FloatingPointError Traceback (most recent call last)
/home/esc/anaconda/lib/python2.7/site-packages/IPython/utils/py3compat.pyc in execfile(fname, *wh
176 else:
177 filename = fname
--> 178 __builtin__.execfile(filename, *where)
/home/esc/physique-cuso-python-2013/scipy-lecture-notes/advanced/debugging/wiener_filtering.py in
55 pl.matshow(noisy_lena[cut], cmap=pl.cm.gray)
56
---> 57 denoised_lena = iterated_wiener(noisy_lena)
58 pl.matshow(denoised_lena[cut], cmap=pl.cm.gray)
59
/home/esc/physique-cuso-python-2013/scipy-lecture-notes/advanced/debugging/wiener_filtering.py in
38 res = noisy_img - denoised_img
39 noise = (res**2).sum()/res.size
---> 40 noise_level = (1 - noise/l_var )
41 noise_level[noise_level<0] = 0
42 denoised_img += noise_level*res
FloatingPointError: divide by zero encountered in divide
Other ways of starting a debugger
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.

Graphical debuggers and alternatives

For stepping through code and inspecting variables, you might find it more convenient to use a graph-
ical debugger such as winpdb.
Alternatively, pudb is a good semi-graphical debugger with a text user interface in the console.
Also, the pydbgr project is probably worth looking at.
9.3.2 Debugger commands and interaction
l(list) Lists the code at the current position

u(p) Walk up the call stack
d(own) Walk down the call stack
n(ext) Execute the next line (does not go down in new functions)
s(tep) Execute the next statement (goes down in new functions)
bt Print the call stack
a Print the local variables
!command Execute the given Python command (by opposition to pdb commands
Warning: Debugger commands are not Python code

You cannot name the variables the way you want. For instance, if in you cannot override the variables in the
current frame with the same name: use different names then your local variable when typing code in the
debugger.
Getting help when in the debugger
Type h or help to access the interactive help:

ipdb> help
Documented commands (type help <topic>):

========================================
EOF bt cont enable jump pdef r tbreak w
a c continue exit l pdoc restart u whatis
alias cl d h list pinfo return unalias where
args clear debug help n pp run unt
b commands disable ignore next q s until
break condition down j p quit step up
Miscellaneous help topics:

==========================
exec pdb
Undocumented commands:
======================
retval rv
9.4 Debugging segmentation faults using gdb
If you have a segmentation fault, you cannot debug it with pdb, as it crashes the Python interpreter before it can
drop in the debugger. Similarly, if you have a bug in C code embedded in Python, pdb is useless. For this we turn
to the gnu debugger, gdb, available on Linux.
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
9.4. Debugging segmentation faults using gdb 201

a simplified version in gdbinit, but feel free to read DebuggingWithGdb.

To debug with gdb the Python script segfault.py, we can run the script in gdb as follows
$ gdb python
...
(gdb) run segfault.py
Starting program: /usr/bin/python segfault.py
[Thread debugging using libthread_db enabled]
Program received signal SIGSEGV, Segmentation fault.

_strided_byte_copy (dst=0x8537478 "\360\343G", outstrides=4, src=
0x86c0690 <Address 0x86c0690 out of bounds>, instrides=32, N=3,
elsize=4)
at numpy/core/src/multiarray/ctors.c:365
365 _FAST_MOVE(Int32);
(gdb)
We get a segfault, and gdb captures it for post-mortem debugging in the C level stack (not the Python call stack).
We can debug the C call stack using gdbs commands:
(gdb) up
#1 0x004af4f5 in _copy_from_same_shape (dest=<value optimized out>,
src=<value optimized out>, myfunc=0x496780 <_strided_byte_copy>,
swap=0)
at numpy/core/src/multiarray/ctors.c:748
748 myfunc(dit->dataptr, dest->strides[maxaxis],
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
#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
3750 ../Python/ceval.c: No such file or directory.
in ../Python/ceval.c
(gdb) up
#9 PyEval_EvalFrameEx (f=
Frame 0x85371ec, for file /home/varoquau/usr/lib/python2.6/site-packages/numpy/core/arrayprint
at ../Python/ceval.c:2412
2412 in ../Python/ceval.c
(gdb)
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:
(gdb) pyframe
/home/varoquau/usr/lib/python2.6/site-packages/numpy/core/arrayprint.py (158): _leading_trailing
(gdb)
This is numpy code, we need to go up until we find code that we have written:
(gdb) up
...
(gdb) up
#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.
in ../Python/ceval.c
(gdb) pyframe
segfault.py (12): print_big_array
The corresponding code is:

def make_big_array(small_array):
big_array = stride_tricks.as_strided(small_array,
shape=(2e6, 2e6), strides=(32, 32))
return big_array
def print_big_array(small_array):
big_array = make_big_array(small_array)
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.
Note: For a list of Python-specific commands defined in the gdbinit, read the source of this file.
Wrap up exercise
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?
Python source code: to_debug.py

CHAPTER 10
Optimizing code
Donald Knuth
Premature optimization is the root of all evil

This chapter deals with strategies to make Python code go faster.
Prerequisites
line_profiler
gprof2dot
dot utility from Graphviz
Chapters contents
Optimization workflow (page 204)
Profiling Python code (page 205)
Timeit (page 205)
Profiler (page 205)
Line-profiler (page 206)
Running cProfile (page 207)
Using gprof2dot (page 207)
Making code go faster (page 208)
Algorithmic optimization (page 208)
* Example of the SVD (page 208)
Writing faster numerical code (page 209)
Additional Links (page 211)
10.1 Optimization workflow
1. Make it work: write the code in a simple legible ways.

2. Make it work reliably: write automated test cases, make really sure that your algorithm is right and that if
you break it, the tests will capture the breakage.
204
3. Optimize the code by profiling simple use-cases to find the bottlenecks and speeding up these bottleneck,
finding a better algorithm or implementation. Keep in mind that a trade off should be found between
profiling on a realistic example and the simplicity and speed of execution of the code. For efficient work, it
is best to work with profiling runs lasting around 10s.
10.2 Profiling Python code
No optimization without measuring!

Measure: profiling, timing
Youll have surprises: the fastest code is not always what you think
10.2.1 Timeit
In IPython, use timeit (http://docs.python.org/library/timeit.html) to time elementary operations:

In [2]: a = np.arange(1000)
In [3]: %timeit a ** 2
In [4]: %timeit a ** 2.1

1000 loops, best of 3: 154 us per loop
In [5]: %timeit a * a
Use this to guide your choice between strategies.
Note: For long running calls, using %time instead of %timeit; it is less precise but faster
10.2.2 Profiler
Useful when you have a large program to profile, for example the following file:
# For this example to run, you also need the ica.py file
import numpy as np
from scipy import linalg
from ica import fastica
def test():
data = np.random.random((5000, 100))
u, s, v = linalg.svd(data)
pca = np.dot(u[:, :10].T, data)
results = fastica(pca.T, whiten=False)
test()
10.2. Profiling Python code 205

Note: This is a combination of two unsupervised learning techniques, principal component analysis (PCA) and
independent component analysis (ICA<http://en.wikipedia.org/wiki/Independent_component_ana lysis>_).
PCA is a technique for dimensionality reduction, i.e. an algorithm to explain the observed variance in your data
using less dimensions. ICA is a source seperation technique, for example to unmix multiple signals that have been
recorded through multiple sensors. Doing a PCA first and then an ICA can be useful if you have more sensors
than signals. For more information see: the FastICA example from scikits-learn.
To run it, you also need to download the ica module. In IPython we can time the script:
In [1]: %run -t demo.py
IPython CPU timings (estimated):

User : 14.3929 s.
System: 0.256016 s.
and profile it:

In [2]: %run -p demo.py
916 function calls in 14.551 CPU seconds
Ordered by: internal time
ncalls tottime percall cumtime percall filename:lineno(function)

1 14.457 14.457 14.479 14.479 decomp.py:849(svd)
1 0.054 0.054 0.054 0.054 {method random_sample of mtrand.RandomState objects
1 0.017 0.017 0.021 0.021 function_base.py:645(asarray_chkfinite)
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}
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)
14 0.001 0.000 0.001 0.000 {numpy.linalg.lapack_lite.dsyevd}
19 0.001 0.000 0.001 0.000 twodim_base.py:204(diag)
1 0.001 0.001 0.008 0.008 ica.py:69(_ica_par)
1 0.001 0.001 14.551 14.551 {execfile}
107 0.000 0.000 0.001 0.000 defmatrix.py:239(__array_finalize__)
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)
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__)
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__)
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)
...
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
make this step go faster, or to avoid this step (algorithmic optimization). Spending time on the rest of the code is
useless.
10.2.3 Line-profiler
The profiler is great: it 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
@profile (no need to import it)

@profile
def test():
data = np.random.random((5000, 100))
u, s, v = linalg.svd(data)
pca = np.dot(u[: , :10], data)
results = fastica(pca.T, whiten=False)
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:
$ kernprof.py -l -v demo.py
Wrote profile results to demo.py.lprof

Timer unit: 1e-06 s
File: demo.py
Function: test at line 5
Total time: 14.2793 s
Line # Hits Time Per Hit % Time Line Contents

==============================================================
5 @profile
6 def test():
7 1 19015 19015.0 0.1 data = np.random.random((5000, 100))
8 1 14242163 14242163.0 99.7 u, s, v = linalg.svd(data)
9 1 10282 10282.0 0.1 pca = np.dot(u[:10, :], data)
10 1 7799 7799.0 0.1 results = fastica(pca.T, whiten=False)
The SVD is taking all the time. We need to optimise this line.
10.2.4 Running cProfile
In the IPython example above, IPython simply calls the built-in Python profilers cProfile and profile. This
can be useful if you wish to process the profiler output with a visualization tool.
$ python -m cProfile -o demo.prof demo.py
Using the -o switch will output the profiler results to the file demo.prof.
10.2.5 Using gprof2dot
In case you want a more visual representation of the profiler output, you can use the gprof2dot tool:
$ gprof2dot -f pstats demo.prof | dot -Tpng -o demo-prof.png
Which will produce the following picture:

Which again paints a similar picture as the previous approaches.
10.3 Making code go faster
Once we have identified the bottlenecks, we need to make the corresponding code go faster.
10.3.1 Algorithmic optimization
The first thing to look for is algorithmic optimization: are there ways to compute less, or better?
For a high-level view of the problem, a good understanding of the maths behind the algorithm helps. However, it
is not uncommon to find simple changes, like moving computation or memory allocation outside a for loop,
that bring in big gains.
Example of the SVD
In both examples above, the SVD - Singular Value Decomposition - is what takes most of the time. Indeed, the
computational cost of this algorithm is roughly n3 in the size of the input matrix.
However, in both of these example, we are not using all the output of the SVD, but only the first few rows of its
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.
10.3. Making code go faster 208

In [3]: %timeit np.linalg.svd(data)

1 loops, best of 3: 14.5 s per loop
In [4]: from scipy import linalg
In [5]: %timeit linalg.svd(data)

In [6]: %timeit linalg.svd(data, full_matrices=False)

1 loops, best of 3: 295 ms per loop
In [7]: %timeit np.linalg.svd(data, full_matrices=False)

We can then use this insight to optimize the previous code:

In [1]: import demo
In [2]: %timeit demo.

demo.fastica demo.np demo.prof.pdf demo.py demo.pyc
demo.linalg demo.prof demo.prof.png demo.py.lprof demo.test
In [2]: %timeit demo.test()

ica.py:65: RuntimeWarning: invalid value encountered in sqrt
W = (u * np.diag(1.0/np.sqrt(s)) * u.T) * W # W = (W * W.T) ^{-1/2} * W
In [3]: import demo_opt
In [4]: %timeit demo_opt.test()

Real incomplete SVDs, e.g. computing only the first 10 eigenvectors, can be computed with arpack, available in
scipy.sparse.linalg.eigsh.
Computational linear algebra

For certain algorithms, many of the bottlenecks will be linear algebra computations. In this case, using the
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
less costly (and more numerically stable) operation.
Know your computational linear algebra. When in doubt, explore scipy.linalg, and use %timeit to
try out different alternatives on your data.
10.4 Writing faster numerical code
A complete discussion on advanced use of numpy is found in chapter Advanced Numpy (page 160), or in the
article The NumPy array: a structure for efficient numerical computation by van der Walt et al. Here we discuss
only some commonly encountered tricks to make code faster.
Vectorizing for loops
Find tricks to avoid for loops using numpy arrays. For this, masks and indices arrays can be useful.
Broadcasting
Use broadcasting (page 59) to do operations on arrays as small as possible before combining them.
In place operations
10.4. Writing faster numerical code 209

In [1]: a = np.zeros(1e7)
In [2]: %timeit global a ; a = 0*a

In [3]: %timeit global a ; a *= 0

note: we need global a in the timeit so that it work, as it is assigning to a, and thus considers it as a
local variable.
Be easy on the memory: use views, and not copies
Copying big arrays is as costly as making simple numerical operations on them:
In [1]: a = np.zeros(1e7)
In [2]: %timeit a.copy()

In [3]: %timeit a + 1
Beware of cache effects

Memory access is cheaper when it is grouped: accessing a big array in a continuous way is much faster than
random access. This implies amongst other things that smaller strides are faster (see CPU cache effects
(page 173)):
In [1]: c = np.zeros((1e4, 1e4), order=C)
In [2]: %timeit c.sum(axis=0)

In [3]: %timeit c.sum(axis=1)

In [4]: c.strides
Out[4]: (80000, 8)
This is the reason why Fortran ordering or C ordering may make a big difference on operations:
In [5]: a = np.random.rand(20, 2**18)
In [6]: b = np.random.rand(20, 2**18)
In [7]: %timeit np.dot(b, a.T)

In [8]: c = np.ascontiguousarray(a.T)
In [9]: %timeit np.dot(b, c)

Note that copying the data to work around this effect may not be worth it:
In [10]: %timeit c = np.ascontiguousarray(a.T)
Using numexpr can be useful to automatically optimize code for such effects.
Use compiled code
The last resort, once you are sure that all the high-level optimizations have been explored, is to transfer the
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,

and with a good use of the numpy support yields efficient code on numpy arrays, for instance by unrolling
loops.
Warning: For all the above: profile and time your choices. Dont base your optimization on theoretical
considerations.
10.4.1 Additional Links
If you need to profile memory usage, you could try the memory_profiler
If you need to profile down into C extensions, you could try using gperftools from Python with yep.
If you would like to track performace of your code across time, i.e. as you make new commits to your
repository, you could try: vbench
If you need some interactive visualization why not try RunSnakeRun

CHAPTER 11
Sparse Matrices in SciPy
author Robert Cimrman
11.1 Introduction
(dense) matrix is:

mathematical object
data structure for storing a 2D array of values
important features:
memory allocated once for all items
usually a contiguous chunk, think NumPy ndarray
fast access to individual items (*)
11.1.1 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)
>>> plt.ylabel(memory [MB])
11.1.2 Sparse Matrices vs. Sparse Matrix Storage Schemes
sparse matrix is a matrix, which is almost empty

storing all the zeros is wasteful -> store only nonzero items
think compression
pros: huge memory savings
212
cons: depends on actual storage scheme, (*) usually does not hold
11.1.3 Typical Applications
solution of partial differential equations (PDEs)

the finite element method
mechanical engineering, electrotechnics, physics, ...
graph theory
nonzero at (i, j) means that node i is connected to node j
...
11.1.4 Prerequisites
recent versions of
numpy
scipy
matplotlib (optional)
ipython (the enhancements come handy)
11.1.5 Sparsity Structure Visualization
spy() from matplotlib

example plots:

11.2 Storage Schemes
seven sparse matrix types in scipy.sparse:

1. csc_matrix: Compressed Sparse Column format
2. csr_matrix: Compressed Sparse Row format
3. bsr_matrix: Block Sparse Row format
4. lil_matrix: List of Lists format
5. dok_matrix: Dictionary of Keys format
6. coo_matrix: COOrdinate format (aka IJV, triplet format)
7. dia_matrix: DIAgonal format
each suitable for some tasks
many employ sparsetools C++ module by Nathan Bell
assume the following is imported:
>>> import scipy.sparse as sps
warning for NumPy users:

the multiplication with * is the matrix multiplication (dot product)
not part of NumPy!
* passing a sparse matrix object to NumPy functions expecting ndarray/matrix does not work
11.2. Storage Schemes 214

11.2.1 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
11.2.2 Sparse Matrix Classes
Diagonal Format (DIA)
very simple scheme

diagonals in dense NumPy array of shape (n_diag, length)
fixed length -> waste space a bit when far from main diagonal
subclass of _data_matrix (sparse matrix classes with data attribute)
offset for each diagonal
0 is the main diagonal
negative offset = below
positive offset = above
fast matrix * vector (sparsetools)
fast and easy item-wise operations
manipulate data array directly (fast NumPy machinery)
constructor accepts:
dense matrix (array)
sparse matrix
shape tuple (create empty matrix)
(data, offsets) tuple
no slicing, no individual item access

use:
rather specialized
solving PDEs by finite differences
with an iterative solver
Examples
create some DIA matrices:

>>> data = np.array([[1, 2, 3, 4]]).repeat(3, axis=0)
>>> data
array([[1, 2, 3, 4],
[1, 2, 3, 4],
[1, 2, 3, 4]])
>>> offsets = np.array([0, -1, 2])
>>> mtx = sparse.dia_matrix((data, offsets), shape=(4, 4))
>>> mtx
<4x4 sparse matrix of type <type numpy.int64>
with 9 stored elements (3 diagonals) in DIAgonal format>
>>> mtx.todense()
matrix([[1, 0, 3, 0],
[1, 2, 0, 4],
[0, 2, 3, 0],
[0, 0, 3, 4]])
>>> data = np.arange(12).reshape((3, 4)) + 1

>>> data
array([[ 1, 2, 3, 4],
[ 5, 6, 7, 8],
[ 9, 10, 11, 12]])
>>> mtx = sparse.dia_matrix((data, offsets), shape=(4, 4))
>>> mtx.data
array([[ 1, 2, 3, 4],
[ 5, 6, 7, 8],
[ 9, 10, 11, 12]])
>>> mtx.offsets
array([ 0, -1, 2], dtype=int32)
>>> print mtx
(0, 0) 1
(1, 1) 2
(2, 2) 3
(3, 3) 4
(1, 0) 5
(2, 1) 6
(3, 2) 7
(0, 2) 11
(1, 3) 12
>>> mtx.todense()
matrix([[ 1, 0, 11, 0],
[ 5, 2, 0, 12],
[ 0, 6, 3, 0],
[ 0, 0, 7, 4]])
explanation with a scheme:

offset: row
2: 9
1: --10------
0: 1 . 11 .

-1: 5 2 . 12
-2: . 6 3 .
-3: . . 7 4
---------8
matrix-vector multiplication
>>> vec = np.ones((4, ))
>>> vec
array([ 1., 1., 1., 1.])
>>> mtx * vec
array([ 12., 19., 9., 11.])
>>> mtx.toarray() * vec
array([[ 1., 0., 11., 0.],
[ 5., 2., 0., 12.],
[ 0., 6., 3., 0.],
[ 0., 0., 7., 4.]])
List of Lists Format (LIL)
row-based linked list

each row is a Python list (sorted) of column indices of non-zero elements
rows stored in a NumPy array (dtype=np.object)
non-zero values data stored analogously
efficient for constructing sparse matrices incrementally
sparse matrix
flexible slicing, changing sparsity structure is efficient
slow arithmetics, slow column slicing due to being row-based
use:
when sparsity pattern is not known apriori or changes
example: reading a sparse matrix from a text file
Examples
create an empty LIL matrix:

>>> mtx = sparse.lil_matrix((4, 5))
prepare random data:

>>> from numpy.random import rand
>>> data = np.round(rand(2, 3))
>>> data
array([[ 1., 1., 1.],
[ 1., 0., 1.]])
assign the data using fancy indexing:

>>> mtx[:2, [1, 2, 3]] = data

>>> mtx
<4x5 sparse matrix of type <type numpy.float64>
with 5 stored elements in LInked List format>
>>> print mtx
(0, 1) 1.0
(0, 2) 1.0
(0, 3) 1.0
(1, 1) 1.0
(1, 3) 1.0
>>> mtx.todense()
matrix([[ 0., 1., 1., 1., 0.],
[ 0., 1., 0., 1., 0.],
[ 0., 0., 0., 0., 0.],
[ 0., 0., 0., 0., 0.]])
>>> mtx.toarray()
array([[ 0., 1., 1., 1., 0.],
[ 0., 1., 0., 1., 0.],
[ 0., 0., 0., 0., 0.],
[ 0., 0., 0., 0., 0.]])
more slicing and indexing:

>>> mtx = sparse.lil_matrix([[0, 1, 2, 0], [3, 0, 1, 0], [1, 0, 0, 1]])
>>> mtx.todense()
matrix([[0, 1, 2, 0],
[3, 0, 1, 0],
[1, 0, 0, 1]])
>>> print mtx
(0, 1) 1
(0, 2) 2
(1, 0) 3
(1, 2) 1
(2, 0) 1
(2, 3) 1
>>> mtx[:2, :]
with 4 stored elements in LInked List format>
>>> mtx[:2, :].todense()
matrix([[0, 1, 2, 0],
[3, 0, 1, 0]])
>>> mtx[1:2, [0,2]].todense()
matrix([[3, 1]])
>>> mtx.todense()
matrix([[0, 1, 2, 0],
[3, 0, 1, 0],
[1, 0, 0, 1]])
Dictionary of Keys Format (DOK)
subclass of Python dict

keys are (row, column) index tuples (no duplicate entries allowed)
values are corresponding non-zero values
efficient for constructing sparse matrices incrementally
sparse matrix


efficient O(1) access to individual elements
flexible slicing, changing sparsity structure is efficient
can be efficiently converted to a coo_matrix once constructed
slow arithmetics (for loops with dict.iteritems())
use:
when sparsity pattern is not known apriori or changes
Examples
create a DOK matrix element by element:

>>> mtx = sparse.dok_matrix((5, 5), dtype=np.float64)
>>> mtx
with 0 stored elements in Dictionary Of Keys format>
>>> for ir in range(5):
... for ic in range(5):
... mtx[ir, ic] = 1.0 * (ir != ic)
>>> mtx
>>> mtx.todense()
matrix([[ 0., 1., 1., 1., 1.],
[ 1., 0., 1., 1., 1.],
[ 1., 1., 0., 1., 1.],
[ 1., 1., 1., 0., 1.],
[ 1., 1., 1., 1., 0.]])
slicing and indexing:

>>> mtx[1, 1]
0.0
>>> mtx[1, 1:3]
>>> mtx[1, 1:3].todense()
matrix([[ 0., 1.]])
>>> mtx[[2,1], 1:3].todense()
...
NotImplementedError: fancy indexing supported over one axis only
Coordinate Format (COO)
also known as the ijv or triplet format

three NumPy arrays: row, col, data
data[i] is value at (row[i], col[i]) position
permits duplicate entries
subclass of _data_matrix (sparse matrix classes with data attribute)
fast format for constructing sparse matrices

sparse matrix
(data, ij) tuple
very fast conversion to and from CSR/CSC formats
fast matrix * vector (sparsetools)
fast and easy item-wise operations
manipulate data array directly (fast NumPy machinery)
no slicing, no arithmetics (directly)
use:
facilitates fast conversion among sparse formats
when converting to other format (usually CSR or CSC), duplicate entries are summed together
* facilitates efficient construction of finite element matrices
Examples
create empty COO matrix:

>>> mtx = sparse.coo_matrix((3, 4), dtype=np.int8)
>>> mtx.todense()
matrix([[0, 0, 0, 0],
[0, 0, 0, 0],
[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])
>>> data = np.array([4, 5, 7, 9])
>>> mtx = sparse.coo_matrix((data, (row, col)), shape=(4, 4))
>>> mtx
with 4 stored elements in COOrdinate format>
>>> mtx.todense()
matrix([[4, 0, 9, 0],
[0, 7, 0, 0],
[0, 0, 0, 0],
[0, 0, 0, 5]])
duplicates entries are summed together:

>>> row = np.array([0, 0, 1, 3, 1, 0, 0])
>>> col = np.array([0, 2, 1, 3, 1, 0, 0])
>>> data = np.array([1, 1, 1, 1, 1, 1, 1])
>>> mtx = sparse.coo_matrix((data, (row, col)), shape=(4, 4))
>>> mtx.todense()
matrix([[3, 0, 1, 0],
[0, 2, 0, 0],
[0, 0, 0, 0],
[0, 0, 0, 1]])
no slicing...:
>>> mtx[2, 3]
...
TypeError: coo_matrix object ...

Compressed Sparse Row Format (CSR)
row oriented
three NumPy arrays: indices, indptr, data
* indices is array of column indices

* data is array of corresponding nonzero values
* indptr points to row starts in indices and data
* length is n_row + 1, last item = number of values = length of both indices and data
* nonzero values of the i-th row are data[indptr[i]:indptr[i+1]] with column
indices indices[indptr[i]:indptr[i+1]]
* item (i, j) can be accessed as data[indptr[i]+k], where k is position of j in

indices[indptr[i]:indptr[i+1]]
subclass of _cs_matrix (common CSR/CSC functionality)
* subclass of _data_matrix (sparse matrix classes with data attribute)

fast matrix vector products and other arithmetics (sparsetools)
sparse matrix
(data, ij) tuple
(data, indices, indptr) tuple
efficient row slicing, row-oriented operations
slow column slicing, expensive changes to the sparsity structure
use:
actual computations (most linear solvers support this format)
Examples
create empty CSR matrix:

>>> mtx = sparse.csr_matrix((3, 4), dtype=np.int8)
>>> mtx.todense()
matrix([[0, 0, 0, 0],
[0, 0, 0, 0],
[0, 0, 0, 0]], dtype=int8)

>>> 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])
>>> mtx = sparse.csr_matrix((data, (row, col)), shape=(3, 3))
>>> mtx
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.indices
array([0, 2, 2, 0, 1, 2], dtype=int32)
>>> mtx.indptr
create using (data, indices, indptr) tuple:

>>> data = np.array([1, 2, 3, 4, 5, 6])
>>> indices = np.array([0, 2, 2, 0, 1, 2])
>>> indptr = np.array([0, 2, 3, 6])
>>> mtx = sparse.csr_matrix((data, indices, indptr), shape=(3, 3))
>>> mtx.todense()
matrix([[1, 0, 2],
[0, 0, 3],
[4, 5, 6]])
Compressed Sparse Column Format (CSC)
column oriented
* indices is array of row indices

* data is array of corresponding nonzero values
* indptr points to column starts in indices and data
* length is n_col + 1, last item = number of values = length of both indices and data
* nonzero values of the i-th column are data[indptr[i]:indptr[i+1]] with row
indices indices[indptr[i]:indptr[i+1]]
* item (i, j) can be accessed as data[indptr[j]+k], where k is position of i in

indices[indptr[j]:indptr[j+1]]

sparse matrix
(data, ij) tuple
efficient column slicing, column-oriented operations
slow row slicing, expensive changes to the sparsity structure
use:
actual computations (most linear solvers support this format)

Examples
create empty CSC matrix:

>>> mtx = sparse.csc_matrix((3, 4), dtype=np.int8)
>>> mtx.todense()
matrix([[0, 0, 0, 0],
[0, 0, 0, 0],
[0, 0, 0, 0]], dtype=int8)

>>> 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])
>>> mtx = sparse.csc_matrix((data, (row, col)), shape=(3, 3))
>>> mtx
with 6 stored elements in Compressed Sparse Column format>
>>> mtx.todense()
matrix([[1, 0, 2],
[0, 0, 3],
[4, 5, 6]])
>>> mtx.data
array([1, 4, 5, 2, 3, 6])
>>> mtx.indices
array([0, 2, 2, 0, 1, 2], dtype=int32)
>>> mtx.indptr
create using (data, indices, indptr) tuple:

>>> data = np.array([1, 4, 5, 2, 3, 6])
>>> indices = np.array([0, 2, 2, 0, 1, 2])
>>> indptr = np.array([0, 2, 3, 6])
>>> mtx = sparse.csc_matrix((data, indices, indptr), shape=(3, 3))
>>> mtx.todense()
matrix([[1, 0, 2],
[0, 0, 3],
[4, 5, 6]])
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)
* indices is array of column indices for each block

* data is array of corresponding nonzero values of shape (nnz, R, C)
* ...

sparse matrix


(data, ij) tuple
many arithmetic operations considerably more efficient than CSR for sparse matrices with dense sub-
matrices
use:
like CSR
vector-valued finite element discretizations
Examples
create empty BSR matrix with (1, 1) block size (like CSR...):
>>> mtx = sparse.bsr_matrix((3, 4), dtype=np.int8)
>>> mtx
with 0 stored elements (blocksize = 1x1) in Block Sparse Row format>
>>> mtx.todense()
matrix([[0, 0, 0, 0],
[0, 0, 0, 0],
[0, 0, 0, 0]], dtype=int8)
create empty BSR matrix with (3, 2) block size:

>>> mtx = sparse.bsr_matrix((3, 4), blocksize=(3, 2), dtype=np.int8)
>>> mtx
>>> mtx.todense()
matrix([[0, 0, 0, 0],
[0, 0, 0, 0],
[0, 0, 0, 0]], dtype=int8)
a bug?
create using (data, ij) tuple with (1, 1) block size (like CSR...):
>>> 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])
>>> mtx = sparse.bsr_matrix((data, (row, col)), shape=(3, 3))
>>> mtx
>>> mtx.todense()
matrix([[1, 0, 2],
[0, 0, 3],
[4, 5, 6]])
>>> mtx.data
array([[[1]],
[[2]],
[[3]],
[[4]],
[[5]],

[[6]]])
>>> mtx.indices
array([0, 2, 2, 0, 1, 2], dtype=int32)
>>> mtx.indptr
create using (data, indices, indptr) tuple with (2, 2) block size:
>>> indptr = np.array([0, 2, 3, 6])
>>> indices = np.array([0, 2, 2, 0, 1, 2])
>>> data = np.array([1, 2, 3, 4, 5, 6]).repeat(4).reshape(6, 2, 2)
>>> mtx = sparse.bsr_matrix((data, indices, indptr), shape=(6, 6))
>>> mtx.todense()
matrix([[1, 1, 0, 0, 2, 2],
[1, 1, 0, 0, 2, 2],
[0, 0, 0, 0, 3, 3],
[0, 0, 0, 0, 3, 3],
[4, 4, 5, 5, 6, 6],
[4, 4, 5, 5, 6, 6]])
>>> data
array([[[1, 1],
[1, 1]],
[[2, 2],
[2, 2]],
[[3, 3],
[3, 3]],
[[4, 4],
[4, 4]],
[[5, 5],
[5, 5]],
[[6, 6],
[6, 6]]])
11.2.3 Summary
Table 11.1: Summary of storage schemes.

for- matrix * get fancy set fancy solvers note
mat vector item get item set
DIA sparsetools . . . . itera- has data array, specialized
tive
LIL via CSR yes yes yes yes itera- arithmetics via CSR,
tive incremental construction
DOK python yes one axis yes yes itera- O(1) item access, incremental
only tive construction
COO sparsetools . . . . itera- has data array, facilitates fast
tive conversion
CSR sparsetools yes yes slow . any has data array, fast row-wise
ops
CSC sparsetools yes yes slow . any has data array, fast column-wise
ops
BSR sparsetools . . . . spe- has data array, specialized
cial-
ized

11.3 Linear System Solvers
sparse matrix/eigenvalue problem solvers live in scipy.sparse.linalg

the submodules:
dsolve: direct factorization methods for solving linear systems
isolve: iterative methods for solving linear systems
eigen: sparse eigenvalue problem solvers
all solvers are accessible from:
>>> import scipy.sparse.linalg as spla
>>> spla.__all__
[LinearOperator, Tester, arpack, aslinearoperator, bicg,
bicgstab, cg, cgs, csc_matrix, csr_matrix, dsolve,
eigen, eigen_symmetric, factorized, gmres, interface,
isolve, iterative, lgmres, linsolve, lobpcg, lsqr,
minres, np, qmr, speigs, spilu, splu, spsolve, svd,
test, umfpack, use_solver, utils, warnings]
11.3.1 Sparse Direct Solvers
default solver: SuperLU 4.0

included in SciPy
real and complex systems
both single and double precision
optional: umfpack
real and complex systems
double precision only
recommended for performance
wrappers now live in scikits.umfpack
check-out the new scikits.suitesparse by Nathaniel Smith
Examples
import the whole module, and see its docstring:

>>> from scipy.sparse.linalg import dsolve
>>> help(dsolve)
both superlu and umfpack can be used (if the latter is installed) as follows:
prepare a linear system:
>>> from scipy import sparse
>>> mtx = sparse.spdiags([[1, 2, 3, 4, 5], [6, 5, 8, 9, 10]], [0, 1], 5, 5)
>>> mtx.todense()
matrix([[ 1, 5, 0, 0, 0],
[ 0, 2, 8, 0, 0],
[ 0, 0, 3, 9, 0],
[ 0, 0, 0, 4, 10],
[ 0, 0, 0, 0, 5]])
>>> rhs = np.array([1, 2, 3, 4, 5])
11.3. Linear System Solvers 226

solve as single precision real:

>>> mtx1 = mtx.astype(np.float32)
>>> x = dsolve.spsolve(mtx1, rhs, use_umfpack=False)
>>> print x
[ 106. -21. 5.5 -1.5 1. ]
>>> print "Error: ", mtx1 * x - rhs
Error: [ 0. 0. 0. 0. 0.]
solve as double precision real:

>>> mtx2 = mtx.astype(np.float64)
>>> x = dsolve.spsolve(mtx2, rhs, use_umfpack=True)
>>> print x
[ 106. -21. 5.5 -1.5 1. ]
Error: [ 0. 0. 0. 0. 0.]
solve as single precision complex:

>>> mtx1 = mtx.astype(np.complex64)
>>> x = dsolve.spsolve(mtx1, rhs, use_umfpack=False)
>>> print x
[ 106.0+0.j -21.0+0.j 5.5+0.j -1.5+0.j 1.0+0.j]
Error: [ 0.+0.j 0.+0.j 0.+0.j 0.+0.j 0.+0.j]
solve as double precision complex:

>>> mtx2 = mtx.astype(np.complex128)
>>> x = dsolve.spsolve(mtx2, rhs, use_umfpack=True)
>>> print x
[ 106.0+0.j -21.0+0.j 5.5+0.j -1.5+0.j 1.0+0.j]
Error: [ 0.+0.j 0.+0.j 0.+0.j 0.+0.j 0.+0.j]
"""
Construct a 1000x1000 lil_matrix and add some values to it, convert it
to CSR format and solve A x = b for x:and solve a linear system with a
direct solver.
"""
import numpy as np
import scipy.sparse as sps
from matplotlib import pyplot as plt
from scipy.sparse.linalg.dsolve import linsolve
rand = np.random.rand
mtx = sps.lil_matrix((1000, 1000), dtype=np.float64)

mtx[0, :100] = rand(100)
mtx[1, 100:200] = mtx[0, :100]
mtx.setdiag(rand(1000))
plt.clf()
plt.spy(mtx, marker=., markersize=2)
plt.show()
mtx = mtx.tocsr()
rhs = rand(1000)
x = linsolve.spsolve(mtx, rhs)
print rezidual:, np.linalg.norm(mtx * x - rhs)
examples/direct_solve.py

11.3.2 Iterative Solvers
the isolve module contains the following solvers:

bicg (BIConjugate Gradient)
bicgstab (BIConjugate Gradient STABilized)
cg (Conjugate Gradient) - symmetric positive definite matrices only
cgs (Conjugate Gradient Squared)
gmres (Generalized Minimal RESidual)
minres (MINimum RESidual)
qmr (Quasi-Minimal Residual)
Common Parameters
mandatory:
A [{sparse matrix, dense matrix, LinearOperator}] The N-by-N matrix of the linear system.
b [{array, matrix}] Right hand side of the linear system. Has shape (N,) or (N,1).
optional:
x0 [{array, matrix}] Starting guess for the solution.
tol [float] Relative tolerance to achieve before terminating.
maxiter [integer] Maximum number of iterations. Iteration will stop after maxiter steps even if the specified
tolerance has not been achieved.
M [{sparse matrix, dense matrix, LinearOperator}] Preconditioner for A. The preconditioner should ap-
proximate the inverse of A. Effective preconditioning dramatically improves the rate of convergence,
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
xk is the current solution vector.
LinearOperator Class
from scipy.sparse.linalg.interface import LinearOperator
common interface for performing matrix vector products

useful abstraction that enables using dense and sparse matrices within the solvers, as well as matrix-free
solutions
has shape and matvec() (+ some optional parameters)
example:
>>> from scipy.sparse.linalg import LinearOperator
>>> def mv(v):
... return np.array([2*v[0], 3*v[1]])
...
>>> A = LinearOperator((2, 2), matvec=mv)
>>> A
<2x2 LinearOperator with unspecified dtype>
>>> A.matvec(np.ones(2))
array([ 2., 3.])
>>> A * np.ones(2)
array([ 2., 3.])

A Few Notes on Preconditioning
problem specific
often hard to develop
if not sure, try ILU
available in dsolve as spilu()
11.3.3 Eigenvalue Problem Solvers
The eigen module
arpack * a collection of Fortran77 subroutines designed to solve large scale eigenvalue problems
lobpcg (Locally Optimal Block Preconditioned Conjugate Gradient Method) * works very well in com-
bination with PyAMG * example by Nathan Bell:
"""
Compute eigenvectors and eigenvalues using a preconditioned eigensolver
In this example Smoothed Aggregation (SA) is used to precondition

the LOBPCG eigensolver on a two-dimensional Poisson problem with
Dirichlet boundary conditions.
"""
import scipy
from scipy.sparse.linalg import lobpcg
from pyamg import smoothed_aggregation_solver

from pyamg.gallery import poisson
N = 100
K = 9
A = poisson((N,N), format=csr)
# create the AMG hierarchy

ml = smoothed_aggregation_solver(A)
# initial approximation to the K eigenvectors

X = scipy.rand(A.shape[0], K)
# preconditioner based on ml
M = ml.aspreconditioner()
# compute eigenvalues and eigenvectors with LOBPCG

W,V = lobpcg(A, X, M=M, tol=1e-8, largest=False)
#plot the eigenvectors

import pylab
pylab.figure(figsize=(9,9))
for i in range(K):
pylab.subplot(3, 3, i+1)
pylab.title(Eigenvector %d % i)
pylab.pcolor(V[:,i].reshape(N,N))
pylab.axis(equal)
pylab.axis(off)
pylab.show()

examples/pyamg_with_lobpcg.py
example by Nils Wagner:
examples/lobpcg_sakurai.py
output:
$ python examples/lobpcg_sakurai.py
Results by LOBPCG for n=2500
[ 0.06250083 0.06250028 0.06250007]
Exact eigenvalues
[ 0.06250005 0.0625002 0.06250044]
Elapsed time 7.01
11.4 Other Interesting Packages
PyAMG
algebraic multigrid solvers
http://code.google.com/p/pyamg
Pysparse
own sparse matrix classes
matrix and eigenvalue problem solvers
11.4. Other Interesting Packages 230

http://pysparse.sourceforge.net/
11.4. Other Interesting Packages 231

CHAPTER 12
Image manipulation and processing

using Numpy and Scipy
authors Emmanuelle Gouillart, Gal Varoquaux
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
a few examples use specialized toolkits working with np.array:

scikit-image
scikit-learn
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
...
More powerful and complete modules:
OpenCV (Python bindings)
CellProfiler
232
ITK with Python bindings

many more...
Chapters contents
Opening and writing to image files (page 233)
Displaying images (page 234)
Basic manipulations (page 236)
Statistical information (page 236)
Geometrical transformations (page 237)
Image filtering (page 238)
Blurring/smoothing (page 238)
Sharpening (page 238)
Denoising (page 239)
Mathematical morphology (page 240)
Feature extraction (page 243)
Edge detection (page 243)
Segmentation (page 243)
Measuring objects properties: ndimage.measurements (page 246)
12.1 Opening and writing to image files
Writing an array to a file:

from scipy import misc
l = misc.lena()
misc.imsave(lena.png, l) # uses the Image module (PIL)
import matplotlib.pyplot as plt

plt.imshow(l)
plt.show()
Creating a numpy array from an image file:

>>> lena = misc.imread(lena.png)
>>> type(lena)
12.1. Opening and writing to image files 233

<type numpy.ndarray>
>>> lena.shape, lena.dtype
((512, 512), dtype(uint8))
dtype is uint8 for 8-bit images (0-255)

Opening raw files (camera, 3-D images)
>>> l.tofile(lena.raw) # Create raw file
>>> lena_from_raw = np.fromfile(lena.raw, dtype=np.int64)
>>> lena_from_raw.shape
(262144,)
>>> lena_from_raw.shape = (512, 512)
>>> import os
>>> os.remove(lena.raw)
Need to know the shape and dtype of the image (how to separate data bytes).
For large data, use np.memmap for memory mapping:
>>> lena_memmap = np.memmap(lena.raw, dtype=np.int64, shape=(512, 512))
(data are read from the file, and not loaded into memory)
Working on a list of image files
>>> for i in range(10):
... im = np.random.random_integers(0, 255, 10000).reshape((100, 100))
... misc.imsave(random_%02d.png % i, im)
>>> from glob import glob
>>> filelist = glob(random*.png)
>>> filelist.sort()
12.2 Displaying images
Use matplotlib and imshow to display an image inside a matplotlib figure:

>>> l = misc.lena()
>>> plt.imshow(l, cmap=plt.cm.gray)
<matplotlib.image.AxesImage object at 0x3c7f710>
Increase contrast by setting min and max values:

>>> plt.imshow(l, cmap=plt.cm.gray, vmin=30, vmax=200)
<matplotlib.image.AxesImage object at 0x33ef750>
>>> # 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.ContourSet instance at 0x33f8c20>
For fine inspection of intensity variations, use interpolation=nearest:

>>> plt.imshow(l[200:220, 200:220], cmap=plt.cm.gray)
>>> plt.imshow(l[200:220, 200:220], cmap=plt.cm.gray, interpolation=nearest)
Other packages sometimes use graphical toolkits for visualization (GTK, Qt):
12.2. Displaying images 234

12.2. Displaying images 235

>>> import skimage.io

>>> skimage.io.use_plugin(gtk, imshow)
>>> im_io.imshow(l)
3-D visualization: Mayavi

See 3D plotting with Mayavi (page 287) and mayavi-voldata-label.
Image plane widgets
Isosurfaces
...
12.3 Basic manipulations
Images are arrays: use the whole numpy machinery.
>>> lena = scipy.misc.lena()

>>> lena[0, 40]
166
>>> # 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
12.3.1 Statistical information

>>> lena.mean()
124.04678344726562
>>> lena.max(), lena.min()
(245, 25)
np.histogram
12.3. Basic manipulations 236

Exercise 1
Open as an array the scikit-image logo (http://scikit-image.org/_static/scikits_image_logo.png),
or an image that you have on your computer.
Crop a meaningful part of the image, for example the python circle in the logo.
Display the image array using matlplotlib. Change the interpolation method and zoom to see the
difference.
Transform your image to greyscale
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)
12.3.2 Geometrical transformations

>>> lx, ly = lena.shape
>>> # Cropping
>>> crop_lena = lena[lx / 4: - lx / 4, ly / 4: - ly / 4]
>>> # up <-> down flip
>>> flip_ud_lena = np.flipud(lena)
>>> # rotation
>>> rotate_lena = ndimage.rotate(lena, 45)
>>> rotate_lena_noreshape = ndimage.rotate(lena, 45, reshape=False)
12.3. Basic manipulations 237

12.4 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.
12.4.1 Blurring/smoothing
Gaussian filter from scipy.ndimage:

>>> blurred_lena = ndimage.gaussian_filter(lena, sigma=3)
>>> very_blurred = ndimage.gaussian_filter(lena, sigma=5)
Uniform filter
>>> local_mean = ndimage.uniform_filter(lena, size=11)
12.4.2 Sharpening
Sharpen a blurred image:

>>> blurred_l = ndimage.gaussian_filter(lena, 3)
increase the weight of edges by adding an approximation of the Laplacian:

>>> filter_blurred_l = ndimage.gaussian_filter(blurred_l, 1)
>>> alpha = 30
>>> sharpened = blurred_l + alpha * (blurred_l - filter_blurred_l)
12.4. Image filtering 238

12.4.3 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:
>>> gauss_denoised = ndimage.gaussian_filter(noisy, 2)
Most local linear isotropic filters blur the image (ndimage.uniform_filter)

A median filter preserves better the edges:
>>> 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)
Other rank filter: ndimage.maximum_filter, ndimage.percentile_filter

Other local non-linear filters: Wiener (scipy.signal.wiener), etc.
Non-local filters

Total-variation (TV) denoising. Find a new image so that the total-variation of the image (integral of the norm
L1 of the gradient) is minimized, while being close to the measured image:
>>> from skimage.filter import tv_denoise
>>> tv_denoised = tv_denoise(noisy, weight=10)
>>> # More denoising (to the expense of fidelity to data)
>>> tv_denoised = tv_denoise(noisy, weight=50)
Exercise 2: denoising
Create a binary image (of 0s and 1s) with several objects (circles, ellipses, squares, or random shapes).
Add some noise (e.g., 20% of noise)
Try three different denoising methods for denoising the image: gaussian filtering, median filtering,
and total variation denoising.
Compare the histograms of the three different denoised images. Which one is the closest to the his-
togram of the original (noise-free) image?
12.4.4 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)
array([[0, 1, 0],
[1, 1, 1],
[0, 1, 0]])
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],
[0, 0, 1, 1, 1, 0, 0],
[0, 0, 0, 0, 0, 0, 0]])
>>> ndimage.binary_erosion(a).astype(a.dtype)
array([[0, 0, 0, 0, 0, 0, 0],
[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],
[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0]])
>>> #Erosion removes objects smaller than the structure
>>> ndimage.binary_erosion(a, structure=np.ones((5,5))).astype(a.dtype)
array([[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],
[0, 0, 0, 0, 0, 0, 0]])
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.]])
>>> ndimage.binary_dilation(a).astype(a.dtype)
array([[ 0., 0., 0., 0., 0.],
[ 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:

>>> x, y = (63*np.random.random((2, 8))).astype(np.int)
>>> im[x, y] = np.arange(8)
>>> bigger_points = ndimage.grey_dilation(im, size=(5, 5), structure=np.ones((5, 5)))

>>> square = np.zeros((16, 16))

>>> square[4:-4, 4:-4] = 1
>>> dist = ndimage.distance_transform_bf(square)
>>> dilate_dist = ndimage.grey_dilation(dist, size=(3, 3), \
... structure=np.ones((3, 3)))
Opening: erosion + dilation:

>>> a[1:4, 1:4] = 1; a[4, 4] = 1
>>> a
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],
[0, 1, 1, 1, 0],
[0, 0, 0, 0, 0]])
>>> # Opening can also smooth corners
>>> ndimage.binary_opening(a).astype(np.int)
array([[0, 0, 0, 0, 0],
[0, 0, 1, 0, 0],
[0, 1, 1, 1, 0],
[0, 0, 1, 0, 0],
[0, 0, 0, 0, 0]])
Application: remove noise:

>>> square = np.zeros((32, 32))
>>> square[10:-10, 10:-10] = 1
>>> x, y = (32*np.random.random((2, 20))).astype(np.int)
>>> square[x, y] = 1
>>> open_square = ndimage.binary_opening(square)
>>> eroded_square = ndimage.binary_erosion(square)

>>> reconstruction = ndimage.binary_propagation(eroded_square, mask=square)
Closing: dilation + erosion

Skeletonization: reduce objects to one-pixel thin lines, keeping the same topology
Many other mathematical morphology operations: hit and miss transform, tophat, etc.

12.5 Feature extraction
12.5.1 Edge detection
Synthetic data:
>>> im = np.zeros((256, 256))
>>> im[64:-64, 64:-64] = 1
>>>
>>> im = ndimage.rotate(im, 15, mode=constant)
>>> im = ndimage.gaussian_filter(im, 8)
Use a gradient operator (Sobel) to find high intensity variations:

>>> sx = ndimage.sobel(im, axis=0, mode=constant)
>>> sy = ndimage.sobel(im, axis=1, mode=constant)
>>> sob = np.hypot(sx, sy)
Canny filter
>>> from skimage.filter import canny
>>> im += 0.1*np.random.random(im.shape)
>>> edges = canny(im, 1, 0.4, 0.2) # not enough smoothing
>>> edges = canny(im, 3, 0.3, 0.2) # better parameters
Several parameters need to be adjusted... risk of overfitting
12.5.2 Segmentation
Histogram-based segmentation (no spatial information)

>>> n = 10
>>> l = 256
>>> im = np.zeros((l, l))
>>> points = l*np.random.random((2, n**2))
>>> im[(points[0]).astype(np.int), (points[1]).astype(np.int)] = 1
>>> im = ndimage.gaussian_filter(im, sigma=l/(4.*n))
>>> mask = (im > im.mean()).astype(np.float)

>>> mask += 0.1 * im
>>> img = mask + 0.2*np.random.randn(*mask.shape)
>>> hist, bin_edges = np.histogram(img, bins=60)

>>> bin_centers = 0.5*(bin_edges[:-1] + bin_edges[1:])
>>> binary_img = img > 0.5
12.5. Feature extraction 243

Automatic thresholding: use Gaussian mixture model:

>>> mask = (im > im.mean()).astype(np.float)
>>> mask += 0.1 * im
>>> img = mask + 0.3*np.random.randn(*mask.shape)
>>> from sklearn.mixture import GMM

>>> classif = GMM(n_components=2)
>>> classif.fit(img.reshape((img.size, 1)))
GMM(...)
>>> classif.means
array([[ 0.9353155 ],
[-0.02966039]])
>>> np.sqrt(classif.covarsi).ravel()
array([ 0.35074631, 0.28225327])
>>> classif.weights
array([ 0.40989799, 0.59010201])
>>> threshold = np.mean(classif.means)
>>> binary_img = img > threshold
Use mathematical morphology to clean up the result:

>>> # Remove small white regions
>>> open_img = ndimage.binary_opening(binary_img)
>>> # Remove small black hole
>>> close_img = ndimage.binary_closing(open_img)

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()
0.014678955078125
>>> np.abs(mask - reconstruct_final).mean()
0.0042572021484375
Exercise
Check how a first denoising step (median filter, total variation) modifies the histogram, and check that the
resulting histogram-based segmentation is more accurate.
Graph-based segmentation: use spatial information.

Watershed segmentation
>>> from skimage.morphology import watershed, is_local_maximum
>>>
>>> # 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)
>>> # Now we want to separate the two objects in image
>>> # Generate the markers as local maxima of the distance
>>> # to the background
>>> distance = ndimage.distance_transform_edt(image)
>>> local_maxi = is_local_maximum(distance, image, np.ones((3, 3)))
>>> markers = ndimage.label(local_maxi)[0]
>>> labels = watershed(-distance, markers, mask=image)
Spectral clustering (normalized cuts) segmentation

>>> from sklearn.feature_extraction import image
>>> from sklearn.cluster import spectral_clustering
>>> l = 100
>>> x, y = np.indices((l, l))

>>> 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
>>> img = circle1 + circle2 + circle3 + circle4
>>> mask = img.astype(bool)
>>> img = img.astype(float)
>>> img += 1 + 0.2*np.random.randn(*img.shape)

>>> # 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, k=4, mode=arpack)

>>> label_im = -np.ones(mask.shape)
>>> label_im[mask] = labels
12.6 Measuring objects properties: ndimage.measurements
Synthetic data:
>>> n = 10
>>> l = 256
>>> mask = im > im.mean()
Analysis of connected components

Label connected components: ndimage.label:
12.6. Measuring objects properties: ndimage.measurements 246

>>> label_im, nb_labels = ndimage.label(mask)

>>> nb_labels # how many regions?
23
>>> 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))
>>> block_mean.shape = (sx/4, sy/6)
When regions are regular blocks, it is more efficient to use stride tricks (Example: fake dimensions with strides
(page 170)).

Non-regularly-spaced blocks: radial mean:

>>> 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))
Exercise: segmentation
Load as an array the coins image from skimage (skimage.data.coins) or from
https://github.com/scikits-image/scikits-image/raw/master/skimage/data/coins.png
Display the histogram and try to perform histogram segmentation.
Try two segmentation methods: an edge-based method using skimage.filter.canny
and scipy.ndimage.binary_fill_holes and a region-based method using
skimage.morphology.watershed and skimage.filter.sobel to compute an ele-
vation map.
Compute the sizes of the coins.

Other measures
Correlation function, Fourier/wavelet spectrum, etc.
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
>>>
>>> mask = im > im.mean()
>>>
>>> granulo = granulometry(mask, sizes=np.arange(2, 19, 4))


CHAPTER 13
Mathematical optimization: finding

minima of functions
authors Gal Varoquaux

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
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.
252
Chapters contents
Knowing your problem (page 253)
Convex versus non-convex optimization (page 254)
Smooth and non-smooth problems (page 254)
Noisy versus exact cost functions (page 255)
Constraints (page 255)
A review of the different optimizers (page 255)
Getting started: 1D optimization (page 255)
Gradient based methods (page 256)
* Some intuitions about gradient descent (page 256)
* Conjugate gradient descent (page 257)
Newton and quasi-newton methods (page 258)
* Newton methods: using the Hessian (2nd differential) (page 258)
* Quasi-Newton methods: approximating the Hessian on the fly (page 260)
Gradient-less methods (page 260)
* A shooting method: the Powell algorithm (page 260)
* Simplex method: the Nelder-Mead (page 261)
Global optimizers (page 262)
* Brute force: a grid search (page 262)
* Simulated annealing (page 262)
Practical guide to optimization with scipy (page 262)
Choosing a method (page 262)
Making your optimizer faster (page 263)
Computing gradients (page 263)
Synthetic exercices (page 264)
Special case: non-linear least-squares (page 264)
Minimizing the norm of a vector function (page 264)
Curve fitting (page 265)
Optimization with constraints (page 266)
Box bounds (page 266)
General constraints (page 266)
13.1 Knowing your problem
Not all optimization problems are equal. Knowing your problem enables you to choose the right tool.
Dimensionality of the problem

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.
13.1. Knowing your problem 253

13.1.1 Convex versus non-convex optimization
A convex function: A non-convex function

f is above all its tangents.
equivalently, for two point A, B, f(C) lies below
the segment [f(A), f(B])], if A < C < B
Optimizing convex functions is easy. Optimizing non-convex functions can be very hard.
Note: A convex function provably has only one minimum, no local minimums
13.1.2 Smooth and non-smooth problems
A smooth function: A non-smooth function

The gradient is defined everywhere, and is a continuous function
Optimizing smooth functions is easier.
13.1. Knowing your problem 254

13.1.3 Noisy versus exact cost functions
Noisy (blue) and non-noisy (green) functions
Noisy gradients
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
noisy,
13.1.4 Constraints
Optimizations under constraints

Here:
1 < x1 < 1
1 < x2 < 1
13.2 A review of the different optimizers
13.2.1 Getting started: 1D optimization
Use scipy.optimize.brent() to minimize 1D functions. It combines a bracketing strategy with a parabolic

approximation.
13.2. A review of the different optimizers 255

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
0.6999999997759...
>>> x_min - .7
-2.1605...e-10
Note: Brents method can be used for optimization constraint to an intervale using
scipy.optimize.fminbound()
Note: In scipy 0.11, scipy.optimize.minimize_scalar() gives a generic interface to 1D scalar mini-

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

Gradient descent basically consists consists in taking small steps in the direction of the gradient.
Table 13.1: Fixed step gradient descent
A well-conditionned quadratic function.
An ill-conditionned quadratic function.

The core problem of gradient-methods on ill-conditioned
problems is that the gradient tends not to point in the direction
of the minimum.
We can see that very anisotropic (ill-conditionned) functions are harder to optimize.

Take home message: conditioning number and preconditioning

If you know natural scaling for your variables, prescale them so that they behave similarly. This is related to
preconditioning.
Also, it clearly can be advantageous to take bigger steps. This is done in gradient descent code using a line search.
Table 13.2: Adaptive step gradient descent
A well-conditionned quadratic
function.
An ill-conditionned quadratic
function.
An ill-conditionned
non-quadratic function.
An ill-conditionned very
The more a function looks like a quadratic function (elliptic iso-curves), the easier it is to optimize.
Conjugate gradient descent
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
it tends to oscillate across a valley, each time following the direction of the gradient, that makes it cross the valley.

The conjugate gradient solves this problem by adding a friction term: each step depends on the two last values of
the gradient and sharp turns are reduced.
Table 13.3: Conjugate gradient descent
An ill-conditionned
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():
>>> def f(x): # The rosenbrock 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
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
array([ 0.99999199, 0.99997536])
Note that the function has only been evaluated 30 times, compared to 120 without the gradient.
13.2.3 Newton and quasi-newton methods
Newton methods: using the Hessian (2nd differential)
Newton methods use a local quadratic approximation to compute the jump direction. For this purpose, they rely
on the 2 first derivative of the function: the gradient and the Hessian.

An ill-conditionned quadratic function:

Note that, as the quadratic approximation is exact, the Newton
method is blazing fast
An ill-conditionned non-quadratic function:

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 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:
... return .5*(1 - x[0])**2 + (x[1] - x[0]**2)**2
>>> def fprime(x):
>>> optimize.fmin_ncg(f, [2, 2], fprime=fprime)
Iterations: 10
Hessian evaluations: 0
array([ 1., 1.])
Note that compared to a conjugate gradient (above), Newtons method has required less function evaluations, but
more gradient evaluations, as it uses it to approximate the Hessian. Lets compute the Hessian and pass it to the
algorithm:
>>> def hessian(x): # Computed with sympy
... return np.array(((1 - 4*x[1] + 12*x[0]**2, -4*x[0]), (-4*x[0], 2)))
>>> optimize.fmin_ncg(f, [2, 2], fprime=fprime, fhess=hessian)
Iterations: 10
Hessian evaluations: 10
array([ 1., 1.])
Note: At very high-dimension, the inversion of the Hessian can be costly and unstable (large scale > 250).
Note: Newton optimizers should not to be confused with Newtons root finding method, based on the same
principles, scipy.optimize.newton().

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.

On a exactly quadratic function, BFGS is not as fast as
Newtons method, but still very fast.
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.

... return .5*(1 - x[0])**2 + (x[1] - x[0]**2)**2
>>> def fprime(x):
>>> optimize.fmin_bfgs(f, [2, 2], fprime=fprime)
Iterations: 16
array([ 1.00000017, 1.00000026])
L-BFGS: Limited-memory BFGS Sits between BFGS and conjugate gradient: in very high dimensions (> 250)
the Hessian matrix is too costly to compute and invert. L-BFGS keeps a low-rank version. In addition, the scipy
version, scipy.optimize.fmin_l_bfgs_b(), includes box bounds:
... return .5*(1 - x[0])**2 + (x[1] - x[0]**2)**2
>>> def fprime(x):
>>> optimize.fmin_l_bfgs_b(f, [2, 2], fprime=fprime)
(array([ 1.00000005, 1.00000009]), 1.4417677473011859e-15, {warnflag: 0, task: CONVERGENCE:
Note: If you do not specify the gradient to the L-BFGS solver, you need to add approx_grad=1
13.2.4 Gradient-less methods
A shooting method: the Powell algorithm
Almost a gradient approach


Powells method isnt too sensitive to local
ill-conditionning in low dimensions
Simplex method: the Nelder-Mead
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
bracket the minimum.
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:
non-quadratic function:
In scipy, scipy.optimize.fmin() implements the Nelder-Mead approach:
... return .5*(1 - x[0])**2 + (x[1] - x[0]**2)**2
>>> optimize.fmin(f, [2, 2])
Iterations: 46
array([ 0.99998568, 0.99996682])

13.2.5 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.
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:
... return .5*(1 - x[0])**2 + (x[1] - x[0]**2)**2
>>> optimize.brute(f, ((-1, 2), (-1, 2)))
array([ 1.00001462, 1.00001547])
Simulated annealing
Simulated annealing does random jumps around the starting point to explore its vicinity, progressively narrowing
the jumps around the minimum points it finds. Its output depends on the random number generator. In scipy, it is
implemented in scipy.optimize.anneal():
... return .5*(1 - x[0])**2 + (x[1] - x[0]**2)**2
>>> optimize.anneal(f, [2, 2])
Warning: Cooled to 5057.768838 at [ 30.27877642 984.84212523] but this
is not the smallest point found.
(array([ -7.70412755, 56.10583526]), 5)
It is a very popular algorithm, but it is not very reliable.
Note: For function of continuous parameters as studied here, a strategy based on grid search for rough exploration
and running optimizers like the Nelder-Mead or gradient-based methods many times with different starting points
should often be preferred to heuristic methods such as simulated annealing.
13.3 Practical guide to optimization with scipy
13.3.1 Choosing a method
13.3. Practical guide to optimization with scipy 262

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-
merically gradients
On well-conditioned problems, Powell (scipy.optimize.fmin_powell()) and
Nelder-Mead (scipy.optimize.fmin()), both gradient-free methods, work well in
high dimension, but they collapse for ill-conditioned problems.
With knowledge of the gradient
BFGS (scipy.optimize.fmin_bfgs()) or L-BFGS
(scipy.optimize.fmin_l_bfgs_b()).
Computational overhead of BFGS is larger than that L-BFGS, itself larger than that of
conjugate gradient. On the other side, BFGS usually needs less function evaluations than
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
(scipy.optimize.fmin_powell()).
13.3.2 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
13.3.3 Computing gradients
Computing gradients, and even more Hessians, is very tedious but worth the effort. Symbolic computation with
Sympy (page 294) may come in handy.
Warning: A very common source of optimization not converging well is human error in the computation of
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:
>>> optimize.check_grad(f, fprime, [2, 2])

2.384185791015625e-07
See also scipy.optimize.approx_fprime() to find your errors.
13.3. Practical guide to optimization with scipy 263

13.3.4 Synthetic exercices
Exercice: A simple (?) quadratic function

Optimize the following function, using K[0] as a starting point:
np.random.seed(0)
K = np.random.normal(size=(100, 100))
def f(x):
return np.sum((np.dot(K, x - 1))**2) + np.sum(x**2)**2
Time your approach. Find the fastest approach. Why is BFGS not working well?
Exercice: A locally flat minimum

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.
13.4 Special case: non-linear least-squares
13.4.1 Minimizing the norm of a vector function
Least square problems, minimizing the norm of a vector function, have a specific structure that can be used in the
LevenbergMarquardt algorithm implemented in scipy.optimize.leastsq().
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)))
>>> x0 = np.zeros(10)
13.4. Special case: non-linear least-squares 264

>>> 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
array([ -7.38998277e-09, 1.11112265e-01, 2.22219893e-01,
3.33331914e-01, 4.44449794e-01, 5.55560493e-01,
6.66672149e-01, 7.77779758e-01, 8.88882036e-01,
1.00001026e+00])
BFGS needs more function calls, and gives a less precise result.
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().
13.4.2 Curve fitting
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():
>>> def f(t, omega, phi):
... return np.cos(omega * t + phi)
>>> x = np.linspace(0, 3, 50)

>>> y = f(x, 1.5, 1) + .1*np.random.normal(size=50)
>>> optimize.curve_fit(f, x, y)
(array([ 1.51854577, 0.92665541]),
array([[ 0.00037994, -0.00056796],
[-0.00056796, 0.00123978]]))
13.4. Special case: non-linear least-squares 265

Exercise
Do the same with omega = 3. What is the difficulty?
13.5 Optimization with constraints
13.5.1 Box bounds
Box bounds correspond to limiting each of the individual parameters of the optimization. Note that some
problems that are not originally written as box bounds can be rewritten as such be a change of variables.
scipy.optimize.fminbound() for 1D-optimization

scipy.optimize.fmin_l_bfgs_b() a quasi-Newton (page 260) 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, 1.5)))
(array([ 1.5, 1.5]), 1.5811388300841898, {warnflag: 0, task: CONVERGENCE: NORM_OF_PROJE
13.5.2 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-
straints:
>>> def f(x):
... return np.sqrt((x[0] - 3)**2 + (x[1] - 2)**2)
>>> def constraint(x):

... return np.atleast_1d(1.5 - np.sum(np.abs(x)))
13.5. Optimization with constraints 266

>>> optimize.fmin_slsqp(f, np.array([0, 0]), ieqcons=[constraint, ])

Optimization terminated successfully. (Exit mode 0)
Iterations: 5
array([ 1.25004696, 0.24995304])
scipy.optimize.fmin_cobyla() Constraints optimization by linear approximation: inequality

constraints only:
>>> optimize.fmin_cobyla(f, np.array([0, 0]), cons=constraint)
Normal return from subroutine COBYLA
NFVALS = 36 F = 2.474874E+00 MAXCV = 0.000000E+00

X = 1.250096E+00 2.499038E-01
array([ 1.25009622, 0.24990378])
Warning: The above problem is known as the Lasso problem in statistics, and there exists very efficient
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.
13.5. Optimization with constraints 267

CHAPTER 14
Traits
author Didrik Pinte

The Traits project allows you to simply add validation, initialization, delegation, notification and a graphical user
interface to Python object attributes. In this tutorial we will explore the Traits toolset and learn how to dramatically
reduce the amount of boilerplate code you write, do rapid GUI application development, and understand the ideas
which underly other parts of the Enthought Tool Suite.
Traits and the Enthought Tool Suite are open source projects licensed under a BSD-style license.
Intended Audience
Intermediate to advanced Python programmers
Requirements
Python 2.6 or 2.7 (www.python.org)
Either wxPython (http://www.wxpython.org/) or PyQt (http://www.riverbankcomputing.co.uk/software/pyqt/intro)
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
(http://www.enthought.com/products/epd.php)
Tutorial content
Example (page 269)
What are Traits (page 270)
Initialisation (page 271)
Validation (page 271)
Documentation (page 272)
Visualisation (page 273)
Deferral (page 274)
Notification (page 279)
Some more advanced traits (page 282)
References (page 285)
268
14.1 Introduction
The Enthought Tool Suite enable the construction of sophisticated application frameworks for data analysis, 2D
plotting and 3D visualization. These powerful, reusable components are released under liberal BSD-style licenses.
The main packages are:
Traits - component based approach to build our applications.
Kiva - 2D primitives supporting path based rendering, affine transforms, alpha blending and more.
Enable - object based 2D drawing canvas.
Chaco - plotting toolkit for building complex interactive 2D plots.
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.
14.2 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
The reservoir has a known behaviour. One part is related to the energy production based on the water released. A
simple formula for approximating electric power production at a hydroelectric plant is P = hrgk, where:
P is Power in watts,
is the density of water (~1000 kg/m3),
h is height in meters,
r is flow rate in cubic meters per second,
g is acceleration due to gravity of 9.8 m/s2,
k is a coefficient of efficiency ranging from 0 to 1.
Annual electric energy production depends on the available water supply. In some installations the water flow rate
can vary by a factor of 10:1 over the course of a year.
The second part of the behaviour is the state of the storage that depends on controlled and uncontrolled parameters
:
storaget+1 = storaget + inf lows release spillage irrigation
Warning: The data used in this tutorial are not real and might even not have sense in the reality.
14.3 What are Traits
A trait is a type definition that can be used for normal Python object attributes, giving the attributes some additional
characteristics:
Standardization:
Initialization
Validation
Deferral
Notification
Visualization
Documentation
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.
The common way of creating a traits class is by extending from the HasTraits base class and defining class traits
:
from traits.api import HasTraits, Str, Float
class Reservoir(HasTraits):
name = Str
max_storage = Float
14.3. What are Traits 270

Warning: For Traits 3.x users

If using Traits 3.x, you need to adapt the namespace of the traits packages:
traits.api should be enthought.traits.api
traitsui.api should be enthought.traits.ui.api
Using a traits class like that is as simple as any other Python class. Note that the trait value are passed using
keyword arguments:
reservoir = Reservoir(name=Lac de Vouglans, max_storage=605)
14.3.1 Initialisation
All the traits do have a default value that initialise the variables. For example, the basic python types do have the
following trait equivalents:
Trait Python Type Built-in Default Value
Bool Boolean False
Complex Complex number 0+0j
Float Floating point number 0.0
Int Plain integer 0
Long Long integer 0L
Str String
Unicode Unicode u
A number of other predefined trait type do exist : Array, Enum, Range, Event, Dict, List, Color, Set, Expression,
Code, Callable, Type, Tuple, etc.
Custom default values can be defined in the code:
name = Str
max_storage = Float(100)
reservoir = Reservoir(name=Lac de Vouglans)
Note: Complex initialisation

When a complex initialisation is required for a trait, a _XXX_default magic method can be implemented. It will
be lazily called when trying to access the XXX trait. For example:
def _name_default(self):
""" Complex initialisation of the reservoir name. """
return Undefined
14.3.2 Validation
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)
/Users/dpinte/projects/scipy-lecture-notes/advanced/traits/<ipython-input-7-979bdff9974a> in <modu

----> 1 reservoir.max_storage = 230
/Users/dpinte/projects/ets/traits/traits/trait_handlers.pyc in error(self, object, name, value)

166 """
167 raise TraitError( object, name, self.full_info( object, name, value ),
--> 168 value )
169
170 def arg_error ( self, method, arg_num, object, name, value ):
TraitError: The max_storage trait of a Reservoir instance must be a float, but a value of 23 <
14.3.3 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
max_storage = Float(100, desc=Maximal storage [hm3])
Lets now define the complete reservoir class:

from traits.api import HasTraits, Str, Float, Range
name = Str
max_storage = Float(1e6, desc=Maximal storage [hm3])
max_release = Float(10, desc=Maximal release [m3/s])
head = Float(10, desc=Hydraulic head [m])
efficiency = Range(0, 1.)
def energy_production(self, release):

Returns the energy production [Wh] for the given release [m3/s]

power = 1000 * 9.81 * self.head * release * self.efficiency
return power * 3600
reservoir = Reservoir(
name = Project A,
max_storage = 30,
max_release = 100.0,
head = 60,
efficiency = 0.8
)
release = 80
print Releasing {} m3/s produces {} kWh.format(

release, reservoir.energy_production(release)
)
14.3.4 Visualisation
The Traits library is also aware of user interfaces and can pop up a default view for the Reservoir class:
reservoir1 = Reservoir()
reservoir1.edit_traits()
TraitsUI simplifies the way user interfaces are created. Every trait on a HasTraits class has a default editor that
will manage the way the trait is rendered to the screen (e.g. the Range trait is displayed as a slider, etc.).
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 traitsui.api import View
name = Str
traits_view = View(
name, max_storage, max_release, head, efficiency,
title = Reservoir,
resizable = True,
)


return power * 3600
name = Project A,
max_storage = 30,

head = 60,
efficiency = 0.8
)
reservoir.configure_traits()
14.3.5 Deferral
Being able to defer the definition of a trait and its value to another object is a powerful feature of Traits.
from traits.api import HasTraits, Instance, DelegatesTo, Float, Range
from reservoir import Reservoir
class ReservoirState(HasTraits):
"""Keeps track of the reservoir state given the initial storage.
"""
reservoir = Instance(Reservoir, ())
min_storage = Float
max_storage = DelegatesTo(reservoir)
min_release = Float
max_release = DelegatesTo(reservoir)
# state attributes
storage = Range(low=min_storage, high=max_storage)
# control attributes
inflows = Float(desc=Inflows [hm3])
release = Range(low=min_release, high=max_release)
spillage = Float(desc=Spillage [hm3])
def print_state(self):
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,
self.spillage)
print - * 79
projectA = Reservoir(
name = Project A,
max_storage = 30,
hydraulic_head = 60,
efficiency = 0.8
)

state = ReservoirState(reservoir=projectA, storage=10)

state.release = 90
state.inflows = 0
state.print_state()
print How do we update the current storage ?
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
For the simplicity of the example, the release is considered in

hm3/timestep and not in m3/s.
"""
min_storage = Float
min_release = Float
# state attributes
update_storage = Event(desc=Updates the storage to the next time step)
def _update_storage_fired(self):
# update storage state
new_storage = self.storage - self.release + self.inflows
self.storage = min(new_storage, self.max_storage)
overflow = new_storage - self.max_storage
self.spillage = max(overflow, 0)
self.spillage)
print - * 79
name = Project A,
max_storage = 30,
max_release = 5.0,
efficiency = 0.8
)

state.release = 5
state.inflows = 0

# release the maximum amount of water during 3 time steps

state.update_storage = True
state.print_state()
state.print_state()
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 :
_get_XXX for the getter of the XXX Property trait
_set_XXX for the setter of the XXX Property trait
from traits.api import Property

"""
min_release = Float
# state attributes
storage = Property(depends_on=inflows, release)
spillage = Property(
desc=Spillage [hm3], depends_on=[storage, inflows, release]
)
### Private traits. ######################################################

_storage = Float
### Traits property implementation. ######################################

def _get_storage(self):
new_storage = self._storage - self.release + self.inflows
return min(new_storage, self.max_storage)
def _set_storage(self, storage_value):

self._storage = storage_value
def _get_spillage(self):
return max(overflow, 0)
self.spillage)
print - * 79

name = Project A,
max_storage = 30,
max_release = 5,
efficiency = 0.8
)

state.release = 4
state.inflows = 0
state.print_state()
Note: Caching property

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
invalidated.
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

"""
name = DelegatesTo(reservoir)
min_release = Float
# state attributes
)
### Traits view ##########################################################

traits_view = View(
Group(
VGroup(Item(name), Item(storage), Item(spillage),
label = State, style = readonly
),
VGroup(Item(inflows), Item(release), label=Control),
)
)
### Private traits. ######################################################

_storage = Float


self.spillage)
print - * 79
name = Project A,
max_storage = 30,
max_release = 5,
efficiency = 0.8
)

state.release = 4
state.inflows = 0
state.print_state()
state.configure_traits()
Some use cases need the delegation mechanism to be broken by the user when setting the value of the trait. The
PrototypeFrom trait implements this behaviour.
from traits.api import HasTraits, Str, Float, Range, PrototypedFrom, Instance
class Turbine(HasTraits):

turbine_type = Str
power = Float(1.0, desc=Maximal power delivered by the turbine [Mw])
name = Str
turbine = Instance(Turbine)
installed_capacity = PrototypedFrom(turbine, power)
turbine = Turbine(turbine_type=type1, power=5.0)
name = Project A,
max_storage = 30,
head = 60,
efficiency = 0.8,
turbine = turbine,
)
print installed capacity is initialised with turbine.power

print reservoir.installed_capacity
print - * 15
print updating the turbine power updates the installed capacity
turbine.power = 10
print reservoir.installed_capacity
print - * 15
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
14.3.6 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:
"""
min_storage = Float
min_release = Float

# state attributes
self.spillage)
print - * 79
### Traits listeners #####################################################

def _release_changed(self, new):
"""When the release is higher than zero, warn all the inhabitants of
the valley.
"""
if new > 0:
print Warning, we are releasing {} hm3 of water.format(new)
name = Project A,
max_storage = 30,
efficiency = 0.8
)

state.release = 90
state.inflows = 0
state.print_state()
The static trait notification signatures can be:

def _release_changed(self): pass
def _release_changed(self, new): pass
def _release_changed(self, old, new): pass
def _release_changed(self, name, old, new pass
Note: 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
to register listeners on the fly with the dynamic listeners
from reservoir_state_property import ReservoirState
def wake_up_watchman_if_spillage(new_value):
if new_value > 0:
print Wake up watchman! Spilling {} hm3.format(new_value)

name = Project A,
max_storage = 30,
efficiency = 0.8
)
#register the dynamic listener

state.on_trait_change(wake_up_watchman_if_spillage, name=spillage)
state.release = 90
state.inflows = 0
state.print_state()
print Forcing spillage

state.inflows = 100
state.release = 0
print Why do we have two executions of the callback ?
The dynamic trait notification signatures are not the same as the static ones :
def wake_up_watchman(): pass
def wake_up_watchman(new): pass
def wake_up_watchman(name, new): pass
def wake_up_watchman(object, name, new): pass
def wake_up_watchman(object, name, old, new): pass
Removing a dynamic listener can be done by:
calling the remove_trait_listener method on the trait with the listener method as argument,
calling the on_trait_change method with listener method and the keyword remove=True,
deleting the instance that holds the listener.
Listeners can also be added to classes using the on_trait_change decorator:
from traits.api import Property, on_trait_change

"""
min_release = Float
# state attributes

)
### Private traits. ######################################################

_storage = Float


@on_trait_change(storage)
self.spillage)
print - * 79
name = Project A,
max_storage = 30,
max_release = 5,
efficiency = 0.8
)

state.release = 4
state.inflows = 0
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.
14.3.7 Some more advanced traits
The following example demonstrate the usage of the Enum and List traits :
from traits.api import HasTraits, Str, Float, Range, Enum, List
from traitsui.api import View, Item
class IrrigationArea(HasTraits):
name = Str
surface = Float(desc=Surface [ha])
crop = Enum(Alfalfa, Wheat, Cotton)
name = Str


irrigated_areas = List(IrrigationArea)


return power * 3600
traits_view = View(
Item(name),
Item(max_storage),
Item(max_release),
Item(head),
Item(efficiency),
Item(irrigated_areas),
resizable = True
)
upper_block = IrrigationArea(name=Section C, surface=2000, crop=Wheat)
name=Project A,
max_storage=30,
max_release=100.0,
head=60,
efficiency=0.8,
irrigated_areas=[upper_block]
)
release = 80
)
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
on linked to a given reservoir.
from traits.api import HasTraits, Str, Float, Range, Enum, List, Property
from traitsui.api import View, Item
class IrrigationArea(HasTraits):
name = Str
surface = Float(desc=Surface [ha])
crop = Enum(Alfalfa, Wheat, Cotton)
name = Str
irrigated_areas = List(IrrigationArea)
total_crop_surface = Property(depends_on=irrigated_areas.surface)
def _get_total_crop_surface(self):
return sum([iarea.surface for iarea in self.irrigated_areas])



return power * 3600
traits_view = View(
Item(name),
Item(max_storage),
Item(max_release),
Item(head),
Item(efficiency),
Item(irrigated_areas),
Item(total_crop_surface),
resizable = True
)
upper_block = IrrigationArea(name=Section C, surface=2000, crop=Wheat)
name=Project A,
max_storage=30,
max_release=100.0,
head=60,
efficiency=0.8,
irrigated_areas=[upper_block],
)
release = 80
)
The next example shows how the Array trait can be used to feed a specialised TraitsUI Item, the ChacoPlotItem:
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)
### Traits view ##########################################################

traits_view = View(
Item(name),
Group(

ChacoPlotItem(month, stock, show_label=False),

),
width = 500,
resizable = True
)
### Traits properties ####################################################

def _get_stock(self):
"""
fixme: should handle cases where we go over the max storage
"""
return self.initial_stock + (self.inflows - self.releases).cumsum()
def _get_month(self):
return np.arange(self.stock.size)
name = Project A,
max_storage = 30,
head = 60,
efficiency = 0.8
)
initial_stock = 10.
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])
view = ReservoirEvolution(
reservoir = reservoir,
inflows = inflows_ts,
releases = releases_ts
)
view.configure_traits()
14.4 References
ETS repositories: http://github.com/enthought
14.4. References 285

Traits manual: http://github.enthought.com/traits/traits_user_manual/index.html

Traits UI manual: http://github.enthought.com/traitsui/traitsui_user_manual/index.html
Mailing list : enthought-dev@enthought.com
14.4. References 286

CHAPTER 15
3D plotting with Mayavi
Chapters contents
Mlab: the scripting interface (page 287)
3D plotting functions (page 288)
* Points (page 288)
* Lines (page 288)
* Elevation surface (page 288)
* Arbitrary regular mesh (page 289)
* Volumetric data (page 289)
Figures and decorations (page 290)
* Figure management (page 290)
* Changing plot properties (page 290)
* Decorations (page 292)
Interactive work (page 293)
The pipeline dialog (page 293)
The script recording button (page 293)
15.1 Mlab: the scripting interface
The mayavi.mlab module provides simple plotting functions to apply to numpy arrays. Try them using in
IPython, by starting IPython with the switch --gui=wx.
287
15.1.1 3D plotting functions
Points
x, y, z, value = np.random.random((4, 40))

mlab.points3d(x, y, z, value)
Lines
mlab.clf() # Clear the figure

mlab.plot3d(np.sin(t), np.cos(t), 0.1*t, t)
Elevation surface
mlab.clf()
x, y = np.mgrid[-10:10:100j, -10:10:100j]
r = np.sqrt(x**2 + y**2)
z = np.sin(r)/r
mlab.surf(z, warp_scale=auto)
15.1. Mlab: the scripting interface 288

Arbitrary regular mesh
mlab.clf()
phi, theta = np.mgrid[0:np.pi:11j, 0:2*np.pi:11j]
x = np.sin(phi) * np.cos(theta)
y = np.sin(phi) * np.sin(theta)
z = np.cos(phi)
mlab.mesh(x, y, z)
mlab.mesh(x, y, z, representation=wireframe,
color=(0, 0, 0))
Note: A surface is defined by points connected to form triangles or polygones. In mlab.surf and mlab.mesh, the
connectivity is implicity given by the layout of the arrays. See also mlab.triangular_mesh.
Our data is often more than points and values: it needs some connectivity information
Volumetric data
mlab.clf()
x, y, z = np.mgrid[-5:5:64j, -5:5:64j, -5:5:64j]
values = x*x*0.5 + y*y + z*z*2.0
mlab.contour3d(values)

This function works with a regular orthogonal grid: the value array is a 3D array that gives the shape of the
grid.
15.1.2 Figures and decorations
Figure management
Here is a list of functions useful to control the current figure

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, 0.5)
Save figure to image file: mlab.savefig(foo.png, size=(300, 300))
Change the view: mlab.view(azimuth=45, elevation=54, distance=1.)
Changing plot properties
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
mask boolean mask array to suppress some data points.
mask_points If supplied, only one out of mask_points data point is displayed.
This option is usefull to reduce the number of points displayed on large datasets
Must be an integer or None.
mode the mode of the glyphs. Must be 2darrow or 2dcircle or 2dcross or
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
cylinder or point or sphere. Default: sphere
name the name of the vtk object created.
representation the representation type used for the surface. Must be surface or
wireframe or points or mesh or fancymesh. Default: surface
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
scalars optional scalar data.
scale_factor scale factor of the glyphs used to represent the vertices, in fancy_mesh
mode. Must be a float. Default: 0.05
scale_mode the scaling mode for the glyphs (vector, scalar, or none).
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.
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
vmin vmin is used to scale the colormap If None, the min of the data will be used
Example:
In [2]: r, theta = np.mgrid[0:10, -np.pi:np.pi:10j]
In [3]: x = r*np.cos(theta)
In [4]: y = r*np.sin(theta)
In [5]: z = np.sin(r)/r
In [6]: from enthought.mayavi import mlab
In [7]: mlab.mesh(x, y, z, colormap=gist_earth, extent=[0, 1, 0, 1, 0, 1])

Out[7]: <enthought.mayavi.modules.surface.Surface object at 0xde6f08c>

In [8]: mlab.mesh(x, y, z, extent=[0, 1, 0, 1, 0, 1],

...: representation=wireframe, line_width=1, color=(0.5, 0.5, 0.5))
Out[8]: <enthought.mayavi.modules.surface.Surface object at 0xdd6a71c>
Decorations
Different items can be added to the figure to carry extra information, such as a colorbar or a title.
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>

Warning: extent: If we specified extents for a plotting object, mlab.outline and mlab.axes dont get them
by default.
mlab_scripting_interface.rst interaction.rst
15.2 Interactive work
The quickest way to create beautiful visualization with Mayavi is probably to interactivly tweak the various set-
tings.
15.2.1 The pipeline dialog
Click on the Mayavi button in the scene, and you can control properties of objects with dialogs.
Set the background of the figure in the Mayavi Scene node

Set the colormap in the Colors and legends node
Right click on the node to add modules or filters
15.2.2 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.
15.2. Interactive work 293

CHAPTER 16
Sympy : Symbolic Mathematics in

Python
author Fabian Pedregosa
Objectives
1. Evaluate expressions with arbitrary precision.
2. Perform algebraic manipulations on symbolic expressions.
3. Perform basic calculus tasks (limits, differentiation and integration) with symbolic expressions.
4. Solve polynomial and transcendental equations.
5. Solve some differential equations.
What is SymPy? SymPy is a Python library for symbolic mathematics. It aims become a full featured computer
algebra system that can compete directly with commercial alternatives (Mathematica, Maple) while keeping the
code as simple as possible in order to be comprehensible and easily extensible. SymPy is written entirely in Python
and does not require any external libraries.
Sympy documentation and packages for installation can be found on http://sympy.org/
294
Chapters contents
First Steps with SymPy (page 295)
Using SymPy as a calculator (page 295)
Exercises (page 296)
Symbols (page 296)
Algebraic manipulations (page 296)
Expand (page 296)
Simplify (page 296)
Calculus (page 297)
Limits (page 297)
Differentiation (page 297)
Series expansion (page 298)
Integration (page 298)
Equation solving (page 298)
Linear Algebra (page 299)
Matrices (page 299)
Differential Equations (page 300)
16.1 First Steps with SymPy
16.1.1 Using SymPy as a calculator
SymPy defines three numerical types: Real, Rational and Integer.

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:
>>> from sympy import *
>>> a = Rational(1,2)
>>> a
1/2
>>> a*2
1
SymPy uses mpmath in the background, which makes it possible to perform computations using arbitrary-
precision arithmetic. That way, some special constants, like e, pi, oo (Infinity), are treated as symbols and can be
evaluated with arbitrary precision:
>>> pi**2
pi**2
>>> pi.evalf()
3.14159265358979
>>> (pi+exp(1)).evalf()
5.85987448204884
as you see, evalf evaluates the expression to a floating-point number.

There is also a class representing mathematical infinity, called oo:
16.1. First Steps with SymPy 295

>>> oo > 99999

True
>>> oo + 1
oo
16.1.2 Exercises

1. Calculate 2 with 100 decimals.
2. Calculate 1/2 + 1/3 in rational arithmetic.
16.1.3 Symbols
In contrast to other Computer Algebra Systems, in SymPy you have to declare symbolic variables explicitly:
>>> from sympy import *
>>> x = Symbol(x)
>>> y = Symbol(y)
Then you can manipulate them:

>>> x+y+x-y
2*x
>>> (x+y)**2
(x + y)**2
Symbols can now be manipulated using some of python operators: +, -, *, ** (arithmetic), &, |, ~ , >>, <<
(boolean).
16.2 Algebraic manipulations
SymPy is capable of performing powerful algebraic manipulations. Well take a look into some of the most
frequently used: expand and simplify.
16.2.1 Expand
Use this to expand an algebraic expression. It will try to denest powers and multiplications:
In [23]: expand((x+y)**3)
Out[23]: 3*x*y**2 + 3*y*x**2 + x**3 + y**3
Further options can be given in form on keywords:

In [28]: expand(x+y, complex=True)
Out[28]: I*im(x) + I*im(y) + re(x) + re(y)
In [30]: expand(cos(x+y), trig=True)

Out[30]: cos(x)*cos(y) - sin(x)*sin(y)
16.2.2 Simplify
Use simplify if you would like to transform an expression into a simpler form:
In [19]: simplify((x+x*y)/x)
Out[19]: 1 + y
16.2. Algebraic manipulations 296

Simplification is a somewhat vague term, and more precises alternatives to simplify exists: powsimp (simplifica-
tion of exponents), trigsimp (for trigonometric expressions) , logcombine, radsimp, together.
16.2.3 Exercises
1. Calculate the expanded form of (x + y)6 .

2. Simplify the trigonometric expression sin(x) / cos(x)
16.3 Calculus
16.3.1 Limits
Limits are easy to use in SymPy, they follow the syntax limit(function, variable, point), so to compute the limit of
f(x) as x -> 0, you would issue limit(f, x, 0):
>>> limit(sin(x)/x, x, 0)
1
you can also calculate the limit at infinity:

>>> limit(x, x, oo)
oo
>>> limit(1/x, x, oo)

0
>>> limit(x**x, x, 0)
1
16.3.2 Differentiation
You can differentiate any SymPy expression using diff(func, var). Examples:
>>> diff(sin(x), x)
cos(x)
>>> diff(sin(2*x), x)
2*cos(2*x)
>>> diff(tan(x), x)
1 + tan(x)**2
You can check, that it is correct by:

>>> limit((tan(x+y)-tan(x))/y, y, 0)
1 + tan(x)**2
Higher derivatives can be calculated using the diff(func, var, n) method:

>>> diff(sin(2*x), x, 1)
2*cos(2*x)
>>> diff(sin(2*x), x, 2)
-4*sin(2*x)
>>> diff(sin(2*x), x, 3)
-8*cos(2*x)
16.3. Calculus 297

16.3.3 Series expansion
SymPy also knows how to compute the Taylor series of an expression at a point. Use series(expr, var):
>>> series(cos(x), x)
1 - x**2/2 + x**4/24 + O(x**6)
>>> series(1/cos(x), x)
1 + x**2/2 + 5*x**4/24 + O(x**6)
16.3.4 Exercises
1. Calculate lim x > 0, sin(x)/x

2. Calulate the derivative of log(x) for x.
16.3.5 Integration
SymPy has support for indefinite and definite integration of transcendental elementary and special functions via
integrate() facility, which uses powerful extended Risch-Norman algorithm and some heuristics and pattern
matching. You can integrate elementary functions:
>>> integrate(6*x**5, x)
x**6
>>> integrate(sin(x), x)
-cos(x)
>>> integrate(log(x), x)
-x + x*log(x)
>>> integrate(2*x + sinh(x), x)
cosh(x) + x**2
Also special functions are handled easily:

>>> integrate(exp(-x**2)*erf(x), x)
pi**(1/2)*erf(x)**2/4
It is possible to compute definite integral:

>>> integrate(x**3, (x, -1, 1))
0
>>> integrate(sin(x), (x, 0, pi/2))
1
>>> integrate(cos(x), (x, -pi/2, pi/2))
2
Also improper integrals are supported as well:

>>> integrate(exp(-x), (x, 0, oo))
1
>>> integrate(exp(-x**2), (x, -oo, oo))
pi**(1/2)
16.3.6 Exercises
16.4 Equation solving
SymPy is able to solve algebraic equations, in one and several variables:

In [7]: solve(x**4 - 1, x)
Out[7]: [I, 1, -1, -I]
16.4. Equation solving 298

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:
In [8]: solve([x + 5*y - 2, -3*x + 6*y - 15], [x, y])
Out[8]: {y: 1, x: -3}
It also has (limited) support for trascendental equations:

In [9]: solve(exp(x) + 1, x)
Out[9]: [pi*I]
Another alternative in the case of polynomial equations is factor. factor returns the polynomial factorized
into irreducible terms, and is capable of computing the factorization over various domains:
In [10]: f = x**4 - 3*x**2 + 1
In [11]: factor(f)
Out[11]: (1 + x - x**2)*(1 - x - x**2)
In [12]: factor(f, modulus=5)

Out[12]: (2 + x)**2*(2 - x)**2
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:
In [13]: satisfiable(x & y)
Out[13]: {x: True, y: True}
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
its arguments can make the expression True, it will return False:
In [14]: satisfiable(x & ~x)
Out[14]: False
16.4.1 Exercises
1. Solve the system of equations x + y = 2, 2 x + y = 0

2. Are there boolean values x, y that make (~x | y) & (~y | x) true?
16.5 Linear Algebra
16.5.1 Matrices
Matrices are created as instances from the Matrix class:

>>> from sympy import Matrix
>>> Matrix([[1,0], [0,1]])
[1, 0]
[0, 1]
unlike a NumPy array, you can also put Symbols in it:

>>> x = Symbol(x)
>>> y = Symbol(y)
>>> A = Matrix([[1,x], [y,1]])
>>> A
[1, x]
[y, 1]
>>> A**2
16.5. Linear Algebra 299

[1 + x*y, 2*x]
[ 2*y, 1 + x*y]
16.5.2 Differential Equations
SymPy is capable of solving (some) Ordinary Differential Equations. sympy.ode.dsolve works like this:
In [4]: f(x).diff(x, x) + f(x)
Out[4]:
2
d
-----(f(x)) + f(x)
dx dx
In [5]: dsolve(f(x).diff(x, x) + f(x), f(x))

Out[5]: C1*sin(x) + C2*cos(x)
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
resolve it as a separable equation.
In [6]: dsolve(sin(x)*cos(f(x)) + cos(x)*sin(f(x))*f(x).diff(x), f(x), hint=separable) Out[6]: -log(1 -
sin(f(x))**2)/2 == C1 + log(1 - sin(x)**2)/2
16.5.3 Exercises
1. Solve the Bernoulli differential equation x*f(x).diff(x) + f(x) - f(x)**2
Warning: TODO: correct this equation and convert to math directive!
2. Solve the same equation using hint=Bernoulli. What do you observe ?
16.5. Linear Algebra 300

CHAPTER 17
scikit-learn: machine learning in

Python
author Fabian Pedregosa, Gael Varoquaux
Prerequisites
Numpy, Scipy
IPython
matplotlib
scikit-learn (http://scikit-learn.org)
Chapters contents
Loading an example dataset (page 302)
Learning and Predicting (page 303)
Classification (page 303)
k-Nearest neighbors classifier (page 303)
Support vector machines (SVMs) for classification (page 304)
Clustering: grouping observations together (page 306)
K-means clustering (page 306)
Dimension Reduction with Principal Component Analysis (page 307)
Putting it all together: face recognition (page 308)
Linear model: from regression to sparsity (page 310)
Sparse models (page 310)
Model selection: choosing estimators and their parameters (page 311)
Grid-search and cross-validated estimators (page 311)
Warning: As of version 0.9 (released in September 2011), the import path for scikit-learn has changed from
scikits.learn to sklearn
301
17.1 Loading an example dataset
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.
We have 150 observations of the iris flower specifying some measurements: sepal length, sepal width, petal length
and petal width together with its subtype: Iris setosa, Iris versicolor, Iris virginica.
To load the dataset into a Python object:
>>> 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])
17.1. Loading an example dataset 302

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)
To use this dataset with the scikit, we transform each 8x8 image into a vector of length 64
>>> data = digits.images.reshape((digits.images.shape[0], -1))
17.1.1 Learning and Predicting
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
>>> clf = svm.LinearSVC()
>>> clf.fit(iris.data, iris.target) # learn from the data
LinearSVC(...)
Once we have learned from the data, we can use our model to predict the most likely outcome on unseen data:
>>> clf.predict([[ 5.0, 3.6, 1.3, 0.25]])
Note: We can access the parameters of the model via its attributes ending with an underscore:
>>> clf.coef_
array([[ 0...]])
17.2 Classification
17.2.1 k-Nearest neighbors classifier
The simplest possible classifier is the nearest neighbor: given a new observation, take the label of the training
samples closest to it in n-dimensional space, where n is the number of features in each sample.
17.2. Classification 303

The k-nearest neighbors classifier internally uses an algorithm based on ball trees to represent the samples it is
trained on.
KNN (k-nearest neighbors) classification example:
>>> # Create and fit a nearest-neighbor classifier
>>> from sklearn import neighbors
>>> knn = neighbors.KNeighborsClassifier()
>>> knn.fit(iris.data, iris.target)
KNeighborsClassifier(...)
>>> knn.predict([[0.1, 0.2, 0.3, 0.4]])
array([0])
Training set and testing set

When experimenting with learning algorithms, it is important not to test the prediction of an estimator on
the data used to fit the estimator. Indeed, with the kNN estimator, we would always get perfect prediction on
the training set.
>>> perm = np.random.permutation(iris.target.size)
>>> iris.data = iris.data[perm]
>>> iris.target = iris.target[perm]
>>> knn.fit(iris.data[:100], iris.target[:100])
KNeighborsClassifier(...)
>>> knn.score(iris.data[100:], iris.target[100:])
0.95999...
Bonus question: why did we use a random permutation?
17.2.2 Support vector machines (SVMs) for classification
Linear Support Vector Machines
SVMs try to construct a hyperplane maximizing the margin between the two classes. It selects a subset of the
input, called the support vectors, which are the observations closest to the separating hyperplane.

>>> from sklearn import svm

>>> svc = svm.SVC(kernel=linear)
>>> svc.fit(iris.data, iris.target)
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.
Using kernels
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:
Linear kernel Polynomial kernel RBF kernel (Radial Basis Func-
tion)
>>> svc = svm.SVC(kernel=linear)

>>> svc = svm.SVC(kernel=poly,
>>> svc = svm.SVC(kernel=rbf)
... degree=3)
>>> # gamma: inverse of size of
>>> # degree: polynomial degree
>>> # radial kernel
Exercise
Which of the kernels noted above has a better prediction performance on the digits dataset?

17.3 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.
17.3.1 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(k=3)
>>> k_means.fit(iris.data)
KMeans(copy_x=True, init=k-means++, k=3, ...
>>> print k_means.labels_[::10]
[1 1 1 1 1 0 0 0 0 0 2 2 2 2 2]
>>> print iris.target[::10]
[0 0 0 0 0 1 1 1 1 1 2 2 2 2 2]
Ground truth K-means (3 clusters) K-means (8 clusters)
17.3. Clustering: grouping observations together 306

Application to Image Compression

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):
>>> lena = misc.lena().astype(np.float32)
>>> X = lena.reshape((-1, 1)) # We need an (n_sample, n_feature) array
>>> 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
Raw image K-means quantization
17.4 Dimension Reduction with Principal Component Analysis
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)
17.4. Dimension Reduction with Principal Component Analysis 307

>>> pca.fit(iris.data)
PCA(copy=True, n_components=2, whiten=False)
>>> X = pca.transform(iris.data)
Now we can visualize the (transformed) iris dataset:

>>> pl.scatter(X[:, 0], X[:, 1], c=iris.target)
<matplotlib.collections...Collection object at ...>
PCA is not just useful for visualization of high dimensional datasets. It can also be used as a preprocessing step
to help speed up supervised methods that are not efficient with high dimensions.
17.5 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.
17.5. Putting it all together: face recognition 308

"""
Stripped-down version of the face recognition example by Olivier Grisel
http://scikit-learn.org/dev/auto_examples/applications/face_recognition.html
## original shape of images: 50, 37

"""
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 = decomposition.RandomizedPCA(n_components=150, whiten=True)
pca.fit(X_train)
X_train_pca = pca.transform(X_train)
X_test_pca = pca.transform(X_test)
# ..
# .. classification ..
clf = svm.SVC(C=5., gamma=0.001)
clf.fit(X_train_pca, y_train)
17.5. Putting it all together: face recognition 309

# ..
# .. 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)
_ = raw_input()
17.6 Linear model: from regression to sparsity
Diabetes dataset
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:
>>> diabetes = datasets.load_diabetes()
>>> diabetes_X_train = diabetes.data[:-20]
>>> diabetes_X_test = diabetes.data[-20:]
>>> diabetes_y_train = diabetes.target[:-20]
>>> diabetes_y_test = diabetes.target[-20:]
The task at hand is to predict disease prediction from physiological variables.
17.6.1 Sparse models
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
are called sparse method, and sparsity can be seen as an application of Occams razor: prefer simpler models to
complex ones.
>>> from sklearn import linear_model
>>> regr = linear_model.Lasso(alpha=.3)
>>> regr.fit(diabetes_X_train, diabetes_y_train)
Lasso(...)
>>> regr.coef_ # very sparse coefficients
array([ 0. , -0. , 497.34075682, 199.17441034,
-0. , -0. , -118.89291545, 0. ,
430.9379595 , 0. ])
>>> regr.score(diabetes_X_test, diabetes_y_test)
0.5510835453...
being the score very similar to linear regression (Least Squares):

>>> lin = linear_model.LinearRegression()
>>> lin.fit(diabetes_X_train, diabetes_y_train)
LinearRegression(...)
>>> lin.score(diabetes_X_test, diabetes_y_test)
0.5850753022...
Different algorithms for a same problem

Different algorithms can be used to solve the same mathematical problem. For instance the Lasso object
in the sklearn solves the lasso regression using a coordinate descent method, that is efficient on large
datasets. However, the sklearn also provides the LassoLARS object, using the LARS which is very
efficient for problems in which the weight vector estimated is very sparse, that is problems with very few
observations.
17.6. Linear model: from regression to sparsity 310

17.7 Model selection: choosing estimators and their parameters
17.7.1 Grid-search and cross-validated estimators
Grid-search
The scikit-learn provides an object that, given data, computes the score during the fit of an estimator on a parameter
grid and chooses the parameters to maximize the cross-validation score. This object takes an estimator during the
construction and exposes an estimator API:
>>> from sklearn import svm, grid_search
>>> gammas = np.logspace(-6, -1, 10)
>>> svc = svm.SVC()
>>> clf = grid_search.GridSearchCV(estimator=svc, param_grid=dict(gamma=gammas),
... n_jobs=-1)
>>> clf.fit(digits.data[:1000], digits.target[:1000])
GridSearchCV(cv=None,
estimator=SVC(C=1.0, ...
>>> clf.best_score
0.98899798001594419
>>> clf.best_estimator.gamma
0.00059948425031894088
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.
Cross-validated estimators
Cross-validation to set a parameter can be done more efficiently on an algorithm-by-algorithm basis. This is
why, for certain estimators, the scikit-learn exposes CV estimators, that set their parameter automatically by
cross-validation:
>>> from sklearn import linear_model, datasets
>>> lasso = linear_model.LassoCV()
>>> diabetes = datasets.load_diabetes()
>>> X_diabetes = diabetes.data
>>> y_diabetes = diabetes.target
>>> lasso.fit(X_diabetes, y_diabetes)
LassoCV(alphas=array([ 2.14804, 2.00327, ..., 0.0023 , 0.00215]),
copy_X=True, cv=None, eps=0.001, fit_intercept=True, max_iter=1000,
n_alphas=100, normalize=False, precompute=auto, tol=0.0001,
verbose=False)
>>> # The estimator chose automatically its lambda:
>>> lasso.alpha
0.013...
These estimators are called similarly to their counterparts, with CV appended to their name.
Exercise
On the diabetes dataset, find the optimal regularization parameter alpha.
17.7. Model selection: choosing estimators and their parameters 311

CHAPTER 18
Interfacing with C
author Valentin Haenel

This chapter contains an introduction to the many different routes for making your native code (primarliy C/C++)
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
want to consult the respective documentation for your selected technique.
Chapters contents
Python-C-Api (page 313)
Ctypes (page 317)
SWIG (page 320)
Cython (page 324)
Summary (page 328)
Further Reading and References (page 328)
18.1 Introduction
This chapter covers the following techniques:

Python-C-Api
Ctypes
SWIG (Simplified Wrapper and Interface Generator)
Cython
These four techniques are perhaps the most well known ones, of which Cython is probably the most advanced one
and the one you should consider using first. The others are also important, if you want to understand the wrapping
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.
The following criteria may be useful when evaluating a technology:
Are additional libraries required?
Is the code autogenerated?
312
Does it need to be compiled?

Is there good support for interacting with Numpy arrays?
Does it support C++?
Before you set out, you should consider your use case. When interfacing with native code, there are usually two
use-cases that come up:
Existing code in C/C++ that needs to be leveraged, either because it already exists, or because it is faster.
Python code too slow, push inner loops to native code
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
includes some form of Numpy support, this is also demonstrated using an example where the cosine is computed
on some kind of array.
Last but not least, two small warnings:
All of these techniques may crash (segmentation fault) the Python interpreter, which is (usually) due to bugs
in the C code.
All the examples have been done on Linux, they should be possible on other operating systems.
You will need a C compiler for most of the examples.
18.2 Python-C-Api
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
compatibility, call any function written in C or C++.
When using the Python-C-API, one usually writes much boilerplate code, first to parse the arguments that were
given to a function, and later to construct the return type.
Advantages
Requires no additional libraries
Lots of low-level control
Entirely usable from C++
Disadvantages
May requires a substantial amount of effort
Much overhead in the code
Must be compiled
High maintenance cost
No forward compatibility across Python versions as C-Api changes
Note: The Python-C-Api example here serves mainly for didactic reasons. Many of the other techniques actually
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
be better off, using an alternative technique.
18.2.1 Example
The following C-extension module, make the cos function from the standard math library available to Python:
18.2. Python-C-Api 313

/* Example of wrapping cos function from math.h with the Python-C-API. */
#include <Python.h>
#include <math.h>
/* wrapped cosine function */

static PyObject* cos_func(PyObject* self, PyObject* args)
{
double value;
double answer;
/* parse the input, from python float to c double */

if (!PyArg_ParseTuple(args, "d", &value))
return NULL;
/* if the above function returns -1, an appropriate Python exception will
* have been set, and the function simply returns NULL
*/
/* call cos from libm */

answer = cos(value);
/* construct the output from cos, from c double to python float */

return Py_BuildValue("f", answer);
}
/* define functions in module */

static PyMethodDef CosMethods[] =
{
{"cos_func", cos_func, METH_VARARGS, "evaluate the cosine"},
{NULL, NULL, 0, NULL}
};
/* module initialization */
PyMODINIT_FUNC
initcos_module(void)
{
(void) Py_InitModule("cos_module", CosMethods);
}
As you can see, there is much boilerplate, both to massage the arguments and return types into place and for
the module initialisation. Although some of this is amortised, as the extension grows, the boilerplate required for
each function(s) remains.
The standard python build system distutils supports compiling C-extensions from a setup.py, which is
rather convenient:
from distutils.core import setup, Extension
# define the extension module

cos_module = Extension(cos_module, sources=[cos_module.c])
# run the setup

setup(ext_modules=[cos_module])
This can be compiled:

$ cd advanced/interfacing_with_c/python_c_api
$ ls
cos_module.c setup.py
$ python setup.py build_ext --inplace

running build_ext
building cos_module extension
creating build
creating build/temp.linux-x86_64-2.7
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
$ ls
build/ cos_module.c cos_module.so setup.py
build_ext is to build extension modules

--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:
In [1]: import cos_module
In [2]: cos_module?
Type: module
String Form:<module cos_module from cos_module.so>
File: /home/esc/git-working/scipy-lecture-notes/advanced/interfacing_with_c/python_c_api/cos
Docstring: <no docstring>
In [3]: dir(cos_module)
Out[3]: [__doc__, __file__, __name__, __package__, cos_func]
In [4]: cos_module.cos_func(1.0)
Out[4]: 0.5403023058681398
Out[5]: 1.0
Out[7]: -1.0
Now lets see how robust this is:

In [10]: cos_module.cos_func(foo)
---------------------------------------------------------------------------
TypeError Traceback (most recent call last)
<ipython-input-10-11bee483665d> in <module>()
----> 1 cos_module.cos_func(foo)
TypeError: a float is required
18.2.2 Numpy Support
Analog to the Python-C-API, Numpy, which is itself implemented as a C-extension, comes with the Numpy-C-
API. This API can be used to create and manipulate Numpy arrays from C, when writing a custom C-extension.
See also: :ref:advanced_numpy_.
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
math.h and returns a resulting new array.
/* Example of wrapping the cos function from math.h using the Numpy-C-API. */
#include <Python.h>
#include <numpy/arrayobject.h>
#include <math.h>
/* wrapped cosine function */

static PyObject* cos_func_np(PyObject* self, PyObject* args)

{
PyArrayObject *in_array;
PyObject *out_array;
PyArrayIterObject *in_iter;
PyArrayIterObject *out_iter;
/* parse single numpy array argument */

if (!PyArg_ParseTuple(args, "O!", &PyArray_Type, &in_array))
return NULL;
/* construct the output array, like the input array */

out_array = PyArray_NewLikeArray(in_array, NPY_ANYORDER, NULL, 0);
if (out_array == NULL)
return NULL;
/* create the iterators */

/* TODO: this iterator API is deprecated since 1.6
* replace in favour of the new NpyIter API */
in_iter = (PyArrayIterObject *)PyArray_IterNew((PyObject*)in_array);
out_iter = (PyArrayIterObject *)PyArray_IterNew(out_array);
if (in_iter == NULL || out_iter == NULL)
goto fail;
/* iterate over the arrays */

while (in_iter->index < in_iter->size
&& out_iter->index < out_iter->size) {
/* get the datapointers */
double * in_dataptr = (double *)in_iter->dataptr;
double * out_dataptr = (double *)out_iter->dataptr;
/* cosine of input into output */
*out_dataptr = cos(*in_dataptr);
/* update the iterator */
PyArray_ITER_NEXT(in_iter);
PyArray_ITER_NEXT(out_iter);
}
/* clean up and return the result */

Py_DECREF(in_iter);
Py_DECREF(out_iter);
Py_INCREF(out_array);
return out_array;
/* in case bad things happen */

fail:
Py_XDECREF(out_array);
Py_XDECREF(in_iter);
Py_XDECREF(out_iter);
return NULL;
}
/* define functions in module */

static PyMethodDef CosMethods[] =
{
{"cos_func_np", cos_func_np, METH_VARARGS,
"evaluate the cosine on a numpy array"},
{NULL, NULL, 0, NULL}
};
/* module initialization */
PyMODINIT_FUNC

initcos_module_np(void)
{
(void) Py_InitModule("cos_module_np", CosMethods);
/* IMPORTANT: this must be called */
import_array();
}
To compile this we can use distutils again. However we need to be sure to include the Numpy headers by using
numpy.get_include().
import numpy
# define the extension module

cos_module_np = Extension(cos_module_np, sources=[cos_module_np.c],
include_dirs=[numpy.get_include()])
# run the setup

setup(ext_modules=[cos_module_np])
To convince ourselves if this does actually works, we run the following test script:
import cos_module_np
import numpy as np
import pylab
x = np.arange(0, 2 * np.pi, 0.1)

y = cos_module_np.cos_func_np(x)
pylab.plot(x, y)
pylab.show()
And this should result in the following figure:
18.3 Ctypes
Ctypes is a foreign function library for Python. It provides C compatible data types, and allows calling functions
in DLLs or shared libraries. It can be used to wrap these libraries in pure Python.
Advantages
Part of the Python standard library
Does not need to be compiled
Wrapping code entirely in Python
Disadvantages
Requires code to be wrapped to be available as a shared library (roughly speaking *.dll in Windows
*.so in Linux and *.dylib in Mac OSX.)
No good support for C++
18.3.1 Example
As advertised, the wrapper code is in pure Python.
18.3. Ctypes 317

""" Example of wrapping cos function from math.h using ctypes. """
import ctypes
from ctypes.util import find_library
# find and load the library

libm = ctypes.cdll.LoadLibrary(find_library(m))
# set the argument type
libm.cos.argtypes = [ctypes.c_double]
# set the return type
libm.cos.restype = ctypes.c_double
def cos_func(arg):
Wrapper for cos from math.h
return libm.cos(arg)
Finding and loading the library may vary depending on your operating system, check the documentation for
details
This may be somewhat deceptive, since the math library exists in compiled form on the system already. If
you were to wrap a in-house library, you would have to compile it first, which may or may not require some
additional effort.
We may now use this, as before:
In [2]: cos_module?
Type: module
String Form:<module cos_module from cos_module.py>
File: /home/esc/git-working/scipy-lecture-notes/advanced/interfacing_with_c/ctypes/cos_modul
Out[3]:
[__builtins__,
__doc__,
__file__,
__name__,
__package__,
cos_func,
ctypes,
find_library,
libm]
Out[4]: 0.5403023058681398
Out[5]: 1.0
Out[6]: -1.0
As with the previous example, this code is somewhat robust, although the error message is not quite as helpful,
since it does not tell us what the type should be.
---------------------------------------------------------------------------
ArgumentError Traceback (most recent call last)
18.3. Ctypes 318

/home/esc/git-working/scipy-lecture-notes/advanced/interfacing_with_c/ctypes/cos_module.py in cos_
12 def cos_func(arg):
13 Wrapper for cos from math.h
---> 14 return libm.cos(arg)
ArgumentError: argument 1: <type exceptions.TypeError>: wrong type
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
and back.
For more information, consult the corresponding section in the Numpy Cookbook and the API documentation for
numpy.ndarray.ctypes and numpy.ctypeslib.
For the following example, lets consider a C function in a library that takes an input and an output array, computes
the cosine of the input array and stores the result in the output_array.
The library consists of the following header file (although this is not strictly needed for this example, we list it for
completeness):
void cos_doubles(double * in_array, double * out_array, int size);
The function implementation resides in the following C source file:

#include <math.h>
/* Compute the cosine of each element in in_array, storing the result in

* out_array. */
void cos_doubles(double * in_array, double * out_array, int size){
int i;
for(i=0;i<size;i++){
out_array[i] = cos(in_array[i]);
}
}
And since the library is pure C, we cant use distutils to compile it, but must use a combination of make and
gcc:
m.PHONY : clean
libcos_doubles.so : cos_doubles.o
gcc -shared -Wl,-soname,libcos_doubles.so -o libcos_doubles.so cos_doubles.o
cos_doubles.o : cos_doubles.c
gcc -c -fPIC cos_doubles.c -o cos_doubles.o
clean :
-rm -vf libcos_doubles.so cos_doubles.o cos_doubles.pyc
We can then compile this (on Linux) into the shared library libcos_doubles.so:
$ ls
cos_doubles.c cos_doubles.h cos_doubles.py makefile test_cos_doubles.py
$ make
gcc -c -fPIC cos_doubles.c -o cos_doubles.o
gcc -shared -Wl,-soname,libcos_doubles.so -o libcos_doubles.so cos_doubles.o
$ ls
cos_doubles.c cos_doubles.o libcos_doubles.so* test_cos_doubles.py
cos_doubles.h cos_doubles.py makefile
Now we can proceed to wrap this library via ctypes with direct support for (certain kinds of) Numpy arrays:
18.3. Ctypes 319

""" Example of wrapping a C library function that accepts a C double array as

input using the numpy.ctypeslib. """
import numpy as np
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
array_1d_double = npct.ndpointer(dtype=np.double, ndim=1, flags=CONTIGUOUS)
# load the library, using numpy mechanisms

libcd = npct.load_library("libcos_doubles", ".")
# setup the return typs and argument types

libcd.cos_doubles.restype = None
libcd.cos_doubles.argtypes = [array_1d_double, array_1d_double, c_int]
def cos_doubles_func(in_array, out_array):

return libcd.cos_doubles(in_array, out_array, len(in_array))
Note the inherent limitation of contiguous single dimensional Numpy arrays, since the C functions requires
this kind of buffer.
Also note that the output array must be preallocated, for example with numpy.zeros() and the function
will write into its buffer.
Although the original signature of the cos_doubles function is ARRAY, ARRAY, int the final
cos_doubles_func takes only two Numpy arrays as arguments.
And, as before, we convince ourselves that it worked:
import numpy as np
import pylab
import cos_doubles

y = np.empty_like(x)
cos_doubles.cos_doubles_func(x, y)
pylab.plot(x, y)
pylab.show()
18.4 SWIG
SWIG, the Simplified Wrapper Interface Generator, is a software development tool that connects programs written
in C and C++ with a variety of high-level programming languages, including Python. The important thing with
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
multiple levels of indirection which are a result of the wrapping process, may be a bit tricky to understand.
Note: The autogenerated C code uses the Python-C-Api.
18.4. SWIG 320

Advantages
Can automatically wrap entire libraries given the headers
Works nicely with C++
Disadvantages
Autogenerates enormous files
Hard to debug if something goes wrong
Steep learning curve
18.4.1 Example
Lets imagine that our cos function lives in a cos_module which has been written in c and consists of the
source file cos_module.c:
#include <math.h>
double cos_func(double arg){

return cos(arg);
}
and the header file cos_module.h:

double cos_func(double arg);
And our goal is to expose the cos_func to Python. To achieve this with SWIG, we must write an interface file
which contains the instructions for SWIG.
/* Example of wrapping cos function from math.h using SWIG. */
%module cos_module
%{
/* the resulting C file should be built as a python extension */
#define SWIG_FILE_WITH_INIT
/* Includes the header in the wrapper code */
#include "cos_module.h"
%}
/* Parse the header file to generate wrappers */
%include "cos_module.h"
As you can see, not too much code is needed here. For this simple example it is enough to simply include the
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:
setup(ext_modules=[Extension("_cos_module",
sources=["cos_module.c", "cos_module.i"])])
$ cd advanced/interfacing_with_c/swig
$ ls
cos_module.c cos_module.h cos_module.i setup.py
18.4. SWIG 321


running build_ext
building _cos_module extension
swigging cos_module.i to cos_module_wrap.c
swig -python -o cos_module_wrap.c cos_module.i
creating build
gcc -pthread -shared build/temp.linux-x86_64-2.7/cos_module.o build/temp.linux-x86_64-2.7/cos_modu
$ ls
build/ cos_module.c cos_module.h cos_module.i cos_module.py _cos_module.so* cos_module_wrap.
We can now load and execute the cos_module as we have done in the previous examples:
In [2]: cos_module?
Type: module
String Form:<module cos_module from cos_module.py>
File: /home/esc/git-working/scipy-lecture-notes/advanced/interfacing_with_c/swig/cos_module.
Out[3]:
[__builtins__,
__doc__,
__file__,
__name__,
__package__,
_cos_module,
_newclass,
_object,
_swig_getattr,
_swig_property,
_swig_repr,
_swig_setattr,
_swig_setattr_nondynamic,
cos_func]
Out[4]: 0.5403023058681398
Out[5]: 1.0
Out[6]: -1.0
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):
---------------------------------------------------------------------------
TypeError: in method cos_func, argument 1 of type double
18.4. SWIG 322

Numpy provides support for SWIG with the numpy.i file. This interface file defines various so-called typemaps
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.
We have the same cos_doubles function as in the ctypes example:
#include <math.h>

* out_array. */
int i;
}
}
This is wrapped as cos_doubles_func using the following SWIG interface file:

/* Example of wrapping a C function that takes a C double array as input using
* numpy typemaps for SWIG. */
%module cos_doubles
%{
/* the resulting C file should be built as a python extension */
#define SWIG_FILE_WITH_INIT
/* Includes the header in the wrapper code */
#include "cos_doubles.h"
%}
/* include the numpy typemaps */

%include "numpy.i"
/* need this for correct module initialization */
%init %{
import_array();
%}
/* typemaps for the two arrays, the second will be modified in-place */
%apply (double* IN_ARRAY1, int DIM1) {(double * in_array, int size_in)}
%apply (double* INPLACE_ARRAY1, int DIM1) {(double * out_array, int size_out)}
/* Wrapper for cos_doubles that massages the types */

%inline %{
/* takes as input two numpy arrays */
void cos_doubles_func(double * in_array, int size_in, double * out_array, int size_out) {
/* calls the original funcion, providing only the size of the first */
cos_doubles(in_array, out_array, size_in);
}
%}
To use the Numpy typemaps, we need include the numpy.i file.

Observe the call to import_array() which we encountered already in the Numpy-C-API example.
Since the type maps only support the signature ARRAY, SIZE we need to wrap the cos_doubles as
cos_doubles_func which takes two arrays including sizes as input.
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.
And, as before we can use distutils to wrap this:
18.4. SWIG 323


import numpy
setup(ext_modules=[Extension("_cos_doubles",
sources=["cos_doubles.c", "cos_doubles.i"],
include_dirs=[numpy.get_include()])])
As previously, we need to use include_dirs to specify the location.

$ ls
cos_doubles.c cos_doubles.h cos_doubles.i numpy.i setup.py test_cos_doubles.py
$ python setup.py build_ext -i
running build_ext
building _cos_doubles extension
swigging cos_doubles.i to cos_doubles_wrap.c
swig -python -o cos_doubles_wrap.c cos_doubles.i
cos_doubles.i:24: Warning(490): Fragment NumPy_Backward_Compatibility not found.
creating build
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
from /home/esc/anaconda/lib/python2.7/site-packages/numpy/core/include/numpy/arra
from cos_doubles_wrap.c:2706:
/home/esc/anaconda/lib/python2.7/site-packages/numpy/core/include/numpy/npy_deprecated_api.h:11:2:
gcc -pthread -shared build/temp.linux-x86_64-2.7/cos_doubles.o build/temp.linux-x86_64-2.7/cos_dou
$ ls
build/ cos_doubles.h cos_doubles.py cos_doubles_wrap.c setup.py
cos_doubles.c cos_doubles.i _cos_doubles.so* numpy.i test_cos_doubles.py

import numpy as np
import pylab
import cos_doubles

pylab.plot(x, y)
pylab.show()
18.5 Cython
Cython is both a Python-like language for writing C-extensions and an advanced compiler for this language. The
Cython language is a superset of Python, which comes with additional constructs that allow you call C functions
and annotate variables and class attributes with c types. In this sense one could also call it a Python with types.
In addition to the basic use case of wrapping native code, Cython supports an additional use-case, namely interac-
tive optimization. Basically, one starts out with a pure-Python script and incrementally adds Cython types to the
18.5. Cython 324

bottleneck code to optimize only those code paths that really matter.
In this sense it is quite similar to SWIG, since the code can be autogenerated but in a sense it also quite similar to
ctypes since the wrapping code can (almost) be written in Python.
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.
Note: The autogenerated C code uses the Python-C-Api.
Advantages
Python like language for writing C-extensions
Autogenerated code
Supports incremental optimization
Includes a GNU debugger extension
Support for C++ (Since version 0.13)
Disadvantages
Must be compiled
Requires an additional library ( but only at build time, at this problem can be overcome by shipping the
generated C files)
18.5.1 Example
The main Cython code for our cos_module is contained in the file cos_module.pyx:
""" Example of wrapping cos function from math.h using Cython. """
cdef extern from "math.h":

double cos(double arg)
def cos_func(arg):
return cos(arg)
Note the additional keywords such as cdef and extern. Also the cos_func is then pure Python.
Again we can use the standard distutils module, but this time we need some additional pieces from the
Cython.Distutils:
from Cython.Distutils import build_ext
setup(
cmdclass={build_ext: build_ext},
ext_modules=[Extension("cos_module", ["cos_module.pyx"])]
)
Compiling this:
$ cd advanced/interfacing_with_c/cython
$ ls
cos_module.pyx setup.py
running build_ext
cythoning cos_module.pyx to cos_module.c
building cos_module extension
creating build
18.5. Cython 325

gcc -pthread -shared build/temp.linux-x86_64-2.7/cos_module.o -L/home/esc/anaconda/lib -lpython2.7
$ ls
build/ cos_module.c cos_module.pyx cos_module.so* setup.py
And running it:

In [2]: cos_module?
Type: module
String Form:<module cos_module from cos_module.so>
File: /home/esc/git-working/scipy-lecture-notes/advanced/interfacing_with_c/cython/cos_modul
Out[3]:
[__builtins__,
__doc__,
__file__,
__name__,
__package__,
__test__,
cos_func]
Out[4]: 0.5403023058681398
Out[5]: 1.0
Out[6]: -1.0
And, testing a little for robustness, we can see that we get good error messages:
---------------------------------------------------------------------------
/home/esc/git-working/scipy-lecture-notes/advanced/interfacing_with_c/cython/cos_module.so in cos_
TypeError: a float is required
Additionally, it is worth noting that Cython ships with complete declarations for the C math library, which
simplifies the code above to become:
""" Simpler example of wrapping cos function from math.h using Cython. """
from libc.math cimport cos
def cos_func(arg):
return cos(arg)
In this case the cimport statement is used to import the cos function.
Cython has support for Numpy via the numpy.pyx file which allows you to add the Numpy array type to your
Cython code. I.e. like specifying that variable i is of type int, you can specify that variable a is of type
18.5. Cython 326

numpy.ndarray with a given dtype. Also certain optimizations such as bounds checking are supported.
Look at the corresponding section in the Cython documentation. In case you want to pass Numpy arrays as C
arrays to your Cython wrapped C functions, there is a section about this in the Cython wiki.
In the following example, we will show how to wrap the familiar cos_doubles function using Cython.
#include <math.h>

* out_array. */
int i;
}
}
This is wrapped as cos_doubles_func using the following Cython code:

""" Example of wrapping a C function that takes C double arrays as input using
the Numpy declarations from Cython """
# import both numpy and the Cython declarations for numpy

import numpy as np
cimport numpy as np
# if you want to use the Numpy-C-API from Cython

# (not strictly necessary for this example)
np.import_array()
# cdefine the signature of our c function

cdef extern from "cos_doubles.h":
void cos_doubles (double * in_array, double * out_array, int size)
# create the wrapper code, with numpy type annotations

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):
cos_doubles(<double*> np.PyArray_DATA(in_array),
<double*> np.PyArray_DATA(out_array),
in_array.shape[0])
And can be compiled using distutils:

import numpy
from Cython.Distutils import build_ext
setup(
cmdclass={build_ext: build_ext},
ext_modules=[Extension("cos_doubles",
sources=["_cos_doubles.pyx", "cos_doubles.c"],
include_dirs=[numpy.get_include()])],
)
As with the previous compiled Numpy examples, we need the include_dirs option.
$ ls
cos_doubles.c cos_doubles.h _cos_doubles.pyx setup.py test_cos_doubles.py
$ python setup.py build_ext -i
running build_ext
cythoning _cos_doubles.pyx to _cos_doubles.c
building cos_doubles extension
18.5. Cython 327

creating build
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
from /home/esc/anaconda/lib/python2.7/site-packages/numpy/core/include/numpy/arra
from _cos_doubles.c:253:
/home/esc/anaconda/lib/python2.7/site-packages/numpy/core/include/numpy/npy_deprecated_api.h:11:2:
/home/esc/anaconda/lib/python2.7/site-packages/numpy/core/include/numpy/__ufunc_api.h:236: warning
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

import numpy as np
import pylab
import cos_doubles

pylab.plot(x, y)
pylab.show()
18.6 Summary
In this section four different techniques for interfacing with native code have been presented. The table below
roughly summarizes some of the aspects of the techniques.
x Part of CPython Compiled Autogenerated Numpy Support
Python-C-Api True True False True
Ctypes True False False True
Swig False True True True
Cython False True True True
Of all three presented techniques, Cython is the most modern and advanced. In particular, the ability to optimize
code incrementally by adding types to your Python code is unique.
18.7 Further Reading and References
Gal Varoquauxs blog post about avoiding data copies provides some insight on how to handle memory
management cleverly. If you ever run into issues with large datasets, this is a reference to come back to for
some inspiration.
18.6. Summary 328

18.8 Exercises
Since this is a brand new section, the exercises are considered more as pointers as to what to look at next, so pick
the ones that you find more interesting. If you have good ideas for exercises, please let us know!
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.
(d) Empty array
(e) Arrays with non-double types
4. Use the %timeit IPython magic to measure the execution time of the various solutions
18.8.1 Python-C-API
1. Modify the Numpy example such that the function takes two input arguments, where the second is the
preallocated output array, making it similar to the other Numpy examples.
2. Modify the example such that the function only takes a single input array and modifies this in place.
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.
18.8.2 Ctypes
1. Modify the Numpy example such that cos_doubles_func handles the preallocation for you, thus mak-
ing it more like the Numpy-C-API example.
18.8.3 SWIG
1. Look at the code that SWIG autogenerates, how much of it do you understand?
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
the output array, so it may be possible to construct a Numpy array from the returned double *.)
18.8. Exercises 329

18.8.4 Cython
1. Look at the code that Cython autogenartes. Take a closer look at some of the comments that Cython inserts.
What do you see?
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.
18.8. Exercises 330

Index
D
diff, 297, 300
differentiation, 297
dsolve, 300
E
equations
algebraic, 298
differential, 300
I
integration, 298
M
Matrix, 299
P
Python Enhancement Proposals
PEP 255, 147
PEP 3118, 183
PEP 3129, 157
PEP 318, 150, 157
PEP 342, 147
PEP 343, 157
PEP 380, 149
PEP 380#id13, 149
PEP 8, 152
S
solve, 298
331

PythonScientific Simple PDF

Uploaded by

PythonScientific Simple PDF

Uploaded by

Python Scientific lecture notes

EuroScipy tutorial team

February 10, 2013 (2013.1-2-g9a33667)

I Getting started with Python for science 2

2 The Python language 9

3 NumPy: creating and manipulating numerical data 42

5 Scipy : high-level scientific computing 104

6 Getting help and finding documentation 139

II Advanced topics 143

8 Advanced Numpy 160

9 Debugging code 193

10 Optimizing code 204

11 Sparse Matrices in SciPy 212

12 Image manipulation and processing using Numpy and Scipy 232

13 Mathematical optimization: finding minima of functions 252

15 3D plotting with Mayavi 287

16 Sympy : Symbolic Mathematics in Python 294

17 scikit-learn: machine learning in Python 301

18 Interfacing with C 312

Getting started with Python for science

Scientific computing with tools and

authors Fernando Perez, Emmanuelle Gouillart, Gal Varoquaux, Valentin Haenel

1.1 Why Python?

1.1.1 The scientists needs

Get data (simulation, experiment control)

1.1.3 Existing solutions

Which solutions do scientists use to work?

Compiled languages: C, C++, Fortran, etc.

1.1. Why Python? 4

1.2 Scientific Python building blocks

IPython, an advanced Python shell http://ipython.scipy.org/moin/

Matplotlib : 2-D visualization, publication-ready plots http://matplotlib.sourceforge.net/

Mayavi : 3-D visualization http://code.enthought.com/projects/mayavi/

1.2. Scientific Python building blocks 5

1.3 The interactive workflow: IPython and a text editor

Note: Reference document for this section:

1.3.1 Command line interaction

Getting help by using the ? operator after an object:

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

1.3.2 Elaboration of the algorithm in an editor

1.3. The interactive workflow: IPython and a text editor 6

From a script to functions

1.3.3 IPython Tips and Tricks

1.3. The interactive workflow: IPython and a text editor 7

1.3. The interactive workflow: IPython and a text editor 8

The Python language

authors Chris Burns, Christophe Combelles, Emmanuelle Gouillart, Gal Varoquaux

Python for scientific computing

2.1 First steps

Start the Ipython shell (an enhanced interactive Python shell):

2.2 Basic types

2.2.1 Numerical types

Python supports the following numerical, scalar types:

2.1. First steps 10

Type conversion (casting):

2.2. Basic types 11

Warning: Integer division

If you explicitly want integer division use //

Indexing: accessing individual objects contained in the list:

Counting from the end with negative indices:

Warning: Indexing starts at 0 (as in C), not at 1 (as in Fortran or Matlab)!

Slicing: obtaining sublists of regularly-spaced elements:

2.2. Basic types 12

Slicing syntax: L[start:stop:stride]

Lists are mutable objects and can be modified:

Note: The elements of a list may have different types:

2.2. Basic types 13

Concatenate and repeat lists:

Note: Methods and Object-Oriented Programming

Note: Discovering methods:

Different string syntaxes (simple, double or triple quotes):

2.2. Basic types 14

In [1]: Hi, whats up?

The newline character is \n, and the tab character is \t.

2.2. Basic types 15

More container types