Scip y Lectures

Contents
SciKits Numpy Matplotlib I Getting started with Python for science 2

1 Python scientific computing ecosystem 4
1.1 Why Python? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2 The Scientific Python ecosystem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3 Before starting: Installing a working environment . . . . . . . . . . . . . . . . . . . . . . 8
2020 1.4 The workflow: interactive environments and text editors . . . . . . . . . . . . . . . . . . 8
SciPy Python
2 The Python language 12
EDITION 2.1 First steps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2 Basic types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.3 Control Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.4 Defining functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.5 Reusing code: scripts and modules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
IP[y]: 2.6 Input and Output . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.7 Standard Library . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
2.8 Exception handling in Python . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
Cython IPython 2.9 Object-oriented programming (OOP) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3 Python 2 and Python 3 47

3.1 A very short summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.2 Breaking changes between Python 2 and Python 3 . . . . . . . . . . . . . . . . . . . . . . 48
Scipy
Edited by 3.3 Some new features in Python 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
4 NumPy: creating and manipulating numerical data 50

Gaël Varoquaux 4.1 The NumPy array object . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
Emmanuelle Gouillart
Lecture Notes
4.2 Numerical operations on arrays . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
4.3 More elaborate arrays . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
Olaf Vahtras 4.4 Advanced operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
Pierre de Buyl 4.5 Some exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
www.scipy-lectures.org 4.6 Full code examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
5 Matplotlib: plotting 98
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
5.2 Simple plot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
5.3 Figures, Subplots, Axes and Ticks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
Gaël Varoquaux • Emmanuelle Gouillart • Olav Vahtras • Pierre de Buyl 5.4 Other Types of Plots: examples and exercises . . . . . . . . . . . . . . . . . . . . . . . . . 109
Christopher Burns • Adrian Chauve • Robert Cimrman • Christophe Combelles 5.5 Beyond this tutorial . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
5.6 Quick references . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
Ralf Gommers • André Espaze • Zbigniew Jędrzejewski-Szmek
Valentin Haenel • Michael Hartmann • Gert-Ludwig Ingold • Fabian Pedregosa
Didrik Pinte • Nicolas P. Rougier • Joris Van den Bossche • Pauli Virtanen i
and man y ot hers . ..

5.7 Full code examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119 13.8 Examples for the image processing chapter . . . . . . . . . . . . . . . . . . . . . . . . . . 370
6 Scipy : high-level scientific computing 185 14 Mathematical optimization: finding minima of functions 395
6.1 File input/output: scipy.io . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186 14.1 Knowing your problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 396
6.2 Special functions: scipy.special . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187 14.2 A review of the different optimizers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 398
6.3 Linear algebra operations: scipy.linalg . . . . . . . . . . . . . . . . . . . . . . . . . . . 187 14.3 Full code examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 403
6.4 Interpolation: scipy.interpolate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189 14.4 Examples for the mathematical optimization chapter . . . . . . . . . . . . . . . . . . . . 403
6.5 Optimization and fit: scipy.optimize . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189 14.5 Practical guide to optimization with scipy . . . . . . . . . . . . . . . . . . . . . . . . . . . 439
6.6 Statistics and random numbers: scipy.stats . . . . . . . . . . . . . . . . . . . . . . . . 194 14.6 Special case: non-linear least-squares . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 441
6.7 Numerical integration: scipy.integrate . . . . . . . . . . . . . . . . . . . . . . . . . . . 197 14.7 Optimization with constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 442
6.8 Fast Fourier transforms: scipy.fftpack . . . . . . . . . . . . . . . . . . . . . . . . . . . 199 14.8 Full code examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 444
6.9 Signal processing: scipy.signal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201 14.9 Examples for the mathematical optimization chapter . . . . . . . . . . . . . . . . . . . . 444
6.10 Image manipulation: scipy.ndimage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
6.11 Summary exercises on scientific computing . . . . . . . . . . . . . . . . . . . . . . . . . . 208 15 Interfacing with C 445
6.12 Full code examples for the scipy chapter . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221 15.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 445
15.2 Python-C-Api . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 446
7 Getting help and finding documentation 259 15.3 Ctypes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 453
15.4 SWIG . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 457
15.5 Cython . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 461
II Advanced topics 262 15.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 466
15.7 Further Reading and References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 466
8 Advanced Python Constructs 264 15.8 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 466
8.1 Iterators, generator expressions and generators . . . . . . . . . . . . . . . . . . . . . . . . 265
8.2 Decorators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269
8.3 Context managers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 277 III Packages and applications 468
9 Advanced NumPy 281 16 Statistics in Python 470
9.1 Life of ndarray . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 282 16.1 Data representation and interaction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 471
9.2 Universal functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 295 16.2 Hypothesis testing: comparing two groups . . . . . . . . . . . . . . . . . . . . . . . . . . 476
9.3 Interoperability features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 304 16.3 Linear models, multiple factors, and analysis of variance . . . . . . . . . . . . . . . . . . . 478
9.4 Array siblings: chararray, maskedarray, matrix . . . . . . . . . . . . . . . . . . . . . . 307 16.4 More visualization: seaborn for statistical exploration . . . . . . . . . . . . . . . . . . . . 483
9.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 309 16.5 Testing for interactions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 487
9.6 Contributing to NumPy/Scipy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 309 16.6 Full code for the figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 488
16.7 Solutions to this chapter’s exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 513
10 Debugging code 313
10.1 Avoiding bugs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 314 17 Sympy : Symbolic Mathematics in Python 516
10.2 Debugging workflow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 316 17.1 First Steps with SymPy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 517
10.3 Using the Python debugger . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 316 17.2 Algebraic manipulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 518
10.4 Debugging segmentation faults using gdb . . . . . . . . . . . . . . . . . . . . . . . . . . . 321 17.3 Calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 519
17.4 Equation solving . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 521
11 Optimizing code 324 17.5 Linear Algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 522
11.1 Optimization workflow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 325
11.2 Profiling Python code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 325 18 Scikit-image: image processing 524
11.3 Making code go faster . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 327 18.1 Introduction and concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 525
11.4 Writing faster numerical code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 329 18.2 Input/output, data types and colorspaces . . . . . . . . . . . . . . . . . . . . . . . . . . . 526
18.3 Image preprocessing / enhancement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 528
12 Sparse Matrices in SciPy 331 18.4 Image segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 532
12.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 331 18.5 Measuring regions’ properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 534
12.2 Storage Schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 333 18.6 Data visualization and interaction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 534
12.3 Linear System Solvers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 345 18.7 Feature extraction for computer vision . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 536
12.4 Other Interesting Packages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 350 18.8 Full code examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 537
18.9 Examples for the scikit-image chapter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 537
13 Image manipulation and processing using Numpy and Scipy 351
13.1 Opening and writing to image files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 352 19 Traits: building interactive dialogs 548
13.2 Displaying images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 354 19.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 549
13.3 Basic manipulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 355 19.2 Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 550
13.4 Image filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 358 19.3 What are Traits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 550
13.5 Feature extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 363
13.6 Measuring objects properties: ndimage.measurements . . . . . . . . . . . . . . . . . . . . 365 20 3D plotting with Mayavi 567
13.7 Full code examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 370
ii iii
Scipy lecture notes, Edition 2022.1
20.1 Mlab: the scripting interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 568

20.2 Interactive work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 574
20.3 Slicing and dicing data: sources, modules and filters . . . . . . . . . . . . . . . . . . . . . 575
20.4 Animating the data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 578
20.5 Making interactive dialogs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 579
20.6 Putting it together . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 581
21 scikit-learn: machine learning in Python 582

21.1 Introduction: problem settings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 583
21.2 Basic principles of machine learning with scikit-learn . . . . . . . . . . . . . . . . . . . . 587
21.3 Supervised Learning: Classification of Handwritten Digits . . . . . . . . . . . . . . . . . . 592
21.4 Supervised Learning: Regression of Housing Data . . . . . . . . . . . . . . . . . . . . . . 597
21.5 Measuring prediction performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 599
21.6 Unsupervised Learning: Dimensionality Reduction and Visualization . . . . . . . . . . . . 605
21.7 The eigenfaces example: chaining PCA and SVMs . . . . . . . . . . . . . . . . . . . . . . 608
21.8 The eigenfaces example: chaining PCA and SVMs . . . . . . . . . . . . . . . . . . . . . . 608
21.9 Parameter selection, Validation, and Testing . . . . . . . . . . . . . . . . . . . . . . . . . 614
21.10 Examples for the scikit-learn chapter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 621
IV About the SciPy lecture notes 665

22 About the SciPy lecture notes 666
22.1 Authors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 666
Index 669
iv Contents 1
This part of the Scipy lecture notes is a self-contained introduction to everything that is needed to use
Python for science, from the language itself, to numerical computing or plotting.
Part I
Getting started with Python for

science
2 3
• Universal Python is a language used for many different problems. Learning Python avoids learning
a new software for each new problem.
1.1.3 How does Python compare to other solutions?
1
Compiled languages: C, C++, Fortran. . .
Pros
• Very fast. For heavy computations, it’s difficult to outperform these languages.
Cons
CHAPTER • Painful usage: no interactivity during development, mandatory compilation steps,

verbose syntax, manual memory management. These are difficult languages for
non programmers.
Matlab scripting language

Pros
• Very rich collection of libraries with numerous algorithms, for many different do-
mains. Fast execution because these libraries are often written in a compiled lan-
Python scientific computing ecosystem guage.

• Pleasant development environment: comprehensive and help, integrated editor, etc.
• Commercial support is available.
Cons
• Base language is quite poor and can become restrictive for advanced users.
• Not free.
Authors: Fernando Perez, Emmanuelle Gouillart, Gaël Varoquaux, Valentin Haenel
Julia
1.1 Why Python? Pros

• Fast code, yet interactive and simple.
1.1.1 The scientist’s needs
• Easily connects to Python or C.
• Get data (simulation, experiment control),
Cons
• Manipulate and process data,
• Ecosystem limited to numerical computing.
• Visualize results, quickly to understand, but also with high quality figures, for reports or publica-
• Still young.
tions.
Other scripting languages: Scilab, Octave, R, IDL, etc.

1.1.2 Python’s strengths
Pros
• Batteries included Rich collection of already existing bricks of classic numerical methods, plot-
ting or data processing tools. We don’t want to re-program the plotting of a curve, a Fourier • Open-source, free, or at least cheaper than Matlab.
transform or a fitting algorithm. Don’t reinvent the wheel! • Some features can be very advanced (statistics in R, etc.)
• Easy to learn Most scientists are not payed as programmers, neither have they been trained so. Cons
They need to be able to draw a curve, smooth a signal, do a Fourier transform in a few minutes.
• Fewer available algorithms than in Matlab, and the language is not more advanced.
• Easy communication To keep code alive within a lab or a company it should be as readable
as a book by collaborators, students, or maybe customers. Python syntax is simple, avoiding • Some software are dedicated to one domain. Ex: Gnuplot to draw curves. These
strange symbols or lengthy routine specifications that would divert the reader from mathematical programs are very powerful, but they are restricted to a single type of usage, such
or scientific understanding of the code. as plotting.
• Efficient code Python numerical modules are computationally efficient. But needless to say that
a very fast code becomes useless if too much time is spent writing it. Python aims for quick
development times and quick execution times.
4 1.1. Why Python? 5

Scipy lecture notes, Edition 2022.1 Scipy lecture notes, Edition 2022.1
Python • Scipy : high-level numerical routines. Optimization, regression, interpolation, etc http://www.
scipy.org/
Pros
See also:
• Very rich scientific computing libraries
chapter on scipy
• Well thought out language, allowing to write very readable and well structured code:
we “code what we think”. • Matplotlib : 2-D visualization, “publication-ready” plots http://matplotlib.org/
• Many libraries beyond scientific computing (web server, serial port access, etc.) See also:
• Free and open-source software, widely spread, with a vibrant community. chapter on matplotlib
• A variety of powerful environments to work in, such as IPython, Spyder, Jupyter
notebooks, Pycharm, Visual Studio Code
Cons
• Not all the algorithms that can be found in more specialized software or toolboxes.
1.2 The Scientific Python ecosystem

Unlike Matlab, 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

• The language: flow control, data types (string, int), data collections (lists, dictionaries), etc.
Advanced interactive environments:
• Modules of the standard library: string processing, file management, simple network protocols.
• IPython, an advanced Python console http://ipython.org/
• A large number of specialized modules or applications written in Python: web framework, etc. . . .
and scientific computing. • Jupyter, notebooks in the browser http://jupyter.org/
• Development tools (automatic testing, documentation generation)
See also:
chapter on Python language
Core numeric libraries

• Numpy: numerical computing with powerful numerical arrays objects, and routines to manip-
ulate them. http://www.numpy.org/
Domain-specific packages,
See also:
• Mayavi for 3-D visualization
chapter on numpy
• pandas, statsmodels, seaborn for statistics
1.2. The Scientific Python ecosystem 6 1.2. The Scientific Python ecosystem 7
• sympy for symbolic computing

In [1]: print('Hello world')
• scikit-image for image processing Hello world
• scikit-learn for machine learning Getting help by using the ? operator after an object:
and much more packages not documented in the scipy lectures.
In [2]: print?
See also: Type: builtin_function_or_method
Base Class: <type 'builtin_function_or_method'>
chapters on advanced topics String Form: <built-in function print>
chapters on packages and applications Namespace: Python builtin
Docstring:
print(value, ..., sep=' ', end='\n', file=sys.stdout)
1.3 Before starting: Installing a working environment
Prints the values to a stream, or to sys.stdout by default.
Python comes in many flavors, and there are many ways to install it. However, we recommend to install Optional keyword arguments:
file: a file-like object (stream); defaults to the current sys.stdout.
a scientific-computing distribution, that comes readily with optimized versions of scientific modules.
sep: string inserted between values, default a space.
end: string appended after the last value, default a newline.
Warning: You should install Python 3
See also:
Python 2.7 is end of life, and will not be maintained past January 1, 2020.
• IPython user manual: https://ipython.readthedocs.io/en/stable/
Working with Python 2.7 is at your own risk. Do not expect much support.
• Jupyter Notebook QuickStart: http://jupyter.readthedocs.io/en/latest/content-quickstart.html
• Official announcement
• The end is nigh 1.4.2 Elaboration of the work in an editor
As you move forward, it will be important to not only work interactively, but also to create and reuse
Under Linux Python files. For this, a powerful code editor will get you far. Here are several good easy-to-use editors:
If you have a recent distribution, most of the tools are probably packaged, and it is recommended to use • Spyder: integrates an IPython console, a debugger, a profiler. . .
your package manager.
• PyCharm: integrates an IPython console, notebooks, a debugger. . . (freely available, but commer-
Other systems cial)
There are several fully-featured Scientific Python distributions: • Visual Studio Code: integrates a Python console, notebooks, a debugger, . . .
• Anaconda • Atom
• EPD Some of these are shipped by the various scientific Python distributions, and you can find them in the
menus.
• WinPython
As an exercise, create a file my_file.py in a code editor, and add the following lines:
1.4 The workflow: interactive environments and text editors s = 'Hello world'
print(s)
Interactive work to test and understand algorithms: In this section, we describe a workflow
combining interactive work and consolidation. Now, you can run it in IPython console or a notebook and explore the resulting variables:
Python is a general-purpose language. As such, there is not one blessed environment to work in, and In [1]: %run my_file.py
not only one way of using it. Although this makes it harder for beginners to find their way, it makes it Hello world
possible for Python to be used for programs, in web servers, or embedded devices.
In [2]: s
Out[2]: 'Hello world'
1.4.1 Interactive work
We recommend an interactive work with the IPython console, or its offspring, the Jupyter notebook. In [3]: %whos
They are handy to explore and understand algorithms. Variable Type Data/Info
----------------------------
s str Hello world
Under the notebook
To execute code, press “shift enter” From a script to functions
Start ipython:
1.3. Before starting: Installing a working environment 8 1.4. The workflow: interactive environments and text editors 9
While it is tempting to work only with scripts, that is a file full of instructions following each other, In [2]: %cpaste
do plan to progressively evolve the script to a set of functions: Pasting code; enter '--' alone on the line to stop or use Ctrl-D.
:>>> for i in range(3):
• A script is not reusable, functions are.
:... print(i)
• Thinking in terms of functions helps breaking the problem in small blocks. :--
0
1
2
1.4.3 IPython and Jupyter Tips and Tricks
The user manuals contain a wealth of information. Here we give a quick introduction to four useful • %timeit allows you to time the execution of short snippets using the timeit module from the
features: history, tab completion, magic functions, and aliases. standard library:
In [3]: %timeit x = 10
10000000 loops, best of 3: 39 ns per loop
See also:
Command history Like a UNIX shell, the IPython console supports command history. Type up and Chapter on optimizing code
down to navigate previously typed commands:
• %debug allows you to enter post-mortem debugging. That is to say, if the code you try to execute,
In [1]: x = 10 raises an exception, using %debug will enter the debugger at the point where the exception was
thrown.
In [2]: <UP>
In [4]: x === 10
In [2]: x = 10 File "<ipython-input-6-12fd421b5f28>", line 1
x === 10
^
SyntaxError: invalid syntax
In [5]: %debug
Tab completion Tab completion, is a convenient way to explore the structure of any object you’re > /.../IPython/core/compilerop.py (87)ast_parse()
86 and are passed to the built-in compile function."""
dealing with. Simply type object_name.<TAB> to view the object’s attributes. Besides Python objects
---> 87 return compile(source, filename, symbol, self.flags | PyCF_ONLY_AST, 1)
and keywords, tab completion also works on file and directory names.* 88
In [1]: x = 10
ipdb>locals()
In [2]: x.<TAB> {'source': u'x === 10\n', 'symbol': 'exec', 'self':
x.bit_length x.denominator x.imag x.real <IPython.core.compilerop.CachingCompiler instance at 0x2ad8ef0>,
x.conjugate x.from_bytes x.numerator x.to_bytes 'filename': '<ipython-input-6-12fd421b5f28>'}
See also:
Chapter on debugging
Magic functions The console and the notebooks support 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. Aliases 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 full list of aliases is shown when
Other useful magic functions are: typing alias).
• %cd to change the current directory.
Getting help
In [1]: cd /tmp
/tmp
• The built-in cheat-sheet is accessible via the %quickref magic function.
• %cpaste allows you to paste code, especially code from websites which has been prefixed with the • A list of all available magic functions is shown when typing %magic.
standard Python prompt (e.g. >>>) or with an ipython prompt, (e.g. in [3]):
1.4. The workflow: interactive environments and text editors 10 1.4. The workflow: interactive environments and text editors 11
• 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-
2
oriented language, with dynamic typing (the same variable can contain objects of different types
during the course of a program).
See https://www.python.org/about/ for more information about distinguishing features of Python.
CHAPTER
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.
The Python language • or by starting the program from a menu, e.g. the Anaconda Navigator, the Python(x,y) menu or
the EPD menu if you have installed one of these scientific-Python suites.
Tip: If you don’t 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.
Authors: Chris Burns, Christophe Combelles, Emmanuelle Gouillart, Gaël Varoquaux Once you have started the interpreter, type
>>> print("Hello, world!")

Python for scientific computing
Hello, world!
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
Tip: The message “Hello, world!” is then displayed. You just executed your first Python instruction,
excellent tutorial https://docs.python.org/tutorial. Dedicated books are also available, such as Dive
congratulations!
into Python 3.
To get yourself started, type the following stack of instructions
>>> a = 3
>>> b = 2*a
>>> type(b)
<type 'int'>
>>> print(b)
6
>>> a*b
Tip: Python is a programming language, as are C, Fortran, BASIC, PHP, etc. Some specific 18
features of Python are as follows: >>> b = 'hello'
>>> type(b)
• an interpreted (as opposed to compiled) language. Contrary to e.g. C or Fortran, one does not <type 'str'>
compile Python code before executing it. In addition, Python can be used interactively: many >>> b + b
Python interpreters are available, from which commands and scripts can be executed. 'hellohello'
• a free software released under an open-source license: Python can be used and distributed free >>> 2*b
'hellohello'
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. Tip: Two variables a and b have been defined above. Note that one does not declare the type of a
• a very readable language with clear non-verbose syntax variable before assigning its value. In C, conversely, one should write:
12 2.1. First steps 13

(continued from previous page)

int a = 3;
>>> 8 % 3
2
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’. Type conversion (casting):
Operations on integers (b=2*a) are coded natively in Python, and so are some operations on strings such >>> float(1)
as additions and multiplications, which amount respectively to concatenation and repetition. 1.0
2.2 Basic types Warning: Integer division

In Python 2:
2.2.1 Numerical types >>> 3 / 2
1
Tip: Python supports the following numerical, scalar types:
In Python 3:
>>> 3 / 2
Integer
1.5
>>> 1 + 1
To be safe: use floats:
2
>>> a = 4 >>> 3 / 2.
>>> type(a) 1.5
<type 'int'>
>>> a = 3
Floats >>> b = 2
>>> a / b # In Python 2
>>> c = 2.1 1
>>> type(c) >>> a / float(b)
<type 'float'> 1.5
Future behavior: to always get the behavior of Python3

Complex
>>> from __future__ import division
>>> a = 1.5 + 0.5j >>> 3 / 2
>>> a.real 1.5
1.5
>>> a.imag
0.5
Tip: If you explicitly want integer division use //:
>>> type(1. + 0j) >>> 3.0 // 2
<type 'complex'> 1.0
Booleans Note: The behaviour of the division operator has changed in Python 3.
>>> 3 > 4
False
>>> test = (3 > 4)
>>> test
False 2.2.2 Containers
>>> type(test)
<type 'bool'>
Tip: Python provides many efficient types of containers, in which collections of objects can be stored.
Tip: A Python shell can therefore replace your pocket calculator, with the basic arithmetic operations
+, -, *, /, % (modulo) natively implemented Lists
Tip: A list is an ordered collection of objects, that may have different types. For example:
>>> 7 * 3.
21.0
>>> 2**10
1024
(continues on next page)
2.2. Basic types 14 2.2. Basic types 15

>>> colors = ['red', 'blue', 'green', 'black', 'white'] >>> colors = [3, -200, 'hello']
>>> type(colors) >>> colors
<type 'list'> [3, -200, 'hello']
>>> colors[1], colors[2]
Indexing: accessing individual objects contained in the list: (-200, 'hello')
>>> colors[2]
'green' Tip: 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
Counting from the end with negative indices: 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.
>>> colors[-1]
'white'
>>> colors[-2]
'black'
Tip: Python offers a large panel of functions to modify lists, or query them. Here are a few examples;
for more details, see https://docs.python.org/tutorial/datastructures.html#more-on-lists
Warning: Indexing starts at 0 (as in C), not at 1 (as in Fortran or Matlab)!
Add and remove elements:
Slicing: obtaining sublists of regularly-spaced elements:
>>> colors = ['red', 'blue', 'green', 'black', 'white']
>>> colors >>> colors.append('pink')
['red', 'blue', 'green', 'black', 'white'] >>> colors
>>> colors[2:4] ['red', 'blue', 'green', 'black', 'white', 'pink']
['green', 'black'] >>> colors.pop() # removes and returns the last item
'pink'
>>> colors
['red', 'blue', 'green', 'black', 'white']
Warning: Note that colors[start:stop] contains the elements with indices i such as start<= >>> colors.extend(['pink', 'purple']) # extend colors, in-place
i < stop (i ranging from start to stop-1). Therefore, colors[start:stop] has (stop - start) >>> colors
elements. ['red', 'blue', 'green', 'black', 'white', 'pink', 'purple']
>>> colors = colors[:-2]
>>> colors
Slicing syntax: colors[start:stop:stride] ['red', 'blue', 'green', 'black', 'white']
Tip: All slicing parameters are optional: Reverse:
>>> colors >>> rcolors = colors[::-1]

['red', 'blue', 'green', 'black', 'white'] >>> rcolors
>>> colors[3:] ['white', 'black', 'green', 'blue', 'red']
['black', 'white'] >>> rcolors2 = list(colors) # new object that is a copy of colors in a different memory area
>>> colors[:3] >>> rcolors2
['red', 'blue', 'green'] ['red', 'blue', 'green', 'black', 'white']
>>> colors[::2] >>> rcolors2.reverse() # in-place; reversing rcolors2 does not affect colors
['red', 'green', 'white'] >>> rcolors2
['white', 'black', 'green', 'blue', 'red']
Lists are mutable objects and can be modified: Concatenate and repeat lists:
>>> colors[0] = 'yellow' >>> rcolors + colors

>>> colors ['white', 'black', 'green', 'blue', 'red', 'red', 'blue', 'green', 'black', 'white']
['yellow', 'blue', 'green', 'black', 'white'] >>> rcolors * 2
>>> colors[2:4] = ['gray', 'purple'] ['white', 'black', 'green', 'blue', 'red', 'white', 'black', 'green', 'blue', 'red']
>>> colors
['yellow', 'blue', 'gray', 'purple', 'white']
Tip: Sort:
>>> sorted(rcolors) # new object

Note: The elements of a list may have different types:
['black', 'blue', 'green', 'red', 'white']
>>> rcolors

(continued from previous page) (continued from previous page)

['white', 'black', 'green', 'blue', 'red'] 'e'
>>> rcolors.sort() # in-place >>> a[-1]
>>> rcolors 'o'
['black', 'blue', 'green', 'red', 'white']
Tip: (Remember that negative indices correspond to counting from the right end.)
Methods and Object-Oriented Programming

Slicing:
The notation rcolors.method() (e.g. rcolors.append(3) and colors.pop()) is our first example >>> a = "hello, world!"
of object-oriented programming (OOP). Being a list, the object rcolors owns the method function >>> a[3:6] # 3rd to 6th (excluded) elements: elements 3, 4, 5
that is called using the notation .. No further knowledge of OOP than understanding the notation . 'lo,'
is necessary for going through this tutorial. >>> a[2:10:2] # Syntax: a[start:stop:step]
'lo o'
>>> a[::3] # every three characters, from beginning to end
'hl r!'
Discovering methods:
Reminder: in Ipython: tab-completion (press tab) Tip: Accents and special characters can also be handled as in Python 3 strings consist of Unicode
In [28]: rcolors.<TAB> characters.
rcolors.append rcolors.index rcolors.remove
rcolors.count rcolors.insert rcolors.reverse
rcolors.extend rcolors.pop rcolors.sort
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!"

Strings In [54]: a[2] = 'z'
---------------------------------------------------------------------------
Different string syntaxes (simple, double or triple quotes): Traceback (most recent call last):
File "<stdin>", line 1, in <module>
s = 'Hello, how are you?' TypeError: 'str' object does not support item assignment
s = "Hi, what's up"
s = '''Hello, In [55]: a.replace('l', 'z', 1)
how are you''' # tripling the quotes allows the Out[55]: 'hezlo, world!'
# string to span more than one line In [56]: a.replace('l', 'z')
s = """Hi, Out[56]: 'hezzo, worzd!'
what's up?"""
In [1]: 'Hi, what's up?' Tip: 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.
File "<ipython console>", line 1
'Hi, what's up?'
^ See also:
SyntaxError: invalid syntax
Python offers advanced possibilities for manipulating strings, looking for patterns or formatting.
The interested reader is referred to https://docs.python.org/library/stdtypes.html#string-methods and
This syntax error can be avoided by enclosing the string in double quotes instead of single quotes. https://docs.python.org/3/library/string.html#format-string-syntax
Alternatively, one can prepend a backslash to the second single quote. Other uses of the backslash are,
e.g., the newline character \n and the tab character \t. String formatting:
>>> 'An integer: %i ; a float: %f ; another string: %s ' % (1, 0.1, 'string') # with more values␣
Tip: Strings are collections like lists. Hence they can be indexed and sliced, using the same syntax and ˓→use tuple after %
rules. 'An integer: 1; a float: 0.100000; another string: string'
>>> i = 102
Indexing: >>> filename = 'processing_of_dataset_%d .txt' % i # no need for tuples with just one value␣
˓→after %
>>> a = "hello" >>> filename
>>> a[0] 'processing_of_dataset_102.txt'
'h'
>>> a[1]

Dictionaries 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
Tip: A dictionary is basically an efficient table that maps keys to values. It is an unordered
container
Things to note:
• a single object can have several names bound to it:
>>> tel = {'emmanuelle': 5752, 'sebastian': 5578}
>>> tel['francis'] = 5915 In [1]: a = [1, 2, 3]
>>> tel In [2]: b = a
{'sebastian': 5578, 'francis': 5915, 'emmanuelle': 5752} In [3]: a
>>> tel['sebastian'] Out[3]: [1, 2, 3]
5578 In [4]: b
>>> tel.keys() Out[4]: [1, 2, 3]
['sebastian', 'francis', 'emmanuelle'] In [5]: a is b
>>> tel.values() Out[5]: True
[5578, 5915, 5752] In [6]: b[1] = 'hi!'
>>> 'francis' in tel In [7]: a
True Out[7]: [1, 'hi!', 3]
• to change a list in place, use indexing/slices:

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

2.3 Control Flow
>>> s = set(('a', 'b', 'c', 'a'))
>>> s Controls the order in which the code is executed.
set(['a', 'c', 'b'])
>>> s.difference(('a', 'b')) 2.3.1 if/elif/else
set(['c'])
>>> if 2**2 == 4:
... print('Obvious!')
2.2.3 Assignment operator ...
Obvious!
Tip: Python library reference says:

Blocks are delimited by indentation
Assignment statements are used to (re)bind names to values and to modify attributes or
items of mutable objects.
Tip: Type the following lines in your Python interpreter, and be careful to respect the indentation
In short, it works as follows (simple assignment): depth. The Ipython shell automatically increases the indentation depth after a colon : sign; to decrease
2.2. Basic types 20 2.3. Control Flow 21

the indentation depth, go four spaces to the left with the Backspace key. Press the Enter key twice to (continued from previous page)
leave the logical block. ... continue
... print(1. / element)
1.0
>>> a = 10 0.5
0.25
>>> if a == 1:
... print(1)
... elif a == 2:
... print(2)
2.3.4 Conditional Expressions
... else: if <OBJECT>
... print('A lot')
A lot Evaluates to False:
• any number equal to zero (0, 0.0, 0+0j)
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. • an empty container (list, tuple, set, dictionary, . . . )
• False, None
2.3.2 for/range Evaluates to True:
Iterating with an index:
• everything else
>>> for i in range(4): a == b Tests equality, with logics:
... print(i)
0 >>> 1 == 1.
1 True
2
3
a is b Tests identity: both sides are the same object:
But most often, it is more readable to iterate over values: >>> 1 is 1.

False
>>> for word in ('cool', 'powerful', 'readable'):
... print('Python is %s ' % word) >>> a = 1
Python is cool >>> b = 1
Python is powerful >>> a is b
Python is readable True
a in b For any collection b: b contains a

2.3.3 while/break/continue
>>> b = [1, 2, 3]
Typical C-style while loop (Mandelbrot problem):
>>> 2 in b
>>> z = 1 + 1j True
>>> while abs(z) < 100: >>> 5 in b
... z = z**2 + 1 False
>>> z
(-134+352j) If b is a dictionary, this tests that a is a key of b.
More advanced features 2.3.5 Advanced iteration

break out of enclosing for/while loop:
Iterate over any sequence
>>> z = 1 + 1j You can iterate over any sequence (string, list, keys in a dictionary, lines in a file, . . . ):
>>> while abs(z) < 100: >>> vowels = 'aeiouy'
... if z.imag == 0:
... break >>> for i in 'powerful':
... z = z**2 + 1 ... if i in vowels:
... print(i)
continue the next iteration of a loop.: o
e
>>> a = [1, 0, 2, 4] u
>>> for element in a:
... if element == 0:
2.3. Control Flow 22 2.3. Control Flow 23

>>> message = "Hello how are you?" >>> [i**2 for i in range(4)]
>>> message.split() # returns a list [0, 1, 4, 9]
['Hello', 'how', 'are', 'you?']
>>> for word in message.split():
... print(word)
...
Exercise
Hello
how
are Compute the decimals of Pi using the Wallis formula:
you?
4𝑖2
∞
∏︁
𝜋=2
𝑖=1
4𝑖2 − 1
Tip: 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 don’t care about. This feature can often be used to make code more readable.
2.4 Defining functions
2.4.1 Function definition
Warning: Not safe to modify the sequence you are iterating over.
In [56]: def test():
....: print('in test function')
Keeping track of enumeration number ....:
....:
Common task is to iterate over a sequence while keeping track of the item number.
In [57]: test()
• Could use while loop with a counter as above. Or a for loop:
in test function
>>> words = ('cool', 'powerful', 'readable')
>>> for i in range(0, len(words)):
... print((i, words[i])) Warning: Function blocks must be indented as other control-flow blocks.
(0, 'cool')
(1, 'powerful')
(2, 'readable')
2.4.2 Return statement
• But, Python provides a built-in function - enumerate - for this: Functions can optionally return values.
>>> for index, item in enumerate(words): In [6]: def disk_area(radius):
... print((index, item)) ...: return 3.14 * radius * radius
(0, 'cool') ...:
(1, 'powerful')
(2, 'readable') In [8]: disk_area(1.5)
Out[8]: 7.0649999999999995
Looping over a dictionary

Use items: Note: By default, functions return None.
>>> d = {'a': 1, 'b':1.2, 'c':1j}
>>> for key, val in sorted(d.items()): Note: Note the syntax to define a function:
... print('Key: %s has value: %s ' % (key, val))
• the def keyword;
Key: a has value: 1
Key: b has value: 1.2 • is followed by the function’s name, then
Key: c has value: 1j
• the arguments of the function are given between parentheses followed by a colon.
• the function body;
Note: The ordering of a dictionary is random, thus we use sorted() which will sort on the keys.
• and return object for optionally returning values.
2.3.6 List Comprehensions

2.4.3 Parameters
Instead of creating a list by means of a loop, one can make use of a list comprehension with a rather
Mandatory parameters (positional arguments)
self-explaining syntax.
2.3. Control Flow 24 2.4. Defining functions 25

In [81]: def double_it(x):

....: return x * 2
....:
Tip: More involved example implementing python’s slicing:
In [82]: double_it(3)
Out[82]: 6 In [98]: def slicer(seq, start=None, stop=None, step=None):
....: """Implement basic python slicing."""
In [83]: double_it() ....: return seq[start:stop:step]
--------------------------------------------------------------------------- ....:
Traceback (most recent call last):
File "<stdin>", line 1, in <module> In [101]: rhyme = 'one fish, two fish, red fish, blue fish'.split()
TypeError: double_it() takes exactly 1 argument (0 given)
In [102]: rhyme
Optional parameters (keyword or named arguments) Out[102]: ['one', 'fish,', 'two', 'fish,', 'red', 'fish,', 'blue', 'fish']
In [84]: def double_it(x=2): In [103]: slicer(rhyme)

....: return x * 2 Out[103]: ['one', 'fish,', 'two', 'fish,', 'red', 'fish,', 'blue', 'fish']
....:
In [104]: slicer(rhyme, step=2)
In [85]: double_it() Out[104]: ['one', 'two', 'red', 'blue']
Out[85]: 4
In [105]: slicer(rhyme, 1, step=2)
In [86]: double_it(3) Out[105]: ['fish,', 'fish,', 'fish,', 'fish']
Out[86]: 6
In [106]: slicer(rhyme, start=1, stop=4, step=2)
Keyword arguments allow you to specify default values. Out[106]: ['fish,', 'fish,']
The order of the keyword arguments does not matter:

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 In [107]: slicer(rhyme, step=2, start=1, stop=4)
function body, since the modifications will be persistent across invocations of the function. Out[107]: ['fish,', 'fish,']
Using an immutable type in a keyword argument:
but it is good practice to use the same ordering as the function’s definition.
In [124]: bigx = 10
In [125]: def double_it(x=bigx): Keyword arguments are a very convenient feature for defining functions with a variable number of argu-
.....: return x * 2 ments, especially when default values are to be used in most calls to the function.
.....:
In [126]: bigx = 1e9 # Now really big

2.4.4 Passing by value
In [128]: double_it() Tip: Can you modify the value of a variable inside a function? Most languages (C, Java, . . . ) distinguish
Out[128]: 20 “passing by value” and “passing by reference”. In Python, such a distinction is somewhat artificial, and
Using an mutable type in a keyword argument (and modifying it inside the function body): it is a bit subtle whether your variables are going to be modified or not. Fortunately, there exist clear
rules.
In [2]: def add_to_dict(args={'a': 1, 'b': 2}):
...: for i in args.keys(): Parameters to functions are references to objects, which are passed by value. When you pass a variable
...: args[i] += 1 to a function, python passes the reference to the object to which the variable refers (the value). Not the
...: print(args) variable itself.
...:
In [3]: add_to_dict If the value passed in a function is immutable, the function does not modify the caller’s variable. If the
Out[3]: <function __main__.add_to_dict> value is mutable, the function may modify the caller’s variable in-place:
In [4]: add_to_dict() >>> def try_to_modify(x, y, z):

{'a': 2, 'b': 3} ... x = 23
... y.append(42)
In [5]: add_to_dict() ... z = [99] # new reference
{'a': 3, 'b': 4} ... print(x)
... print(y)
In [6]: add_to_dict() ... print(z)
{'a': 4, 'b': 5} (continues on next page)
2.4. Defining functions 26 2.4. Defining functions 27


... In [123]: x
>>> a = 77 # immutable variable Out[123]: 10
>>> b = [99] # mutable variable
>>> c = [28]
>>> try_to_modify(a, b, c) 2.4.6 Variable number of parameters
23
[99, 42] Special forms of parameters:
[99]
• *args: any number of positional arguments packed into a tuple
>>> print(a)
77 • **kwargs: any number of keyword arguments packed into a dictionary
>>> print(b)
[99, 42] In [35]: def variable_args(*args, **kwargs):
>>> print(c) ....: print('args is', args)
[28] ....: print('kwargs is', kwargs)
....:
Functions have a local variable table called a local namespace.
In [36]: variable_args('one', 'two', x=1, y=2, z=3)
The variable x only exists within the function try_to_modify. args is ('one', 'two')
kwargs is {'x': 1, 'y': 2, 'z': 3}
2.4.5 Global variables
Variables declared outside the function can be referenced within the function: 2.4.7 Docstrings
In [114]: x = 5 Documentation about what the function does and its parameters. General convention:
In [115]: def addx(y): In [67]: def funcname(params):

.....: return x + y ....: """Concise one-line sentence describing the function.
.....: ....:
....: Extended summary which can contain multiple paragraphs.
In [116]: addx(10) ....: """
Out[116]: 15 ....: # function body
....: pass
....:
But these “global” variables cannot be modified within the function, unless declared global in the
function. In [68]: funcname?
This doesn’t work: Type: function
Base Class: type 'function'>
In [117]: def setx(y): String Form: <function funcname at 0xeaa0f0>
.....: x = y Namespace: Interactive
.....: print('x is %d ' % x) File: <ipython console>
.....: Definition: funcname(params)
.....: Docstring:
Concise one-line sentence describing the function.
In [118]: setx(10)
x is 10 Extended summary which can contain multiple paragraphs.
In [120]: x
Out[120]: 5 Note: Docstring guidelines
This works: For the sake of standardization, the Docstring Conventions webpage documents the semantics and con-
ventions associated with Python docstrings.
In [121]: def setx(y):
Also, the Numpy and Scipy modules have defined a precise standard for documenting scientific functions,
.....: global x
.....: x = y
that you may want to follow for your own functions, with a Parameters section, an Examples section,
.....: print('x is %d ' % x) etc. See https://numpydoc.readthedocs.io/en/latest/format.html#docstring-standard
.....:
.....:
2.4.8 Functions are objects
In [122]: setx(10)
x is 10 Functions are first-class objects, which means they can be:
• assigned to a variable
2.4. Defining functions 28 2.4. Defining functions 29

• an item in a list (or any collection)

message = "Hello how are you?"
• passed as an argument to another function. for word in message.split():
print(word)
In [38]: va = variable_args
In [39]: va('three', x=1, y=2) Tip: Let us now execute the script interactively, that is inside the Ipython interpreter. This is maybe
args is ('three',)
the most common use of scripts in scientific computing.
kwargs is {'x': 1, 'y': 2}
2.4.9 Methods Note: in Ipython, the syntax to execute a script is %run script.py. For example,
Methods are functions attached to objects. You’ve seen these in our examples on lists, dictionaries,
strings, etc. . . In [1]: %run test.py
Hello
2.4.10 Exercises how
are
you?
Exercise: Fibonacci sequence
In [2]: message
Write a function that displays the n first terms of the Fibonacci sequence, defined by: Out[2]: 'Hello how are you?'
⎧
⎨ 𝑈0 = 0 The script has been executed. Moreover the variables defined in the script (such as message) are now
𝑈1 = 1 available inside the interpreter’s namespace.
⎩
𝑈𝑛+2 = 𝑈𝑛+1 + 𝑈𝑛
Tip: Other interpreters also offer the possibility to execute scripts (e.g., execfile in the plain Python
interpreter, etc.).
Exercise: Quicksort
It is also possible In order to execute this script as a standalone program, by executing the script inside
Implement the quicksort algorithm, as defined by wikipedia 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:
function quicksort(array)
$ python test.py
var list less, greater
Hello
if length(array) < 2
how
return array
are
select and remove a pivot value pivot from array
you?
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)) Tip: Standalone scripts may also take command-line arguments
In file.py:
2.5 Reusing code: scripts and modules import sys

print(sys.argv)
For now, we have typed all instructions in the interpreter. For longer sets of instructions we need to
change track and write the code in text files (using a text editor), that we will call either scripts or $ python file.py test arguments
modules. Use your favorite text editor (provided it offers syntax highlighting for Python), or the editor ['file.py', 'test', 'arguments']
that comes with the Scientific Python Suite you may be using.
2.5.1 Scripts Warning: Don’t implement option parsing yourself. Use a dedicated module such as argparse.
Tip: 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!). 2.5.2 Importing objects from modules
In [1]: import os
The extension for Python files is .py. Write or copy-and-paste the following lines in a file called test.py
2.5. Reusing code: scripts and modules 30 2.5. Reusing code: scripts and modules 31

"A demo module."
In [2]: os
Out[2]: <module 'os' from '/usr/lib/python2.6/os.pyc'> def print_b():
"Prints b."
In [3]: os.listdir('.') print('b')
Out[3]:
['conf.py', def print_a():
'basic_types.rst', "Prints a."
'control_flow.rst', print('a')
'functions.rst',
'python_language.rst',
'reusing.rst', c = 2
'file_io.rst', d = 2
'exceptions.rst',
'workflow.rst',
'index.rst']
Tip: In this file, we defined two functions print_a and print_b. Suppose we want to call the print_a
function from the interpreter. We could execute the file as a script, but since we just want to have access
And also: to the function print_a, we are rather going to import it as a module. The syntax is as follows.
In [4]: from os import listdir
In [1]: import demo
Importing shorthands:
In [5]: import numpy as np In [2]: demo.print_a()

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

b
from os import *
This is called the star import and please, Do not use it Importing the module gives access to its objects, using the module.object syntax. Don’t forget to put
the module’s name before the object’s name, otherwise Python won’t recognize the instruction.
• Makes the code harder to read and understand: where do symbols come from?
Introspection
• 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. In [4]: demo?
Type: module
• Restricts the variable names you can use: os.name might override name, or vise-versa. Base Class: <type 'module'>
String Form: <module 'demo' from 'demo.py'>
• Creates possible name clashes between modules.
Namespace: Interactive
• Makes the code impossible to statically check for undefined symbols. File: /home/varoquau/Projects/Python_talks/scipy_2009_tutorial/source/demo.py
Docstring:
A demo module.
Tip: 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: In [5]: who
demo
>>> import numpy as np # data arrays
>>> np.linspace(0, 10, 6) In [6]: whos
array([ 0., 2., 4., 6., 8., 10.]) Variable Type Data/Info
>>> import scipy # scientific computing ------------------------------
demo module <module 'demo' from 'demo.py'>
In [7]: dir(demo)
2.5.3 Creating modules Out[7]:
['__builtins__',
'__doc__',
Tip: If we want to write larger and better organized programs (compared to simple scripts), where '__file__',
some objects are defined, (variables, functions, classes) and that we want to reuse several times, we have '__name__',
to create our own modules. '__package__',
'c',
'd',
Let us create a module demo contained in the file demo.py: 'print_a',

In [11]: import demo2
'print_b'] b
In [12]: import demo2

In [8]: demo.<TAB>
demo.c demo.print_a demo.py
Running it:
demo.d demo.print_b demo.pyc
In [13]: %run demo2
Importing objects from modules into the main namespace b
a
In [9]: from demo import print_a, print_b
In [10]: whos 2.5.5 Scripts or modules? How to organize your code

Variable Type Data/Info
--------------------------------
demo module <module 'demo' from 'demo.py'> Note: Rule of thumb
print_a function <function print_a at 0xb7421534>
print_b function <function print_b at 0xb74214c4> • Sets of instructions that are called several times should be written inside functions for better code
reusability.
In [11]: print_a()
• Functions (or other bits of code) that are called from several scripts should be written inside a
a
module, so that only the module is imported in the different scripts (do not copy-and-paste your
functions in the different scripts!).
Warning: Module caching

Modules are cached: if you modify demo.py and re-import it in the old session, you will How modules are found and imported
get the old one.
When the import mymodule statement is executed, the module mymodule is searched in a given list of
Solution: directories. This list includes a list of installation-dependent default path (e.g., /usr/lib/python) as
In [10]: reload(demo) 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 Python3 instead reload is not builtin, so you have to import the importlib module first and then
do: In [1]: import sys
In [10]: importlib.reload(demo)
In [2]: sys.path
Out[2]:
['',
2.5.4 ‘__main__’ and module loading '/home/varoquau/.local/bin',
'/usr/lib/python2.7',
'/home/varoquau/.local/lib/python2.7/site-packages',
Tip: Sometimes we want code to be executed when a module is run directly, but not when it is imported '/usr/lib/python2.7/dist-packages',
by another module. if __name__ == '__main__' allows us to check whether the module is being run '/usr/local/lib/python2.7/dist-packages',
directly. ...]
File demo2.py: 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. $HOME/.local/
def print_b():
"Prints b."
lib/python2.7/dist-packages). You may use symbolic links (on Linux) to keep the code some-
print('b') where else.
• modify the environment variable PYTHONPATH to include the directories containing the user-defined
def print_a(): modules.
"Prints a."
print('a')
Tip: On Linux/Unix, add the following line to a file read by the shell at startup (e.g. /etc/profile,
# print_b() runs on import .profile)
print_b()
export PYTHONPATH=$PYTHONPATH:/home/emma/user_defined_modules
if __name__ == '__main__':
# print_a() is only executed when the module is run directly. On Windows, http://support.microsoft.com/kb/310519 explains how to handle environment vari-
print_a() ables.
Importing it:
• or modify the sys.path variable itself within a Python script. (continued from previous page)
Type: function
Base Class: <type 'function'>
Tip:
String Form: <function binary_dilation at 0x9bedd84>
import sys Namespace: Interactive
new_path = '/home/emma/user_defined_modules' File: /usr/lib/python2.6/dist-packages/scipy/ndimage/morphology.py
if new_path not in sys.path: Definition: morphology.binary_dilation(input, structure=None,
sys.path.append(new_path) iterations=1, mask=None, output=None, border_value=0, origin=0,
brute_force=False)
Docstring:
This method is not very robust, however, because it makes the code less portable (user-dependent
Multi-dimensional binary dilation with the given structure.
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. An output array can optionally be provided. The origin parameter
controls the placement of the filter. If no structuring element is
See also: provided an element is generated with a squared connectivity equal
to one. The dilation operation is repeated iterations times. If
See https://docs.python.org/tutorial/modules.html for more information about modules. 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
2.5.6 Packages modified at each iteration.
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. 2.5.7 Good practices
$ ls • Use meaningful object names
cluster/ io/ README.txt@ stsci/ • Indentation: no choice!
__config__.py@ LATEST.txt@ setup.py@ __svn_version__.py@
__config__.pyc lib/ setup.pyc __svn_version__.pyc
constants/ linalg/ setupscons.py@ THANKS.txt@ Tip: Indenting is compulsory in Python! Every command block following a colon bears an
fftpack/ linsolve/ setupscons.pyc TOCHANGE.txt@ additional indentation level with respect to the previous line with a colon. One must therefore
__init__.py@ maxentropy/ signal/ version.py@ indent after def f(): or while:. At the end of such logical blocks, one decreases the indentation
__init__.pyc misc/ sparse/ version.pyc depth (and re-increases it if a new block is entered, etc.)
INSTALL.txt@ ndimage/ spatial/ weave/
integrate/ odr/ special/ Strict respect of indentation is the price to pay for getting rid of { or ; characters that delineate
interpolate/ optimize/ stats/ logical blocks in other languages. Improper indentation leads to errors such as
$ cd ndimage
$ ls ------------------------------------------------------------
doccer.py@ fourier.pyc interpolation.py@ morphology.pyc setup.pyc IndentationError: unexpected indent (test.py, line 2)
doccer.pyc info.py@ interpolation.pyc _nd_image.so
setupscons.py@ All this indentation business can be a bit confusing in the beginning. However, with the clear
filters.py@ info.pyc measurements.py@ _ni_support.py@ indentation, and in the absence of extra characters, the resulting code is very nice to read compared
setupscons.pyc to other languages.
filters.pyc __init__.py@ measurements.pyc _ni_support.pyc tests/
fourier.py@ __init__.pyc morphology.py@ setup.py@
• Indentation depth: Inside your text editor, you may choose to indent with any positive number
From Ipython: 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 [1]: import scipy
• Style guidelines
In [2]: scipy.__file__
Long lines: you should not write very long lines that span over more than (e.g.) 80 characters.
Out[2]: '/usr/lib/python2.6/dist-packages/scipy/__init__.pyc'
Long lines can be broken with the \ character
In [3]: import scipy.version >>> long_line = "Here is a very very long line \
... that we break in two parts."
In [4]: scipy.version.version
Out[4]: '0.7.0'
Spaces
In [5]: import scipy.ndimage.morphology Write well-spaced code: put whitespaces after commas, around arithmetic operators, etc.:
In [6]: from scipy.ndimage import morphology >>> a = 1 # yes
>>> a=1 # too cramped
In [17]: morphology.binary_dilation?
A certain number of rules for writing “beautiful” code (and more importantly using the same • Append a file: a
conventions as anybody else!) are given in the Style Guide for Python Code.
• Read and Write: r+
• Binary mode: b
Quick read – Note: Use for binary files, especially on Windows.
If you want to do a first quick pass through the Scipy lectures to learn the ecosystem, you can directly
skip to the next chapter: NumPy: creating and manipulating numerical data. 2.7 Standard Library
The remainder of this chapter is not necessary to follow the rest of the intro part. But be sure to
come back and finish this chapter later. Note: Reference document for this section:
• The Python Standard Library documentation: https://docs.python.org/library/index.html
• Python Essential Reference, David Beazley, Addison-Wesley Professional
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. 2.7.1 os module: operating system functionality
We write or read strings to/from files (other types must be converted to strings). To write in a file: “A portable way of using operating system dependent functionality.”
>>> f = open('workfile', 'w') # opens the workfile file Directory and file manipulation
>>> type(f)
<type 'file'> Current directory:
>>> f.write('This is a test \nand another test')
>>> f.close() In [17]: os.getcwd()
Out[17]: '/Users/cburns/src/scipy2009/scipy_2009_tutorial/source'
To read from a file
List a directory:
In [1]: f = open('workfile', 'r')
In [31]: os.listdir(os.curdir)
In [2]: s = f.read() Out[31]:
['.index.rst.swo',
In [3]: print(s) '.python_language.rst.swp',
This is a test '.view_array.py.swp',
and another test '_static',
'_templates',
In [4]: f.close() 'basic_types.rst',
'conf.py',
'control_flow.rst',
See also:
'debugging.rst',
For more details: https://docs.python.org/tutorial/inputoutput.html ...
2.6.1 Iterating over a file Make a directory:

In [32]: os.mkdir('junkdir')
In [6]: f = open('workfile', 'r')
In [33]: 'junkdir' in os.listdir(os.curdir)
In [7]: for line in f:
Out[33]: True
...: print(line)
...:
This is a test Rename the directory:
In [36]: os.rename('junkdir', 'foodir')
and another test
In [37]: 'junkdir' in os.listdir(os.curdir)
In [8]: f.close()
Out[37]: False
File modes In [38]: 'foodir' in os.listdir(os.curdir)

Out[38]: True
• Read-only: r
In [41]: os.rmdir('foodir')
• Write-only: w
– Note: Create a new file or overwrite existing file. (continues on next page)
2.6. Input and Output 38 2.7. Standard Library 39

(continued from previous page) Running an external command

In [42]: 'foodir' in os.listdir(os.curdir)
In [8]: os.system('ls')
Out[42]: False
basic_types.rst demo.py functions.rst python_language.rst standard_library.rst
control_flow.rst exceptions.rst io.rst python-logo.png
Delete a file: demo2.py first_steps.rst oop.rst reusing_code.rst
In [44]: fp = open('junk.txt', 'w')
In [45]: fp.close() Note: Alternative to os.system

A noteworthy alternative to os.system is the sh module. Which provides much more convenient ways
In [46]: 'junk.txt' in os.listdir(os.curdir)
Out[46]: True
to obtain the output, error stream and exit code of the external command.
In [20]: import sh
In [47]: os.remove('junk.txt') In [20]: com = sh.ls()
In [48]: 'junk.txt' in os.listdir(os.curdir) In [21]: print(com)
Out[48]: False 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
os.path: path manipulations demo.py io.rst reusing_code.rst
os.path provides common operations on pathnames.
In [22]: print(com.exit_code)
In [70]: fp = open('junk.txt', 'w') 0
In [23]: type(com)
In [71]: fp.close() Out[23]: sh.RunningCommand
In [72]: a = os.path.abspath('junk.txt')
In [73]: a Walking a directory

Out[73]: '/Users/cburns/src/scipy2009/scipy_2009_tutorial/source/junk.txt'
os.path.walk generates a list of filenames in a directory tree.
In [74]: os.path.split(a)
In [10]: for dirpath, dirnames, filenames in os.walk(os.curdir):
Out[74]: ('/Users/cburns/src/scipy2009/scipy_2009_tutorial/source',
....: for fp in filenames:
'junk.txt')
....: print(os.path.abspath(fp))
....:
In [78]: os.path.dirname(a)
....:
Out[78]: '/Users/cburns/src/scipy2009/scipy_2009_tutorial/source'
/Users/cburns/src/scipy2009/scipy_2009_tutorial/source/.index.rst.swo
/Users/cburns/src/scipy2009/scipy_2009_tutorial/source/.view_array.py.swp
In [79]: os.path.basename(a)
/Users/cburns/src/scipy2009/scipy_2009_tutorial/source/basic_types.rst
Out[79]: 'junk.txt'
/Users/cburns/src/scipy2009/scipy_2009_tutorial/source/conf.py
/Users/cburns/src/scipy2009/scipy_2009_tutorial/source/control_flow.rst
In [80]: os.path.splitext(os.path.basename(a))
...
Out[80]: ('junk', '.txt')
In [84]: os.path.exists('junk.txt')
Environment variables:
Out[84]: True
In [9]: import os
In [86]: os.path.isfile('junk.txt')
Out[86]: True In [11]: os.environ.keys()
Out[11]:
In [87]: os.path.isdir('junk.txt') ['_',
Out[87]: False 'FSLDIR',
'TERM_PROGRAM_VERSION',
In [88]: os.path.expanduser('~/local') 'FSLREMOTECALL',
Out[88]: '/Users/cburns/local' 'USER',
'HOME',
In [92]: os.path.join(os.path.expanduser('~'), 'local', 'bin') 'PATH',
Out[92]: '/Users/cburns/local/bin' 'PS1',
'SHELL',
'EDITOR',
'WORKON_HOME',
2.7. Standard Library 40 2.7. Standard Library 41


'PYTHONPATH', ['',
... '/Users/cburns/local/bin',
'/Users/cburns/local/lib/python2.5/site-packages/grin-1.1-py2.5.egg',
In [12]: os.environ['PYTHONPATH'] '/Users/cburns/local/lib/python2.5/site-packages/argparse-0.8.0-py2.5.egg',
Out[12]: '.:/Users/cburns/src/utils:/Users/cburns/src/nitools: '/Users/cburns/local/lib/python2.5/site-packages/urwid-0.9.7.1-py2.5.egg',
/Users/cburns/local/lib/python2.5/site-packages/: '/Users/cburns/local/lib/python2.5/site-packages/yolk-0.4.1-py2.5.egg',
/usr/local/lib/python2.5/site-packages/: '/Users/cburns/local/lib/python2.5/site-packages/virtualenv-1.2-py2.5.egg',
/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: 2.7.5 pickle: easy persistence
/Users/cburns/local/lib/python2.5/site-packages/:
/usr/local/lib/python2.5/site-packages/: Useful to store arbitrary objects to a file. Not safe or fast!
/Library/Frameworks/Python.framework/Versions/2.5/lib/python2.5'
In [1]: import pickle
In [2]: l = [1, None, 'Stan']

2.7.2 shutil: high-level file operations
The shutil provides useful file operations: In [3]: pickle.dump(l, file('test.pkl', 'w'))
• shutil.rmtree: Recursively delete a directory tree. In [4]: pickle.load(file('test.pkl'))

Out[4]: [1, None, 'Stan']
• shutil.move: Recursively move a file or directory to another location.
• shutil.copy: Copy files or directories.
Exercise
2.7.3 glob: Pattern matching on files
Write a program to search your PYTHONPATH for the module site.py.
The glob module provides convenient file pattern matching.
Find all files ending in .txt: path_site
In [18]: import glob
In [19]: glob.glob('*.txt')
2.8 Exception handling in Python
Out[19]: ['holy_grail.txt', 'junk.txt', 'newfile.txt']
It is likely that you have 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.
2.7.4 sys module: system-specific information Exceptions are raised by different kinds of errors arising when executing Python code. In your own code,
System-specific information related to the Python interpreter. 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.
• Which version of python are you running and where is it installed:
In [117]: sys.platform 2.8.1 Exceptions

Out[117]: 'darwin'
Exceptions are raised by errors in Python:
In [118]: sys.version In [1]: 1/0
Out[118]: '2.5.2 (r252:60911, Feb 22 2008, 07:57:53) \n ---------------------------------------------------------------------------
[GCC 4.0.1 (Apple Computer, Inc. build 5363)]' ZeroDivisionError: integer division or modulo by zero
In [119]: sys.prefix In [2]: 1 + 'e'

Out[119]: '/Library/Frameworks/Python.framework/Versions/2.5' ---------------------------------------------------------------------------
TypeError: unsupported operand type(s) for +: 'int' and 'str'
• List of command line arguments passed to a Python script:
In [3]: d = {1:1, 2:2}
In [100]: sys.argv
Out[100]: ['/Users/cburns/local/bin/ipython'] In [4]: d[3]
---------------------------------------------------------------------------
sys.path is a list of strings that specifies the search path for modules. Initialized from PYTHONPATH: KeyError: 3
In [121]: sys.path In [5]: l = [1, 2, 3]

Out[121]:
(continues on next page) (continues on next page)
2.7. Standard Library 42 2.8. Exception handling in Python 43


In [6]: l[4]
--------------------------------------------------------------------------- In [14]: print_sorted('132')
IndexError: list index out of range 132
In [7]: l.foobar
--------------------------------------------------------------------------- 2.8.3 Raising exceptions
AttributeError: 'list' object has no attribute 'foobar'
• Capturing and reraising an exception:
As you can see, there are different types of exceptions for different errors. In [15]: def filter_name(name):
....: try:
2.8.2 Catching exceptions ....: name = name.encode('ascii')
....: except UnicodeError as e:
try/except ....: if name == 'Gaël':
....: print('OK, Gaël')
In [10]: while True: ....: else:
....: try: ....: raise e
....: x = int(raw_input('Please enter a number: ')) ....: return name
....: break ....:
....: except ValueError:
....: print('That was no valid number. Try again...') In [16]: filter_name('Gaël')
....: OK, Gaël
Please enter a number: a Out[16]: 'Ga\xc3\xabl'
That was no valid number. Try again...
Please enter a number: 1 In [17]: filter_name('Stéfan')
---------------------------------------------------------------------------
In [9]: x UnicodeDecodeError: 'ascii' codec can't decode byte 0xc3 in position 2: ordinal not in␣
Out[9]: 1 ˓→range(128)
• Exceptions to pass messages between parts of the code:

try/finally
In [17]: def achilles_arrow(x):
In [10]: try: ....: if abs(x - 1) < 1e-3:
....: x = int(raw_input('Please enter a number: ')) ....: raise StopIteration
....: finally: ....: x = 1 - (1-x)/2.
....: print('Thank you for your input') ....: return x
....: ....:
....:
Please enter a number: a In [18]: x = 0
Thank you for your input
--------------------------------------------------------------------------- In [19]: while True:
ValueError: invalid literal for int() with base 10: 'a' ....: try:
....: x = achilles_arrow(x)
Important for resource management (e.g. closing a file) ....: except StopIteration:
....: break
Easier to ask for forgiveness than for permission ....:
....:
In [11]: def print_sorted(collection):
....: try: In [20]: x
....: collection.sort() Out[20]: 0.9990234375
....: except AttributeError:
....: pass # The pass statement does nothing Use exceptions to notify certain conditions are met (e.g. StopIteration) or not (e.g. custom error raising)
....: print(collection)
....:
....: 2.9 Object-oriented programming (OOP)
In [12]: print_sorted([1, 3, 2]) Python supports object-oriented programming (OOP). The goals of OOP are:
[1, 2, 3]
• to organize the code, and
In [13]: print_sorted(set((1, 3, 2)))
set([1, 2, 3]) • to re-use code in similar contexts.
2.8. Exception handling in Python 44 2.9. Object-oriented programming (OOP) 45

Here is a small example: we create a Student class, which is an object gathering several custom functions
(methods) and variables (attributes), we will be able to use:
>>> class Student(object):

... def __init__(self, name):
... self.name = name
3
... def set_age(self, age):
... self.age = age
... def set_major(self, major):
... self.major = major
...
>>> anna = Student('anna')
>>>
>>>
anna.set_age(21)
anna.set_major('physics')
CHAPTER
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 won’t copy the previous class, but
inherit from it:
>>> class MasterStudent(Student):

Python 2 and Python 3
... internship = 'mandatory, from March to June'
...
>>> james = MasterStudent('james')
>>> james.internship
'mandatory, from March to June'
>>> james.set_age(23)
>>> james.age
Author: Pierre de Buyl
23
The MasterStudent class inherited from the Student attributes and methods. Python 2 / 3
Thanks to classes and object-oriented programming, we can organize code with different classes corre-
Two major versions of Python exist, Python 2 and Python 3. Python 3 is the only supported version
sponding to different objects we encounter (an Experiment class, an Image class, a Flow class, etc.), with
since january 2020 but the two versions coexisted for about a decade of transition from
their own methods and attributes. Then we can use inheritance to consider variations around a base
Python 2 to Python 3. The transition has come to and end as most software libraries drop Python
class and re-use code. Ex : from a Flow base class, we can create derived StokesFlow, TurbulentFlow,
2 support.
PotentialFlow, etc.
3.1 A very short summary

• Python 2 is not supported by the Python Software Foundation since January 1st 2020.
There will be no more security patches for Python 2.7. See Sunsetting Python 2 and the Python
3 Q & A.
• The default choice for everyone should be Python 3. Choosing Python 2 should remain
motivated by specific circumstances such as the dependency on unported libraries, with the under-
standing of the lack of official and community support.
• Python 2 and Python 3 share most of their syntax, enabling many programmers to port their
programs. It is even possible to make many codes Python 2/3 compatible, even though there are
limitations. This strategy was important in making the transition but is no longer recommended.
• The division of integers, 1/2 for instance, returns 0 under Python 2 (integer division, preserving
type) and 0.5 under Python 3 (real division, promoting the integer to a floating point value). A
line of code can thus execute with no visible warning in both Python 2 and Python 3
but result in different outcomes.
2.9. Object-oriented programming (OOP) 46 47

• Most scientific libraries have moved to Python 3. NumPy and many scientific software libraries
>>> name = 'SciPy'
dropped Python 2 support or will do so soon, see the Python 3 statement. >>> print(f"Hello, { name} !")
The SciPy Lecture Notes dropped Python 2 support in 2020. The release 2020.1 is almost entirely Python Hello, SciPy!
2 compatible, so you may use it as a reference if necessary. Know that installing suitable packages will
probably be challenging. • In Python 2, range(N) return a list. For large value of N (for a loop iterating many times), this
implies the creation of a large list in memory even though it is not necessary. Python 2 provided
the alternative xrange, that you will find in many scientific programs.
3.2 Breaking changes between Python 2 and Python 3 In Python 3, range() return a dedicated type and does not allocate the memory for the corre-
Python 3 differs from Python 2 in several ways. We list the most relevant ones for scientific users below. sponding list.
>>> type(range(8))
3.2.1 Print function <class 'range'>
>>> range(8)
The most visible change is that print is not a “statement” anymore but a function. range(0, 8)
Whereas in Python 2 you could write
You can transform the output of range into a list if necessary:
>>> print 'hello, world'
hello, world >>> list(range(8))
[0, 1, 2, 3, 4, 5, 6, 7]
in Python 3 you must write
>>> print('hello, world')

hello, world
By making print() a function, one can pass arguments such a file identifier where the output will be
sent.
3.2.2 Division
In Python 2, the division of two integers with a single slash character results in floor-based integer
division:
>>> 1/2
0
In Python 3, the default behavior is to use real-valued division:
>>> 1/2
0.5
Integer division is given by a double slash operator:
>>> 1//2
0
3.3 Some new features in Python 3

Changing print to a function and changing the result of the division operator were only two of the
motivations for Python 3. An incomplete list of the changes follows (there are many more).
• By default, strings are in unicode. Sequence of arbitrary bytes use the type bytes. This change
leads to heavy porting for applications dealing with text.
• Since Python 3.5 and NumPy 1.10, there is a matrix multiplication operator:
>>> np.eye(2) @ np.array([3, 4])

array([3., 4.])
• Since Python 3.6, there is a new string formatting method, the “f-string”:
3.2. Breaking changes between Python 2 and Python 3 48 3.3. Some new features in Python 3 49
• high-level number objects: integers, floating point

• containers: lists (costless insertion and append), dictionaries (fast lookup)
NumPy provides
• extension package to Python for multi-dimensional arrays
4
• closer to hardware (efficiency)
• designed for scientific computation (convenience)
• Also known as array oriented computing
CHAPTER
>>> import numpy as np

>>> a = np.array([0, 1, 2, 3])
>>> a
array([0, 1, 2, 3])
NumPy: creating and manipulating Tip: For example, An array containing:
numerical data
• 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.

Authors: Emmanuelle Gouillart, Didrik Pinte, Gaël Varoquaux, and Pauli Virtanen
In [1]: L = range(1000)
This chapter gives an overview of NumPy, the core tool for performant numerical computing with Python.
In [2]: %timeit [i**2 for i in L]
1000 loops, best of 3: 403 us per loop
4.1 The NumPy array object In [3]: a = np.arange(1000)
In [4]: %timeit a**2

Section contents 100000 loops, best of 3: 12.7 us per loop
• What are NumPy and NumPy arrays?

NumPy Reference documentation
• Creating arrays
• On the web: https://numpy.org/doc/
• Basic data types
• Interactive help:
• Basic visualization
In [5]: np.array?
• Indexing and slicing String Form:<built-in function array>
Docstring:
• Copies and views array(object, dtype=None, copy=True, order=None, subok=False, ndmin=0, ...
• Fancy indexing
• Looking for something:
>>> np.lookfor('create array')
4.1.1 What are NumPy and NumPy arrays? Search results for 'create array'
---------------------------------
NumPy arrays numpy.array
Python objects (continues on next page)
50 4.1. The NumPy array object 51

(continued from previous page) • Create a simple two dimensional array. First, redo the examples from above. And then create
Create an array. your own: how about odd numbers counting backwards on the first row, and even numbers on
numpy.memmap the second?
Create a memory-map to an array stored in a *binary* file on disk.
• Use the functions len(), numpy.shape() on these arrays. How do they relate to each other?
And to the ndim attribute of the arrays?
In [6]: np.con*?
np.concatenate
np.conj
np.conjugate Functions for creating arrays
np.convolve
Tip: In practice, we rarely enter items one by one. . .
Import conventions
The recommended convention to import numpy is: • Evenly spaced:
>>> import numpy as np >>> 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 (exclusive), step
4.1.2 Creating arrays >>> b
Manual construction of arrays array([1, 3, 5, 7])
• 1-D: • or by number of points:

>>> a = np.array([0, 1, 2, 3]) >>> c = np.linspace(0, 1, 6) # start, end, num-points
>>> a >>> c
array([0, 1, 2, 3]) array([0. , 0.2, 0.4, 0.6, 0.8, 1. ])
>>> a.ndim >>> d = np.linspace(0, 1, 5, endpoint=False)
1 >>> d
>>> a.shape array([0. , 0.2, 0.4, 0.6, 0.8])
(4,)
>>> len(a)
• Common arrays:
4
>>> a = np.ones((3, 3)) # reminder: (3, 3) is a tuple
• 2-D, 3-D, . . . : >>> a
array([[1., 1., 1.],
>>> b = np.array([[0, 1, 2], [3, 4, 5]]) # 2 x 3 array [1., 1., 1.],
>>> b [1., 1., 1.]])
array([[0, 1, 2], >>> b = np.zeros((2, 2))
[3, 4, 5]]) >>> b
>>> b.ndim array([[0., 0.],
2 [0., 0.]])
>>> b.shape >>> c = np.eye(3)
(2, 3) >>> c
>>> len(b) # returns the size of the first dimension array([[1., 0., 0.],
2 [0., 1., 0.],
[0., 0., 1.]])
>>> c = np.array([[[1], [2]], [[3], [4]]]) >>> d = np.diag(np.array([1, 2, 3, 4]))
>>> c >>> d
array([[[1], array([[1, 0, 0, 0],
[2]], [0, 2, 0, 0],
[0, 0, 3, 0],
[[3], [0, 0, 0, 4]])
[4]]])
>>> c.shape
• np.random: random numbers (Mersenne Twister PRNG):
(2, 2, 1)
>>> a = np.random.rand(4) # uniform in [0, 1]
>>> a
Exercise: Simple arrays array([ 0.95799151, 0.14222247, 0.08777354, 0.51887998])
>>> b = np.random.randn(4) # Gaussian

4.1. The NumPy array object 52 4.1. The NumPy array object 53
(continued from previous page) Strings

>>> b
>>> f = np.array(['Bonjour', 'Hello', 'Hallo'])
array([ 0.37544699, -0.11425369, -0.47616538, 1.79664113])
>>> f.dtype # <--- strings containing max. 7 letters
dtype('S7')
>>> np.random.seed(1234) # Setting the random seed
Much more
Exercise: Creating arrays using functions • int32
• int64
• Experiment with arange, linspace, ones, zeros, eye and diag.
• uint32
• Create different kinds of arrays with random numbers.
• uint64
• Try setting the seed before creating an array with random values.
• Look at the function np.empty. What does it do? When might this be useful? 4.1.4 Basic visualization
Now that we have our first data arrays, we are going to visualize them.
4.1.3 Basic data types Start by launching IPython:
You may have noticed that, in some instances, array elements are displayed with a trailing dot (e.g. 2. $ ipython # or ipython3 depending on your install
vs 2). This is due to a difference in the data-type used:
>>> a = np.array([1, 2, 3]) Or the notebook:
>>> a.dtype
$ jupyter notebook
dtype('int64')
>>> b = np.array([1., 2., 3.]) Once IPython has started, enable interactive plots:
>>> b.dtype
>>> %matplotlib
dtype('float64')
Or, from the notebook, enable plots in the notebook:

Tip: Different data-types allow us to store data more compactly in memory, but most of the time we >>> %matplotlib inline
simply work with floating point numbers. Note that, in the example above, NumPy auto-detects the
data-type from the input. The inline is important for the notebook, so that plots are displayed in the notebook and not in a new
window.
Matplotlib is a 2D plotting package. We can import its functions as below:
You can explicitly specify which data-type you want:
>>> import matplotlib.pyplot as plt # the tidy way
>>> c = np.array([1, 2, 3], dtype=float)
>>> c.dtype And then use (note that you have to use show explicitly if you have not enabled interactive plots with
dtype('float64') %matplotlib):
The default data type is floating point: >>> plt.plot(x, y) # line plot
>>> plt.show() # <-- shows the plot (not needed with interactive plots)
>>> a = np.ones((3, 3))
>>> a.dtype Or, if you have enabled interactive plots with %matplotlib:
dtype('float64')
>>> plt.plot(x, y) # line plot
There are also other types:
• 1D plotting:
Complex
>>> x = np.linspace(0, 3, 20)
>>> d = np.array([1+2j, 3+4j, 5+6*1j])
>>> y = np.linspace(0, 9, 20)
>>> d.dtype
>>> plt.plot(x, y) # line plot
dtype('complex128')
[<matplotlib.lines.Line2D object at ...>]
>>> plt.plot(x, y, 'o') # dot plot
Bool [<matplotlib.lines.Line2D object at ...>]
>>> e = np.array([True, False, False, True])
>>> e.dtype
dtype('bool')
Warning: Indices begin at 0, like other Python sequences (and C/C++). In contrast, in Fortran
or Matlab, indices begin at 1.
The usual python idiom for reversing a sequence is supported:
>>> a[::-1]
array([9, 8, 7, 6, 5, 4, 3, 2, 1, 0])
For multidimensional arrays, indices are tuples of integers:
>>> a = np.diag(np.arange(3))
• 2D arrays (such as images): >>> a
array([[0, 0, 0],
>>> image = np.random.rand(30, 30) [0, 1, 0],
>>> plt.imshow(image, cmap=plt.cm.hot) [0, 0, 2]])
<matplotlib.image.AxesImage object at ...> >>> a[1, 1]
>>> plt.colorbar() 1
<matplotlib.colorbar.Colorbar object at ...> >>> 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:
• 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 = np.arange(10)
See also: >>> a
array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9])
More in the: matplotlib chapter >>> a[2:9:3] # [start:end:step]
array([2, 5, 8])
Exercise: Simple visualizations

Note that the last index is not included! :
• Plot some simple arrays: a cosine as a function of time and a 2D matrix. >>> a[:4]
array([0, 1, 2, 3])
• Try using the gray colormap on the 2D matrix.
All three slice components are not required: by default, start is 0, end is the last and step is 1:
4.1.5 Indexing and slicing >>> a[1:3]

array([1, 2])
The items of an array can be accessed and assigned to the same way as other Python sequences (e.g. >>> a[::2]
lists): array([0, 2, 4, 6, 8])
>>> a[3:]
>>> a = np.arange(10) array([3, 4, 5, 6, 7, 8, 9])
>>> a
array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9])
>>> a[0], a[2], a[-1] A small illustrated summary of NumPy indexing and slicing. . .
(0, 2, 9)
>>> a[0, 3:5] Par on course: 3 statements for each

array([3, 4]) 0 1 2 3 4 5 Hint: Individual array elements can be accessed similarly to a list, e.g. a[1] or a[1, 2].
>>> a[4:, 4:] Hint: Examine the docstring for diag.

10 11 12 13 14 15
array([[44, 55],
[54, 55]]) 20 21 22 23 24 25 Exercise: Tiling for array creation
>>> a[:, 2]
30 31 32 33 34 35 Skim through the documentation for np.tile, and use this function to construct the array:
a([2, 12, 22, 32, 42, 52])
[[4, 3, 4, 3, 4, 3],
40 41 42 43 44 45 [2, 1, 2, 1, 2, 1],
>>> a[2::2, ::2]
[4, 3, 4, 3, 4, 3],
array([[20, 22, 24], [2, 1, 2, 1, 2, 1]]
50 51 52 53 54 55
[40, 42, 44]])
You can also combine assignment and slicing:

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

True, True, False, True, False, False])
>>> 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 = np.arange(0, 100, 10)

>>> a
array([ 0, 10, 20, 30, 40, 50, 60, 70, 80, 90])
Compute prime numbers in 0–99, with a sieve

Indexing can be done with an array of integers, where the same index is repeated several time:
• Construct a shape (100,) boolean array is_prime, filled with True in the beginning:
>>> a[[2, 3, 2, 4, 2]] # note: [2, 3, 2, 4, 2] is a Python list
>>> is_prime = np.ones((100,), dtype=bool) array([20, 30, 20, 40, 20])
• Cross out 0 and 1 which are not primes:

New values can be assigned with this kind of indexing:
>>> is_prime[:2] = 0
>>> a[[9, 7]] = -100
• For each integer j starting from 2, cross out its higher multiples: >>> a
array([ 0, 10, 20, 30, 40, 50, 60, -100, 80, -100])
>>> N_max = int(np.sqrt(len(is_prime) - 1))
>>> for j in range(2, N_max + 1):
... is_prime[2*j::j] = False
Tip: When a new array is created by indexing with an array of integers, the new array has the same
• Skim through help(np.nonzero), and print the prime numbers shape as the array of integers:
• Follow-up: >>> a = np.arange(10)
>>> idx = np.array([[3, 4], [9, 7]])
– Move the above code into a script file named prime_sieve.py >>> idx.shape
– Run it to check it works (2, 2)
>>> a[idx]
– Use the optimization suggested in the sieve of Eratosthenes: array([[3, 4],
[9, 7]])
1. Skip j which are already known to not be primes
2. The first number to cross out is 𝑗 2
The image below illustrates various fancy indexing applications

4.1.7 Fancy indexing
>>> a[(0,1,2,3,4), (1,2,3,4,5)]
Tip: NumPy arrays can be indexed with slices, but also with boolean or integer arrays (masks). This 0 1 2 3 4 5
array([1, 12, 23, 34, 45])
method is called fancy indexing. It creates copies not views.
>>> a[3:, [0,2,5]] 10 11 12 13 14 15
array([[30, 32, 35],
Using boolean masks 20 21 22 23 24 25
[40, 42, 45],
>>> np.random.seed(3) [50, 52, 55]])
>>> a = np.random.randint(0, 21, 15)
30 31 32 33 34 35
>>> a >>> mask = np.array([1,0,1,0,0,1], dtype=bool)
array([10, 3, 8, 0, 19, 10, 11, 9, 10, 6, 0, 20, 12, 7, 14]) >>> a[mask, 2] 40 41 42 43 44 45
>>> (a % 3 == 0)
array([False, True, False, True, False, False, False, True, False, array([2, 22, 52])
50 51 52 53 54 55
Exercise: Fancy indexing >>> c = np.ones((3, 3))

>>> c * c # NOT matrix multiplication!
• Again, reproduce the fancy indexing shown in the diagram above. array([[1., 1., 1.],
[1., 1., 1.],
• Use fancy indexing on the left and array creation on the right to assign values into an array, for [1., 1., 1.]])
instance by setting parts of the array in the diagram above to zero.
Note: Matrix multiplication:

4.2 Numerical operations on arrays
>>> c.dot(c)
array([[3., 3., 3.],
Section contents [3., 3., 3.],
[3., 3., 3.]])
• Elementwise operations
• Basic reductions
• Broadcasting Exercise: Elementwise operations
• Array shape manipulation
• Try simple arithmetic elementwise operations: add even elements with odd elements
• Sorting data
• Time them against their pure python counterparts using %timeit.
• Summary
• Generate:
– [2**0, 2**1, 2**2, 2**3, 2**4]
4.2.1 Elementwise operations – a_j = 2^(3*j) - j
Basic operations
With scalars: Other operations
>>> a = np.array([1, 2, 3, 4]) Comparisons:
>>> a + 1
array([2, 3, 4, 5]) >>> a = np.array([1, 2, 3, 4])
>>> 2**a >>> b = np.array([4, 2, 2, 4])
array([ 2, 4, 8, 16]) >>> a == b
array([False, True, False, True])
All arithmetic operates elementwise: >>> a > b
array([False, False, True, False])
>>> b = np.ones(4) + 1
>>> a - b
array([-1., 0., 1., 2.]) Tip: Array-wise comparisons:
>>> a * b
array([2., 4., 6., 8.]) >>> a = np.array([1, 2, 3, 4])
>>> b = np.array([4, 2, 2, 4])
>>> j = np.arange(5) >>> c = np.array([1, 2, 3, 4])
>>> 2**(j + 1) - j >>> np.array_equal(a, b)
array([ 2, 3, 6, 13, 28]) False
>>> np.array_equal(a, c)
These operations are of course much faster than if you did them in pure python: True
>>> a = np.arange(10000)
>>> %timeit a + 1 Logical operations:
10000 loops, best of 3: 24.3 us per loop
>>> l = range(10000) >>> a = np.array([1, 1, 0, 0], dtype=bool)
>>> %timeit [i+1 for i in l] >>> b = np.array([1, 0, 1, 0], dtype=bool)
1000 loops, best of 3: 861 us per loop >>> np.logical_or(a, b)
array([ True, True, True, False])
>>> np.logical_and(a, b)
array([ True, False, False, False])
Warning: Array multiplication is not matrix multiplication:
Transcendental functions:
4.2. Numerical operations on arrays 62 4.2. Numerical operations on arrays 63

4.2.2 Basic reductions
>>> np.sin(a) Computing sums
array([ 0. , 0.84147098, 0.90929743, 0.14112001, -0.7568025 ])
>>> np.log(a) >>> x = np.array([1, 2, 3, 4])
array([ -inf, 0. , 0.69314718, 1.09861229, 1.38629436]) >>> np.sum(x)
>>> np.exp(a) 10
array([ 1. , 2.71828183, 7.3890561 , 20.08553692, 54.59815003]) >>> x.sum()
10
Shape mismatches
>>> a + np.array([1, 2])
ValueError: operands could not be broadcast together with shapes (4) (2)
Broadcasting? We’ll return to that later.

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.], Sum by rows and by columns:
[1., 1., 0.]])
>>> x = np.array([[1, 1], [2, 2]])
>>> x
array([[1, 1],
Note: The transposition is a view
[2, 2]])
The transpose returns a view of the original array: >>> x.sum(axis=0) # columns (first dimension)
array([3, 3])
>>> a = np.arange(9).reshape(3, 3) >>> x[:, 0].sum(), x[:, 1].sum()
>>> a.T[0, 2] = 999 (3, 3)
>>> a.T >>> x.sum(axis=1) # rows (second dimension)
array([[ 0, 3, 999], array([2, 4])
[ 1, 4, 7], >>> x[0, :].sum(), x[1, :].sum()
[ 2, 5, 8]]) (2, 4)
>>> a
array([[ 0, 1, 2],
[ 3, 4, 5], Tip: Same idea in higher dimensions:
[999, 7, 8]])
>>> x = np.random.rand(2, 2, 2)
>>> x.sum(axis=2)[0, 1]
1.14764...
Note: Linear algebra >>> x[0, 1, :].sum()
1.14764...
The sub-module numpy.linalg implements basic linear algebra, such as solving linear systems, singular
value decomposition, etc. However, it is not guaranteed to be compiled using efficient routines, and thus
we recommend the use of scipy.linalg, as detailed in section Linear algebra operations: scipy.linalg
Other reductions
— works the same way (and take axis=)
Exercise other operations
Extrema:
• Look at the help for np.allclose. When might this be useful? >>> x = np.array([1, 3, 2])
>>> x.min()
• Look at the help for np.triu and np.tril.
1
>>> x.max()


Tip: Let us consider a simple 1D random walk process: at each time step a walker jumps right or
3
left with equal probability.
>>> x.argmin() # index of minimum
0 We are interested in finding the typical distance from the origin of a random walker after t left or
>>> x.argmax() # index of maximum right jumps? We are going to simulate many “walkers” to find this law, and we are going to do so
1
using array computing tricks: we are going to create a 2D array with the “stories” (each walker has a
story) in one direction, and the time in the other:
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() >>> n_stories = 1000 # number of walkers
True >>> t_max = 200 # time during which we follow the walker
We randomly choose all the steps 1 or -1 of the walk:

Statistics:
>>> x = np.array([1, 2, 3, 1]) >>> t = np.arange(t_max)

>>> y = np.array([[1, 2, 3], [5, 6, 1]]) >>> steps = 2 * np.random.randint(0, 1 + 1, (n_stories, t_max)) - 1 # +1 because the high␣
˓→value is exclusive
>>> x.mean()
1.75 >>> np.unique(steps) # Verification: all steps are 1 or -1
>>> np.median(x) array([-1, 1])
1.5
>>> np.median(y, axis=-1) # last axis
array([2., 5.]) We build the walks by summing steps along the time:
>>> x.std() # full population standard dev. >>> positions = np.cumsum(steps, axis=1) # axis = 1: dimension of time
0.82915619758884995 >>> sq_distance = positions**2
. . . and many more (best to learn as you go).

We get the mean in the axis of the stories:
Exercise: Reductions >>> mean_sq_distance = np.mean(sq_distance, axis=0)
• Given there is a sum, what other function might you expect to see?
Plot the results:
• What is the difference between sum and cumsum?
>>> plt.figure(figsize=(4, 3))
<Figure size ... with 0 Axes>
>>> plt.plot(t, np.sqrt(mean_sq_distance), 'g.', t, np.sqrt(t), 'y-')
Worked Example: diffusion using a random walk algorithm [<matplotlib.lines.Line2D object at ...>, <matplotlib.lines.Line2D object at ...>]
>>> plt.xlabel(r"$t$")
Text(...'$t$')
>>> plt.ylabel(r"$\sqrt{\langle (\delta x)^2 \rangle}$")
Text(...'$\\sqrt{\\langle (\\delta x)^2 \\rangle}$')
>>> plt.tight_layout() # provide sufficient space for labels

>>> 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]])
We have already used broadcasting without knowing it!:

We find a well-known result in physics: the RMS
distance grows as the square root of the time! >>> 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.],
4.2.3 Broadcasting [1., 1., 1., 1., 1.],
[1., 1., 1., 1., 1.],
• Basic operations on numpy arrays (addition, etc.) are elementwise
[1., 1., 1., 1., 1.]])
• This works on arrays of the same size.
A useful trick:
Nevertheless, It’s also possible to do operations on arrays of different
sizes if NumPy can transform these arrays so that they all have >>> a = np.arange(0, 40, 10)
the same size: this conversion is called broadcasting. >>> a.shape
(4,)
The image below gives an example of broadcasting: >>> 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]])
Tip: 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.
Worked Example: Broadcasting
Let’s 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],
Let’s verify:
[ 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],
4.2. Numerical operations on arrays 68 [1913, 1715,
4.2. Numerical 1610, on
operations 1177, 1042, 738, 438, 369,
arrays 0, 535], 69
[2448, 2250, 2145, 1712, 1577, 1273, 973, 904, 535, 0]])

>>> x.shape, y.shape
((5, 1), (1, 5))
>>> distance = np.sqrt(x ** 2 + y ** 2)
Tip: So, np.ogrid is very useful as soon as we have to handle computations on a grid. On the other
hand, np.mgrid directly provides matrices full of indices for cases where we can’t (or don’t 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]])
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 5x5 grid, we can do
>>> x, y = np.arange(5), np.arange(5)[:, np.newaxis] See also:
>>> distance = np.sqrt(x ** 2 + y ** 2)
Broadcasting: discussion of broadcasting in the Advanced NumPy chapter.
>>> distance
array([[0. , 1. , 2. , 3. , 4. ],
[1. , 1.41421356, 2.23606798, 3.16227766, 4.12310563], 4.2.4 Array shape manipulation
[2. , 2.23606798, 2.82842712, 3.60555128, 4.47213595],
[3. , 3.16227766, 3.60555128, 4.24264069, 5. ], Flattening
[4. , 4.12310563, 4.47213595, 5. , 5.65685425]])
>>> a = np.array([[1, 2, 3], [4, 5, 6]])
>>> a.ravel()
Or in color:
array([1, 2, 3, 4, 5, 6])
>>> plt.pcolor(distance) >>> a.T
>>> plt.colorbar() 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 = b.reshape((2, 3))
>>> b
Remark : the numpy.ogrid() function allows to array([[1, 2, 3],
directly create vectors x and y of the previous example, with two “significant dimensions”: [4, 5, 6]])
>>> x, y = np.ogrid[0:5, 0:5]

>>> x, y
Or,
(array([[0], >>> a.reshape((2, -1)) # unspecified (-1) value is inferred
[1], array([[1, 2, 3],
[2], [4, 5, 6]])
[3],
[4]]), array([[0, 1, 2, 3, 4]]))

Resizing
Warning: ndarray.reshape may return a view (cf help(np.reshape))), or copy
Size of an array can be changed with ndarray.resize:
Tip: >>> a.resize((8,))
>>> a
>>> b[0, 0] = 99 array([0, 1, 2, 3, 0, 0, 0, 0])
>>> a
array([[99, 2, 3], However, it must not be referred to somewhere else:
[ 4, 5, 6]])
>>> b = a
Beware: reshape may also return a copy!: >>> a.resize((4,))
>>> a = np.zeros((3, 2)) File "<stdin>", line 1, in <module>
>>> b = a.T.reshape(3*2) ValueError: cannot resize an array that has been referenced or is
>>> b[0] = 9 referencing another array in this way. Use the resize function
>>> a
array([[0., 0.],
[0., 0.], Exercise: Shape manipulations
[0., 0.]])
• Look at the docstring for reshape, especially the notes section which has some more information
To understand this you need to learn more about the memory layout of a numpy array. about copies and views.
• Use flatten as an alternative to ravel. What is the difference? (Hint: check which one returns
a view and which a copy)
Adding a dimension
• Experiment with transpose for dimension shuffling.
Indexing with the np.newaxis object allows us to add an axis to an array (you have seen this already
above in the broadcasting section):
>>> z = np.array([1, 2, 3]) 4.2.5 Sorting data

>>> z
array([1, 2, 3]) Sorting along an axis:
>>> z[:, np.newaxis] >>> a = np.array([[4, 3, 5], [1, 2, 1]])

array([[1], >>> b = np.sort(a, axis=1)
[2], >>> b
[3]]) array([[3, 4, 5],
[1, 1, 2]])
>>> z[np.newaxis, :]
array([[1, 2, 3]])
Note: Sorts each row separately!
Dimension shuffling
In-place sort:
>>> a = np.arange(4*3*2).reshape(4, 3, 2)
>>> a.shape >>> a.sort(axis=1)
(4, 3, 2) >>> a
>>> a[0, 2, 1] array([[3, 4, 5],
5 [1, 1, 2]])
>>> b = a.transpose(1, 2, 0)
>>> b.shape Sorting with fancy indexing:
(3, 2, 4)
>>> b[2, 1, 0] >>> a = np.array([4, 3, 1, 2])
5 >>> j = np.argsort(a)
>>> j
array([2, 3, 1, 0])
Also creates a view:
>>> a[j]
>>> b[2, 1, 0] = -1 array([1, 2, 3, 4])
>>> a[0, 2, 1]
-1 Finding minima and maxima:

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

4.3.1 More data types
>>> j_max = np.argmax(a) Casting
>>> j_min = np.argmin(a)
>>> j_max, j_min “Bigger” type wins in mixed-type operations:
(0, 2)
>>> np.array([1, 2, 3]) + 1.5
array([2.5, 3.5, 4.5])
Exercise: Sorting
Assignment never changes the type!
• Try both in-place and out-of-place sorting. >>> a = np.array([1, 2, 3])
• Try creating arrays with different dtypes and sorting them. >>> a.dtype
dtype('int64')
• Use all or array_equal to check the results. >>> a[0] = 1.9 # <-- float is truncated to integer
>>> a
• Look at np.random.shuffle for a way to create sortable input quicker.
array([1, 2, 3])
• Combine ravel, sort and reshape.
Forced casts:
• Look at the axis keyword for sort and rewrite the previous exercise.
>>> a = np.array([1.7, 1.2, 1.6])
>>> b = a.astype(int) # <-- truncates to integer
4.2.6 Summary >>> b
array([1, 1, 1])
What do you need to know to get started?
• Know how to create arrays : array, arange, ones, zeros. Rounding:
• Know the shape of the array with array.shape, then use slicing to obtain different views of the >>> a = np.array([1.2, 1.5, 1.6, 2.5, 3.5, 4.5])
array: array[::2], etc. Adjust the shape of the array using reshape or flatten it with ravel. >>> b = np.around(a)
>>> b # still floating-point
• Obtain a subset of the elements of an array and/or modify their values with masks array([1., 2., 2., 2., 4., 4.])
>>> c = np.around(a).astype(int)
>>> a[a < 0] = 0 >>> c
array([1, 2, 2, 2, 4, 4])
• 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())!! Different data type sizes
• For advanced use: master the indexing with arrays of integers, as well as broadcasting. Know more Integers (signed):
NumPy functions to handle various array operations.
int8 8 bits
int16 16 bits
Quick read
int32 32 bits (same as int on 32-bit platform)
If you want to do a first quick pass through the Scipy lectures to learn the ecosystem, you can directly int64 64 bits (same as int on 64-bit platform)
skip to the next chapter: Matplotlib: plotting.
The remainder of this chapter is not necessary to follow the rest of the intro part. But be sure to >>> np.array([1], dtype=int).dtype
dtype('int64')
come back and finish this chapter, as well as to do some more exercices.
>>> np.iinfo(np.int32).max, 2**31 - 1
(2147483647, 2147483647)
4.3 More elaborate arrays Unsigned integers:
uint8 8 bits
Section contents uint16 16 bits
uint32 32 bits
• More data types uint64 64 bits
• Structured data types
• maskedarray: dealing with (propagation of) missing data >>> np.iinfo(np.uint32).max, 2**32 - 1
(4294967295, 4294967295)
4.3. More elaborate arrays 74 4.3. More elaborate arrays 75

Long integers 4.3.2 Structured data types
Python 2 has a specific type for ‘long’ integers, that cannot overflow, represented with an ‘L’ at the sensor_code (4-character string)
end. In Python 3, all integers are long, and thus cannot overflow. position (float)
>>> np.iinfo(np.int64).max, 2**63 - 1 value (float)
(9223372036854775807, 9223372036854775807L)
>>> samples = np.zeros((6,), dtype=[('sensor_code', 'S4'),
... ('position', float), ('value', float)])
Floating-point numbers: >>> samples.ndim
1
float16 16 bits >>> samples.shape
float32 32 bits (6,)
float64 64 bits (same as float) >>> samples.dtype.names
('sensor_code', 'position', 'value')
float96 96 bits, platform-dependent (same as np.longdouble)
float128 128 bits, platform-dependent (same as np.longdouble) >>> 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)]
>>> np.finfo(np.float32).eps >>> samples
1.1920929e-07 array([('ALFA', 1.0, 0.37), ('BETA', 1.0, 0.11), ('TAU', 1.0, 0.13),
>>> np.finfo(np.float64).eps ('ALFA', 1.5, 0.37), ('ALFA', 3.0, 0.11), ('TAU', 1.2, 0.13)],
2.2204460492503131e-16 dtype=[('sensor_code', 'S4'), ('position', '<f8'), ('value', '<f8')])
>>> np.float32(1e-8) + np.float32(1) == 1 Field access works by indexing with field names:
True
>>> np.float64(1e-8) + np.float64(1) == 1 >>> samples['sensor_code']
False array(['ALFA', 'BETA', 'TAU', 'ALFA', 'ALFA', 'TAU'],
dtype='|S4')
>>> samples['value']
Complex floating-point numbers:
array([0.37, 0.11, 0.13, 0.37, 0.11, 0.13])
>>> samples[0]
complex64 two 32-bit floats ('ALFA', 1.0, 0.37)
complex128 two 64-bit floats
complex192 two 96-bit floats, platform-dependent >>> samples[0]['sensor_code'] = 'TAU'
complex256 two 128-bit floats, platform-dependent >>> samples[0]
('TAU', 1.0, 0.37)
Smaller data types Multiple fields at once:
>>> samples[['position', 'value']]

If you don’t know you need special data types, then you probably don’t.
array([(1. , 0.37), (1. , 0.11), (1. , 0.13), (1.5, 0.37),
Comparison on using float32 instead of float64: (3. , 0.11), (1.2, 0.13)],
dtype=[('position', '<f8'), ('value', '<f8')])
• Half the size in memory and on disk
• Half the memory bandwidth required (may be a bit faster in some operations) Fancy indexing works, as usual:
In [1]: a = np.zeros((int(1e6),), dtype=np.float64) >>> samples[samples['sensor_code'] == b'ALFA']
array([(b'ALFA', 1.5, 0.37), (b'ALFA', 3. , 0.11)],
In [2]: b = np.zeros((int(1e6),), dtype=np.float32) dtype=[('sensor_code', 'S4'), ('position', '<f8'), ('value', '<f8')])
In [3]: %timeit a*a

1000 loops, best of 3: 1.78 ms per loop Note: There are a bunch of other syntaxes for constructing structured arrays, see here and here.
In [4]: %timeit b*b
1000 loops, best of 3: 1.07 ms per loop
4.3.3 maskedarray: dealing with (propagation of) missing data
• But: bigger rounding errors — sometimes in surprising places (i.e., don’t use them unless you
really need them) • For floats one could use NaN’s, 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, --],
4.3. More elaborate arrays 76 4.3. More elaborate arrays 77


>>> x = np.linspace(0, 1, 20)
mask=[False, True, False, True], >>> y = np.cos(x) + 0.3*np.random.rand(20)
fill_value=999999) >>> p = np.poly1d(np.polyfit(x, y, 3))
>>> t = np.linspace(0, 1, 200) # use a larger number of points for smoother plotting
>>> y = np.ma.array([1, 2, 3, 4], mask=[0, 1, 1, 1]) >>> plt.plot(x, y, 'o', t, p(t), '-')
>>> x + y [<matplotlib.lines.Line2D object at ...>, <matplotlib.lines.Line2D object at ...>]
masked_array(data=[2, --, --, --],
mask=[False, True, True, True],
fill_value=999999)
• Masking versions of common functions:

>>> np.ma.sqrt([1, -1, 2, -2])
masked_array(data=[1.0, --, 1.41421356237... --],
mask=[False, True, False, True],
fill_value=1e+20)
Note: There are other useful array siblings
While it is off topic in a chapter on numpy, let’s take a moment to recall good coding practice, which See http://numpy.org/doc/stable/reference/
really do pay off in the long run: routines.polynomials.poly1d.html for more.
More polynomials (with more bases)

Good practices
NumPy also has a more sophisticated polynomial interface, which supports e.g. the Chebyshev basis.
• Explicit variable names (no need of a comment to explain what is in the variable) 3𝑥2 + 2𝑥 − 1:
• Style: spaces after commas, around =, etc.
>>> p = np.polynomial.Polynomial([-1, 2, 3]) # coefs in different order!
A certain number of rules for writing “beautiful” code (and, more importantly, using the same >>> p(0)
conventions as everybody else!) are given in the Style Guide for Python Code and the Docstring -1.0
Conventions page (to manage help strings). >>> p.roots()
array([-1. , 0.33333333])
• Except some rare cases, variable names and comments in English. >>> p.degree() # In general polynomials do not always expose 'order'
2
Example using polynomials in Chebyshev basis, for polynomials in range [-1, 1]:
4.4 Advanced operations
>>> x = np.linspace(-1, 1, 2000)
>>> y = np.cos(x) + 0.3*np.random.rand(2000)
Section contents >>> p = np.polynomial.Chebyshev.fit(x, y, 90)
• Polynomials >>> plt.plot(x, y, 'r.')

• Loading data files >>> plt.plot(x, p(x), 'k-', lw=3)
4.4.1 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
4.4. Advanced operations 78 4.4. Advanced operations 79

>>> img = plt.imread('data/elephant.png')

>>> img.shape, img.dtype
((200, 300, 3), dtype('float32'))
>>> plt.imshow(img)
<matplotlib.image.AxesImage object at ...>
>>> plt.savefig('plot.png')
>>> plt.imsave('red_elephant.png', img[:,:,0], cmap=plt.cm.gray)
The Chebyshev polynomials have some advantages

in interpolation.
4.4.2 Loading data files

Text files
Example: populations.txt:
# year hare lynx carrot
1900 30e3 4e3 48300 This saved only one channel (of RGB):
1901 47.2e3 6.1e3 48200
>>> plt.imshow(plt.imread('red_elephant.png'))
1902 70.2e3 9.8e3 41500 <matplotlib.image.AxesImage object at ...>
1903 77.4e3 35.2e3 38200
>>> data = np.loadtxt('data/populations.txt')

>>> 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 Other libraries:
• Using Python’s I/O functions and e.g. regexps for parsing (Python is quite well suited for this) >>> import imageio
>>> imageio.imsave('tiny_elephant.png', img[::6,::6])
>>> plt.imshow(plt.imread('tiny_elephant.png'), interpolation='nearest')
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:
4.4. Advanced operations 80 4.4. Advanced operations 81

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.)
4.5.2 Picture manipulation: Framing a Face

NumPy’s own format
Let’s do some manipulations on numpy arrays by starting with an image of a racoon. scipy provides a
NumPy has its own binary format, not portable but with efficient I/O:
2D array of this image with the scipy.misc.face function:
>>> data = np.ones((3, 3))
>>> from scipy import misc
>>> np.save('pop.npy', data)
>>> face = misc.face(gray=True) # 2D grayscale image
>>> data3 = np.load('pop.npy')
Here are a few images we will be able to obtain with our manipulations: use different colormaps, crop
Well-known (& more obscure) file formats the image, change some parts of the image.
• HDF5: h5py, PyTables
• NetCDF: scipy.io.netcdf_file, netcdf4-python, . . .
• Matlab: scipy.io.loadmat, scipy.io.savemat
• MatrixMarket: scipy.io.mmread, scipy.io.mmwrite
• IDL: scipy.io.readsav
. . . if somebody uses it, there’s probably also a Python library for it. • Let’s use the imshow function of matplotlib to display the image.
>>> import matplotlib.pyplot as plt

Exercise: Text data files
>>> face = misc.face(gray=True)
>>> plt.imshow(face)
Write a Python script that loads data from populations.txt:: and drop the last column and the first <matplotlib.image.AxesImage object at 0x...>
5 rows. Save the smaller dataset to pop2.txt.
• The face is displayed in false colors. A colormap must be specified for it to be displayed
in grey.
NumPy internals
>>> plt.imshow(face, cmap=plt.cm.gray)
<matplotlib.image.AxesImage object at 0x...>
If you are interested in the NumPy internals, there is a good discussion in Advanced NumPy.
• Create an array of the image with a narrower centering [for example,] remove 100 pixels
from all the borders of the image. To check the result, display this new array with imshow.
4.5 Some exercises >>> crop_face = face[100:-100, 100:-100]
4.5.1 Array manipulations • We will now frame the face with a black locket. For this, we need to create a mask cor-
1. Form the 2-D array (without typing it in explicitly): responding to the pixels we want to be black. The center of the face is around (660, 330), so
we defined the mask by this condition (y-300)**2 + (x-660)**2
[[1, 6, 11],
[2, 7, 12], >>> sy, sx = face.shape
[3, 8, 13], >>> y, x = np.ogrid[0:sy, 0:sx] # x and y indices of pixels
[4, 9, 14], >>> y.shape, x.shape
[5, 10, 15]] ((768, 1), (1, 1024))
4.5. Some exercises 82 4.5. Some exercises 83

(continued from previous page) . . . all without for-loops.

>>> centerx, centery = (660, 300) # center of the image Solution: Python source file
>>> 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 4.5.4 Crude integral approximations
is extremely simple and intuitive: Write a function f(a, b, c) that returns 𝑎𝑏 −𝑐. Form a 24x12x6 array containing its values in parameter
>>> face[mask] = 0 ranges [0,1] x [0,1] x [0,1].
>>> plt.imshow(face) Approximate the 3-d integral
∫︁ 1 ∫︁ 1 ∫︁ 1
• Follow-up: copy all instructions of this exercise in a script called face_locket.py then (𝑎𝑏 − 𝑐)𝑑𝑎 𝑑𝑏 𝑑𝑐
0 0 0
execute this script in IPython with %run face_locket.py.
over this volume with the mean. The exact result is: ln 2 − 12 ≈ 0.1931 . . . — what is your relative error?
Change the circle to an ellipsoid.
(Hints: use elementwise operations and broadcasting. You can make np.ogrid give a number of points
4.5.3 Data statistics in given range with np.ogrid[0:1:20j].)
Reminder Python functions:
The data in populations.txt describes the populations of hares and lynxes (and carrots) in northern
Canada during 20 years: def f(a, b, c):
return some_result
>>> data = np.loadtxt('data/populations.txt')
>>> year, hares, lynxes, carrots = data.T # trick: columns to variables
Solution: Python source file
>>> plt.axes([0.2, 0.1, 0.5, 0.8]) 4.5.5 Mandelbrot set
<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 ...>
Write a script that computes the Mandelbrot frac-

tal. The Mandelbrot iteration:
N_max = 50
some_threshold = 50
Computes and print, based on the data in
populations.txt. . . c = x + 1j*y
1. The mean and std of the populations of each species for the years in the period. z = 0
for j in range(N_max):
2. Which year each species had the largest population.
z = z**2 + c
3. Which species has the largest population for each year. (Hint: argsort & fancy indexing of
np.array(['H', 'L', 'C'])) Point (x, y) belongs to the Mandelbrot set if |𝑧| < some_threshold.
4. Which years any of the populations is above 50000. (Hint: comparisons and np.any) Do this computation by:
5. The top 2 years for each species when they had the lowest populations. (Hint: argsort, fancy 1. Construct a grid of c = x + 1j*y values in range [-2, 1] x [-1.5, 1.5]
indexing)
2. Do the iteration
6. Compare (plot) the change in hare population (see help(np.gradient)) and the number of lynxes.
3. Form the 2-d boolean mask indicating which points are in the set
Check correlation (see help(np.corrcoef)).
4. Save the result to an image with:
4.5. Some exercises 84 4.5. Some exercises 85

2D plotting
>>> plt.imshow(mask.T, extent=[-2, 1, -1.5, 1.5]) Plot a basic 2D figure
>>> plt.gray()
>>> plt.savefig('mandelbrot.png')
4.5.6 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 import numpy as np
import matplotlib.pyplot as plt
Write a script that works with 5 states, and:
• Constructs a random matrix, and normalizes each row so that it is a transition matrix. image = np.random.rand(30, 30)
plt.imshow(image, cmap=plt.cm.hot)
• Starts from a random (normalized) probability distribution p and takes 50 steps => p_50 plt.colorbar()
plt.show()
• Computes the stationary distribution: the eigenvector of P.T with eigenvalue 1 (numerically: closest
to 1) => p_stationary
Total running time of the script: ( 0 minutes 0.025 seconds)
Remember to normalize the eigenvector — I didn’t. . .
• Checks if p_50 and p_stationary are equal to tolerance 1e-5 Note: Click here to download the full example code
Toolbox: np.random.rand, .dot(), np.linalg.eig, reductions, abs(), argmin, comparisons, all, np.
linalg.norm, etc.
1D plotting
Plot a basic 1D figure
4.6 Full code examples

4.6.1 Full code examples for the numpy chapter
Note: Click here to download the full example code
4.6. Full code examples 86 4.6. Full code examples 87

import numpy as np import numpy as np

import matplotlib.pyplot as plt import matplotlib.pyplot as plt
x = np.linspace(0, 3, 20) x, y = np.arange(5), np.arange(5)[:, np.newaxis]

y = np.linspace(0, 9, 20) distance = np.sqrt(x ** 2 + y ** 2)
plt.plot(x, y) plt.pcolor(distance)
plt.plot(x, y, 'o') plt.colorbar()
plt.show() plt.show()
Total running time of the script: ( 0 minutes 0.013 seconds) Total running time of the script: ( 0 minutes 0.084 seconds)
Note: Click here to download the full example code Note: Click here to download the full example code
Distances exercise Fitting to polynomial

Plot distances in a grid Plot noisy data and their polynomial fit


np.random.seed(12) np.random.seed(0)
x = np.linspace(0, 1, 20) x = np.linspace(-1, 1, 2000)

y = np.cos(x) + 0.3*np.random.rand(20) y = np.cos(x) + 0.3*np.random.rand(2000)
p = np.poly1d(np.polyfit(x, y, 3)) p = np.polynomial.Chebyshev.fit(x, y, 90)
t = np.linspace(0, 1, 200) plt.plot(x, y, 'r.')

plt.plot(x, y, 'o', t, p(t), '-') plt.plot(x, p(x), 'k-', lw=3)
Fitting in Chebyshev basis Population exercise

Plot noisy data and their polynomial fit in a Chebyshev basis Plot populations of hares, lynxes, and carrots

import numpy as np
red channel displayed in grey
data = np.loadtxt('../../../data/populations.txt') plt.figure()

year, hares, lynxes, carrots = data.T img_red = img[:, :, 0]
plt.imshow(img_red, cmap=plt.cm.gray)
plt.axes([0.2, 0.1, 0.5, 0.8])
plt.plot(year, hares, year, lynxes, year, carrots)
plt.legend(('Hare', 'Lynx', 'Carrot'), loc=(1.05, 0.5))
plt.show()
Reading and writing an elephant

Read and write images
import numpy as np
original figure
plt.figure()
img = plt.imread('../../../data/elephant.png')
plt.imshow(img)


lower resolution
plt.figure() Note: Click here to download the full example code

img_tiny = img[::6, ::6]
plt.imshow(img_tiny, interpolation='nearest')
Mandelbrot set
plt.show()
Compute the Mandelbrot fractal and plot it

Random walk exercise

Plot distance as a function of time for a random walk together with the theoretical result
import numpy as np
import numpy as np
import matplotlib.pyplot as plt # We create 1000 realizations with 200 steps each
from numpy import newaxis n_stories = 1000
t_max = 200
def compute_mandelbrot(N_max, some_threshold, nx, ny):
# A grid of c-values t = np.arange(t_max)
x = np.linspace(-2, 1, nx) # Steps can be -1 or 1 (note that randint excludes the upper limit)
y = np.linspace(-1.5, 1.5, ny) steps = 2 * np.random.randint(0, 1 + 1, (n_stories, t_max)) - 1
c = x[:,newaxis] + 1j*y[newaxis,:] # The time evolution of the position is obtained by successively
# summing up individual steps. This is done for each of the
# Mandelbrot iteration # realizations, i.e. along axis 1.
positions = np.cumsum(steps, axis=1)
z = c
# Determine the time evolution of the mean square distance.
# The code below overflows in many regions of the x-y grid, suppress sq_distance = positions**2
# warnings temporarily mean_sq_distance = np.mean(sq_distance, axis=0)
with np.warnings.catch_warnings():
np.warnings.simplefilter("ignore") # Plot the distance d from the origin as a function of time and
for j in range(N_max): # compare with the theoretically expected result where d(t)
z = z**2 + c # grows as a square root of time t.
mandelbrot_set = (abs(z) < some_threshold) plt.figure(figsize=(4, 3))
plt.plot(t, np.sqrt(mean_sq_distance), 'g.', t, np.sqrt(t), 'y-')
return mandelbrot_set plt.xlabel(r"$t$")
plt.ylabel(r"$\sqrt{\langle (\delta x)^2 \rangle}$")
mandelbrot_set = compute_mandelbrot(50, 50., 601, 401) plt.tight_layout()
plt.show()
plt.imshow(mandelbrot_set.T, extent=[-2, 1, -1.5, 1.5])
plt.gray()
plt.show() Total running time of the script: ( 0 minutes 0.057 seconds)

5.1.1 IPython, Jupyter, and matplotlib modes
Tip: The Jupyter notebook and the IPython enhanced interactive Python, are tuned for the scientific-
computing workflow in Python, in combination with Matplotlib:
5
For interactive matplotlib sessions, turn on the matplotlib mode
IPython console When using the IPython console, use:
In [1]: %matplotlib
CHAPTER Jupyter notebook In the notebook, insert, at the beginning of the notebook the
following magic:
%matplotlib inline
5.1.2 pyplot
Tip: pyplot provides a procedural interface to the matplotlib object-oriented plotting library. It is
Matplotlib: plotting modeled closely after Matlab™. Therefore, the majority of plotting commands in pyplot have Matlab™
analogs with similar arguments. Important commands are explained with interactive examples.
from matplotlib import pyplot as plt
5.2 Simple plot
Thanks Tip: In this section, we want to draw the cosine and sine functions on the same plot. Starting from
the default settings, we’ll enrich the figure step by step to make it nicer.
Many thanks to Bill Wing and Christoph Deil for review and corrections.
First step is to get the data for the sine and cosine functions:
Authors: Nicolas Rougier, Mike Müller, Gaël Varoquaux

import numpy as np
Chapter contents X = np.linspace(-np.pi, np.pi, 256)

C, S = np.cos(X), np.sin(X)
• Introduction
• Simple plot X is now a numpy array with 256 values ranging from −𝜋 to +𝜋 (included). C is the cosine (256 values)
and S is the sine (256 values).
• Figures, Subplots, Axes and Ticks
To run the example, you can type them in an IPython interactive session:
• Other Types of Plots: examples and exercises
$ ipython --matplotlib
• Beyond this tutorial
• Quick references This brings us to the IPython prompt:
• Full code examples IPython 0.13 -- An enhanced Interactive Python.
? -> Introduction to IPython's features.
%magic -> Information about IPython's 'magic' % functions.
help -> Python's own help system.
5.1 Introduction object? -> Details about 'object'. ?object also works, ?? prints more.
Tip: Matplotlib is probably the most used Python package for 2D-graphics. It provides both a quick Tip: You can also download each of the examples and run it using regular python, but you will lose
way to visualize data from Python and publication-quality figures in many formats. We are going to interactive data manipulation:
explore matplotlib in interactive mode covering most common cases.
98 5.2. Simple plot 99

$ python plot_exercise_1.py
Hint: Documentation
You can get source for each step by clicking on the corresponding figure. • Customizing matplotlib
In the script below, we’ve instantiated (and commented) all the figure settings that influence the ap-
5.2.1 Plotting with default settings pearance of the plot.
Tip: The settings have been explicitly set to their default values, but now you can interactively play
with the values to explore their affect (see Line properties and Line styles below).
import numpy as np
# Create a figure of size 8x6 inches, 80 dots per inch

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

plt.subplot(1, 1, 1)
Hint: Documentation
X = np.linspace(-np.pi, np.pi, 256)
• plot tutorial
• plot() command
# Plot cosine with a blue continuous line of width 1 (pixels)
plt.plot(X, C, color="blue", linewidth=1.0, linestyle="-")
Tip: Matplotlib comes with a set of default settings that allow customizing all kinds of properties. You # Plot sine with a green continuous line of width 1 (pixels)
can control the defaults of almost every property in matplotlib: figure size and dpi, line width, color and plt.plot(X, S, color="green", linewidth=1.0, linestyle="-")
style, axes, axis and grid properties, text and font properties and so on.
# Set x limits
plt.xlim(-4.0, 4.0)
import numpy as np
import matplotlib.pyplot as plt # Set x ticks
plt.xticks(np.linspace(-4, 4, 9))
C, S = np.cos(X), np.sin(X) # Set y limits
plt.ylim(-1.0, 1.0)
plt.plot(X, C)
plt.plot(X, S) # Set y ticks
plt.yticks(np.linspace(-1, 1, 5))
plt.show()
# Save figure using 72 dots per inch
# plt.savefig("exercise_2.png", dpi=72)
5.2.2 Instantiating defaults
# Show result on screen
plt.show()
5.2. Simple plot 100 5.2. Simple plot 101

5.2.3 Changing colors and line widths (continued from previous page)
plt.ylim(C.min() * 1.1, C.max() * 1.1)
...
5.2.5 Setting ticks
Hint: Documentation
• Controlling line properties
• Line2D API
Hint: Documentation
Tip: First step, we want to have the cosine in blue and the sine in red and a slighty thicker line for • xticks() command
both of them. We’ll also slightly alter the figure size to make it more horizontal.
• yticks() command
• Tick container
...
plt.figure(figsize=(10, 6), dpi=80) • Tick locating and formatting
plt.plot(X, S, color="red", linewidth=2.5, linestyle="-")
...
Tip: Current ticks are not ideal because they do not show the interesting values (±𝜋,:math:pm pi/2)
for sine and cosine. We’ll change them such that they show only these values.
5.2.4 Setting limits
...
plt.xticks([-np.pi, -np.pi/2, 0, np.pi/2, np.pi])
plt.yticks([-1, 0, +1])
...
5.2.6 Setting tick labels
Hint: Documentation
• xlim() command
• ylim() command
Tip: Current limits of the figure are a bit too tight and we want to make some space in order to clearly
Hint: Documentation
see all data points.
• Working with text
... • xticks() command
plt.xlim(X.min() * 1.1, X.max() * 1.1)
• yticks() command

• set_xticklabels() 5.2.8 Adding a legend

• set_yticklabels()
Tip: Ticks are now properly placed but their label is not very explicit. We could guess that 3.142 is 𝜋
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 we’ll use latex to allow for nice rendering of the label.
...
plt.xticks([-np.pi, -np.pi/2, 0, np.pi/2, np.pi],
[r'$-\pi$', r'$-\pi/2$', r'$0$', r'$+\pi/2$', r'$+\pi$'])
plt.yticks([-1, 0, +1],
[r'$-1$', r'$0$', r'$+1$']) Hint: Documentation
... • Legend guide
• legend() command
5.2.7 Moving spines • legend API
Tip: Let’s 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.
...
plt.plot(X, C, color="blue", linewidth=2.5, linestyle="-", label="cosine")
plt.plot(X, S, color="red", linewidth=2.5, linestyle="-", label="sine")
plt.legend(loc='upper left')
...
Hint: Documentation
• spines API 5.2.9 Annotate some points
• Axis container
• Transformations tutorial
Tip: Spines are the lines connecting the axis tick marks and noting the boundaries of the data area.
They can be placed at arbitrary positions and until now, they were on the border of the axis. We’ll change
that since we want to have them in the middle. Since there are four of them (top/bottom/left/right),
we’ll discard the top and right by setting their color to none and we’ll move the bottom and left ones to
coordinate 0 in data space coordinates.
...
Hint: Documentation
ax = plt.gca() # gca stands for 'get current axis'
ax.spines['right'].set_color('none') • Annotating axis
ax.spines['top'].set_color('none')
ax.xaxis.set_ticks_position('bottom') • annotate() command
ax.spines['bottom'].set_position(('data',0))
ax.yaxis.set_ticks_position('left')
ax.spines['left'].set_position(('data',0))
...
Tip: Let’s annotate some interesting points using the annotate command. We chose the 2𝜋/3 value
and we want to annotate both the sine and the cosine. We’ll first draw a marker on the curve as well as
a straight dotted line. Then, we’ll use the annotate command to display some text with an arrow.

more control over the display using figure, subplot, and axes explicitly. While subplot positions the plots
...
in a regular grid, axes allows free placement within the figure. Both can be useful depending on your
t = 2 * np.pi / 3 intention. We’ve already worked with figures and subplots without explicitly calling them. When we
plt.plot([t, t], [0, np.cos(t)], color='blue', linewidth=2.5, linestyle="--") call plot, matplotlib calls gca() to get the current axes and gca in turn calls gcf() to get the current
plt.scatter([t, ], [np.cos(t), ], 50, color='blue') figure. If there is none it calls figure() to make one, strictly speaking, to make a subplot(111). Let’s
look at the details.
plt.annotate(r'$cos(\frac{2\pi}{3} )=-\frac{1} {2} $',
xy=(t, np.cos(t)), xycoords='data',
xytext=(-90, -50), textcoords='offset points', fontsize=16,
arrowprops=dict(arrowstyle="->", connectionstyle="arc3,rad=.2")) 5.3.1 Figures
plt.plot([t, t],[0, np.sin(t)], color='red', linewidth=2.5, linestyle="--") Tip: A figure is the windows in the GUI that has “Figure #” as title. Figures are numbered starting
plt.scatter([t, ],[np.sin(t), ], 50, color='red')
from 1 as opposed to the normal Python way starting from 0. This is clearly MATLAB-style. There are
plt.annotate(r'$sin(\frac{2\pi}{3} )=\frac{\sqrt{3} }{2} $',
several parameters that determine what the figure looks like:
xy=(t, np.sin(t)), xycoords='data',
xytext=(+10, +30), textcoords='offset points', fontsize=16,
arrowprops=dict(arrowstyle="->", connectionstyle="arc3,rad=.2")) Argument Default Description
... num 1 number of figure
figsize figure.figsize figure size in inches (width, height)
dpi figure.dpi resolution in dots per inch
5.2.10 Devil is in the details 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
Tip: The defaults can be specified in the resource file and will be used most of the time. Only the
number of the figure is frequently changed.
As with other objects, you can set figure properties also setp or with the set_something methods.
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).
Hint: Documentation
• artist API plt.close(1) # Closes figure 1
• set_bbox() method
5.3.2 Subplots
Tip: The tick labels are now hardly visible because of the blue and red lines. We can make them Tip: With subplot you can arrange plots in a regular grid. You need to specify the number of rows and
bigger and we can also adjust their properties such that they’ll be rendered on a semi-transparent white columns and the number of the plot. Note that the gridspec command is a more powerful alternative.
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))
...
5.3 Figures, Subplots, Axes and Ticks 5.3.3 Axes

Axes are very similar to subplots but allow placement of plots at any location in the fig-
A “figure” in matplotlib means the whole window in the user interface. Within this figure there can be ure. So if we want to put a smaller plot inside a bigger one we do so with axes.
“subplots”.
Tip: So far we have used implicit figure and axes creation. This is handy for fast plots. We can have
5.3. Figures, Subplots, Axes and Ticks 106 5.3. Figures, Subplots, Axes and Ticks 107
5.4 Other Types of Plots: examples and exercises
5.3.4 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 = plt.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.
5.3. Figures, Subplots, Axes and Ticks 108 5.4. Other Types of Plots: examples and exercises 109
5.4.2 Scatter Plots
Starting from the code below, try to reproduce the graphic taking
care of marker size, color and transparency.
Hint: Color is given by angle of (X,Y).
n = 1024
X = np.random.normal(0,1,n)
Y = np.random.normal(0,1,n)
plt.scatter(X,Y)
Click on figure for solution.

5.4.1 Regular Plots
5.4.3 Bar Plots
care of filled areas: Starting from the code below, try to reproduce the graphic by
adding labels for red bars.
Hint: You need to use the fill_between() command.
Hint: You need to take care of text alignment.
n = 256
X = np.linspace(-np.pi, np.pi, n) n = 12
Y = np.sin(2 * X) X = np.arange(n)
Y1 = (1 - X / float(n)) * np.random.uniform(0.5, 1.0, n)
plt.plot(X, Y + 1, color='blue', alpha=1.00) Y2 = (1 - X / float(n)) * np.random.uniform(0.5, 1.0, n)
plt.plot(X, Y - 1, color='blue', alpha=1.00)
plt.bar(X, +Y1, facecolor='#9999ff', edgecolor='white')
Click on the figure for solution. plt.bar(X, -Y2, facecolor='#ff9999', edgecolor='white')
for x, y in zip(X, Y1):

plt.text(x + 0.4, y + 0.05, '%.2f ' % y, ha='center', va='bottom')
plt.ylim(-1.25, +1.25)
5.4. Other Types of Plots: examples and exercises 110 5.4. Other Types of Plots: examples and exercises 111
5.4.4 Contour Plots 5.4.6 Pie Charts
Starting from the code below, try to reproduce the graphic taking Starting from the code below, try to reproduce the graphic taking
care of the colormap (see Colormaps below). care of colors and slices size.
Hint: You need to use the clabel() command. Hint: You need to modify Z.
def f(x, y): Z = np.random.uniform(0, 1, 20)

return (1 - x / 2 + x ** 5 + y ** 3) * np.exp(-x ** 2 -y ** 2) plt.pie(Z)
n = 256 Click on the figure for the solution.

x = np.linspace(-3, 3, n)
y = np.linspace(-3, 3, n)
X, Y = np.meshgrid(x, y) 5.4.7 Quiver Plots
plt.contourf(X, Y, f(X, Y), 8, alpha=.75, cmap='jet')

C = plt.contour(X, Y, f(X, Y), 8, colors='black', linewidth=.5)
5.4.5 Imshow
care of colors and orientations.
Hint: You need to draw arrows twice.
n = 8
Starting from the code below, try to reproduce the graphic taking X, Y = np.mgrid[0:n, 0:n]
care of colormap, image interpolation and origin. plt.quiver(X, Y)

Hint: You need to take care of the origin of the image in the imshow command and use a colorbar()
5.4.8 Grids
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)
plt.imshow(f(X, Y))
Click on the figure for the solution.

care of line styles. (continued from previous page)

for r, bar in zip(radii, bars):
axes = plt.gca()
bar.set_facecolor(plt.cm.jet(r / 10.))
axes.set_xlim(0, 4)
bar.set_alpha(0.5)
axes.set_ylim(0, 3)
axes.set_xticklabels([])
axes.set_yticklabels([]) Click on figure for solution.
Click on figure for solution. 5.4.11 3D Plots
5.4.9 Multi Plots
Starting from the code below, try to reproduce the graphic.
Starting from the code below, try to reproduce the graphic. Hint: You need to use contourf()
Hint: You can use several subplots with different partition. from mpl_toolkits.mplot3d import Axes3D
fig = plt.figure()
ax = Axes3D(fig)
X = np.arange(-4, 4, 0.25)
Y = np.arange(-4, 4, 0.25)
X, Y = np.meshgrid(X, Y)
Click on figure for solution. R = np.sqrt(X**2 + Y**2)
Z = np.sin(R)
5.4.10 Polar Axis
ax.plot_surface(X, Y, Z, rstride=1, cstride=1, cmap='hot')

See also:
3D plotting with Mayavi
5.4.12 Text
Hint: You only need to modify the axes line
Starting from the code below, try to reproduce the graphic.

plt.axes([0, 0, 1, 1])
Try to do the same from scratch !
N = 20
theta = np.arange(0., 2 * np.pi, 2 * np.pi / N)
radii = 10 * np.random.rand(N) Hint: Have a look at the matplotlib logo.
width = np.pi / 4 * np.random.rand(N)
bars = plt.bar(theta, radii, width=width, bottom=0.0)
• Using offset transforms to create a shadow effect

• The transformation pipeline
5.5.2 Matplotlib documentation

• User guide
• FAQ
• Installation
• Usage
Quick read • How-To
• Troubleshooting
If you want to do a first quick pass through the Scipy lectures to learn the ecosystem, you can directly • Environment Variables
skip to the next chapter: Scipy : high-level scientific computing. • Screenshots
The remainder of this chapter is not necessary to follow the rest of the intro part. But be sure to
come back and finish this chapter later. 5.5.3 Code documentation
The code is well documented and you can quickly access a specific command from within a python
session:
5.5 Beyond this tutorial >>> import matplotlib.pyplot as plt
>>> help(plt.plot)
Matplotlib benefits from extensive documentation as well as a large community of users and developers. Help on function plot in module matplotlib.pyplot:
Here are some links of interest:
plot(*args,...)
Plot y versus x as lines and/or markers.
5.5.1 Tutorials
• Pyplot tutorial Call signatures::
• Introduction
plot([x], y, [fmt],...data=None, **kwargs)
• Controlling line properties
plot([x], y, [fmt], [x2], y2, [fmt2], ..., **kwargs)
• Working with multiple figures and axes
...
• Working with text
• Image tutorial
• Startup commands 5.5.4 Galleries
• Importing image data into Numpy arrays
• Plotting numpy arrays as images The matplotlib gallery is also incredibly useful when you search how to render a given graphic. Each
• Text tutorial example comes with its source.
• Text introduction
• Basic text commands 5.5.5 Mailing lists
• Text properties and layout
• Writing mathematical expressions Finally, there is a user mailing list where you can ask for help and a developers mailing list that is more
• Text rendering With LaTeX technical.
• Annotating text
• Artist tutorial
• Introduction
5.6 Quick references
• Customizing your objects
Here is a set of tables that show main properties and styles.
• Object containers
• Figure container
• Axes container
• Axis containers
• Tick containers
• Path tutorial
• Introduction
• Bézier example
• Compound paths
• Transforms tutorial
• Introduction
• Data coordinates
• Axes coordinates
• Blended transformations
5.5. Beyond this tutorial 116 5.6. Quick references 117

5.6.1 Line properties 5.6.3 Markers
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
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
markeredgewidth line width around the
(mew) marker symbol 5.6.4 Colormaps
markeredgecolor edge color if a marker is used All colormaps can be reversed by appending _r. For instance, gray_r is the reverse of gray.
(mec)
If you want to know more about colormaps, check the documentation on Colormaps in matplotlib.
markerfacecolor face color if a marker is used
(mfc)
markersize (ms) size of the marker in points
5.6.2 Line styles

5.7.1 Code samples for Matplotlib
The examples here are only examples relevant to the points raised in this chapter. The matplotlib
documentation comes with a much more exhaustive gallery.
5.6. Quick references 118 5.7. Full code examples 119

Pie chart
A simple pie chart example with matplotlib.
import numpy as np
import matplotlib
matplotlib.use('Agg')
fig = plt.figure(figsize=(5, 4), dpi=72)

axes = fig.add_axes([0.01, 0.01, .98, 0.98])
X = np.linspace(0, 2, 200)
import numpy as np Y = np.sin(2*np.pi*X)
import matplotlib.pyplot as plt plt.plot(X, Y, lw=2)
plt.ylim(-1.1, 1.1)
n = 20 plt.grid()
Z = np.ones(n)
Z[-1] *= 2 plt.show()
plt.axes([0.025, 0.025, 0.95, 0.95]) Total running time of the script: ( 0 minutes 0.011 seconds)
plt.pie(Z, explode=Z*.05, colors = ['%f ' % (i/float(n)) for i in range(n)])
plt.axis('equal') Note: Click here to download the full example code
plt.xticks([])
plt.yticks()
Plotting a scatter of points
plt.show()
A simple example showing how to plot a scatter of points with matplotlib.
A simple, good-looking plot

Demoing some simple features of matplotlib
import numpy as np import matplotlib.pyplot as plt

fig = plt.figure()
n = 1024 fig.subplots_adjust(bottom=0.025, left=0.025, top = 0.975, right=0.975)
X = np.random.normal(0, 1, n)
Y = np.random.normal(0, 1, n) plt.subplot(2, 1, 1)
T = np.arctan2(Y, X) plt.xticks([]), plt.yticks([])
plt.axes([0.025, 0.025, 0.95, 0.95]) plt.subplot(2, 3, 4)

plt.scatter(X, Y, s=75, c=T, alpha=.5) plt.xticks([])
plt.yticks([])
plt.xlim(-1.5, 1.5)
plt.xticks([]) plt.subplot(2, 3, 5)
plt.ylim(-1.5, 1.5) plt.xticks([])
plt.yticks([]) plt.yticks([])
plt.show() plt.subplot(2, 3, 6)
plt.xticks([])
Total running time of the script: ( 0 minutes 0.013 seconds) plt.yticks([])
plt.show()
Subplots
Show multiple subplots in matplotlib.
A simple plotting example

A plotting example with a few simple tweaks
import numpy as np
import matplotlib
fig = plt.figure(figsize=(5, 4), dpi=72) import matplotlib.pyplot as plt

axes = fig.add_axes([0.01, 0.01, .98, 0.98])
x = np.linspace(0, 2, 200) plt.axes([.1, .1, .8, .8])
y = np.sin(2 * np.pi * x) plt.xticks([])
plt.plot(x, y, lw=.25, c='k') plt.yticks([])
plt.xticks(np.arange(0.0, 2.0, 0.1)) plt.text(.6, .6, 'axes([0.1, 0.1, 0.8, 0.8])', ha='center', va='center',
plt.yticks(np.arange(-1.0, 1.0, 0.1)) size=20, alpha=.5)
plt.grid()
plt.show() plt.axes([.2, .2, .3, .3])
plt.xticks([])
plt.yticks([])
Total running time of the script: ( 0 minutes 0.035 seconds) plt.text(.5, .5, 'axes([0.2, 0.2, 0.3, 0.3])', ha='center', va='center',
size=16, alpha=.5)
plt.show()

Simple axes example
This example shows a couple of simple usage of axes.
Horizontal arrangement of subplots

An example showing horizontal arrangement of subplots with matplotlib.
plt.figure(figsize=(6, 4))
plt.subplot(2, 1, 1) plt.figure(figsize=(6, 4))
plt.yticks([]) plt.xticks([])
plt.text(0.5, 0.5, 'subplot(2,1,1)', ha='center', va='center', plt.yticks([])
size=24, alpha=.5) plt.text(0.5, 0.5, 'subplot(1,2,1)', ha='center', va='center',
size=24, alpha=.5)
size=24, alpha=.5) plt.text(0.5, 0.5, 'subplot(1,2,2)', ha='center', va='center',
size=24, alpha=.5)
plt.tight_layout()
plt.show() plt.tight_layout()
plt.show()
Subplot plot arrangement vertical

3D plotting
An example showing vertical arrangement of subplots with matplotlib.
A simple example of 3D plotting.

from mpl_toolkits.mplot3d import Axes3D
def f(x, y):
fig = plt.figure() return (1 - x / 2 + x ** 5 + y ** 3 ) * np.exp(-x ** 2 - y ** 2)
ax = Axes3D(fig)
X = np.arange(-4, 4, 0.25) n = 10
Y = np.arange(-4, 4, 0.25) x = np.linspace(-3, 3, 3.5 * n)
X, Y = np.meshgrid(X, Y) y = np.linspace(-3, 3, 3.0 * n)
R = np.sqrt(X ** 2 + Y ** 2) X, Y = np.meshgrid(x, y)
Z = np.sin(R) Z = f(X, Y)
ax.plot_surface(X, Y, Z, rstride=1, cstride=1, cmap=plt.cm.hot) plt.axes([0.025, 0.025, 0.95, 0.95])

ax.contourf(X, Y, Z, zdir='z', offset=-2, cmap=plt.cm.hot) plt.imshow(Z, interpolation='nearest', cmap='bone', origin='lower')
ax.set_zlim(-2, 2) plt.colorbar(shrink=.92)
plt.show() plt.xticks([])
plt.yticks([])
Total running time of the script: ( 0 minutes 0.122 seconds) plt.show()


Imshow elaborate
An example demoing imshow and styling the figure. Plotting a vector field: quiver
A simple example showing how to plot a vector field (quiver) with matplotlib.

n = 8 ax = plt.axes([0.025, 0.025, 0.95, 0.95], polar=True)

X, Y = np.mgrid[0:n, 0:n]
T = np.arctan2(Y - n / 2., X - n/2.) N = 20
R = 10 + np.sqrt((Y - n / 2.0) ** 2 + (X - n / 2.0) ** 2) theta = np.arange(0.0, 2 * np.pi, 2 * np.pi / N)
U, V = R * np.cos(T), R * np.sin(T) radii = 10 * np.random.rand(N)
plt.axes([0.025, 0.025, 0.95, 0.95]) bars = plt.bar(theta, radii, width=width, bottom=0.0)
plt.quiver(X, Y, U, V, R, alpha=.5)
plt.quiver(X, Y, U, V, edgecolor='k', facecolor='None', linewidth=.5) for r,bar in zip(radii, bars):
bar.set_facecolor(plt.cm.jet(r/10.))
plt.xlim(-1, n) bar.set_alpha(0.5)
plt.xticks([])
plt.ylim(-1, n) ax.set_xticklabels([])
plt.yticks([]) ax.set_yticklabels([])
plt.show()
plt.show()
Displaying the contours of a function

Plotting in polar coordinnates
An example showing how to display the contours of a function with matplotlib.
A simple example showing how to plot in polar coordinnates with matplotlib.
import numpy as np
import matplotlib
import numpy as np matplotlib.rc('grid', color='black', linestyle='-', linewidth=1)
fig = plt.figure(figsize=(5,4),dpi=72)
def f(x,y): axes = fig.add_axes([0.01, 0.01, .98, 0.98], facecolor='.75')
return (1 - x / 2 + x**5 + y**3) * np.exp(-x**2 -y**2) X = np.linspace(0, 2, 40)
Y = np.sin(2 * np.pi * X)
n = 256 plt.plot(X, Y, lw=.05, c='b', antialiased=False)
y = np.linspace(-3, 3, n) plt.xticks([])
X,Y = np.meshgrid(x, y) plt.yticks(np.arange(-1., 1., 0.2))
plt.grid()
plt.axes([0.025, 0.025, 0.95, 0.95]) ax = plt.gca()
plt.contourf(X, Y, f(X, Y), 8, alpha=.75, cmap=plt.cm.hot) plt.show()
C = plt.contour(X, Y, f(X, Y), 8, colors='black', linewidth=.5)
plt.clabel(C, inline=1, fontsize=10)
plt.xticks([])
plt.yticks([]) Note: Click here to download the full example code
plt.show()

Plot and filled plots
Simple example of plots and filling between them with matplotlib.
A example of plotting not quite right

An “ugly” example of plotting.

n = 256 n = 12
X = np.linspace(-np.pi, np.pi, n) X = np.arange(n)
Y = np.sin(2 * X) Y1 = (1 - X / float(n)) * np.random.uniform(0.5, 1.0, n)
plt.axes([0.025, 0.025, 0.95, 0.95])
plt.axes([0.025, 0.025, 0.95, 0.95])
plt.plot(X, Y + 1, color='blue', alpha=1.00) plt.bar(X, +Y1, facecolor='#9999ff', edgecolor='white')
plt.fill_between(X, 1, Y + 1, color='blue', alpha=.25) plt.bar(X, -Y2, facecolor='#ff9999', edgecolor='white')
plt.plot(X, Y - 1, color='blue', alpha=1.00) for x, y in zip(X, Y1):

plt.fill_between(X, -1, Y - 1, (Y - 1) > -1, color='blue', alpha=.25) plt.text(x + 0.4, y + 0.05, '%.2f ' % y, ha='center', va= 'bottom')
plt.fill_between(X, -1, Y - 1, (Y - 1) < -1, color='red', alpha=.25)
for x, y in zip(X, Y2):
plt.xlim(-np.pi, np.pi) plt.text(x + 0.4, -y - 0.05, '%.2f ' % y, ha='center', va= 'top')
plt.xticks([])
plt.ylim(-2.5, 2.5) plt.xlim(-.5, n)
plt.ylim(-1.25, 1.25)
plt.show() plt.yticks([])


Bar plots
An example of bar plots with matplotlib.
Subplot grid
An example showing the subplot grid in matplotlib.
Axes
This example shows various axes command to position matplotlib axes.
plt.xticks([])
plt.yticks([])
plt.text(0.5, 0.5, 'subplot(2,2,1)', ha='center', va='center',
size=20, alpha=.5)
plt.subplot(2, 2, 2) import matplotlib.pyplot as plt

plt.xticks([])
plt.yticks([]) plt.axes([.1, .1, .5, .5])
plt.text(0.5, 0.5, 'subplot(2,2,2)', ha='center', va='center', plt.xticks([])
size=20, alpha=.5) plt.yticks([])
plt.text(.1, .1, 'axes([0.1, 0.1, 0.5, 0.5])', ha='left', va='center',
plt.subplot(2, 2, 3) size=16, alpha=.5)
plt.xticks([])
plt.yticks([]) plt.axes([.2, .2, .5, .5])
plt.xticks([])
size=20, alpha=.5) plt.text(.1, .1, 'axes([0.2, 0.2, 0.5, 0.5])', ha='left', va='center',
size=16, alpha=.5)
plt.xticks([]) plt.axes([.3, .3, .5, .5])
size=20, alpha=.5) plt.text(.1, .1, 'axes([0.3, 0.3, 0.5, 0.5])', ha='left', va='center',
size=16, alpha=.5)
plt.tight_layout()
plt.show() plt.axes([.4, .4, .5, .5])
plt.xticks([])
Total running time of the script: ( 0 minutes 0.045 seconds) plt.yticks([])

plt.text(.1, .1, 'axes([0.4, 0.4, 0.5, 0.5])', ha='left', va='center', ax.set_xticklabels([])
size=16, alpha=.5) ax.set_yticklabels([])
Grid 3D plotting
Displaying a grid on the axes in matploblib. Demo 3D plotting with matplotlib and style the figure.

from mpl_toolkits.mplot3d import axes3d
ax = plt.axes([0.025, 0.025, 0.95, 0.95])
ax = plt.gca(projection='3d')
ax.set_xlim(0,4) X, Y, Z = axes3d.get_test_data(0.05)
ax.set_ylim(0,3) cset = ax.contourf(X, Y, Z)
ax.xaxis.set_major_locator(plt.MultipleLocator(1.0)) ax.clabel(cset, fontsize=9, inline=1)
ax.xaxis.set_minor_locator(plt.MultipleLocator(0.1))
ax.yaxis.set_major_locator(plt.MultipleLocator(1.0)) plt.xticks([])
ax.yaxis.set_minor_locator(plt.MultipleLocator(0.1)) plt.yticks([])
ax.grid(which='major', axis='x', linewidth=0.75, linestyle='-', color='0.75') ax.set_zticks([])
ax.grid(which='minor', axis='x', linewidth=0.25, linestyle='-', color='0.75')
ax.grid(which='major', axis='y', linewidth=0.75, linestyle='-', color='0.75')
ax.grid(which='minor', axis='y', linewidth=0.25, linestyle='-', color='0.75') ax.text2D(-0.05, 1.05, " 3D plots \n",

horizontalalignment='left', plt.yticks([])
verticalalignment='top', plt.text(0.5, 0.5, 'Axes 1', ha='center', va='center', size=24, alpha=.5)
bbox=dict(facecolor='white', alpha=1.0),
family='Lint McCree Intl BB', axes_2 = plt.subplot(G[1, :-1])
size='x-large', plt.xticks([])
transform=plt.gca().transAxes) plt.yticks([])
plt.text(0.5, 0.5, 'Axes 2', ha='center', va='center', size=24, alpha=.5)
ax.text2D(-0.05, .975, " Plot 2D or 3D data",
horizontalalignment='left', axes_3 = plt.subplot(G[1:, -1])
verticalalignment='top', plt.xticks([])
family='Lint McCree Intl BB', plt.yticks([])
size='medium', plt.text(0.5, 0.5, 'Axes 3', ha='center', va='center', size=24, alpha=.5)
transform=plt.gca().transAxes)
axes_4 = plt.subplot(G[-1, 0])
plt.show() plt.xticks([])
plt.yticks([])
Total running time of the script: ( 0 minutes 0.053 seconds) plt.text(0.5, 0.5, 'Axes 4', ha='center', va='center', size=24, alpha=.5)
axes_5 = plt.subplot(G[-1, -2])

Note: Click here to download the full example code plt.xticks([])
plt.yticks([])
plt.text(0.5, 0.5, 'Axes 5', ha='center', va='center', size=24, alpha=.5)
GridSpec
plt.tight_layout()
An example demoing gridspec plt.show()
Demo text printing

A example showing off elaborate text printing with matplotlib.

import matplotlib.gridspec as gridspec
G = gridspec.GridSpec(3, 3)
axes_1 = plt.subplot(G[0, :])

plt.xticks([])
5.7.2 Code for the chapter’s exercises
Exercise 1
Solution of the exercise 1 with matplotlib.
import numpy as np
eqs = []
eqs.append((r"$W^{3\beta}_{\delta_1 \rho_1 \sigma_2} = U^{3\beta}_{\delta_1 \rho_1} + \frac{1}
˓→{8 \pi 2} \int^{\alpha_2}_{\alpha_2} d \alpha^\prime_2 \left[\frac{ U^{2\beta}_{\delta_1␣
˓→\rho_1} - \alpha^\prime_2U^{1\beta}_{\rho_1 \sigma_2} }{U^{0\beta}_{\rho_1 \sigma_2}}\right]$
˓→"))
eqs.append((r"$\frac{d\rho}{d t} + \rho \vec{v} \cdot\nabla\vec{v} = -\nabla p + \mu\nabla^2␣

˓→\vec{v} + \rho \vec{g} $"))
import numpy as np
eqs.append((r"$\int_{-\infty}^\infty e^{-x^2}dx=\sqrt{\pi}$")) import matplotlib.pyplot as plt
eqs.append((r"$E = mc^2 = \sqrt{{m_0}^2c^4 + p^2c^2}$"))
eqs.append((r"$F_G = G\frac{m_1m_2} {r^2}$")) n = 256
plt.axes([0.025, 0.025, 0.95, 0.95]) C,S = np.cos(X), np.sin(X)
plt.plot(X, C)
for i in range(24): plt.plot(X,S)
index = np.random.randint(0, len(eqs))
eq = eqs[index] plt.show()
size = np.random.uniform(12, 32)
x,y = np.random.uniform(0, 1, 2) Total running time of the script: ( 0 minutes 0.012 seconds)
alpha = np.random.uniform(0.25, .75)
plt.text(x, y, eq, ha='center', va='center', color="#11557c", alpha=alpha,
transform=plt.gca().transAxes, fontsize=size, clip_on=True) Note: Click here to download the full example code
plt.xticks([])
plt.yticks([])
Exercise 4
plt.show()
Exercise 4 with matplotlib.

plt.figure(figsize=(8, 5), dpi=80) plt.figure(figsize=(8, 5), dpi=80)

plt.subplot(111) plt.subplot(111)
X = np.linspace(-np.pi, np.pi, 256) X = np.linspace(-np.pi, np.pi, 256)

S = np.sin(X) C, S = np.cos(X), np.sin(X)
C = np.cos(X)
plt.plot(X, C, color="blue", linewidth=2.5, linestyle="-") plt.plot(X, S, color="red", linewidth=2.5, linestyle="-")
plt.plot(X, S, color="red", linewidth=2.5, linestyle="-")
plt.xlim(-4.0, 4.0)
plt.xlim(X.min() * 1.1, X.max() * 1.1) plt.xticks(np.linspace(-4, 4, 9))
plt.ylim(-1.0, 1.0)
plt.show() plt.yticks(np.linspace(-1, 1, 5))


Exercise 3
Exercise 3 with matplotlib. Exercise 5

plt.figure(figsize=(8, 5), dpi=80) plt.figure(figsize=(8, 5), dpi=80)

X = np.linspace(-np.pi, np.pi, 256) X = np.linspace(-np.pi, np.pi, 256)

S = np.sin(X) C = np.cos(X)
C = np.cos(X) S = np.sin(X)
plt.plot(X, C, color="blue", linewidth=2.5, linestyle="-") plt.plot(X, C, color="blue", linewidth=2.5, linestyle="-")

plt.plot(X, S, color="red", linewidth=2.5, linestyle="-") plt.plot(X, S, color="red", linewidth=2.5, linestyle="-")
plt.xlim(X.min() * 1.1, X.max() * 1.1) plt.xlim(X.min() * 1.1, X.max() * 1.1)

plt.xticks([-np.pi, -np.pi/2, 0, np.pi/2, np.pi]) plt.xticks([-np.pi, -np.pi/2, 0, np.pi/2, np.pi],
[r'$-\pi$', r'$-\pi/2$', r'$0$', r'$+\pi/2$', r'$+\pi$'])
plt.yticks([-1, 0, +1]) plt.ylim(C.min() * 1.1, C.max() * 1.1)
plt.yticks([-1, 0, +1],
plt.show() [r'$-1$', r'$0$', r'$+1$'])


Exercise 6
Exercise 6 with matplotlib. Exercise 2

plt.show()
Exercise 7
Exercise 7 with matplotlib
import numpy as np
# Create a new figure of size 8x6 points, using 100 dots per inch
# Create a new subplot from a grid of 1x1

plt.subplot(111)
import numpy as np
X = np.linspace(-np.pi, np.pi, 256) import matplotlib.pyplot as plt
plt.figure(figsize=(8,5), dpi=80)
# Plot cosine using blue color with a continuous line of width 1 (pixels) plt.subplot(111)
X = np.linspace(-np.pi, np.pi, 256,endpoint=True)
# Plot sine using green color with a continuous line of width 1 (pixels) C = np.cos(X)
plt.plot(X, S, color="green", linewidth=1.0, linestyle="-") S = np.sin(X)
# Set x limits plt.plot(X, C, color="blue", linewidth=2.5, linestyle="-")

plt.xlim(-4., 4.) plt.plot(X, S, color="red", linewidth=2.5, linestyle="-")
# Set x ticks ax = plt.gca()

plt.xticks(np.linspace(-4, 4, 9)) ax.spines['right'].set_color('none')
# Set y limits ax.xaxis.set_ticks_position('bottom')
plt.ylim(-1.0, 1.0) ax.spines['bottom'].set_position(('data',0))
# Set y ticks ax.spines['left'].set_position(('data',0))
plt.yticks(np.linspace(-1, 1, 5))
# Show result on screen plt.xticks([-np.pi, -np.pi/2, 0, np.pi/2, np.pi],

[r'$-\pi$', r'$-\pi/2$', r'$0$', r'$+\pi/2$', r'$+\pi$']) ax.spines['bottom'].set_position(('data',0))
plt.ylim(C.min() * 1.1, C.max() * 1.1) ax.spines['left'].set_position(('data',0))
plt.yticks([-1, 0, +1],
[r'$-1$', r'$0$', r'$+1$']) plt.xlim(X.min() * 1.1, X.max() * 1.1)
plt.show() [r'$-\pi$', r'$-\pi/2$', r'$0$', r'$+\pi/2$', r'$+\pi$'])
Total running time of the script: ( 0 minutes 0.020 seconds) plt.ylim(C.min() * 1.1, C.max() * 1.1)
plt.yticks([-1, +1],
[r'$-1$', r'$+1$'])
Exercise 8 plt.show()
Exercise 8 with matplotlib. Total running time of the script: ( 0 minutes 0.023 seconds)
Exercise 9
import numpy as np
plt.figure(figsize=(8,5), dpi=80)
plt.subplot(111)
X = np.linspace(-np.pi, np.pi, 256,endpoint=True)

C = np.cos(X)
S = np.sin(X) import numpy as np
plt.plot(X, S, color="red", linewidth=2.5, linestyle="-", label="sine") plt.figure(figsize=(8, 5), dpi=80)
plt.subplot(111)
ax = plt.gca()
ax.spines['right'].set_color('none') X = np.linspace(-np.pi, np.pi, 256,endpoint=True)
ax.spines['top'].set_color('none') C = np.cos(X)
ax.xaxis.set_ticks_position('bottom') S = np.sin(X)

plt.plot(X, S, color="red", linewidth=2.5, linestyle="-", label="sine")
ax = plt.gca()
ax.spines['right'].set_color('none')
ax.xaxis.set_ticks_position('bottom')
ax.spines['left'].set_position(('data',0))

[r'$-\pi$', r'$-\pi/2$', r'$0$', r'$+\pi/2$', r'$+\pi$'])

plt.yticks([-1, +1],
[r'$-1$', r'$+1$'])
t = 2*np.pi/3
plt.plot([t, t], [0, np.cos(t)],
color='blue', linewidth=1.5, linestyle="--")
plt.scatter([t, ], [np.cos(t), ], 50, color='blue')
xy=(t, np.sin(t)), xycoords='data', import numpy as np
xytext=(+10, +30), textcoords='offset points', fontsize=16, import matplotlib.pyplot as plt
arrowprops=dict(arrowstyle="->", connectionstyle="arc3,rad=.2"))
plt.plot([t, t], [0, np.sin(t)], plt.subplot(111)
color='red', linewidth=1.5, linestyle="--")
plt.scatter([t, ], [np.sin(t), ], 50, color='red') X = np.linspace(-np.pi, np.pi, 256)
plt.annotate(r'$cos(\frac{2\pi}{3} )=-\frac{1} {2} $', xy=(t, np.cos(t)), C, S = np.cos(X), np.sin(X)
xycoords='data', xytext=(-90, -50), textcoords='offset points',
fontsize=16, plt.plot(X, C, color="blue", linewidth=2.5, linestyle="-", label="cosine")
arrowprops=dict(arrowstyle="->", connectionstyle="arc3,rad=.2")) plt.plot(X, S, color="red", linewidth=2.5, linestyle="-", label="sine")
plt.legend(loc='upper left') ax = plt.gca()

plt.show() ax.spines['top'].set_color('none')
Total running time of the script: ( 0 minutes 0.025 seconds) ax.spines['bottom'].set_position(('data', 0))
ax.spines['left'].set_position(('data', 0))
plt.xticks([-np.pi, -np.pi / 2, 0, np.pi / 2, np.pi],
Exercise [r'$-\pi$', r'$-\pi/2$', r'$0$', r'$+\pi/2$', r'$+\pi$'])
Exercises with matplotlib. plt.ylim(C.min() * 1.1, C.max() * 1.1)

plt.yticks([-1, 1],
[r'$-1$', r'$+1$'])
t = 2*np.pi/3
plt.plot([t, t], [0, np.cos(t)],
color='blue', linewidth=1.5, linestyle="--")
plt.scatter([t, ], [np.cos(t), ], 50, color='blue')
xy=(t, np.sin(t)), xycoords='data',

xytext=(10, 30), textcoords='offset points', fontsize=16,
arrowprops=dict(arrowstyle="->", connectionstyle="arc3,rad=.2")) size = 256, 16
dpi = 72.0
plt.plot([t, t], [0, np.sin(t)], figsize = size[0] / float(dpi), size[1] / float(dpi)
color='red', linewidth=1.5, linestyle="--") fig = plt.figure(figsize=figsize, dpi=dpi)
plt.scatter([t, ], [np.sin(t), ], 50, color ='red') fig.patch.set_alpha(0)
plt.annotate(r'$cos(\frac{2\pi}{3} )=-\frac{1} {2} $', xy=(t, np.cos(t)), plt.axes([0, .1, 1, .8], frameon=False)
xycoords='data', xytext=(-90, -50),
textcoords='offset points', fontsize=16, for i in range(1, 11):
arrowprops=dict(arrowstyle="->", connectionstyle="arc3,rad=.2")) plt.plot([i, i], [0, 1], color='b', lw=i/2.)
for label in ax.get_xticklabels() + ax.get_yticklabels(): plt.xlim(0, 11)

label.set_fontsize(16) plt.ylim(0, 1)
label.set_bbox(dict(facecolor='white', edgecolor='None', alpha=0.65 )) plt.xticks([])
plt.yticks([])
plt.show()
plt.show()
5.7.3 Example demoing choices for an option
Alpha: transparency
The colors matplotlib line plots This example demonstrates using alpha for transparency.
An example demoing the various colors taken by matplotlib’s plot.

size = 256,16
size = 256, 16 dpi = 72.0
dpi = 72.0 figsize= size[0] / float(dpi), size[1] / float(dpi)
figsize = size[0] / float(dpi), size[1] / float(dpi) fig = plt.figure(figsize=figsize, dpi=dpi)
fig = plt.figure(figsize=figsize, dpi=dpi) fig.patch.set_alpha(0)
fig.patch.set_alpha(0) plt.axes([0, 0.1, 1, .8], frameon=False)
plt.axes([0, 0.1, 1, .8], frameon=False)
for i in range(1, 11):
for i in range(1,11): plt.axvline(i, linewidth=1, color='blue', alpha= .25 + .75 * i / 10.)
plt.plot([i, i], [0, 1], lw=1.5)
plt.xlim(0, 11)
plt.xlim(0, 11) plt.xticks([])
plt.xticks([]) plt.yticks([])
plt.yticks([]) plt.show()
plt.show()
Aliased versus anti-aliased

Linewidth This example demonstrates aliased versus anti-aliased text.
Plot various linewidth with matplotlib.
size = 128, 16 size = 256, 16

dpi = 72.0 dpi = 72.0
figsize= size[0] / float(dpi), size[1] / float(dpi) figsize = size[0] / float(dpi), size[1] / float(dpi)
fig = plt.figure(figsize=figsize, dpi=dpi) fig = plt.figure(figsize=figsize, dpi=dpi)
fig.patch.set_alpha(0) fig.patch.set_alpha(0)
plt.axes([0, 0, 1, 1], frameon=False)
plt.rcParams['text.antialiased'] = False plt.plot([i, ], [1, ], 's', markersize=i, markerfacecolor='w',
plt.text(0.5, 0.5, "Aliased", ha='center', va='center') markeredgewidth=.5, markeredgecolor='k')
plt.xlim(0, 1) plt.xlim(0, 11)

plt.ylim(0, 1) plt.xticks([])
plt.xticks([]) plt.yticks([])
plt.yticks([])
plt.show()
plt.show()
Marker edge width

Aliased versus anti-aliased
Demo the marker edge widths of matplotlib’s markers.
The example shows aliased versus anti-aliased text.

size = 256, 16
size = 128, 16 dpi = 72.0
dpi = 72.0 figsize= size[0] / float(dpi), size[1] / float(dpi)
figsize= size[0] / float(dpi), size[1] / float(dpi) fig = plt.figure(figsize=figsize, dpi=dpi)
fig = plt.figure(figsize=figsize, dpi=dpi) fig.patch.set_alpha(0)
fig.patch.set_alpha(0) plt.axes([0, 0, 1, 1], frameon=False)
for i in range(1,11):
plt.rcParams['text.antialiased'] = True plt.plot([i, ], [1, ], 's', markersize=5,
plt.text(0.5, 0.5, "Anti-aliased", ha='center', va='center') markeredgewidth=1 + i/10., markeredgecolor='k', markerfacecolor='w')
plt.xlim(0, 11)
plt.xlim(0, 1) plt.xticks([])
plt.ylim(0, 1) plt.yticks([])
plt.xticks([])


Marker edge color
Marker size Demo the marker edge color of matplotlib’s markers.
Demo the marker size control in matplotlib.
import numpy as np
size = 256,16
dpi = 72.0
figsize= size[0] / float(dpi), size[1] / float(dpi)
fig = plt.figure(figsize=figsize, dpi=dpi)
fig.patch.set_alpha(0)

r, g, b = np.random.uniform(0, 1, 3)
plt.plot([i, ], [1, ], 's', markersize=5, markerfacecolor='w',
markeredgewidth=1.5, markeredgecolor=(r, g, b, 1))
plt.xlim(0, 11)
plt.xticks([])
plt.yticks([])
plt.show()
Total running time of the script: ( 0 minutes 0.021 seconds) import numpy as np
Note: Click here to download the full example code plt.rc('text', usetex=False)
a = np.outer(np.arange(0, 1, 0.01), np.ones(10))
Marker face color
plt.subplots_adjust(top=0.8, bottom=0.05, left=0.01, right=0.99)
Demo the marker face color of matplotlib’s markers. maps = [m for m in plt.cm.datad if not m.endswith("_r")]
maps.sort()
l = len(maps) + 1
import numpy as np for i, m in enumerate(maps):

import matplotlib.pyplot as plt plt.subplot(1, l, i+1)
plt.axis("off")
size = 256, 16 plt.imshow(a, aspect='auto', cmap=plt.get_cmap(m), origin="lower")
dpi = 72.0 plt.title(m, rotation=90, fontsize=10, va='bottom')
figsize = size[0] / float(dpi), size[1] / float(dpi)
fig = plt.figure(figsize=figsize, dpi=dpi) plt.show()
plt.axes([0, 0, 1, 1], frameon=False) Total running time of the script: ( 0 minutes 0.770 seconds)

r, g, b = np.random.uniform(0, 1, 3) Note: Click here to download the full example code
plt.plot([i, ], [1, ], 's', markersize=8, markerfacecolor=(r, g, b, 1),
markeredgewidth=.1, markeredgecolor=(0, 0, 0, .5))
plt.xlim(0, 11) Solid cap style
plt.xticks([])
plt.yticks([]) An example demoing the solide cap style in matplotlib.
plt.show()

import numpy as np
size = 256, 16
dpi = 72.0
figsize= size[0] / float(dpi), size[1] / float(dpi)
Colormaps fig = plt.figure(figsize=figsize, dpi=dpi)
An example plotting the matplotlib colormaps. fig.patch.set_alpha(0)
plt.plot(np.arange(4), np.ones(4), color="blue", linewidth=8, import numpy as np

solid_capstyle='butt') import matplotlib.pyplot as plt
plt.plot(5 + np.arange(4), np.ones(4), color="blue", linewidth=8, size = 256, 16

solid_capstyle='round') dpi = 72.0
plt.plot(10 + np.arange(4), np.ones(4), color="blue", linewidth=8, fig = plt.figure(figsize=figsize, dpi=dpi)
solid_capstyle='projecting') fig.patch.set_alpha(0)
plt.xlim(0, 14)
plt.xticks([]) plt.plot(np.arange(4), np.ones(4), color="blue", dashes=[15, 15],
plt.yticks([]) linewidth=8, dash_capstyle='butt')
plt.show() plt.plot(5 + np.arange(4), np.ones(4), color="blue", dashes=[15, 15],

linewidth=8, dash_capstyle='round')
plt.plot(10 + np.arange(4), np.ones(4), color="blue", dashes=[15, 15],
linewidth=8, dash_capstyle='projecting')
plt.xlim(0, 14)
plt.xticks([])
Solid joint style plt.yticks([])
An example showing the differen solid joint styles in matplotlib. plt.show()
import numpy as np
import matplotlib.pyplot as plt Note: Click here to download the full example code
size = 256, 16
dpi = 72.0 Dash join style
fig = plt.figure(figsize=figsize, dpi=dpi) Example demoing the dash join style.
plt.plot(np.arange(3), [0, 1, 0], color="blue", linewidth=8, import numpy as np

solid_joinstyle='miter') import matplotlib.pyplot as plt
plt.plot(4 + np.arange(3), [0, 1, 0], color="blue", linewidth=8,
solid_joinstyle='bevel') size = 256, 16
plt.plot(8 + np.arange(3), [0, 1, 0], color="blue", linewidth=8, dpi = 72.0
solid_joinstyle='round') figsize= size[0] / float(dpi), size[1] / float(dpi)
plt.xlim(0, 12) fig.patch.set_alpha(0)
plt.ylim(-1, 2) plt.axes([0, 0, 1, 1], frameon=False)
plt.xticks([])
plt.yticks([]) plt.plot(np.arange(3), [0, 1, 0], color="blue", dashes=[12, 5], linewidth=8,
dash_joinstyle='miter')
plt.show() plt.plot(4 + np.arange(3), [0, 1, 0], color="blue", dashes=[12, 5],
linewidth=8, dash_joinstyle='bevel')
Total running time of the script: ( 0 minutes 0.013 seconds) plt.plot(8 + np.arange(3), [0, 1, 0], color="blue", dashes=[12, 5],
linewidth=8, dash_joinstyle='round')
Note: Click here to download the full example code plt.xlim(0, 12)
plt.ylim(-1, 2)
plt.xticks([])
Dash capstyle plt.yticks([])
An example demoing the dash capstyle. plt.show()
Linestyles Markers
Plot the different line styles. Show the different markers of matplotlib.
import numpy as np
def marker(m, i):

X = i * .5 * np.ones(11)
import numpy as np Y = np.arange(11)
plt.plot(X, Y, lw=1, marker=m, ms=10, mfc=(.75, .75, 1, 1),
def linestyle(ls, i): mec=(0, 0, 1, 1))
X = i * .5 * np.ones(11) plt.text(.5 * i, 10.25, repr(m), rotation=90, fontsize=15, va='bottom')
Y = np.arange(11)
plt.plot(X, Y, ls, color=(.0, .0, 1, 1), lw=3, ms=8, markers = [0, 1, 2, 3, 4, 5, 6, 7, 'o', 'h', '_', '1', '2', '3', '4',
mfc=(.75, .75, 1, 1), mec=(0, 0, 1, 1)) '8', 'p', '^', 'v', '<', '>', '|', 'd', ',', '+', 's', '*',
plt.text(.5 * i, 10.25, ls, rotation=90, fontsize=15, va='bottom') '|', 'x', 'D', 'H', '.']
linestyles = ['-', '--', ':', '-.', '.', ',', 'o', '^', 'v', '<', '>', 's', n_markers = len(markers)
'+', 'x', 'd', '1', '2', '3', '4', 'h', 'p', '|', '_', 'D', 'H']
n_lines = len(linestyles) size = 20 * n_markers, 300
dpi = 72.0
size = 20 * n_lines, 300 figsize= size[0] / float(dpi), size[1] / float(dpi)
dpi = 72.0 fig = plt.figure(figsize=figsize, dpi=dpi)
figsize= size[0] / float(dpi), size[1] / float(dpi) plt.axes([0, 0.01, 1, .9], frameon=False)
plt.axes([0, 0.01, 1, .9], frameon=False) for i, m in enumerate(markers):
marker(m, i)
for i, ls in enumerate(linestyles):
linestyle(ls, i) plt.xlim(-.2, .2 + .5 * n_markers)
plt.xticks([])
plt.xlim(-.2, .2 + .5*n_lines) plt.yticks([])
plt.xticks([])

Note: Click here to download the full example code figsize = size[0] / float(dpi), size[1] / float(dpi)
Locators for tick on axis
An example demoing different locators to position ticks on axis for matplotlib. for i, locator in enumerate(locators):
plt.subplot(n_locators, 1, i + 1)
ax = tickline()
ax.xaxis.set_major_locator(eval(locator))
plt.text(5, 0.3, locator[3:], ha='center')
plt.subplots_adjust(bottom=.01, top=.99, left=.01, right=.99)

plt.show()
5.7.4 Code generating the summary figures with a title
Plotting in polar, decorated

An example showing how to plot in polar coordinnate, and some decorations.
import numpy as np
def tickline():
plt.xlim(0, 10), plt.ylim(-1, 1), plt.yticks([])
ax = plt.gca()
ax.spines['left'].set_color('none')
ax.yaxis.set_ticks_position('none')
ax.xaxis.set_minor_locator(plt.MultipleLocator(0.1))
ax.plot(np.arange(11), np.zeros(11))
return ax
locators = [
'plt.NullLocator()',
'plt.MultipleLocator(1.0)',
'plt.FixedLocator([0, 2, 8, 9, 10])',
'plt.IndexLocator(3, 1)',
'plt.LinearLocator(5)',
'plt.LogLocator(2, [1.0])',
'plt.AutoLocator()',
]
n_locators = len(locators)
import numpy as np
size = 512, 40 * n_locators import matplotlib.pyplot as plt
dpi = 72.0

plt.subplot(1, 1, 1, polar=True)
N = 20
theta = np.arange(0.0, 2 * np.pi, 2 * np.pi / N)
radii = 10 * np.random.rand(N)
bars = plt.bar(theta, radii, width=width, bottom=0.0)
for r, bar in zip(radii, bars):
bar.set_facecolor(plt.cm.jet(r / 10.))
bar.set_alpha(0.5)
plt.gca().set_xticklabels([])
plt.gca().set_yticklabels([])
plt.text(-0.2, 1.02, " Polar Axis \n",

horizontalalignment='left',
verticalalignment='top',
size='xx-large',
plt.text(-0.2, 1.01, "\n\n Plot anything using polar axis ",

size='large',
plt.show()

import numpy as np
fig = plt.figure()
3D plotting vignette ax = Axes3D(fig)
X = np.arange(-4, 4, 0.25)
Demo 3D plotting with matplotlib and decorate the figure. 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=plt.cm.hot)

ax.contourf(X, Y, Z, zdir='z', offset=-2, cmap=plt.cm.hot)
ax.set_zlim(-2, 2)
plt.xticks([])
plt.yticks([])
ax.set_zticks([])
ax.text2D(0.05, .93, " 3D plots \n",

size='xx-large',
ax.text2D(0.05, .87, " Plot 2D or 3D data",

size='large',

ax = plt.gca()
plt.show() ax.add_patch(FancyBboxPatch((-0.05, .87),
width=.66, height=.165, clip_on=False,
Total running time of the script: ( 0 minutes 0.062 seconds) boxstyle="square,pad=0", zorder=3,
facecolor='white', alpha=1.0,
transform=plt.gca().transAxes))
plt.text(-0.05, 1.02, " Regular Plot: plt.plot(...)\n",
Plot example vignette
size='xx-large',
An example of plots with matplotlib, and added annotations. transform=plt.gca().transAxes)
plt.text(-0.05, 1.01, "\n\n Plot lines and/or markers ",

size='large',
plt.show()
Multiple plots vignette

Demo multiple plots and style the figure.
import numpy as np
n = 256
X = np.linspace(0, 2, n)
Y = np.sin(2 * np.pi * X)
plt.plot (X, Y, lw=2, color='violet')

plt.xlim(-0.2, 2.2)
plt.xticks([])
plt.ylim(-1.2, 1.2)
plt.yticks([])
# Add a title and a box around it

from matplotlib.patches import FancyBboxPatch
ax = plt.subplot(2, 1, 1)
ax.set_xticklabels([])
ax.set_yticklabels([])

ax = plt.gca()
ax.add_patch(FancyBboxPatch((-0.05, .72),
boxstyle="square,pad=0", zorder=3,
plt.text(-0.05, 1.02, " Multiplot: plt.subplot(...)\n",

size='xx-large',
transform=ax.transAxes)
plt.text(-0.05, 1.01, "\n\n Plot several plots at once ",
size='large',
transform=ax.transAxes)
import numpy as np
ax = plt.subplot(2, 2, 3) import matplotlib.pyplot as plt
ax.set_yticklabels([])
fig = plt.figure(figsize=(8, 5))
ax = plt.subplot(2, 2, 4) axes = plt.subplot(111)
ax.set_yticklabels([]) n = 5
Z = np.zeros((n, 4))
plt.show() X = np.linspace(0, 2, n)
Y = np.random.random((n, 4))
Total running time of the script: ( 0 minutes 0.055 seconds) plt.boxplot(Y)
plt.xticks([])
Note: Click here to download the full example code plt.yticks([])
Boxplot with matplotlib # Add a title and a box around it

An example of doing box plots with matplotlib ax = plt.gca()
plt.text(-0.05, 1.02, " Box Plot: plt.boxplot(...)\n ",

size='xx-large',
transform=axes.transAxes)
plt.text(-0.04, .98, "\n Make a box and whisker plot ",

size='large',

plt.show() from matplotlib.patches import FancyBboxPatch
ax = plt.gca()
Total running time of the script: ( 0 minutes 0.057 seconds) ax.add_patch(FancyBboxPatch((-0.05, .87),
Note: Click here to download the full example code facecolor='white', alpha=1.0,
plt.text(-0.05, 1.02, " Scatter Plot: plt.scatter(...)\n",

Plot scatter decorated
An example showing the scatter function, with decorations. verticalalignment='top',
size='xx-large',
plt.text(-0.05, 1.01, "\n\n Make a scatter plot of x versus y ",

size='large',
plt.show()
Pie chart vignette

Demo pie chart with matplotlib and style the figure.
import numpy as np
n = 1024
X = np.random.normal(0, 1, n)
Y = np.random.normal(0, 1, n)
T = np.arctan2(Y,X)
plt.scatter(X, Y, s=75, c=T, alpha=.5)

plt.xlim(-1.5, 1.5)
plt.xticks([])
plt.ylim(-1.5, 1.5)
plt.yticks([])

plt.text(-0.05, 1.01, "\n\n Make a pie chart of an array ",

size='large',
plt.show()
Bar plot advanced

An more elaborate bar plot example
import numpy as np
n = 20
X = np.ones(n)
X[-1] *= 2
plt.pie(X, explode=X*.05, colors = ['%f ' % (i/float(n)) for i in range(n)])
fig = plt.gcf()
w, h = fig.get_figwidth(), fig.get_figheight()
r = h / float(w)
plt.xlim(-1.5, 1.5)
plt.ylim(-1.5 * r, 1.5 * r)
plt.xticks([])
plt.yticks([])

ax = plt.gca()
ax.add_patch(FancyBboxPatch((-0.05, .87), import numpy as np
width=.66, height=.165, clip_on=False, import matplotlib.pyplot as plt
facecolor='white', alpha=1.0, n = 16
transform=plt.gca().transAxes)) X = np.arange(n)
plt.text(-0.05, 1.02, " Pie Chart: plt.pie(...)\n", Y2 = (1 - X / float(n)) * np.random.uniform(0.5, 1.0, n)
horizontalalignment='left', plt.bar(X, Y1, facecolor='#9999ff', edgecolor='white')
verticalalignment='top', plt.bar(X, -Y2, facecolor='#ff9999', edgecolor='white')
size='xx-large', (continues on next page)

plt.xlim(-.5, n)
plt.xticks([])
plt.ylim(-1, 1)
plt.yticks([])

ax = plt.gca()
plt.text(-0.05, 1.02, " Bar Plot: plt.bar(...)\n",

size='xx-large',
plt.text(-0.05, 1.01, "\n\n Make a bar plot with rectangles ",

size='large',
plt.show()

import numpy as np
n = 8
X, Y = np.mgrid[0:n, 0:n]
Plotting quiver decorated T = np.arctan2(Y - n/ 2., X - n / 2.)
R = 10 + np.sqrt((Y - n / 2.) ** 2 + (X - n / 2.) ** 2)
An example showing quiver with decorations. U, V = R * np.cos(T), R * np.sin(T)
plt.quiver(X, Y, U, V, R, alpha=.5)
plt.quiver(X, Y, U, V, edgecolor='k', facecolor='None', linewidth=.5)
plt.xlim(-1, n)
plt.xticks([])
plt.ylim(-1, n)
plt.yticks([])

ax = plt.gca()
plt.text(-0.05, 1.02, " Quiver Plot: plt.quiver(...)\n",

size='xx-large',

transform=plt.gca().transAxes) y = np.linspace(-3, 3, 6 * n)
plt.text(-0.05, 1.01, "\n\n Plot a 2-D field of arrows ", Z = f(X, Y)
horizontalalignment='left', plt.imshow(Z, interpolation='nearest', cmap='bone', origin='lower')
verticalalignment='top', plt.xticks([])
size='large', plt.yticks([])

plt.show() from matplotlib.patches import FancyBboxPatch
ax = plt.gca()
Total running time of the script: ( 0 minutes 0.014 seconds) ax.add_patch(FancyBboxPatch((-0.05, .87),
Note: Click here to download the full example code facecolor='white', alpha=1.0,
plt.text(-0.05, 1.02, " Imshow: plt.imshow(...)\n",

Imshow demo
Demoing imshow verticalalignment='top',
size='xx-large',
plt.text(-0.05, 1.01, "\n\n Display an image to current axes ",

family='Lint McCree Intl BB',
size='large',
plt.show()
Display the contours of a function

An example demoing how to plot the contours of a function, with additional layout tweeks.
import numpy as np
def f(x, y):

n = 10
x = np.linspace(-3, 3, 8 * n)

plt.text(-0.05, 1.01, "\n\n Draw contour lines and filled contours ",
size='large',
plt.show()
Grid elaborate
An example displaying a grid on the axes and tweaking the layout.
import numpy as np
def f(x,y):
n = 256
y = np.linspace(-3, 3, n)
plt.contourf(X, Y, f(X, Y), 8, alpha=.75, cmap=plt.cm.hot)

C = plt.contour(X, Y, f(X,Y), 8, colors='black', linewidth=.5)
plt.clabel(C, inline=1, fontsize=10)
plt.xticks([])
plt.yticks([])

ax = plt.gca()
ax.add_patch(FancyBboxPatch((-0.05, .87), import matplotlib.pyplot as plt
width=.66, height=.165, clip_on=False, from matplotlib.ticker import MultipleLocator
facecolor='white', alpha=1.0, fig = plt.figure(figsize=(8, 6), dpi=72, facecolor="white")
transform=plt.gca().transAxes)) axes = plt.subplot(111)
axes.set_xlim(0, 4)
plt.text(-0.05, 1.02, " Contour Plot: plt.contour(..)\n", axes.set_ylim(0, 3)
verticalalignment='top', axes.xaxis.set_major_locator(MultipleLocator(1.0))
size='xx-large', (continues on next page)

axes.xaxis.set_minor_locator(MultipleLocator(0.1))
axes.yaxis.set_major_locator(MultipleLocator(1.0))
axes.yaxis.set_minor_locator(MultipleLocator(0.1))
axes.grid(which='major', axis='x', linewidth=0.75, linestyle='-', color='0.75')
axes.grid(which='minor', axis='x', linewidth=0.25, linestyle='-', color='0.75')
axes.grid(which='major', axis='y', linewidth=0.75, linestyle='-', color='0.75')
axes.grid(which='minor', axis='y', linewidth=0.25, linestyle='-', color='0.75')
axes.set_xticklabels([])
axes.set_yticklabels([])

ax = plt.gca()
plt.text(-0.05, 1.02, " Grid: plt.grid(...)\n",

size='xx-large',
plt.text(-0.05, 1.01, "\n\n Draw ticks and grid ",

size='large',
import numpy as np
fig = plt.figure()
plt.xticks([])
plt.yticks([])
eqs = []
Text printing decorated eqs.append((r"$W^{3\beta}_{\delta_1 \rho_1 \sigma_2} = U^{3\beta}_{\delta_1 \rho_1} + \frac{1}
˓→{8 \pi 2} \int^{\alpha_2}_{\alpha_2} d \alpha^\prime_2 \left[\frac{ U^{2\beta}_{\delta_1␣
An example showing text printing and decorating the resulting figure. ˓→\rho_1} - \alpha^\prime_2U^{1\beta}_{\rho_1 \sigma_2} }{U^{0\beta}_{\rho_1 \sigma_2}}\right]$
˓→"))
eqs.append((r"$\frac{d\rho}{d t} + \rho \vec{v} \cdot\nabla\vec{v} = -\nabla p + \mu\nabla^2␣

˓→\vec{v} + \rho \vec{g} $"))
eqs.append((r"$\int_{-\infty}^\infty e^{-x^2}dx=\sqrt{\pi}$"))
eqs.append((r"$E = mc^2 = \sqrt{{m_0}^2c^4 + p^2c^2}$"))
eqs.append((r"$F_G = G\frac{m_1m_2} {r^2}$"))
for i in range(24):
index = np.random.randint(0,len(eqs))
eq = eqs[index]
size = np.random.uniform(12,32)
x,y = np.random.uniform(0,1,2)
alpha = np.random.uniform(0.25,.75)
plt.text(x, y, eq, ha='center', va='center', color="#11557c", alpha=alpha,
transform=plt.gca().transAxes, fontsize=size, clip_on=True)

ax = plt.gca()

6
plt.text(-0.05, 1.02, " Text: plt.text(...)\n",

size='xx-large',
CHAPTER
plt.text(-0.05, 1.01, "\n\n Draw any kind of text ",
size='large',
plt.show()
Scipy : high-level scientific computing

Authors: Gaël Varoquaux, Adrien Chauve, Andre Espaze, Emmanuelle Gouillart, 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, opti-
mization, image processing, statistics, special functions, etc.
Tip: scipy can be compared to other standard scientific-computing libraries, such as the GSL (GNU
Scientific Library for C and C++), or Matlab’s 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, Scipy’s routines
are optimized and tested, and should therefore be used when possible.
Chapters contents
• File input/output: scipy.io

• Special functions: scipy.special
• Linear algebra operations: scipy.linalg
• Interpolation: scipy.interpolate
5.7. Full code examples 184 185

• Optimization and fit: scipy.optimize (continued from previous page)

>>> data = spio.loadmat('file.mat')
• Statistics and random numbers: scipy.stats >>> data['a']
• Numerical integration: scipy.integrate array([[1., 1., 1.],
[1., 1., 1.],
• Fast Fourier transforms: scipy.fftpack [1., 1., 1.]])
• Signal processing: scipy.signal
• Image manipulation: scipy.ndimage
Warning: Python / Matlab mismatches, eg matlab does not represent 1D arrays
• Summary exercises on scientific computing >>> a = np.ones(3)
• Full code examples for the scipy chapter >>> a
array([1., 1., 1.])
>>> spio.savemat('file.mat', {'a': a})
>>> spio.loadmat('file.mat')['a']
Warning: This tutorial is far from an introduction to numerical computing. As enumerating the array([[1., 1., 1.]])
different submodules and functions in scipy would be very boring, we concentrate instead on a few
Notice the difference?
examples to give a general idea of how to use scipy for scientific computing.
scipy is composed of task-specific sub-modules:
scipy.cluster Vector quantization / Kmeans

scipy.constants Physical and mathematical constants Image files: Reading images:
scipy.fftpack Fourier transform
scipy.integrate Integration routines >>> import imageio
scipy.interpolate Interpolation >>> imageio.imread('fname.png')
scipy.io Data input and output Array(...)
>>> # Matplotlib also has a similar function
scipy.linalg Linear algebra routines
scipy.ndimage n-dimensional image package >>> plt.imread('fname.png')
scipy.odr Orthogonal distance regression array(...)
scipy.optimize Optimization
scipy.signal Signal processing See also:
scipy.sparse Sparse matrices
scipy.spatial Spatial data structures and algorithms • Load text files: numpy.loadtxt()/numpy.savetxt()
scipy.special Any special mathematical functions • Clever loading of text/csv files: numpy.genfromtxt()/numpy.recfromcsv()
scipy.stats Statistics
• Fast and efficient, but numpy-specific, binary format: numpy.save()/numpy.load()
• More advanced input/output of images in scikit-image: skimage.io
Tip: They all depend on numpy, but are mostly independent of each other. The standard way of
importing Numpy and these Scipy modules is:
6.2 Special functions: scipy.special
>>> from scipy import stats # same for other sub-modules Special functions are transcendental functions. The docstring of the scipy.special module is well-
written, so we won’t list all functions here. Frequently used ones are:
The main scipy namespace mostly contains functions that are really numpy functions (try scipy.cos • Bessel function, such as scipy.special.jn() (nth integer order Bessel function)
is np.cos). Those are exposed for historical reasons; there’s no reason to use import scipy in your
code. • 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.
6.1 File input/output: scipy.io • Erf, the area under a Gaussian curve: scipy.special.erf()
Matlab files: Loading and saving:

6.3 Linear algebra operations: scipy.linalg
>>> from scipy import io as spio
>>> a = np.ones((3, 3))
>>> spio.savemat('file.mat', {'a': a}) # savemat expects a dictionary
6.1. File input/output: scipy.io 186 6.2. Special functions: scipy.special 187
6.4 Interpolation: scipy.interpolate

Tip: The scipy.linalg module provides standard linear algebra operations, relying on an underlying
efficient implementation (BLAS, LAPACK). 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.
• The scipy.linalg.det() function computes the determinant of a square matrix: By imagining experimental data close to a sine function:
>>> from scipy import linalg >>> measured_time = np.linspace(0, 1, 10)
>>> arr = np.array([[1, 2], >>> noise = (np.random.random(10)*2 - 1) * 1e-1
... [3, 4]]) >>> measures = np.sin(2 * np.pi * measured_time) + noise
>>> linalg.det(arr)
-2.0 scipy.interpolate.interp1d can build a linear interpolation function:
>>> arr = np.array([[3, 2],
... [6, 4]]) >>> from scipy.interpolate import interp1d
>>> linalg.det(arr) >>> linear_interp = interp1d(measured_time, measures)
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
Then the result can be evaluated at the time
Finally computing the inverse of a singular matrix (its determinant is zero) will raise LinAlgError: of interest:
>>> arr = np.array([[3, 2], >>> interpolation_time = np.linspace(0, 1, 50)
... [6, 4]]) >>> linear_results = linear_interp(interpolation_time)
>>> linalg.inv(arr)
Traceback (most recent call last): A cubic interpolation can also be selected by providing the kind optional keyword argument:
...
...LinAlgError: singular matrix >>> cubic_interp = interp1d(measured_time, measures, kind='cubic')
>>> cubic_results = cubic_interp(interpolation_time)
• More advanced operations are available, for example singular-value decomposition (SVD):
scipy.interpolate.interp2d is similar to scipy.interpolate.interp1d, but for 2-D arrays. Note
>>> arr = np.arange(9).reshape((3, 3)) + np.diag([1, 0, 1]) that for the interp family, the interpolation points must stay within the range of given data points. See
>>> uarr, spec, vharr = linalg.svd(arr)
the summary exercise on Maximum wind speed prediction at the Sprogø station for a more advanced
spline interpolation example.
The resulting array spectrum is:
>>> spec
array([14.88982544, 0.45294236, 0.29654967])
6.5 Optimization and fit: scipy.optimize
Optimization is the problem of finding a numerical solution to a minimization or equality.
The original matrix can be re-composed by matrix multiplication of the outputs of svd with np.dot:
>>> sarr = np.diag(spec) Tip: The scipy.optimize module provides algorithms for function minimization (scalar or multi-
>>> svd_mat = uarr.dot(sarr).dot(vharr) dimensional), curve fitting and root finding.
>>> np.allclose(svd_mat, arr)
True >>> from scipy import optimize
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.
6.3. Linear algebra operations: scipy.linalg 188 6.4. Interpolation: scipy.interpolate 189
6.5.1 Curve fitting 5. Is the time offset for min and max temperatures the same within the fit accuracy?
solution
6.5.2 Finding the minimum of a scalar function
Suppose we have data on a sine wave, with some noise:
>>> x_data = np.linspace(-5, 5, num=50)

>>> y_data = 2.9 * np.sin(1.5 * x_data) + np.random.normal(size=50)
If we know that the data lies on a sine wave, but not the amplitudes or the period, we can find those
by least squares curve fitting. First we have to define the test function to fit, here a sine with unknown
amplitude and period:
Let’s define the following function:
>>> def test_func(x, a, b):
... return a * np.sin(b * x) >>> 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()
This function has a global minimum around -1.3 and a local minimum around 3.8.
Searching for minimum can be done with scipy.optimize.minimize(), given a starting point x0, it
returns the location of the minimum that it has found:
We then use scipy.optimize.curve_fit() to find 𝑎 result type

and 𝑏:
The result of scipy.optimize.minimize() is a compound object comprising all information on the
>>> params, params_covariance = optimize.curve_fit(test_func, x_data, y_data, p0=[2, 2]) convergence
>>> print(params)
[3.05931973 1.45754553]
>>> result = optimize.minimize(f, x0=0)
>>> result
Exercise: Curve fitting of temperature data fun: -7.9458233756...
hess_inv: array([[0.0858...]])
The temperature extremes in Alaska for each month, starting in January, are given by (in jac: array([-1.19209...e-06])
degrees Celcius): message: 'Optimization terminated successfully.'
nfev: 18
max: 17, 19, 21, 28, 33, 38, 37, 37, 31, 23, 19, 18 nit: 5
min: -62, -59, -56, -46, -32, -18, -9, -13, -25, -46, -52, -58 njev: 6
status: 0
1. Plot these temperature extremes. success: True
x: array([-1.30644...])
2. Define a function that can describe min and max temperatures. Hint: this function
>>> result.x # The coordinate of the minimum
has to have a period of 1 year. Hint: include a time offset. array([-1.30644...])
3. Fit this function to the data with scipy.optimize.curve_fit().
4. Plot the result. Is the fit reasonable? If not, why?
6.5. Optimization and fit: scipy.optimize 190 6.5. Optimization and fit: scipy.optimize 191
Methods: As the function is a smooth function, gradient-descent based methods are good options. The
lBFGS algorithm is a good choice in general: Tip: What has happened? Why are we finding 0, which is not a mimimum of our function.
>>> optimize.minimize(f, x0=0, method="L-BFGS-B")

fun: array([-7.94582338])
hess_inv: <1x1 LbfgsInvHessProduct with dtype=float64> Minimizing functions of several variables
jac: array([-1.42108547e-06])
message: ...'CONVERGENCE: NORM_OF_PROJECTED_GRADIENT_<=_PGTOL' To minimize over several variables, the trick is to turn them into a function of a multi-dimensional
nfev: 12 variable (a vector). See for instance the exercise on 2D minimization below.
nit: 5
status: 0
success: True
x: array([-1.30644013]) Note: scipy.optimize.minimize_scalar() is a function with dedicated methods to minimize func-
tions of only one variable.
Note how it cost only 12 functions evaluation above to find a good value for the minimum.
See also:
Finding minima of function is discussed in more details in the advanced chapter: Mathematical opti-
mization: finding minima of functions.
Global minimum: A possible issue with this approach is that, if the function has local minima, the
Exercise: 2-D minimization
algorithm may find these local minima instead of the global minimum depending on the initial point x0:
>>> res = optimize.minimize(f, x0=3, method="L-BFGS-B")

>>> res.x
array([3.83746709])
If we don’t 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, we use scipy.optimize.basinhopping()
(added in version 0.12.0 of Scipy). It combines a local optimizer with sampling of starting points:
>>> optimize.basinhopping(f, 0)
nfev: 1725
minimization_failures: 0
fun: -7.9458233756152845
x: array([-1.30644001])
message: ['requested number of basinhopping iterations completed successfully']
njev: 575 The six-hump camelback function
nit: 100
4
𝑥
𝑓 (𝑥, 𝑦) = (4 − 2.1𝑥2 + )𝑥2 + 𝑥𝑦 + (4𝑦 2 − 4)𝑦 2
3
Note: scipy used to contain the routine anneal, it has been removed in SciPy 0.16.0.
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 matplotlib.pyplot.imshow() to find visually the re-
gions.
Constraints: We can constrain the variable to the interval (0, 10) using the “bounds” argument:
• Use scipy.optimize.minimize(), optionally trying out several of its methods.
A list of bounds 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) ?
As minimize() works in general with x multidimensionsal, the “bounds” argument is a list of bound solution
on each dimension.
>>> res = optimize.minimize(f, x0=1, 6.5.3 Finding the roots of a scalar function
... bounds=((0, 10), ))
>>> res.x
To find a root, i.e. a point where 𝑓 (𝑥) = 0, of the function 𝑓 above we can use scipy.optimize.root():
array([0.])
6.5. Optimization and fit: scipy.optimize 192 6.5. Optimization and fit: scipy.optimize 193
>>> root = optimize.root(f, x0=1) # our initial guess is 1

6.6.1 Distributions: histogram and probability density function
>>> root # The full result Given observations of a random process, their histogram is an estimator of the random process’s PDF
fjac: array([[-1.]]) (probability density function):
fun: array([0.])
message: 'The solution converged.' >>> samples = np.random.normal(size=1000)
nfev: 10 >>> bins = np.arange(-4, 5)
qtf: array([1.33310463e-32]) >>> bins
r: array([-10.]) array([-4, -3, -2, -1, 0, 1, 2, 3, 4])
status: 1 >>> histogram = np.histogram(samples, bins=bins, density=True)[0]
success: True >>> bins = 0.5*(bins[1:] + bins[:-1])
x: array([0.]) >>> bins
>>> root.x # Only the root found array([-3.5, -2.5, -1.5, -0.5, 0.5, 1.5, 2.5, 3.5])
array([0.]) >>> from scipy import stats
>>> pdf = stats.norm.pdf(bins) # norm is a distribution object
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: >>> plt.plot(bins, histogram)
>>> root2 = optimize.root(f, x0=-2.5) >>> plt.plot(bins, pdf)
>>> root2.x [<matplotlib.lines.Line2D object at ...>]
array([-2.47948183])
Note: scipy.optimize.root() also comes with a variety of algorithms, set via the “method” argument.
The distribution objects
scipy.stats.norm is a distribution object: each distribution in scipy.stats is represented as an

object. Here it’s the normal distribution, and it comes with a PDF, a CDF, and much more.
Now that we have found the minima and roots of f and used curve fitting on it, we put all those results
If we know that the random process belongs to a given family of random processes, such as normal
together in a single plot:
processes, we can do a maximum-likelihood fit of the observations to estimate the parameters of the
See also: underlying distribution. Here we fit a normal process to the observed data:
You can find all algorithms and functions with similar functionalities in the documentation of scipy. >>> loc, std = stats.norm.fit(samples)
optimize. >>> loc
-0.045256707...
See the summary exercise on Non linear least squares curve fitting: application to point extraction in
>>> std
topographical lidar data for another, more advanced example. 0.9870331586...
6.6 Statistics and random numbers: scipy.stats

Exercise: Probability distributions
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. 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)?
6.6. Statistics and random numbers: scipy.stats 194 6.6. Statistics and random numbers: scipy.stats 195
Extra: the distributions have many useful methods. Explore them by reading the docstring or by using >>> a = np.random.normal(0, 1, size=100)
tab completion. Can you recover the shape parameter 1 by using the fit method on your random >>> b = np.random.normal(1, 1, size=10)
variates? >>> stats.ttest_ind(a, b)
(array(-3.177574054...), 0.0019370639...)
6.6.2 Mean, median and percentiles

Tip: The resulting output is composed of:
The mean is an estimator of the center of the distribution:
• The T statistic value: it is a number the sign of which is proportional to the difference between
>>> np.mean(samples) the two random processes and the magnitude is related to the significance of this difference.
-0.0452567074...
• the p value: the probability of both processes being identical. If it is close to 1, the two process
The median another estimator of the center. It is the value with half of the observations below, and half are almost certainly identical. The closer it is to zero, the more likely it is that the processes have
above: different means.
>>> np.median(samples)
-0.0580280347... See also:
The chapter on statistics introduces much more elaborate tools for statistical testing and statistical data
loading and visualization outside of scipy.
Tip: Unlike the mean, the median is not sensitive to the tails of the distribution. It is “robust”.
6.7 Numerical integration: scipy.integrate

Exercise: Compare mean and median on samples of a Gamma distribution
6.7.1 Function integrals
Which one seems to be the best estimator of the center for the Gamma distribution? ∫︀ 𝜋/2
The most generic integration routine is scipy.integrate.quad(). To compute 0
𝑠𝑖𝑛(𝑡)𝑑𝑡:
The median is also the percentile 50, because 50% of the observation are below it: >>> from scipy.integrate import quad
>>> res, err = quad(np.sin, 0, np.pi/2)
>>> stats.scoreatpercentile(samples, 50) >>> np.allclose(res, 1) # res is the result, is should be close to 1
-0.0580280347... True
>>> np.allclose(err, 1 - res) # err is an estimate of the err
Similarly, we can calculate the percentile 90: True
>>> stats.scoreatpercentile(samples, 90) Other integration schemes are available: scipy.integrate.fixed_quad(), scipy.integrate.
1.2315935511... quadrature(), scipy.integrate.romberg(). . .
6.7.2 Integrating differential equations

Tip: The percentile is an estimator of the CDF: cumulative distribution function.
scipy.integrate also features routines for integrating Ordinary Differential Equations (ODE). In par-
ticular, scipy.integrate.odeint() solves ODE of the form:
6.6.3 Statistical tests 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):

... return -2 * ypos
A statistical test is a decision indicator. For

instance, if we have two sets of observations, that we assume are generated from Gaussian processes, we
can use a T-test to decide whether the means of two sets of observations are significantly different:
6.6. Statistics and random numbers: scipy.stats 196 6.7. Numerical integration: scipy.integrate 197
Tip: scipy.integrate.odeint() uses the LSODA (Livermore Solver for Ordinary Differential equa-
tions with Automatic method switching for stiff and non-stiff problems), see the ODEPACK Fortran
library for more details.
See also:
Partial Differental Equations
There is no Partial Differential Equations (PDE) solver in Scipy. Some Python packages for solving
PDE’s are available, such as fipy or SfePy.
Then, to compute y as a function of time:

6.8 Fast Fourier transforms: scipy.fftpack
>>> from scipy.integrate import odeint
>>> time_vec = np.linspace(0, 4, 40) The scipy.fftpack module computes fast Fourier transforms (FFTs) and offers utilities to handle them.
>>> y = odeint(calc_derivative, y0=1, t=time_vec) The main functions are:
• scipy.fftpack.fft() to compute the FFT
Let us integrate a more complex ODE: a damped spring-mass oscillator. The position of a mass attached
to a spring obeys the 2nd order ODE 𝑦 ′′ + 2𝜀𝜔0 𝑦 ′ + 𝜔02 𝑦 = 0 with 𝜔02 = 𝑘/𝑚 with 𝑘 the spring constant, • scipy.fftpack.fftfreq() to generate the sampling frequencies
𝑚 the mass and 𝜀 = 𝑐/(2𝑚𝜔0 ) with 𝑐 the damping coefficient. We set: • scipy.fftpack.ifft() computes the inverse FFT, from frequency space to signal space
>>> mass = 0.5 # kg
>>> kspring = 4 # N/m
>>> cviscous = 0.4 # N s/m
Hence:
As an illustration, a (noisy) input signal (sig), and its FFT:
>>> eps = cviscous / (2 * mass * np.sqrt(kspring/mass))
>>> omega = np.sqrt(kspring / mass) >>> from scipy import fftpack
>>> sig_fft = fftpack.fft(sig)
The system is underdamped, as: >>> freqs = fftpack.fftfreq(sig.size, d=time_step)
>>> eps < 1

True
For odeint(), the 2nd order equation needs to be transformed in a system of two first-order equations
for the vector 𝑌 = (𝑦, 𝑦 ′ ): the function computes the velocity and acceleration:
>>> def calc_deri(yvec, time, eps, omega):

... return (yvec[1], -2.0 * eps * omega * yvec[1] - omega **2 * yvec[0])
Signal FFT
As the signal comes from a real function, the Fourier transform is symmetric.
The peak signal frequency can be found with freqs[power.argmax()]
Integration of the system follows:
>>> time_vec = np.linspace(0, 10, 100)

>>> yinit = (1, 0)
>>> yarr = odeint(calc_deri, yinit, time_vec, args=(eps, omega))
6.7. Numerical integration: scipy.integrate 198 6.8. Fast Fourier transforms: scipy.fftpack 199
Exercise: Denoise moon landing image
Setting the Fourrier component above

this frequency to zero and inverting the FFT with scipy.fftpack.ifft(), gives a filtered signal.
Note: The code of this example can be found here
numpy.fft
Numpy also has an implementation of FFT (numpy.fft). However, the scipy one should be preferred,
as it uses more efficient underlying implementations.
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 matplotlib.pyplot.imread().
Fully worked examples:
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?
Crude periodicity finding (link) Gaussian image blur (link)
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.
Solution
6.9 Signal processing: scipy.signal
Tip: scipy.signal is for typical signal processing: 1D, regularly-sampled signals.
6.8. Fast Fourier transforms: scipy.fftpack 200 6.9. Signal processing: scipy.signal 201

>>> plt.plot(t, x_detrended)
Filtering: For non-linear filtering, scipy.signal has filtering (median filter scipy.signal.medfilt(),
Wiener scipy.signal.wiener()), but we will discuss this in the image section.
Tip: scipy.signal also has a full-blown set of tools for the design of linear filter (finite and infinite
response filters), but this is out of the scope of this tutorial.
Spectral analysis: scipy.signal.spectrogram() compute a spectrogram –frequency spectrums over

consecutive time windows–, while scipy.signal.welch() comptes a power spectrum density (PSD).
Resampling scipy.signal.resample(): re-

sample a signal to n points using FFT.
>>> t = np.linspace(0, 5, 100)
>>> x = np.sin(t)
>>> from scipy import signal

>>> x_resampled = signal.resample(x, 25)
>>> plt.plot(t, x)
>>> plt.plot(t[::4], x_resampled, 'ko')
Tip: Notice how on the side of the window the resampling is less accurate and has a rippling effect.
This resampling is different from the interpolation provided by scipy.interpolate as it only applies to
regularly sampled data.
6.10 Image manipulation: scipy.ndimage

scipy.ndimage provides manipulation of n-dimensional arrays as images.
6.10.1 Geometrical transformations on images

Changing orientation, resolution, ..
Detrending scipy.signal.detrend(): remove >>> from scipy import misc # Load an image
linear trend from signal: >>> face = misc.face(gray=True)
>>> t = np.linspace(0, 5, 100) >>> from scipy import ndimage # Shift, roate and zoom it
>>> x = t + np.random.normal(size=100) >>> shifted_face = ndimage.shift(face, (50, 50))
>>> shifted_face2 = ndimage.shift(face, (50, 50), mode='nearest')
>>> from scipy import signal >>> rotated_face = ndimage.rotate(face, 30)
>>> x_detrended = signal.detrend(x) >>> cropped_face = face[50:-50, 50:-50]
>>> zoomed_face = ndimage.zoom(face, 2)
>>> plt.plot(t, x) (continues on next page)
6.9. Signal processing: scipy.signal 202 6.10. Image manipulation: scipy.ndimage 203
(continued from previous page) 6.10.3 Mathematical morphology

>>> zoomed_face.shape
(1536, 2048)
Tip: Mathematical morphology 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.
>>> plt.subplot(151)
<matplotlib.axes._subplots.AxesSubplot object at 0x...>
>>> plt.imshow(shifted_face, cmap=plt.cm.gray)

>>> plt.axis('off')
(-0.5, 1023.5, 767.5, -0.5)
>>> # etc. Mathematical-morphology operations use a structuring element in order to modify geometrical structures.
Let us first generate a structuring element:
6.10.2 Image filtering >>> el = ndimage.generate_binary_structure(2, 1)
>>> el
Generate a noisy face:
array([[False, True, False],
>>> from scipy import misc [...True, True, True],
>>> face = misc.face(gray=True) [False, True, False]])
>>> face = face[:512, -512:] # crop out square on right >>> el.astype(np.int)
>>> import numpy as np array([[0, 1, 0],
>>> noisy_face = np.copy(face).astype(np.float) [1, 1, 1],
>>> noisy_face += face.std() * 0.5 * np.random.standard_normal(face.shape) [0, 1, 0]])
Apply a variety of filters on it: • Erosion scipy.ndimage.binary_erosion()
>>> blurred_face = ndimage.gaussian_filter(noisy_face, sigma=3) >>> a = np.zeros((7, 7), dtype=np.int)

>>> median_face = ndimage.median_filter(noisy_face, size=5) >>> a[1:6, 2:5] = 1
>>> from scipy import signal >>> a
>>> wiener_face = signal.wiener(noisy_face, (5, 5)) 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]])
Other filters in scipy.ndimage.filters and scipy.signal can be applied to images. >>> #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],
Exercise
[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0],
Compare histograms for the different filtered images. [0, 0, 0, 0, 0, 0, 0],
6.10. Image manipulation: scipy.ndimage 204 6.10. Image manipulation: scipy.ndimage 205

[0, 0, 0, 0, 0, 0, 0], >>> mask = a>=0.5
[0, 0, 0, 0, 0, 0, 0], >>> opened_mask = ndimage.binary_opening(mask)
[0, 0, 0, 0, 0, 0, 0]]) >>> closed_mask = ndimage.binary_closing(opened_mask)
• Dilation scipy.ndimage.binary_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.], Exercise
[0., 0., 0., 0., 0.]])
Check that the area of the reconstructed square is smaller than the area of the initial
• Opening scipy.ndimage.binary_opening() square. (The opposite would occur if the closing step was performed before the opening).
>>> a = np.zeros((5, 5), dtype=np.int)

>>> a[1:4, 1:4] = 1 For gray-valued images, eroding (resp. dilating) amounts to replacing a pixel by the minimal (resp.
>>> a[4, 4] = 1 maximal) value among pixels covered by the structuring element centered on the pixel of interest.
>>> a
array([[0, 0, 0, 0, 0], >>> a = np.zeros((7, 7), dtype=np.int)
[0, 1, 1, 1, 0], >>> a[1:6, 1:6] = 3
[0, 1, 1, 1, 0], >>> a[4, 4] = 2; a[2, 3] = 1
[0, 1, 1, 1, 0], >>> a
[0, 0, 0, 0, 1]]) array([[0, 0, 0, 0, 0, 0, 0],
>>> # Opening removes small objects [0, 3, 3, 3, 3, 3, 0],
>>> ndimage.binary_opening(a, structure=np.ones((3, 3))).astype(np.int) [0, 3, 3, 1, 3, 3, 0],
array([[0, 0, 0, 0, 0], [0, 3, 3, 3, 3, 3, 0],
[0, 1, 1, 1, 0], [0, 3, 3, 3, 2, 3, 0],
[0, 1, 1, 1, 0], [0, 3, 3, 3, 3, 3, 0],
[0, 1, 1, 1, 0], [0, 0, 0, 0, 0, 0, 0]])
[0, 0, 0, 0, 0]]) >>> ndimage.grey_erosion(a, size=(3, 3))
>>> # Opening can also smooth corners array([[0, 0, 0, 0, 0, 0, 0],
>>> ndimage.binary_opening(a).astype(np.int) [0, 0, 0, 0, 0, 0, 0],
array([[0, 0, 0, 0, 0], [0, 0, 1, 1, 1, 0, 0],
[0, 0, 1, 0, 0], [0, 0, 1, 1, 1, 0, 0],
[0, 1, 1, 1, 0], [0, 0, 3, 2, 2, 0, 0],
[0, 0, 1, 0, 0], [0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0]]) [0, 0, 0, 0, 0, 0, 0]])
• Closing: scipy.ndimage.binary_closing()
6.10.4 Connected components and measurements on images
Exercise Let us first generate a nice synthetic binary image.
>>> x, y = np.indices((100, 100))

Check that opening amounts to eroding, then dilating. >>> sig = np.sin(2*np.pi*x/50.) * np.sin(2*np.pi*y/50.) * (1+x*y/50.**2)**2
>>> mask = sig > 1
An opening operation removes small structures, while a closing operation fills small holes. Such opera-
tions 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)
6.10. Image manipulation: scipy.ndimage 206 6.10. Image manipulation: scipy.ndimage 207
6.11.1 Maximum wind speed prediction at the Sprogø 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 Sprogø 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
scipy.ndimage.label() assigns a different label to each connected component:
it. In the current model, it is supposed that the maximum wind speed occurring every 50 years is defined
>>> labels, nb = ndimage.label(mask) as the upper 2% quantile.
>>> nb By definition, the quantile function is the inverse of the cumulative distribution function. The latter
8
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
Now compute measurements on each connected component: it will be possible to calculate the cumulative probability of every measured wind speed maxima. From
>>> areas = ndimage.sum(mask, labels, range(1, labels.max()+1)) those experimental points, the scipy.interpolate module will be very useful for fitting the quantile function.
>>> areas # The number of pixels in each connected component Finally the 50 years maxima is going to be evaluated from the cumulative probability of the 2% quantile.
array([190., 45., 424., 278., 459., 190., 549., 424.])
>>> maxima = ndimage.maximum(sig, labels, range(1, labels.max()+1)) Computing the cumulative probabilities
>>> maxima # The maximum signal in each connected component
array([ 1.80238238, 1.13527605, 5.51954079, 2.49611818, 6.71673619, The annual wind speeds maxima have already been computed and saved in the numpy format in the file
1.80238238, 16.76547217, 5.51954079]) 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)
Extract the 4th connected component, and crop the array around it:
Prediction with UnivariateSpline
>>> ndimage.find_objects(labels==4)
[(slice(30L, 48L, None), slice(30L, 48L, None))] In this section the quantile function will be estimated by using the UnivariateSpline class which can
>>> sl = ndimage.find_objects(labels==4) represent a spline from points. The default behavior is to build a spline of degree 3 and points can
>>> from matplotlib import pyplot as plt have different weights according to their reliability. Variants are InterpolatedUnivariateSpline and
>>> plt.imshow(sig[sl[0]]) LSQUnivariateSpline on which errors checking is going to change. In case a 2D spline is wanted,
<matplotlib.image.AxesImage object at ...> the BivariateSpline class family is provided. All those classes for 1D and 2D splines use the FIT-
PACK Fortran subroutines, that’s why a lower library access is available through the splrep and
See the summary exercise on Image processing application: counting bubbles and unmolten grains for a splev functions for respectively representing and evaluating a spline. Moreover interpolation functions
more advanced example. without the use of FITPACK parameters are also provided for simpler use (see interp1d, interp2d,
barycentric_interpolate and so on).
6.11 Summary exercises on scientific computing For the Sprogø maxima wind speeds, the UnivariateSpline will be used because a spline of degree 3
seems to correctly fit the data:
The summary exercises use mainly Numpy, Scipy and Matplotlib. They provide some real-life examples
>>> from scipy.interpolate import UnivariateSpline
of scientific computing with Python. Now that the basics of working with Numpy and Scipy have been
>>> quantile_func = UnivariateSpline(cprob, sorted_max_speeds)
introduced, the interested user is invited to try these exercises.
The quantile function is now going to be evaluated from the full range of probabilities:
6.11. Summary exercises on scientific computing 208 6.11. Summary exercises on scientific computing 209
>>> nprob = np.linspace(0, 1, 1e2)

>>> fitted_max_speeds = quantile_func(nprob)
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
array(32.97989825...)
The results are now gathered on a Matplotlib figure:
Fig. 2: Solution: Python source file

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.
• 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 contains peaks whose center and amplitude permit to compute the position and some characteristics of
-log( -log(p_i) ) for fitting a linear quantile function (remember that you can define the degree the hit target. When the footprint of the laser beam is around 1m on the Earth surface, the beam can
of the UnivariateSpline). Plotting the annual maxima versus the Gumbell distribution should hit multiple targets during the two-way propagation (for example the ground and the top of a tree or
give you the following figure. 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('intro/summary-exercises/examples/waveform_1.npy')
and visualize it:

>>> t = np.arange(len(waveform_1))
>>> plt.plot(t, waveform_1)
>>> plt.show()
As shown below, this waveform is a 80-bin-length signal with a single peak with an amplitude of ap-
proximately 30 in the 15 nanosecond bin. Additionally, the base level of noise is approximately 3. These
values can be used in the initial solution.
• The last step will be to find 34.23 m/s for the maximum wind speed occurring every 50 years.
6.11.2 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 you’re impatient and want to
practice now, please skip it and go directly to Loading and visualization.
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 Earth’s topography (see1 for
more details).
In this tutorial, the goal is to analyze the waveform recorded by the lidar system2 . Such a signal
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

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

fig, ax = plt.subplots(figsize=(8, 6))
The signal is very simple and can be modeled as a single Gaussian function and an offset corresponding plt.plot(t, waveform_1, t, model(t, x))
to the background noise. To fit the signal with the function, we must: plt.xlabel('Time [ns]')
plt.ylabel('Amplitude [bins]')
• define the model plt.legend(['Waveform', 'Model'])
plt.show()
• propose an initial solution
• call scipy.optimize.leastsq
Model
A Gaussian function defined by

{︃ (︂ )︂2 }︃
𝑡−𝜇
𝐵 + 𝐴 exp −
𝜎
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 𝜇 (center)
• coeffs[3] is 𝜎 (width)
Initial solution
One possible initial solution that we determine by inspection is:
>>> 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, Remark: from scipy v0.8 and above, you should rather use scipy.optimize.curve_fit() which takes
the function to minimize is the residuals (the difference between the data and the model): the model and the data as arguments, so you don’t need to define the residuals any more.
>>> def residuals(coeffs, y, t):

Going further
... return y - model(t, coeffs)
• Try with a more complex waveform (for instance waveform_2.npy) that contains three significant
So let’s get our solution by calling scipy.optimize.leastsq() with the following arguments: peaks. You must adapt the model which is now a sum of Gaussian functions instead of only one
Gaussian peak.
• the function to minimize
• In some cases, writing an explicit function to compute the Jacobian is faster than letting leastsq
• an initial solution
estimate it numerically. Create a function to compute the Jacobian of the residuals and use it as
• the additional arguments to pass to the function an input for leastsq.
>>> from scipy.optimize import leastsq • When we want to detect very small peaks in the signal, or when the initial guess is too far from a
>>> t = np.arange(len(waveform_1)) good solution, the result given by the algorithm is often not satisfying. Adding constraints to the
>>> x, flag = leastsq(residuals, x0, args=(waveform_1, t)) parameters of the model enables to overcome such limitations. An example of a priori knowledge
>>> print(x) we can add is the sign of our variables (which are all positive).
[ 2.70363341 27.82020742 15.47924562 3.05636228]
• See the solution.
And visualize the solution: • Further exercise: compare the result of scipy.optimize.leastsq() and what you can get with
scipy.optimize.fmin_slsqp() when adding boundary constraints.
6.11.3 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.
6.11.4 Example of solution for the image processing exercise: unmolten grains in
glass
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)
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 = plt.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))

(1699.875, 65.0)
6.12 Full code examples for the scipy chapter
6.12.1 Finding the minimum of a smooth function

Demos various methods to find the minimum of a function.
import numpy as np
def f(x):
return x**2 + 10*np.sin(x)
x = np.arange(-10, 10, 0.1)

plt.plot(x, f(x))
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.

Now find the minimum with a few methods
>>> bubbles_labels, bubbles_nb = ndimage.label(void)
>>> bubbles_areas = np.bincount(bubbles_labels.ravel())[1:] from scipy import optimize
>>> mean_bubble_size = bubbles_areas.mean()
>>> median_bubble_size = np.median(bubbles_areas) # The default (Nelder Mead)
>>> mean_bubble_size, median_bubble_size print(optimize.minimize(f, x0=0))
6.11. Summary exercises on scientific computing 220 6.12. Full code examples for the scipy chapter 221
Out:
fun: -7.945823375615215
hess_inv: array([[0.08589237]])
jac: array([-1.1920929e-06])
message: 'Optimization terminated successfully.'
nfev: 18
nit: 5
njev: 6
status: 0
success: True
x: array([-1.30644012])
print(optimize.minimize(f, x0=0, method="L-BFGS-B"))
Out:
fun: array([-7.94582338])
hess_inv: <1x1 LbfgsInvHessProduct with dtype=float64>
jac: array([-1.42108547e-06])
message: b'CONVERGENCE: NORM_OF_PROJECTED_GRADIENT_<=_PGTOL'
nfev: 12
nit: 5
status: 0
success: True
x: array([-1.30644013])

plt.show()
Total running time of the script: ( 0 minutes 0.016 seconds) Note: Click here to download the full example code

6.12.3 Resample a signal with scipy.signal.resample
scipy.signal.resample() uses FFT to resample a 1D signal.
6.12.2 Detrending a signal Generate a signal with 100 data point
scipy.signal.detrend() removes a linear trend. import numpy as np
Generate a random signal with a trend t = np.linspace(0, 5, 100)
x = np.sin(t)
import numpy as np
t = np.linspace(0, 5, 100) Downsample it by a factor of 4
x = t + np.random.normal(size=100)
from scipy import signal
Detrend x_resampled = signal.resample(x, 25)
from scipy import signal Plot

x_detrended = signal.detrend(x)
Plot plt.figure(figsize=(5, 4))
plt.plot(t, x, label='Original signal')
from matplotlib import pyplot as plt plt.plot(t[::4], x_resampled, 'ko', label='Resampled signal')
plt.plot(t, x, label="x") plt.legend(loc='best')
plt.plot(t, x_detrended, label="x_detrended") plt.show()
plt.legend(loc='best')
plt.show()
6.12. Full code examples for the scipy chapter 222 6.12. Full code examples for the scipy chapter 223

return -2*ypos
time_vec = np.linspace(0, 4, 40)

yvec = odeint(calc_derivative, 1, time_vec)
plt.plot(time_vec, yvec)
plt.xlabel('t: Time')
plt.ylabel('y: Position')
plt.tight_layout()
6.12.5 Comparing 2 sets of samples from Gaussians
6.12.4 Integrating a simple ODE

Solve the ODE dy/dt = -2y between t = 0..4, with the initial condition y(t=0) = 1.
import numpy as np
# Generates 2 sets of observations

samples1 = np.random.normal(0, size=1000)
samples2 = np.random.normal(1, size=1000)
# Compute a histogram of the sample

bins = np.linspace(-4, 4, 30)
import numpy as np histogram1, bins = np.histogram(samples1, bins=bins, density=True)
from scipy.integrate import odeint histogram2, bins = np.histogram(samples2, bins=bins, density=True)
def calc_derivative(ypos, time): plt.hist(samples1, bins=bins, density=True, label="Samples 1")

plt.hist(samples2, bins=bins, density=True, label="Samples 2") Note: Click here to download the full example code
plt.show()
6.12.7 Normal distribution: histogram and PDF
Explore the normal distribution: a histogram built from samples and the PDF (probability density
function).
6.12.6 Integrate the Damped spring-mass oscillator
import numpy as np
from scipy.integrate import odeint
from matplotlib import pyplot as plt import numpy as np
mass = 0.5 # kg # Sample from a normal distribution using numpy's random number generator
kspring = 4 # N/m samples = np.random.normal(size=10000)
cviscous = 0.4 # N s/m
# Compute a histogram of the sample
bins = np.linspace(-5, 5, 30)
eps = cviscous / (2 * mass * np.sqrt(kspring/mass)) histogram, bins = np.histogram(samples, bins=bins, density=True)
omega = np.sqrt(kspring / mass)
bin_centers = 0.5*(bins[1:] + bins[:-1])
def calc_deri(yvec, time, eps, omega): # Compute the PDF on the bin centers from scipy distribution object
return (yvec[1], -eps * omega * yvec[1] - omega **2 * yvec[0]) from scipy import stats
pdf = stats.norm.pdf(bin_centers)
time_vec = np.linspace(0, 10, 100)
yinit = (1, 0) from matplotlib import pyplot as plt
yarr = odeint(calc_deri, yinit, time_vec, args=(eps, omega)) plt.figure(figsize=(6, 4))
plt.plot(bin_centers, histogram, label="Histogram of samples")
plt.figure(figsize=(4, 3)) plt.plot(bin_centers, pdf, label="PDF")
plt.plot(time_vec, yarr[:, 0], label='y') plt.legend()
plt.plot(time_vec, yarr[:, 1], label="y'") plt.show()

6.12.8 Curve fitting plt.figure(figsize=(6, 4))

Demos a simple curve fitting plt.scatter(x_data, y_data, label='Data')
plt.plot(x_data, test_func(x_data, params[0], params[1]),
First generate some data label='Fitted function')
import numpy as np plt.legend(loc='best')
# Seed the random number generator for reproducibility plt.show()

np.random.seed(0)
x_data = np.linspace(-5, 5, num=50)

y_data = 2.9 * np.sin(1.5 * x_data) + np.random.normal(size=50)
# And plot it
plt.scatter(x_data, y_data)
6.12.9 Spectrogram, power spectral density

Demo spectrogram and power spectral density on a frequency chirp.
Now fit a simple sine function to the data
import numpy as np
from scipy import optimize from matplotlib import pyplot as plt
def test_func(x, a, b):

return a * np.sin(b * x) Generate a chirp signal
params, params_covariance = optimize.curve_fit(test_func, x_data, y_data, # Seed the random number generator
p0=[2, 2]) np.random.seed(0)
print(params) time_step = .01

time_vec = np.arange(0, 70, time_step)
Out:
# A signal with a small frequency chirp
[3.05931973 1.45754553] sig = np.sin(0.5 * np.pi * time_vec * (1 + .1 * time_vec))
And plot the resulting curve on the data plt.figure(figsize=(8, 5))

plt.plot(time_vec, sig)
Compute and plot the spectrogram

Compute and plot the power spectral density (PSD)
The spectrum of the signal on consecutive time windows
The power of the signal per frequency band
freqs, psd = signal.welch(sig)
freqs, times, spectrogram = signal.spectrogram(sig)
plt.semilogx(freqs, psd)
plt.imshow(spectrogram, aspect='auto', cmap='hot_r', origin='lower')
plt.title('PSD: power spectral density')
plt.title('Spectrogram')
plt.xlabel('Frequency')
plt.ylabel('Frequency band')
plt.ylabel('Power')
plt.xlabel('Time window')
plt.tight_layout()
plt.tight_layout()
6.12.10 A demo of 1D interpolation
plt.show()
Total running time of the script: ( 0 minutes 0.200 seconds) # Generate data
import numpy as np
np.random.seed(0)
Note: Click here to download the full example code measured_time = np.linspace(0, 1, 10)
noise = 1e-1 * (np.random.random(10)*2 - 1)
measures = np.sin(2 * np.pi * measured_time) + noise
# Interpolate it to new time points

from scipy.interpolate import interp1d
linear_interp = interp1d(measured_time, measures)
interpolation_time = np.linspace(0, 1, 50)
linear_results = linear_interp(interpolation_time)
cubic_interp = interp1d(measured_time, measures, kind='cubic')
cubic_results = cubic_interp(interpolation_time)
# Plot the data and the interpolation

plt.plot(measured_time, measures, 'o', ms=6, label='measures')
plt.plot(interpolation_time, linear_results, label='linear interp')
plt.plot(interpolation_time, cubic_results, label='cubic interp')
plt.legend()
plt.show()
6.12.11 Demo mathematical morphology

A basic demo of binary opening and closing.
# Load some data

from scipy import misc
face = misc.face(gray=True)
# Generate some binary data # Apply a variety of transformations

import numpy as np from scipy import ndimage
np.random.seed(0) from matplotlib import pyplot as plt
a = np.zeros((50, 50)) shifted_face = ndimage.shift(face, (50, 50))
a[10:-10, 10:-10] = 1 shifted_face2 = ndimage.shift(face, (50, 50), mode='nearest')
a += 0.25 * np.random.standard_normal(a.shape) rotated_face = ndimage.rotate(face, 30)
mask = a>=0.5 cropped_face = face[50:-50, 50:-50]
zoomed_face = ndimage.zoom(face, 2)
# Apply mathematical morphology zoomed_face.shape
from scipy import ndimage
opened_mask = ndimage.binary_opening(mask) plt.figure(figsize=(15, 3))
closed_mask = ndimage.binary_closing(opened_mask) plt.subplot(151)
plt.imshow(shifted_face, cmap=plt.cm.gray)
# Plot plt.axis('off')
plt.subplot(152)
plt.figure(figsize=(12, 3.5)) plt.imshow(shifted_face2, cmap=plt.cm.gray)
plt.subplot(141) plt.axis('off')
plt.imshow(a, cmap=plt.cm.gray)
plt.axis('off') plt.subplot(153)
plt.title('a') plt.imshow(rotated_face, cmap=plt.cm.gray)
plt.axis('off')
plt.subplot(142)
plt.imshow(mask, cmap=plt.cm.gray) plt.subplot(154)
plt.axis('off') plt.imshow(cropped_face, cmap=plt.cm.gray)
plt.title('mask') plt.axis('off')
plt.imshow(opened_mask, cmap=plt.cm.gray) plt.imshow(zoomed_face, cmap=plt.cm.gray)
plt.axis('off') plt.axis('off')
plt.title('opened_mask')
plt.subplots_adjust(wspace=.05, left=.01, bottom=.01, right=.99, top=.99)
plt.subplot(144)
plt.imshow(closed_mask, cmap=plt.cm.gray) plt.show()
plt.title('closed_mask')
plt.axis('off')
plt.show()

6.12.13 Demo connected components
Extracting and labeling connected components in a 2D array
import numpy as np
6.12.12 Plot geometrical transformations on images
Generate some binary data
Demo geometrical transformations of images.
np.random.seed(0)
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
plt.figure(figsize=(7, 3.5))
plt.imshow(sig)
plt.axis('off')
plt.title('sig')
plt.imshow(mask, cmap=plt.cm.gray)
plt.axis('off')
plt.title('mask')
Extract the 4th connected component, and crop the array around it
sl = ndimage.find_objects(labels==4)
plt.figure(figsize=(3.5, 3.5))
plt.imshow(sig[sl[0]])
plt.title('Cropped connected component')
plt.axis('off')
plt.show()
Label connected components

labels, nb = ndimage.label(mask)
plt.imshow(labels)
plt.title('label')
plt.axis('off')

Note: Click here to download the full example code # Plot the roots
roots = np.array([root.x, root2.x])
ax.plot(roots, f(roots), 'kv', label="Roots")
6.12.14 Minima and roots of a function
# Decorate the figure
Demos finding minima and roots of a function. ax.legend(loc='best')
ax.set_xlabel('x')
ax.set_ylabel('f(x)')
Define the function
ax.axhline(0, color='gray')
import numpy as np plt.show()
x = np.arange(-10, 10, 0.1)

def f(x):
return x**2 + 10*np.sin(x)
Find minima
from scipy import optimize
# Global optimization
grid = (-10, 10, 0.1)
xmin_global = optimize.brute(f, (grid, ))
print("Global minima found %s " % xmin_global)
# Constrain optimization
xmin_local = optimize.fminbound(f, 0, 10)
print("Local minimum found %s " % xmin_local)
Out:
Global minima found [-1.30641113]

Local minimum found 3.8374671194983834
Root finding
root = optimize.root(f, 1) # our initial guess is 1

print("First root found %s " % root.x) Total running time of the script: ( 0 minutes 0.084 seconds)
root2 = optimize.root(f, -2.5)
print("Second root found %s " % root2.x)
Out:
First root found [0.] 6.12.15 Plot filtering on images

Second root found [-2.47948183]
Demo filtering for denoising of images.
Plot function, minima, and roots

ax = fig.add_subplot(111)
# Plot the function

ax.plot(x, f(x), 'b-', label="f(x)")
# Plot the minima

xmins = np.array([xmin_global[0], xmin_local])
ax.plot(xmins, f(xmins), 'go', label="Minima")
A 2D image plot of the function

# Load some data
from scipy import misc Simple visualization in 2D
face = misc.face(gray=True)
face = face[:512, -512:] # crop out square on right import matplotlib.pyplot as plt
plt.figure()
# Apply a variety of filters plt.imshow(sixhump([xg, yg]), extent=[-2, 2, -1, 1], origin="lower")
from scipy import ndimage plt.colorbar()
import numpy as np
noisy_face = np.copy(face).astype(np.float)
noisy_face += face.std() * 0.5 * np.random.standard_normal(face.shape)
blurred_face = ndimage.gaussian_filter(noisy_face, sigma=3)
median_face = ndimage.median_filter(noisy_face, size=5)
wiener_face = signal.wiener(noisy_face, (5, 5))
plt.subplot(141)
plt.imshow(noisy_face, cmap=plt.cm.gray)
plt.axis('off')
plt.title('noisy')
plt.subplot(142)
plt.imshow(blurred_face, cmap=plt.cm.gray)
plt.axis('off')
plt.title('Gaussian filter')
plt.subplot(143)
plt.imshow(median_face, cmap=plt.cm.gray)
plt.axis('off')
plt.title('median filter')
plt.subplot(144)
plt.imshow(wiener_face, cmap=plt.cm.gray)
plt.title('Wiener filter')
plt.axis('off')
plt.show()
A 3D surface plot of the function
Total running time of the script: ( 0 minutes 0.382 seconds) from mpl_toolkits.mplot3d import Axes3D
fig = plt.figure()
ax = fig.add_subplot(111, projection='3d')
Note: Click here to download the full example code surf = ax.plot_surface(xg, yg, sixhump([xg, yg]), rstride=1, cstride=1,
cmap=plt.cm.jet, linewidth=0, antialiased=False)
6.12.16 Optimization of a two-parameter function ax.set_xlabel('x')

ax.set_ylabel('y')
import numpy as np ax.set_zlabel('f(x, y)')
ax.set_title('Six-hump Camelback function')
# Define the function that we are interested in

def sixhump(x):
return ((4 - 2.1*x[0]**2 + x[0]**4 / 3.) * x[0]**2 + x[0] * x[1]
+ (-4 + 4*x[1]**2) * x[1] **2)
# Make a grid to evaluate the function (for plotting)

x = np.linspace(-2, 2)
y = np.linspace(-1, 1)
xg, yg = np.meshgrid(x, y)

Find the minima
from scipy import optimize Note: Click here to download the full example code
x_min = optimize.minimize(sixhump, x0=[0, 0])
plt.figure() 6.12.17 Plotting and manipulating FFTs for filtering

# Show the function in 2D
Plot the power of the FFT of a signal and inverse FFT back to reconstruct a signal.
plt.imshow(sixhump([xg, yg]), extent=[-2, 2, -1, 1], origin="lower")
plt.colorbar() This example demonstrate scipy.fftpack.fft(), scipy.fftpack.fftfreq() and scipy.fftpack.
# And the minimum that we've found: ifft(). It implements a basic filter that is very suboptimal, and should not be used.
plt.scatter(x_min.x[0], x_min.x[1])
import numpy as np
plt.show() from scipy import fftpack
Generate the signal
# Seed the random number generator

np.random.seed(1234)
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))
plt.plot(time_vec, sig, label='Original signal')

plt.title('Peak frequency')
plt.plot(freqs[:8], power[pos_mask][:8])
plt.setp(axes, yticks=[])
# scipy.signal.find_peaks_cwt can also be used for more advanced

# peak detection
Compute and plot the power
# The FFT of the signal

sig_fft = fftpack.fft(sig)
# And the power (sig_fft is of complex dtype)

power = np.abs(sig_fft)**2
# The corresponding frequencies

sample_freq = fftpack.fftfreq(sig.size, d=time_step)
Remove all the high frequencies
We now remove all the high frequencies and transform back from frequencies to signal.
# Plot the FFT power
plt.figure(figsize=(6, 5)) high_freq_fft = sig_fft.copy()
plt.plot(sample_freq, power) high_freq_fft[np.abs(sample_freq) > peak_freq] = 0
plt.xlabel('Frequency [Hz]') filtered_sig = fftpack.ifft(high_freq_fft)
plt.ylabel('plower')
# Find the peak frequency: we can focus on only the positive frequencies plt.plot(time_vec, sig, label='Original signal')
pos_mask = np.where(sample_freq > 0) plt.plot(time_vec, filtered_sig, linewidth=3, label='Filtered signal')
freqs = sample_freq[pos_mask] plt.xlabel('Time [s]')
peak_freq = freqs[power[pos_mask].argmax()] plt.ylabel('Amplitude')
# Check that it does indeed correspond to the frequency that we generate plt.legend(loc='best')
# the signal with
np.allclose(peak_freq, 1./period)
# An inner plot to show the peak frequency

axes = plt.axes([0.55, 0.3, 0.3, 0.5])
Plot the data

plt.figure()
plt.plot(years, populations * 1e-3)
plt.xlabel('Year')
plt.ylabel('Population number ($\cdot10^3$)')
plt.legend(['hare', 'lynx', 'carrot'], loc=1)
Note This is actually a bad way of creating a filter: such brutal cut-off in frequency space does not
control distorsion on the signal.
Filters should be created using the scipy filter design code
plt.show()

Plot its periods
6.12.18 Solutions of the exercises for scipy
from scipy import fftpack
Note: Click here to download the full example code ft_populations = fftpack.fft(populations, axis=0)
frequencies = fftpack.fftfreq(populations.shape[0], years[1] - years[0])
periods = 1 / frequencies
Crude periodicity finding
plt.figure()
Discover the periods in evolution of animal populations (../../../../data/populations.txt) plt.plot(periods, abs(ft_populations) * 1e-3, 'o')
plt.xlim(0, 22)
Load the data plt.xlabel('Period')
plt.ylabel('Power ($\cdot10^3$)')
import numpy as np
plt.show()
data = np.loadtxt('../../../../data/populations.txt')
years = data[:, 0]
populations = data[:, 1:]
There’s probably a period of around 10 years (obvious from the plot), but for this crude a method,
there’s not enough data to say much more. Fitting it to a periodic function
Total running time of the script: ( 0 minutes 0.027 seconds) from scipy import optimize
def yearly_temps(times, avg, ampl, time_offset):
return (avg
+ ampl * np.cos((times + time_offset) * 2 * np.pi / times.max()))
res_max, cov_max = optimize.curve_fit(yearly_temps, months,

Curve fitting: temperature as a function of month of the year temp_max, [20, 10, 0])
res_min, cov_min = optimize.curve_fit(yearly_temps, months,
We have the min and max temperatures in Alaska for each months of the year. We would like to find a temp_min, [-40, 20, 0])
function to describe this yearly evolution.
For this, we will fit a periodic function.
Plotting the fit
The data days = np.linspace(0, 12, num=365)
import numpy as np plt.figure()

plt.plot(months, temp_max, 'ro')
temp_max = np.array([17, 19, 21, 28, 33, 38, 37, 37, 31, 23, 19, 18]) plt.plot(days, yearly_temps(days, *res_max), 'r-')
temp_min = np.array([-62, -59, -56, -46, -32, -18, -9, -13, -25, -46, -52, -58]) plt.plot(months, temp_min, 'bo')
plt.plot(days, yearly_temps(days, *res_min), 'b-')
import matplotlib.pyplot as plt plt.xlabel('Month')
months = np.arange(12) plt.ylabel('Temperature ($^\circ$C)')
plt.plot(months, temp_max, 'ro')
plt.plot(months, temp_min, 'bo') plt.show()
plt.xlabel('Month')
plt.ylabel('Min and max temperature')

Prepare an Gaussian convolution kernel
Note: Click here to download the full example code # First a 1-D Gaussian
t = np.linspace(-10, 10, 30)
bump = np.exp(-0.1*t**2)
Simple image blur by convolution with a Gaussian kernel bump /= np.trapz(bump) # normalize the integral to 1
Blur an an image (../../../../data/elephant.png) using a Gaussian kernel. # make a 2-D kernel out of it
kernel = bump[:, np.newaxis] * bump[np.newaxis, :]
Convolution is easy to perform with FFT: convolving two signals boils down to multiplying their FFTs
(and performing an inverse FFT).
Implement convolution via FFT
import numpy as np
from scipy import fftpack
import matplotlib.pyplot as plt # Padded fourier transform, with the same shape as the image
# We use :func:`scipy.signal.fftpack.fft2` to have a 2D FFT
kernel_ft = fftpack.fft2(kernel, shape=img.shape[:2], axes=(0, 1))
The original image
# convolve
img_ft = fftpack.fft2(img, axes=(0, 1))
# read image
# the 'newaxis' is to match to color direction
img = plt.imread('../../../../data/elephant.png')
img2_ft = kernel_ft[:, :, np.newaxis] * img_ft
plt.figure()
img2 = fftpack.ifft2(img2_ft, axes=(0, 1)).real
plt.imshow(img)
# clip values to range
img2 = np.clip(img2, 0, 1)
# plot output
plt.figure()
plt.imshow(img2)
Further exercise (only if you are familiar with this stuff): Note that we still have a decay to zero at the border of the image. Using scipy.ndimage.
gaussian_filter() would get rid of this artifact
A “wrapped border” appears in the upper left and top edges of the image. This is because the padding
is not done correctly, and does not take the kernel size into account (so the convolution “flows out of plt.show()
bounds of the image”). Try to remove this artifact.
A function to do it: scipy.signal.fftconvolve()
The above exercise was only for didactic reasons: there exists a function in scipy that will do Note: Click here to download the full example code
this for us, and probably do a better job: scipy.signal.fftconvolve()
from scipy import signal Image denoising by FFT

# mode='same' is there to enforce the same output shape as input arrays
# (ie avoid border effects) Denoise an image (../../../../data/moonlanding.png) by implementing a blur with an FFT.
img3 = signal.fftconvolve(img, kernel[:, :, np.newaxis], mode='same')
Implements, via FFT, the following convolution:
plt.figure() ∫︁
plt.imshow(img3)
𝑓1 (𝑡) = 𝑑𝑡′ 𝐾(𝑡 − 𝑡′ )𝑓0 (𝑡′ )
𝑓˜1 (𝜔) = 𝐾(𝜔)

˜ 𝑓˜0 (𝜔)
Read and plot the image
import numpy as np
im = plt.imread('../../../../data/moonlanding.png').astype(float)
plt.figure()
plt.imshow(im, plt.cm.gray)
plt.title('Original image')
Compute the 2d FFT of the input image Filter in FFT
from scipy import fftpack # In the lines following, we'll make a copy of the original spectrum and
im_fft = fftpack.fft2(im) # truncate coefficients.
# Show the results # Define the fraction of coefficients (in each direction) we keep
keep_fraction = 0.1
def plot_spectrum(im_fft):
from matplotlib.colors import LogNorm # Call ff a copy of the original transform. Numpy arrays have a copy
# A logarithmic colormap # method for this purpose.
plt.imshow(np.abs(im_fft), norm=LogNorm(vmin=5)) im_fft2 = im_fft.copy()
plt.colorbar()
# Set r and c to be the number of rows and columns of the array.
plt.figure() r, c = im_fft2.shape
plot_spectrum(im_fft)
plt.title('Fourier transform') # Set to zero all rows with indices between r*keep_fraction and
# r*(1-keep_fraction):
im_fft2[int(r*keep_fraction):int(r*(1-keep_fraction))] = 0
# Similarly with the columns:

im_fft2[:, int(c*keep_fraction):int(c*(1-keep_fraction))] = 0
plt.figure()
plot_spectrum(im_fft2)
plt.title('Filtered Spectrum')
Reconstruct the final image Easier and better: scipy.ndimage.gaussian_filter()
# Reconstruct the denoised image from the filtered spectrum, keep only the Implementing filtering directly with FFTs is tricky and time consuming. We can use the
# real part for display. Gaussian filter from scipy.ndimage
im_new = fftpack.ifft2(im_fft2).real
plt.figure() im_blur = ndimage.gaussian_filter(im, 4)
plt.imshow(im_new, plt.cm.gray)
plt.title('Reconstructed Image') plt.figure()
plt.imshow(im_blur, plt.cm.gray)
plt.title('Blurred image')
plt.show()
CHAPTER 7
Getting help and finding documentation
Total running time of the script: ( 0 minutes 0.211 seconds) Author: Emmanuelle Gouillart
See also: 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:
References to go further
• In Ipython, help function opens the docstring of the function. Only type the beginning of the
• Some chapters of the advanced and the packages and applications parts of the scipy lectures
function’s name and use tab completion to display the matching functions.
• The scipy cookbook
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; how-
ever one can always open a second Ipython shell just to display help and docstrings. . .
• Numpy’s and Scipy’s 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://
numpy.org/doc/stable/reference/ and http://docs.scipy.org/doc/scipy/reference/).
6.12. Full code examples for the scipy chapter 258 259
Tutorials on various topics as well as the complete API with all docstrings are found on this website. (continued from previous page)
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.
• Numpy’s and Scipy’s documentation is enriched and updated on a regular basis by users on a
wiki http://numpy.org/doc/stable/. As a result, some docstrings are clearer or more detailed on • If everything listed above fails (and Google doesn’t have the answer). . . don’t despair! Write to the
the wiki, and you may want to read directly the documentation on the wiki instead of the official mailing-list suited to your problem: you should have a quick answer if you describe your problem
documentation website. Note that anyone can create an account on the wiki and write better well. Experts on scientific python often give very enlightening explanations on the mailing-list.
documentation; this is an easy way to contribute to an open-source project and improve the tools
– Numpy discussion ([email protected]): all about numpy arrays, manipulating
you are using!
them, indexation questions, etc.
• The SciPy Cookbook https://scipy-cookbook.readthedocs.io gives recipes on many com-
– SciPy Users List ([email protected]): scientific computing with Python, high-level data
mon problems frequently encountered, such as fitting data points, solving ODE, etc.
processing, in particular with the scipy package.
– [email protected] for plotting with matplotlib.
• Matplotlib’s website http://matplotlib.org/ 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.
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 [3]: import numpy as np

In [4]: %psearch np.diag*
np.diag
np.diagflat
np.diagonal
• numpy.lookfor looks for keywords inside the docstrings of specified modules.
In [45]: numpy.lookfor('convolution')
Search results for 'convolution'
--------------------------------
numpy.convolve
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')
260 261
This part of the Scipy lecture notes is dedicated to advanced usage. It strives to educate the proficient
Python coder to be an expert and tackles various specific topics.
Part II
Advanced topics
262 263
– Decorators implemented as classes and as functions

– Copying the docstring and other attributes of the original function
– Examples in the standard library
– Deprecation of functions
8
– A while-loop removing decorator
– A plugin registration system
• Context managers
– Catching exceptions
CHAPTER
– Using generators to define context managers
8.1 Iterators, generator expressions and generators

8.1.1 Iterators
Advanced Python Constructs 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
Author Zbigniew Jędrzejewski-Szmek method, which, when called, returns the next item in the sequence, and when there’s nothing to return,
raises the StopIteration exception.
This section covers 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 An iterator object allows to loop just once. It holds the state (position) of a single iteration, or from the
complicated programs or libraries, but not in the sense of being particularly specialized, or particularly other side, each loop over a sequence requires a single iterator object. This means that we can iterate
complicated. 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.
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 Calling the __iter__ method on a container to create an iterator object is the most straightforward way
be implemented through clever external modules. to get hold of an iterator. The iter function does that for us, saving a few keystrokes.
The process of developing the Python programming language, its syntax, is very transparent; proposed >>> nums = [1, 2, 3] # note that ... varies: these are different objects
changes are evaluated from various angles and discussed via Python Enhancement Proposals — PEPs. >>> iter(nums)
As a result, features described in this chapter were added after it was shown that they indeed solve real <...iterator object at ...>
>>> nums.__iter__()
problems and that their use is as simple as possible.
<...iterator object at ...>
>>> nums.__reversed__()
Chapter contents <...reverseiterator object at ...>
• Iterators, generator expressions and generators >>> it = iter(nums)

>>> next(it)
– Iterators 1
>>> next(it)
– Generator expressions 2
– Generators >>> next(it)
3
– Bidirectional communication >>> next(it)
– Chaining generators File "<stdin>", line 1, in <module>
• Decorators StopIteration
– Replacing or tweaking the original object When used in a loop, StopIteration is swallowed and causes the loop to finish. But with explicit
264 8.1. Iterators, generator expressions and generators 265

invocation, we can see that once the iterator is exhausted, accessing it raises an exception. 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
Using the for..in loop also uses the __iter__ method. This allows us to transparently start the iteration
this function is suspended.
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 >>> def f():
called __iter__ which returns the iterator (self). ... yield 1
... yield 2
Support for iteration is pervasive in Python: all sequences and unordered containers in the standard
>>> f()
library allow this. The concept is also stretched to other things: e.g. file objects support iteration over <generator object f at 0x...>
lines. >>> gen = f()
>>> next(gen)
>>> f = open('/etc/fstab')
1
>>> f is f.__iter__()
>>> next(gen)
True
2
>>> next(gen)
The file is an iterator itself and it’s __iter__ method doesn’t create a separate object: only a single Traceback (most recent call last):
thread of sequential access is allowed. File "<stdin>", line 1, in <module>
StopIteration
8.1.2 Generator expressions
Let’s go over the life of the single invocation of the generator function.
A second way in which iterator objects are created is through generator expressions, the basis for list
comprehensions. To increase clarity, a generator expression must always be enclosed in parentheses >>> def f():
or an expression. If round parentheses are used, then a generator iterator is created. If rectangular ... print("-- start --")
parentheses are used, the process is short-circuited and we get a list. ... yield 3
... print("-- middle --")
>>> (i for i in nums) ... yield 4
<generator object <genexpr> at 0x...> ... print("-- finished --")
>>> [i for i in nums] >>> gen = f()
[1, 2, 3] >>> next(gen)
>>> list(i for i in nums) -- start --
[1, 2, 3] 3
>>> next(gen)
The list comprehension syntax also extends to dictionary and set comprehensions. A set is cre- -- middle --
4
ated when the generator expression is enclosed in curly braces. A dict is created when the generator
>>> next(gen)
expression contains “pairs” of the form key:value: -- finished --
>>> {i for i in range(3)} Traceback (most recent call last):
set([0, 1, 2]) ...
>>> {i:i**2 for i in range(3)} StopIteration
{0: 0, 1: 1, 2: 4}
Contrary to a normal function, where executing f() would immediately cause the first print to be
One gotcha should be mentioned: in old Pythons the index variable (i) would leak, and in versions >= executed, gen is assigned without executing any statements in the function body. Only when gen.
3 this is fixed. 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.
8.1.3 Generators
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
Generators 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
A generator is a function that produces a sequence of results instead of a single value. value.
David Beazley — A Curious Course on Coroutines and Concurrency 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
A third way to create iterator objects is to call a generator function. A generator is a function containing done with next methods. Nevertheless, using a function and having the interpreter perform its magic
the keyword yield. It must be noted that the mere presence of this keyword completely changes the to create an iterator has advantages. A function can be much shorter than the definition of a class with
nature of the function: this yield statement doesn’t have to be invoked, or even reachable, but causes the required next and __iter__ methods. What is more important, it is easier for the author of the
the function to be marked as a generator. When a normal function is called, the instructions contained in generator to understand the state which is kept in local variables, as opposed to instance attributes,
the body start to be executed. When a generator is called, the execution stops before the first instruction which have to be used to pass data between consecutive invocations of next on an iterator object.
in the body. An invocation of a generator function creates a generator object, adhering to the iterator A broader question is why are iterators useful? When an iterator is used to power a loop, the loop
protocol. As with normal function invocations, concurrent and recursive invocations are allowed. becomes very simple. The code to initialise the state, to decide if the loop is finished, and to find the
8.1. Iterators, generator expressions and generators 266 8.1. Iterators, generator expressions and generators 267
next value is extracted into a separate place. This highlights the body of the loop — the interesting (continued from previous page)
part. In addition, it is possible to reuse the iterator code in other places. --yielding 1--
1
8.1.4 Bidirectional communication >>> it.throw(IndexError)
--yield raised IndexError()--
Each yield statement causes a value to be passed to the caller. This is the reason for the introduction --yielding 2--
of generators by PEP 255 (implemented in Python 2.2). But communication in the reverse direction is 2
also useful. One obvious way would be some external state, either a global variable or a shared mutable >>> it.close()
object. Direct communication is possible thanks to PEP 342 (implemented in 2.5). It is achieved --closing--
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 next or __next__?
an exception into the generator.
In Python 2.x, the iterator method to retrieve the next value is called next. It is invoked implicitly
The first of the new methods is send(value), which is similar to next(), but passes value into the through the global function next, which means that it should be called __next__. Just like the global
generator to be used for the value of the yield expression. In fact, g.next() and g.send(None) are function iter calls __iter__. This inconsistency is corrected in Python 3.x, where it.next becomes
equivalent. it.__next__. For other generator methods — send and throw — the situation is more complicated,
The second of the new methods is throw(type, value=None, traceback=None) which is equivalent to: because they are not called implicitly by the interpreter. Nevertheless, there’s a proposed syntax
extension to allow continue to take an argument which will be passed to send of the loop’s iterator.
raise type, value, traceback If this extension is accepted, it’s 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.
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 8.1.5 Chaining generators
is suggestive of putting the exception in another location, and is associated with exceptions in other
languages.
Note: This is a preview of PEP 380 (not yet implemented, but accepted for Python 3.3).
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
Let’s say we are writing a generator and we want to yield a number of values generated by a second
the throw() method. In either case, such an exception propagates in the standard manner: it can
generator, a subgenerator. If yielding of values is the only concern, this can be performed without
be intercepted by an except or finally clause, or otherwise it causes the execution of the generator
much difficulty using a loop such as
function to be aborted and propagates in the caller.
For completeness’ sake, it’s worth mentioning that generator iterators also have a close() method, subgen = some_other_generator()
for v in subgen:
which can be used to force a generator that would otherwise be able to provide more values to finish
yield v
immediately. It allows the generator __del__ method to destroy objects holding the state of generator.
Let’s define a generator which just prints what is passed in through send and throw. 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
>>> import itertools
try..except..finally structure similar to the one defined in the previous section to “debug” the generator
>>> def g():
... print('--start--') function. Such code is provided in PEP 380#id13, here it suffices to say that new syntax to properly
... for i in itertools.count(): yield from a subgenerator is being introduced in Python 3.3:
... print('--yielding %i --' % i)
yield from some_other_generator()
... try:
... ans = yield i
... except GeneratorExit: This behaves like the explicit loop above, repeatedly yielding values from some_other_generator until
... print('--closing--') it is exhausted, but also forwards send, throw and close to the subgenerator.
... raise
... except Exception as e:
... print('--yield raised %r --' % e) 8.2 Decorators
... else:
... print('--yield returned %s --' % ans)
Summary
>>> it = g()
>>> next(it) This amazing feature appeared in the language almost apologetically and with concern that it might
--start-- not be that useful.
--yielding 0--
0 Bruce Eckel — An Introduction to Python Decorators
>>> it.send(11)
--yield returned 11--
Since functions and classes are objects, they can be passed around. Since they are mutable objects, they
8.1. Iterators, generator expressions and generators 268 8.2. Decorators 269
can be modified. The act of altering a function or class object after it has been constructed but before 8.2.2 Decorators implemented as classes and as functions
is is bound to its name is called decorating.
The only requirement on decorators is that they can be called with a single argument. This means that
There are two things hiding behind the name “decorator” — one is the function which does the work of decorators can be implemented as normal functions, or as classes with a __call__ method, or in theory,
decorating, i.e. performs the real work, and the other one is the expression adhering to the decorator even as lambda functions.
syntax, i.e. an at-symbol and the name of the decorating function.
Let’s compare the function and class approaches. The decorator expression (the part after @) can be
Function can be decorated by using the decorator syntax for functions: 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
@decorator # · can be used in those two cases:
def function(): # ¶
pass >>> def simple_decorator(function):
... print("doing decoration")
• A function is defined in the standard way. ¶ ... return function
>>> @simple_decorator
• An expression starting with @ placed before the function definition is the decorator ·. The part ... def function():
after @ must be a simple expression, usually this is just the name of a function or class. This part is ... print("inside function")
evaluated first, and after the function defined below is ready, the decorator is called with the newly doing decoration
defined function object as the single argument. The value returned by the decorator is attached to >>> function()
the original name of the function. inside function
Decorators can be applied to functions and to classes. For classes the semantics are identical — the >>> def decorator_with_arguments(arg):
original class definition is used as an argument to call the decorator and whatever is returned is assigned ... print("defining the decorator")
under the original name. ... def _decorator(function):
... # in this inner function, arg is available too
Before the decorator syntax was implemented (PEP 318), it was possible to achieve the same effect by
... print("doing decoration, %r " % arg)
assigning the function or class object to a temporary variable and then invoking the decorator explicitly ... return function
and then assigning the return value to the name of the function. This sounds like more typing, and it ... return _decorator
is, and also the name of the decorated function doubling as a temporary variable must be used at least >>> @decorator_with_arguments("abc")
three times, which is prone to errors. Nevertheless, the example above is equivalent to: ... def function():
... print("inside function")
def function(): # ¶ defining the decorator
pass doing decoration, 'abc'
function = decorator(function) # · >>> function()
inside function
Decorators can be stacked — the order of application is bottom-to-top, or inside-out. The semantics
are such that the originally defined function is used as an argument for the first decorator, whatever is The two trivial decorators above fall into the category of decorators which return the original function.
returned by the first decorator is used as an argument for the second decorator, . . . , and whatever is If they were to return a new function, an extra level of nestedness would be required. In the worst case,
returned by the last decorator is attached under the name of the original function. three levels of nested functions.
The decorator syntax was chosen for its readability. Since the decorator is specified before the header >>> def replacing_decorator_with_args(arg):
of the function, it is obvious that its is not a part of the function body and its clear that it can only ... print("defining the decorator")
operate on the whole function. Because the expression is prefixed with @ is stands out and is hard to ... def _decorator(function):
miss (“in your face”, according to the PEP :) ). When more than one decorator is applied, each one is ... # in this inner function, arg is available too
placed on a separate line in an easy to read way. ... print("doing decoration, %r " % arg)
... def _wrapper(*args, **kwargs):
... print("inside wrapper, %r %r " % (args, kwargs))
8.2.1 Replacing or tweaking the original object ... return function(*args, **kwargs)
Decorators can either return the same function or class object or they can return a completely different ... return _wrapper
object. In the first case, the decorator can exploit the fact that function and class objects are mutable ... return _decorator
and add attributes, e.g. add a docstring to a class. A decorator might do something useful even without >>> @replacing_decorator_with_args("abc")
... def function(*args, **kwargs):
modifying the object, for example register the decorated class in a global registry. In the second case,
... print("inside function, %r %r " % (args, kwargs))
virtually anything is possible: when something different is substituted for the original function or class,
... return 14
the new object can be completely different. Nevertheless, such behaviour is not the purpose of decorators: defining the decorator
they are intended to tweak the decorated object, not do something unpredictable. Therefore, when a doing decoration, 'abc'
function is “decorated” by replacing it with a different function, the new function usually calls the original >>> function(11, 12)
function, after doing some preparatory work. Likewise, when a class is “decorated” by replacing if with inside wrapper, (11, 12) {}
a new class, the new class is usually derived from the original class. When the purpose of the decorator inside function, (11, 12) {}
is to do something “every time”, like to log every call to a decorated function, only the second type of 14
decorators can be used. On the other hand, if the first type is sufficient, it is better to use it, because it
is simpler. The _wrapper function is defined to accept all positional and keyword arguments. In general we cannot
know what arguments the decorated function is supposed to accept, so the wrapper function just passes
8.2. Decorators 270 8.2. Decorators 271

everything to the wrapped function. One unfortunate consequence is that the apparent argument list is 8.2.3 Copying the docstring and other attributes of the original function
misleading.
When a new function is returned by the decorator to replace the original function, an unfortunate
Compared to decorators defined as functions, complex decorators defined as classes are simpler. When consequence is that the original function name, the original docstring, the original argument list are
an object is created, the __init__ method is only allowed to return None, and the type of the created lost. Those attributes of the original function can partially be “transplanted” to the new function
object cannot be changed. This means that when a decorator is defined as a class, it doesn’t make by setting __doc__ (the docstring), __module__ and __name__ (the full name of the function), and
much sense to use the argument-less form: the final decorated object would just be an instance of the __annotations__ (extra information about arguments and the return value of the function available in
decorating class, returned by the constructor call, which is not very useful. Therefore it’s enough to Python 3). This can be done automatically by using functools.update_wrapper.
discuss class-based decorators where arguments are given in the decorator expression and the decorator
__init__ method is used for decorator construction.
functools.update_wrapper(wrapper, wrapped)
>>> class decorator_class(object):
... def __init__(self, arg): “Update a wrapper function to look like the wrapped function.”
... # this method is called in the decorator expression
>>> import functools
... print("in decorator init, %s " % arg)
>>> def replacing_decorator_with_args(arg):
... self.arg = arg
... print("defining the decorator")
... def __call__(self, function):
... def _decorator(function):
... # this method is called to do the job
... print("doing decoration, %r " % arg)
... print("in decorator call, %s " % self.arg)
... def _wrapper(*args, **kwargs):
... return function
... print("inside wrapper, %r %r " % (args, kwargs))
>>> deco_instance = decorator_class('foo')
... return function(*args, **kwargs)
in decorator init, foo
... return functools.update_wrapper(_wrapper, function)
>>> @deco_instance
... return _decorator
... def function(*args, **kwargs):
>>> @replacing_decorator_with_args("abc")
... print("in function, %s %s " % (args, kwargs))
... def function():
in decorator call, foo
... "extensive documentation"
>>> function()
... print("inside function")
in function, () {}
... return 14
defining the decorator
Contrary to normal rules (PEP 8) decorators written as classes behave more like functions and therefore doing decoration, 'abc'
their name often starts with a lowercase letter. >>> function
<function function at 0x...>
In reality, it doesn’t make much sense to create a new class just to have a decorator which returns the
>>> print(function.__doc__)
original function. Objects are supposed to hold state, and such decorators are more useful when the extensive documentation
decorator returns a new object.
>>> class replacing_decorator_class(object):

... def __init__(self, arg): One important thing is missing from the list of attributes which can be copied to the replacement
... # this method is called in the decorator expression function: the argument list. The default values for arguments can be modified through the __defaults__,
... print("in decorator init, %s " % arg) __kwdefaults__ attributes, but unfortunately the argument list itself cannot be set as an attribute. This
... self.arg = arg means that help(function) will display a useless argument list which will be confusing for the user of
... def __call__(self, function): the function. An effective but ugly way around this problem is to create the wrapper dynamically, using
... # this method is called to do the job eval. This can be automated by using the external decorator module. It provides support for the
... print("in decorator call, %s " % self.arg) decorator decorator, which takes a wrapper and turns it into a decorator which preserves the function
... self.function = function signature.
... return self._wrapper
... def _wrapper(self, *args, **kwargs): To sum things up, decorators should always use functools.update_wrapper or some other means of
... print("in the wrapper, %s %s " % (args, kwargs)) copying function attributes.
... return self.function(*args, **kwargs)
>>> deco_instance = replacing_decorator_class('foo')
in decorator init, foo
8.2.4 Examples in the standard library
>>> @deco_instance First, it should be mentioned that there’s a number of useful decorators available in the standard library.
... def function(*args, **kwargs): There are three decorators which really form a part of the language:
... print("in function, %s %s " % (args, kwargs))
in decorator call, foo • classmethod causes a method to become a “class method”, which means that it can be invoked
>>> function(11, 12) without creating an instance of the class. When a normal method is invoked, the interpreter inserts
in the wrapper, (11, 12) {} the instance object as the first positional parameter, self. When a class method is invoked, the
in function, (11, 12) {} class itself is given as the first parameter, often called cls.
A decorator like this can do pretty much anything, since it can modify the original function object and Class methods are still accessible through the class’ namespace, so they don’t pollute the module’s
mangle the arguments, call the original function or not, and afterwards mangle the return value. namespace. Class methods can be used to provide alternative constructors:

To make everything crystal clear, let’s define a “debug” example:

class Array(object):
def __init__(self, data): >>> class D(object):
self.data = data ... @property
... def a(self):
@classmethod ... print("getting 1")
def fromfile(cls, file): ... return 1
data = numpy.load(file) ... @a.setter
return cls(data) ... def a(self, value):
... print("setting %r " % value)
This is cleaner than using a multitude of flags to __init__. ... @a.deleter
... def a(self):
• staticmethod is applied to methods to make them “static”, i.e. basically a normal function, but
... print("deleting")
accessible through the class namespace. This can be useful when the function is only needed inside >>> D.a
this class (its name would then be prefixed with _), or when we want the user to think of the <property object at 0x...>
method as connected to the class, despite an implementation which doesn’t require this. >>> D.a.fget
<function ...>
• property is the pythonic answer to the problem of getters and setters. A method decorated with
>>> D.a.fset
property becomes a getter which is automatically called on attribute access.
<function ...>
>>> class A(object): >>> D.a.fdel
... @property <function ...>
... def a(self): >>> d = D() # ... varies, this is not the same à` function
... "an important attribute" >>> d.a
... return "a value" getting 1
>>> A.a 1
<property object at 0x...> >>> d.a = 2
>>> A().a setting 2
'a value' >>> del d.a
deleting
>>> d.a
In this example, A.a is an read-only attribute. It is also documented: help(A) includes the getting 1
docstring for attribute a taken from the getter method. Defining a as a property allows it to be a 1
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 Properties are a bit of a stretch for the decorator syntax. One of the premises of the decorator
syntax is preferred: 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.
class Rectangle(object):
def __init__(self, edge): Some newer examples include:
self.edge = edge • functools.lru_cache memoizes an arbitrary function maintaining a limited cache of argu-
ments:answer pairs (Python 3.2)
@property
def area(self): • functools.total_ordering is a class decorator which fills in missing ordering methods (__lt__,
"""Computed area. __gt__, __le__, . . . ) based on a single available one (Python 2.7).
Setting this updates the edge length to the proper value.

""" 8.2.5 Deprecation of functions
return self.edge**2 Let’s say we want to print a deprecation warning on stderr on the first invocation of a function we don’t
like anymore. If we don’t want to modify the function, we can use a decorator:
@area.setter
def area(self, area): class deprecated(object):
self.edge = area ** 0.5 """Print a deprecation warning once on first use of the function.
The way that this works, is that the property decorator replaces the getter method with a property >>> @deprecated() # doctest: +SKIP
object. This object in turn has three methods, getter, setter, and deleter, which can be used ... def f():
as decorators. Their job is to set the getter, setter and deleter of the property object (stored as ... pass
attributes fget, fset, and fdel). The getter can be set like in the example above, when creating >>> f() # doctest: +SKIP
the object. When defining the setter, we already have the property object under area, and we add f is deprecated
"""
the setter to it by using the setter method. All this happens when we are creating the class.
def __call__(self, func):
Afterwards, when an instance of the class has been created, the property object is special. When the self.func = func
interpreter executes attribute access, assignment, or deletion, the job is delegated to the methods self.count = 0
of the property object. return self._wrapper

(continued from previous page) 8.2.7 A plugin registration system

def _wrapper(self, *args, **kwargs): This is a class decorator which doesn’t modify the class, but just puts it in a global registry. It falls into
self.count += 1
the category of decorators returning the original object:
if self.count == 1:
print(self.func.__name__, 'is deprecated') class WordProcessor(object):
return self.func(*args, **kwargs) PLUGINS = []
def process(self, text):
It can also be implemented as a function: for plugin in self.PLUGINS:
text = plugin().cleanup(text)
def deprecated(func): return text
"""Print a deprecation warning once on first use of the function.
@classmethod
>>> @deprecated # doctest: +SKIP def plugin(cls, plugin):
... def f(): cls.PLUGINS.append(plugin)
... pass
>>> f() # doctest: +SKIP @WordProcessor.plugin
f is deprecated class CleanMdashesExtension(object):
""" def cleanup(self, text):
count = [0] return text.replace('—', u'\N{em dash}')
def wrapper(*args, **kwargs):
count[0] += 1
Here we use a decorator to decentralise the registration of plugins. We call our decorator with a noun,
if count[0] == 1:
print(func.__name__, 'is deprecated')
instead of a verb, because we use it to declare that our class is a plugin for WordProcessor. Method
return func(*args, **kwargs) plugin simply appends the class to the list of plugins.
return wrapper A word about the plugin itself: it replaces HTML entity for em-dash with a real Unicode em-dash
character. It exploits the unicode literal notation to insert a character by using its name in the unicode
database (“EM DASH”). If the Unicode character was inserted directly, it would be impossible to
8.2.6 A while-loop removing decorator distinguish it from an en-dash in the source of a program.
Let’s say we have function which returns a lists of things, and this list created by running a loop. If we See also:
don’t know how many objects will be needed, the standard way to do this is something like:
More examples and reading
def find_answers():
answers = [] • PEP 318 (function and method decorator syntax)
while True: • PEP 3129 (class decorator syntax)
ans = look_for_next_answer()
if ans is None: • http://wiki.python.org/moin/PythonDecoratorLibrary
break
answers.append(ans) • https://docs.python.org/dev/library/functools.html
return answers • http://pypi.python.org/pypi/decorator
This is fine, as long as the body of the loop is fairly compact. Once it becomes more complicated, as • Bruce Eckel
often happens in real code, this becomes pretty unreadable. We could simplify this by using yield – Decorators I: Introduction to Python Decorators
statements, but then the user would have to explicitly call list(find_answers()).
– Python Decorators II: Decorator Arguments
We can define a decorator which constructs the list for us:
– Python Decorators III: A Decorator-Based Build System
def vectorized(generator_func):
def wrapper(*args, **kwargs):
return list(generator_func(*args, **kwargs)) 8.3 Context managers
return functools.update_wrapper(wrapper, generator_func)
A context manager is an object with __enter__ and __exit__ methods which can be used in the with
Our function then becomes: statement:
@vectorized with manager as var:

def find_answers(): do_something(var)
while True:
ans = look_for_next_answer() is in the simplest case equivalent to
if ans is None:
break var = manager.__enter__()
yield ans try:
do_something(var)
8.2. Decorators 276 8.3. Context managers 277

(continued from previous page) • contextlib.closing å the same as the example above, call close
finally: • parallel programming
manager.__exit__()
– concurrent.futures.ThreadPoolExecutor å invoke in parallel then kill thread pool (py
In other words, the context manager protocol defined in PEP 343 permits the extraction of the boring >= 3.2)
part of a try..except..finally structure into a separate class leaving only the interesting do_something – concurrent.futures.ProcessPoolExecutor å invoke in parallel then kill process pool (py
block. >= 3.2)
1. The __enter__ method is called first. It can return a value which will be assigned to var. The – nogil å solve the GIL problem temporarily (cython only :( )
as-part is optional: if it isn’t 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 8.3.1 Catching exceptions
successfully 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, When an exception is thrown in the with-block, it is passed as arguments to __exit__. Three arguments
the information about the exception is passed to __exit__, which is described below in the next are used, the same as returned by sys.exc_info(): type, value, traceback. When no exception is thrown,
subsection. In the normal case, exceptions can be ignored, just like in a finally clause, and will None is used for all three arguments. The context manager can “swallow” the exception by returning a
be rethrown after __exit__ is finished. true value from __exit__. Exceptions can be easily ignored, because if __exit__ doesn’t use return
and just falls of the end, None is returned, a false value, and therefore the exception is rethrown after
Let’s say we want to make sure that a file is closed immediately after we are done writing to it: __exit__ is finished.
>>> class closing(object): The ability to catch exceptions opens interesting possibilities. A classic example comes from unit-tests
... def __init__(self, obj): — we want to make sure that some code throws the right kind of exception:
... self.obj = obj
... def __enter__(self): class assert_raises(object):
... return self.obj # based on pytest and unittest.TestCase
... def __exit__(self, *args): def __init__(self, type):
... self.obj.close() self.type = type
>>> with closing(open('/tmp/file', 'w')) as f: def __enter__(self):
... f.write('the contents\n') pass
def __exit__(self, type, value, traceback):
Here we have made sure that the f.close() is called when the with block is exited. Since closing files is if type is None:
such a common operation, the support for this is already present in the file class. It has an __exit__ raise AssertionError('exception expected')
if issubclass(type, self.type):
method which calls close and can be used as a context manager itself:
return True # swallow the expected exception
>>> with open('/tmp/file', 'a') as f: raise AssertionError('wrong exception type')
... f.write('more contents\n')
with assert_raises(KeyError):
{}['foo']
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, there’s often a natural operation to perform after the 8.3.2 Using generators to define context managers
object has been used and it is most convenient to have the support built in. With each release, Python
provides support in more places: When discussing generators, 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
• all file-like objects:
hand, as described in Bidirectional communication, the flow of data between the generator and its caller
– file å automatically closed 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
– fileinput, tempfile (py >= 3.2)
to support this use case.
– bz2.BZ2File, gzip.GzipFile, tarfile.TarFile, zipfile.ZipFile
@contextlib.contextmanager
– ftplib, nntplib å close connection (py >= 3.2 or 3.3) def some_generator(<arguments>):
<setup>
• locks try:
– multiprocessing.RLock å lock and unlock yield <value>
finally:
– multiprocessing.Semaphore <cleanup>
– memoryview å automatically release (py >= 3.2 and 2.7)
The contextlib.contextmanager helper takes a generator and turns it into a context manager. The
• decimal.localcontext å modify precision of computations temporarily generator has to obey some rules which are enforced by the wrapper function — most importantly it
• _winreg.PyHKEY å open and close hive key must yield exactly once. The part before the yield is executed from __enter__, the block of code
protected by the context manager is executed when the generator is suspended in yield, and the rest
• warnings.catch_warnings å kill warnings temporarily is executed in __exit__. If an exception is thrown, the interpreter hands it to the wrapper through
8.3. Context managers 278 8.3. Context managers 279

__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.
Let’s rewrite the closing example as a generator:
9
def closing(obj):
try:
yield obj
finally:
obj.close()
Let’s rewrite the assert_raises example as a generator:

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

NumPy is at the base of Python’s scientific stack of tools. Its purpose to implement 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.
This section covers:
• 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 what’s in them: PEP 3118 buffers, generalized ufuncs, . . .
Prerequisites
• NumPy
• Cython
• Pillow (Python imaging library, used in a couple of examples)
Chapter contents
• Life of ndarray
– It’s. . .
8.3. Context managers 280 281

– Block of memory
– Data types
– Indexing scheme: strides
– Findings in dissection
• Universal functions
– What they are?
– Exercise: building an ufunc from scratch
– Solution: building an ufunc from scratch
– Generalized ufuncs
• Interoperability features
typedef struct PyArrayObject {
– Sharing multidimensional, typed data PyObject_HEAD
– The old buffer protocol
/* Block of memory */
– The old buffer protocol char *data;
– Array interface protocol /* Data type descriptor */

PyArray_Descr *descr;
• Array siblings: chararray, maskedarray, matrix
– chararray: vectorized string operations /* Indexing scheme */
int nd;
– masked_array missing data npy_intp *dimensions;
– recarray: purely convenience npy_intp *strides;
– matrix: convenience? /* Other stuff */

PyObject *base;
• Summary int flags;
• Contributing to NumPy/Scipy PyObject *weakreflist;
} PyArrayObject;
– Why
– Reporting bugs
9.1.2 Block of memory
– Contributing to documentation
>>> x = np.array([1, 2, 3], dtype=np.int32)
– Contributing features >>> x.data
– How to help, in general <... at ...>
>>> bytes(x.data)
'\x01\x00\x00\x00\x02\x00\x00\x00\x03\x00\x00\x00'
Tip: In this section, numpy will be imported as follows: Memory address of the data:
>>> import numpy as np >>> x.__array_interface__['data'][0]
64803824
The whole __array_interface__:

9.1 Life of ndarray >>> x.__array_interface__
{'data': (35828928, False),
9.1.1 It’s. . . 'descr': [('', '<i4')],
'shape': (4,),
ndarray = 'strides': None,
block of memory + indexing scheme + data type descriptor 'typestr': '<i4',
'version': 3}
• raw data
• how to locate an element Reminder: two ndarrays may share the same memory:
• how to interpret an element
9.1. Life of ndarray 282 9.1. Life of ndarray 283

>>> x = np.array([1, 2, 3, 4]) chunk_id "RIFF"

>>> y = x[:-1]
chunk_size 4-byte unsigned little-endian integer
>>> x[0] = 9
>>> y
format "WAVE"
array([9, 2, 3]) fmt_id "fmt "
fmt_size 4-byte unsigned little-endian integer
Memory does not need to be owned by an ndarray: audio_fmt 2-byte unsigned little-endian integer
num_channels 2-byte unsigned little-endian integer
>>> x = b'1234' # The 'b' is for "bytes", necessary in Python 3 sample_rate 4-byte unsigned little-endian integer
byte_rate 4-byte unsigned little-endian integer
x is a string (in Python 3 a bytes), we can represent its data as an array of ints: block_align 2-byte unsigned little-endian integer
bits_per_sample 2-byte unsigned little-endian integer
>>> y = np.frombuffer(x, dtype=np.int8)
>>> y.data
data_id "data"
<... at ...> data_size 4-byte unsigned little-endian integer
>>> y.base is x
True • 44-byte block of raw data (in the beginning of the file)
• . . . followed by data_size bytes of actual sound data.
>>> y.flags
C_CONTIGUOUS : True The .wav file header as a NumPy structured data type:
F_CONTIGUOUS : True
OWNDATA : False >>> wav_header_dtype = np.dtype([
WRITEABLE : False ... ("chunk_id", (bytes, 4)), # flexible-sized scalar type, item size 4
ALIGNED : True ... ("chunk_size", "<u4"), # little-endian unsigned 32-bit integer
WRITEBACKIFCOPY : False ... ("format", "S4"), # 4-byte string
UPDATEIFCOPY : False ... ("fmt_id", "S4"),
... ("fmt_size", "<u4"),
... ("audio_fmt", "<u2"), #
The owndata and writeable flags indicate status of the memory block.
... ("num_channels", "<u2"), # .. more of the same ...
See also: ... ("sample_rate", "<u4"), #
... ("byte_rate", "<u4"),
array interface ... ("block_align", "<u2"),
... ("bits_per_sample", "<u2"),
9.1.3 Data types ... ("data_id", ("S1", (2, 2))), # sub-array, just for fun!
... ("data_size", "u4"),
The descriptor ... #
... # the sound data itself cannot be represented here:
dtype describes a single item in the array: ... # it does not have a fixed size
... ])
type scalar type of the data, one of:
int8, int16, float64, et al. (fixed size) See also:
str, unicode, void (flexible size)
wavreader.py
itemsize size of the data block
byte- byte order: big-endian > / little-endian < / not applicable | >>> wav_header_dtype['format']
order dtype('S4')
fields sub-dtypes, if it’s a structured data type >>> wav_header_dtype.fields
shape shape of the array, if it’s a sub-array dict_proxy({'block_align': (dtype('uint16'), 32), 'format': (dtype('S4'), 8), 'data_id':␣
˓→(dtype(('S1', (2, 2))), 36), 'fmt_id': (dtype('S4'), 12), 'byte_rate': (dtype('uint32'), 28),
˓→ 'chunk_id': (dtype('S4'), 0), 'num_channels': (dtype('uint16'), 22), 'sample_rate': (dtype(
>>> np.dtype(int).type ˓→'uint32'), 24), 'bits_per_sample': (dtype('uint16'), 34), 'chunk_size': (dtype('uint32'), 4),
<type 'numpy.int64'> ˓→ 'fmt_size': (dtype('uint32'), 16), 'data_size': (dtype('uint32'), 40), 'audio_fmt': (dtype(
>>> np.dtype(int).itemsize ˓→'uint16'), 20)})
8 >>> wav_header_dtype.fields['format']
>>> np.dtype(int).byteorder (dtype('S4'), 8)
'='
• The first element is the sub-dtype in the structured data, corresponding to the name format
Example: reading .wav files • The second one is its offset (in bytes) from the beginning of the item
The .wav file header:

Exercise • Casting in general copies data:
>>> x = np.array([1, 2, 3, 4], dtype=np.float)

Mini-exercise, make a “sparse” dtype by using offsets, and only some of the fields: >>> x
>>> wav_header_dtype = np.dtype(dict( array([1., 2., 3., 4.])
... names=['format', 'sample_rate', 'data_id'], >>> y = x.astype(np.int8)
... offsets=[offset_1, offset_2, offset_3], # counted from start of structure in bytes >>> y
... formats=list of dtypes for each of the fields, array([1, 2, 3, 4], dtype=int8)
... )) >>> y + 1
array([2, 3, 4, 5], dtype=int8)
and use that to read the sample rate, and data_id (as sub-array). >>> y + 256
array([257, 258, 259, 260], dtype=int16)
>>> y + 256.0
>>> f = open('data/test.wav', 'r') array([257., 258., 259., 260.])
>>> wav_header = np.fromfile(f, dtype=wav_header_dtype, count=1) >>> y + np.array([256], dtype=np.int32)
>>> f.close() array([257, 258, 259, 260], dtype=int32)
>>> print(wav_header)
[ ('RIFF', 17402L, 'WAVE', 'fmt ', 16L, 1, 1, 16000L, 32000L, 2, 16, [['d', 'a'], ['t', 'a']],␣ • Casting on setitem: dtype of the array is not changed on item assignment:
˓→17366L)]
>>> wav_header['sample_rate'] >>> y[:] = y + 1.5

array([16000], dtype=uint32) >>> y
Let’s try accessing the sub-array:
>>> wav_header['data_id'] Note: Exact rules: see numpy documentation

array([[['d', 'a'],
['t', 'a']]],
dtype='|S1')
>>> wav_header.shape Re-interpretation / viewing
(1,)
>>> wav_header['data_id'].shape • Data block in memory (4 bytes)
(1, 2, 2)
0x01 || 0x02 || 0x03 || 0x04
When accessing sub-arrays, the dimensions get added to the end!
– 4 of uint8, OR,
Note: There are existing modules such as wavfile, audiolab, etc. for loading sound data. . . – 4 of int8, OR,
– 2 of int16, OR,
Casting and re-interpretation/views – 1 of int32, OR,
casting – 1 of float32, OR,
• on assignment – ...
• on array construction How to switch from one to another?
• on arithmetic 1. Switch the dtype:
• etc. >>> x = np.array([1, 2, 3, 4], dtype=np.uint8)

>>> x.dtype = "<i2"
• and manually: .astype(dtype) >>> x
data re-interpretation array([ 513, 1027], dtype=int16)
>>> 0x0201, 0x0403
• manually: .view(dtype) (513, 1027)
Casting
0x01 0x02 || 0x03 0x04
• Casting in arithmetic, in nutshell:
– only type (not value!) of operands matters Note: little-endian: least significant byte is on the left in memory
– largest “safe” type able to represent both is picked
2. Create a new view:
– scalars can “lose” to arrays in some situations

>>> y = x.view("<i4")
>>> y >>> y = np.array([[1, 3], [2, 4]], dtype=np.uint8).transpose()
array([67305985], dtype=int32) >>> x = y.copy()
>>> 0x04030201 >>> x
67305985 array([[1, 2],
[3, 4]], dtype=uint8)
>>> y
0x01 0x02 0x03 0x04 array([[1, 2],
[3, 4]], dtype=uint8)
>>> x.view(np.int16)
array([[ 513],
Note: [1027]], dtype=int16)
>>> 0x0201, 0x0403
• .view() makes views, does not copy (or alter) the memory block (513, 1027)
• only changes the dtype (and adjusts array shape): >>> y.view(np.int16)
array([[ 769, 1026]], dtype=int16)
>>> x[1] = 5
>>> y • What happened?
array([328193], dtype=int32)
• . . . we need to look into what x[0,1] actually means
>>> y.base is x
True >>> 0x0301, 0x0402
(769, 1026)
Mini-exercise: data re-interpretation

9.1.4 Indexing scheme: strides
See also:
Main point
view-colors.py
The question:
You have RGBA data in an array:
>>> x = np.array([[1, 2, 3],
>>> x = np.zeros((10, 10, 4), dtype=np.int8) ... [4, 5, 6],
>>> x[:, :, 0] = 1 ... [7, 8, 9]], dtype=np.int8)
>>> x[:, :, 1] = 2 >>> x.tobytes('A')
>>> x[:, :, 2] = 3 b'\x01\x02\x03\x04\x05\x06\x07\x08\t'
>>> x[:, :, 3] = 4
At which byte in ``x.data`` does the item ``x[1, 2]`` begin?
where the last three dimensions are the R, B, and G, and alpha channels.
The answer (in NumPy)
How to make a (10, 10) structured array with field names ‘r’, ‘g’, ‘b’, ‘a’ without copying data?
• strides: the number of bytes to jump to find the next element
>>> y = ...
• 1 stride per dimension
>>> assert (y['r'] == 1).all()
>>> x.strides
>>> assert (y['g'] == 2).all()
(3, 1)
>>> assert (y['b'] == 3).all()
>>> byte_offset = 3*1 + 1*2 # to find x[1, 2]
>>> assert (y['a'] == 4).all()
>>> x.flat[byte_offset]
6
Solution >>> x[1, 2]
6
>>> y = x.view([('r', 'i1'),
... ('g', 'i1'),
- simple, **flexible**
... ('b', 'i1'),
... ('a', 'i1')]
... )[:, :, 0]
C and Fortran order
Warning: Another array taking exactly 4 bytes of memory: Note: The Python built-in bytes returns bytes in C-order by default which can cause confusion when
trying to inspect memory layout. We use numpy.ndarray.tobytes() with order=A instead, which
preserves the C or F ordering of the bytes in memory.

Slicing with integers

>>> x = np.array([[1, 2, 3],
... [4, 5, 6]], dtype=np.int16, order='C')
• Everything can be represented by changing only shape, strides, and possibly adjusting the data
>>> x.strides
(6, 2) pointer!
>>> x.tobytes('A') • Never makes copies of the data
b'\x01\x00\x02\x00\x03\x00\x04\x00\x05\x00\x06\x00'
>>> x = np.array([1, 2, 3, 4, 5, 6], dtype=np.int32)
• Need to jump 6 bytes to find the next row >>> y = x[::-1]
>>> y
• Need to jump 2 bytes to find the next column array([6, 5, 4, 3, 2, 1], dtype=int32)
>>> y.strides
>>> y = np.array(x, order='F') (-4,)
>>> y.strides
(2, 4) >>> y = x[2:]
>>> y.tobytes('A') >>> y.__array_interface__['data'][0] - x.__array_interface__['data'][0]
b'\x01\x00\x04\x00\x02\x00\x05\x00\x03\x00\x06\x00' 8
• Need to jump 2 bytes to find the next row >>> x = np.zeros((10, 10, 10), dtype=np.float)
>>> x.strides
• Need to jump 4 bytes to find the next column
(800, 80, 8)
• Similarly to higher dimensions: >>> x[::2,::3,::4].strides
(1600, 240, 32)
– C: last dimensions vary fastest (= smaller strides)
– F: first dimensions vary fastest • Similarly, transposes never make copies (it just swaps strides):
shape = (𝑑1 , 𝑑2 , ..., 𝑑𝑛 ) >>> x = np.zeros((10, 10, 10), dtype=np.float)
strides = (𝑠1 , 𝑠2 , ..., 𝑠𝑛 ) >>> x.strides
(800, 80, 8)
𝑗 = 𝑑𝑗+1 𝑑𝑗+2 ...𝑑𝑛 × itemsize
𝑠𝐶 >>> x.T.strides
(8, 80, 800)
𝑗 = 𝑑1 𝑑2 ...𝑑𝑗−1 × itemsize
𝑠𝐹
But: not all reshaping operations can be represented by playing with strides:
Note: Now we can understand the behavior of .view():
>>> a = np.arange(6, dtype=np.int8).reshape(3, 2)
>>> y = np.array([[1, 3], [2, 4]], dtype=np.uint8).transpose() >>> b = a.T
>>> x = y.copy() >>> b.strides
(1, 2)
Transposition does not affect the memory layout of the data, only strides
So far, so good. However:
>>> x.strides
(2, 1) >>> bytes(a.data)
>>> y.strides b'\x00\x01\x02\x03\x04\x05'
(1, 2) >>> b
array([[0, 2, 4],
>>> x.tobytes('A') [1, 3, 5]], dtype=int8)
b'\x01\x02\x03\x04' >>> c = b.reshape(3*2)
>>> y.tobytes('A') >>> c
b'\x01\x03\x02\x04' array([0, 2, 4, 1, 3, 5], dtype=int8)
• the results are different when interpreted as 2 of int16 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.
• .copy() creates new arrays in the C order (by default)
Example: fake dimensions with strides
Stride manipulation
Note: In-place operations with views
Prior to NumPy version 1.13, in-place operations with views could result in incorrect results for large >>> from numpy.lib.stride_tricks import as_strided
arrays. Since version 1.13, NumPy includes checks for memory overlap to guarantee that results are >>> help(as_strided)
consistent with the non in-place version (e.g. a = a + a.T produces the same result as a += a.T). as_strided(x, shape=None, strides=None)
Note however that this may result in the data being copied (as if using a += a.T.copy()), ultimately Make an ndarray from the given array with the given shape and strides
resulting in more memory being used than might otherwise be expected for in-place operations!

. . . seems somehow familiar . . .

Warning: as_strided does not check that you stay inside the memory block bounds. . .
>>> x = np.array([1, 2, 3, 4], dtype=np.int16)
>>> y = np.array([5, 6, 7], dtype=np.int16)
>>> x = np.array([1, 2, 3, 4], dtype=np.int16) >>> x[np.newaxis,:] * y[:,np.newaxis]
>>> as_strided(x, strides=(2*2, ), shape=(2, )) array([[ 5, 10, 15, 20],
array([1, 3], dtype=int16) [ 6, 12, 18, 24],
>>> x[::2] [ 7, 14, 21, 28]], dtype=int16)
array([1, 3], dtype=int16)
• Internally, array broadcasting is indeed implemented using 0-strides.
See also:
stride-fakedims.py More tricks: diagonals
Exercise See also:
array([1, 2, 3, 4], dtype=np.int8) stride-diagonals.py
-> array([[1, 2, 3, 4],

Challenge
[1, 2, 3, 4], • Pick diagonal entries of the matrix: (assume C memory order):
[1, 2, 3, 4]], dtype=np.int8)
>>> x = np.array([[1, 2, 3],
using only as_strided.: ... [4, 5, 6],
... [7, 8, 9]], dtype=np.int32)
Hint: byte_offset = stride[0]*index[0] + stride[1]*index[1] + ...
>>> x_diag = as_strided(x, shape=(3,), strides=(???,))
Spoiler
• Pick the first super-diagonal entries [2, 6].
Stride can also be 0 :
• And the sub-diagonals?
>>> y = as_strided(x, strides=(0, 1), shape=(3, 4)) (Hint to the last two: slicing first moves the point where striding starts from.)
>>> y Solution
array([[1, 2, 3, 4],
[1, 2, 3, 4], Pick diagonals:
[1, 2, 3, 4]], dtype=int8)
>>> y.base.base is x >>> x_diag = as_strided(x, shape=(3, ), strides=((3+1)*x.itemsize, ))
True >>> x_diag
array([1, 5, 9], dtype=int32)
Broadcasting Slice first, to adjust the data pointer:

• Doing something useful with it: outer product of [1, 2, 3, 4] and [5, 6, 7] >>> as_strided(x[0, 1:], shape=(2, ), strides=((3+1)*x.itemsize, ))
>>> x2 = as_strided(x, strides=(0, 1*2), shape=(3, 4))
>>> as_strided(x[1:, 0], shape=(2, ), strides=((3+1)*x.itemsize, ))
>>> x2
array([[1, 2, 3, 4],
[1, 2, 3, 4],
[1, 2, 3, 4]], dtype=int16)
Note: Using np.diag
>>> y = np.array([5, 6, 7], dtype=np.int16) >>> y = np.diag(x, k=1)
>>> y2 = as_strided(y, strides=(1*2, 0), shape=(3, 4)) >>> y
>>> y2 array([2, 6], dtype=int32)
array([[5, 5, 5, 5],
[6, 6, 6, 6],
However,
[7, 7, 7, 7]], dtype=int16)
>>> y.flags.owndata
>>> x2 * y2 False
array([[ 5, 10, 15, 20],
[ 6, 12, 18, 24], Note This behavior has changed: before numpy 1.9, np.diag would make a copy.
[ 7, 14, 21, 28]], dtype=int16)

See also: – ⇒ faster

stride-diagonals.py See also:
Challenge • numexpr is designed to mitigate cache effects when evaluating array expressions.
Compute the tensor trace: • numba is a compiler for Python code, that is aware of numpy arrays.
>>> x = np.arange(5*5*5*5).reshape(5, 5, 5, 5)
>>> s = 0 9.1.5 Findings in dissection
>>> for i in range(5):
... for j in range(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()
• memory block: may be shared, .base, .data

CPU cache effects
• data type descriptor: structured data, sub-arrays, byte order, casting, viewing, .astype(), .view()
Memory layout can affect performance:
• strided indexing: strides, C/F-order, slicing w/ integers, as_strided, broadcasting, stride tricks,
In [1]: x = np.zeros((20000,)) diag, CPU cache coherence
In [2]: y = np.zeros((20000*67,))[::67]
9.2 Universal functions
In [3]: x.shape, y.shape
((20000,), (20000,)) 9.2.1 What they are?
In [4]: %timeit x.sum() • Ufunc performs and elementwise operation on all elements of an array.
Examples:
In [5]: %timeit y.sum()
np.add, np.subtract, scipy.special.*, ...
In [6]: x.strides, y.strides • Automatically support: broadcasting, casting, . . .

((8,), (536,)) • The author of an ufunc only has to supply the elementwise operation, NumPy takes care of the
rest.
Smaller strides are faster? • The elementwise operation needs to be implemented in C (or, e.g., Cython)
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.
• 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): char *input_1 = (char*)args[0];
char *input_2 = (char*)args[1];
– ⇒ fewer transfers needed (continues on next page)
9.1. Life of ndarray 294 9.2. Universal functions 295


z = z0
char *output = (char*)args[2]; for k in range(iterations):
int i; z = z*z + c
for (i = 0; i < dimensions[0]; ++i) {

say, 100 iterations or until z.real**2 + z.imag**2 > 1000. Use it to determine which c are in
*output = elementwise_function(*input_1, *input_2);
the Mandelbrot set.
input_1 += steps[0];
input_2 += steps[1]; • Our function is a simple one, so make use of the PyUFunc_* helpers.
output += steps[2];
} • Write it in Cython
}
See also:
2. The NumPy part, built by mandel.pyx, mandelplot.py
char types[3] #
# Fix the parts marked by TODO
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 */ #
# Compile this file by (Cython >= 0.12 required because of the complex vars)
PyObject *python_ufunc = PyUFunc_FromFuncAndData( #
ufunc_loop, # cython mandel.pyx
NULL, # python setup.py build_ext -i
types, #
1, /* ntypes */ # and try it out with, in this directory,
2, /* num_inputs */ #
1, /* num_outputs */ # >>> import mandel
identity_element, # >>> mandel.mandel(0, 1 + 2j)
name, #
docstring, #
unused)
# The elementwise function
# ------------------------
• A ufunc can also support multiple different input-output type combinations.
cdef void mandel_single_point(double complex *z_in,
Making it easier double complex *c_in,
double complex *z_out) nogil:
3. ufunc_loop is of very generic form, and NumPy provides pre-made ones
#
# The Mandelbrot iteration
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) # Some points of note:
PyUfunc_D_D elementwise_func(npy_cdouble *input, npy_cdouble* output) #
# - It's *NOT* allowed to call any Python functions here.
PyUfunc_DD_D elementwise_func(npy_cdouble *in1, npy_cdouble *in2, npy_cdouble*
#
out) # The Ufunc loop runs with the Python Global Interpreter Lock released.
# Hence, the ``nogil``.
• Only elementwise_func needs to be supplied #
• . . . except when your elementwise function is not in one of the above forms # - And so all local variables must be declared with ``cdef``
#
# - Note also that this function receives *pointers* to the data
9.2.2 Exercise: building an ufunc from scratch #
The Mandelbrot fractal is defined by the iteration
cdef double complex z = z_in[0]
cdef double complex c = c_in[0]
𝑧 ← 𝑧2 + 𝑐 cdef int k # the integer we use in the for loop
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. # TODO: write the Mandelbrot iteration for one point here,
# as you would write it in Python.
• Make ufunc called mandel(z0, c) that computes:
#
9.2. Universal functions 296 9.2. Universal functions 297


# Say, use 100 as the maximum number of iterations, and 1000 # ``PyUfunc_ff_f`` ``float elementwise_func(float input_1, float input_2)``
# as the cutoff for z.real**2 + z.imag**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`` `èlementwise_func(complex_double *input, complex_double* complex_double)``
TODO: mandelbrot iteration should go here # ``PyUfunc_DD_D`` `èlementwise_func(complex_double *in1, complex_double *in2, complex_
˓→double* out)``
# Return the answer for this point # ================ =======================================================

z_out[0] = z #
# The full list is above.
#
# Boilerplate Cython definitions #
# # Type codes:
# Pulls definitions from the Numpy C headers. #
# ------------------------------------------- # 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,
from numpy cimport import_array, import_ufunc # NPY_LONGDOUBLE, NPY_CFLOAT, NPY_CDOUBLE, NPY_CLONGDOUBLE, NPY_DATETIME,
from numpy cimport (PyUFunc_FromFuncAndData, # NPY_TIMEDELTA, NPY_OBJECT, NPY_STRING, NPY_UNICODE, NPY_VOID
PyUFuncGenericFunction) #
from numpy cimport NPY_CDOUBLE, NP_DOUBLE, NPY_LONG
loop_func[0] = ... TODO: suitable PyUFunc_* ...
# Import all pre-defined loop functions input_output_types[0] = ... TODO ...
# you won't need all of them - keep the relevant ones ... TODO: fill in rest of input_output_types ...
from numpy cimport ( # This thing is passed as the ``data`` parameter for the generic
PyUFunc_f_f_As_d_d, # PyUFunc_* loop, to let it know which function it should call.
PyUFunc_d_d, elementwise_funcs[0] = <void*>mandel_single_point
PyUFunc_f_f,
PyUFunc_g_g, # Construct the ufunc:
PyUFunc_F_F_As_D_D,
PyUFunc_F_F, mandel = PyUFunc_FromFuncAndData(
PyUFunc_D_D, loop_func,
PyUFunc_G_G, elementwise_funcs,
PyUFunc_ff_f_As_dd_d, input_output_types,
PyUFunc_ff_f, 1, # number of supported input types
PyUFunc_dd_d, TODO, # number of input args
PyUFunc_gg_g, TODO, # number of output args
PyUFunc_FF_F_As_DD_D, 0, # ìdentity` element, never mind this
PyUFunc_DD_D, "mandel", # function name
PyUFunc_FF_F, "mandel(z, c) -> computes z*z + c", # docstring
PyUFunc_GG_G) 0 # unused
)
# Required module initialization Reminder: some pre-made Ufunc loops:

# ------------------------------
import_array() PyUfunc_f_f float elementwise_func(float input_1)

import_ufunc() 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)
# The actual ufunc declaration PyUfunc_D_D elementwise_func(complex_double *input, complex_double* output)
# ---------------------------- PyUfunc_DD_D elementwise_func(complex_double *in1, complex_double *in2,
complex_double* out)
cdef PyUFuncGenericFunction loop_func[1]
cdef char input_output_types[3]
cdef void *elementwise_funcs[1] Type codes:
# NPY_BOOL, NPY_BYTE, NPY_UBYTE, NPY_SHORT, NPY_USHORT, NPY_INT, NPY_UINT,

# Reminder: some pre-made Ufunc loops: 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
# ``PyUfunc_f_f`` ``float elementwise_func(float input_1)``

9.2.3 Solution: building an ufunc from scratch (continued from previous page)
cdef PyUFuncGenericFunction loop_func[1]
# The elementwise function
cdef char input_output_types[3]
# ------------------------
cdef void *elementwise_funcs[1]
loop_func[0] = PyUFunc_DD_D
double complex *c_in,
input_output_types[0] = NPY_CDOUBLE
#
# The Mandelbrot iteration
#
elementwise_funcs[0] = <void*>mandel_single_point
#
# Some points of note:
mandel = PyUFunc_FromFuncAndData(
#
loop_func,
# - It's *NOT* allowed to call any Python functions here.
elementwise_funcs,
#
input_output_types,
# The Ufunc loop runs with the Python Global Interpreter Lock released.
1, # number of supported input types
# Hence, the ``nogil``.
2, # number of input args
#
1, # number of output args
# - And so all local variables must be declared with ``cdef``
0, # ìdentity` element, never mind this
#
"mandel", # function name
# - Note also that this function receives *pointers* to the data;
"mandel(z, c) -> computes iterated z*z + c", # docstring
# the "traditional" solution to passing complex variables around
0 # unused
#
)
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 Plot Mandelbrot
================
# Straightforward iteration
Plot the Mandelbrot ensemble.
for k in range(100):
z = z*z + c """
if z.real**2 + z.imag**2 > 1000:
break import numpy as np
import mandel
# Return the answer for this point x = np.linspace(-1.7, 0.6, 1000)
z_out[0] = z y = np.linspace(-1.4, 1.4, 1000)
c = x[None,:] + 1j*y[:,None]
z = mandel.mandel(c, c)
# Boilerplate Cython definitions
# import matplotlib.pyplot as plt
# Pulls definitions from the Numpy C headers. plt.imshow(abs(z)**2 < 1000, extent=[-1.7, 0.6, -1.4, 1.4])
# ------------------------------------------- plt.gray()
plt.show()
from numpy cimport import_array, import_ufunc
from numpy cimport (PyUFunc_FromFuncAndData,
PyUFuncGenericFunction)
from numpy cimport NPY_CDOUBLE
from numpy cimport PyUFunc_DD_D
# Required module initialization

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

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

input shape = (n, n)

Note: Most of the boilerplate could be automated by these Cython modules: output shape = () i.e. scalar
https://github.com/cython/cython/wiki/MarkLodato-CreatingUfuncs
(n, n) -> ()
Matrix product:
Several accepted input types
input_1 shape = (m, n)
E.g. supporting both single- and double-precision versions input_2 shape = (n, p)
output shape = (m, p)
double complex *c_in,
(m, n), (n, p) -> (m, p)
...
• This is called the “signature” of the generalized ufunc
cdef void mandel_single_point_singleprec(float complex *z_in, • The dimensions on which the g-ufunc acts, are “core dimensions”
float complex *c_in,
float complex *z_out) nogil:
... Status in NumPy
cdef PyUFuncGenericFunction loop_funcs[2] • g-ufuncs are in NumPy already . . .

cdef char input_output_types[3*2] • new ones can be created with PyUFunc_FromFuncAndDataAndSignature
cdef void *elementwise_funcs[1*2]
• most linear-algebra functions are implemented as g-ufuncs to enable working with stacked arrays:
loop_funcs[0] = PyUFunc_DD_D
input_output_types[0] = NPY_CDOUBLE >>> import numpy as np
input_output_types[1] = NPY_CDOUBLE >>> np.linalg.det(np.random.rand(3, 5, 5))
input_output_types[2] = NPY_CDOUBLE array([ 0.00965823, -0.13344729, 0.04583961])
elementwise_funcs[0] = <void*>mandel_single_point >>> np.linalg._umath_linalg.det.signature
'(m,m)->()'
loop_funcs[1] = PyUFunc_FF_F
input_output_types[3] = NPY_CFLOAT • we also ship with a few g-ufuncs for testing, ATM:
input_output_types[4] = NPY_CFLOAT
input_output_types[5] = NPY_CFLOAT >>> import numpy.core.umath_tests as ut
elementwise_funcs[1] = <void*>mandel_single_point_singleprec >>> ut.matrix_multiply.signature
'(m,n),(n,p)->(m,p)'
mandel = PyUFunc_FromFuncAndData(
loop_func, >>> x = np.ones((10, 2, 4))
elementwise_funcs, >>> y = np.ones((10, 4, 5))
input_output_types, >>> ut.matrix_multiply(x, y).shape
2, # number of supported input types <---------------- (10, 2, 5)
2, # number of input args
1, # number of output args • in both examples the last two dimensions became core dimensions, and are modified as per the
0, # ìdentity` element, never mind this signature
"mandel", # function name
"mandel(z, c) -> computes iterated z*z + c", # docstring • otherwise, the g-ufunc operates “elementwise”
0 # unused • matrix multiplication this way could be useful for operating on many small matrices at once
)
Generalized ufunc loop

9.2.4 Generalized ufuncs Matrix multiplication (m,n),(n,p) -> (m,p)
ufunc
void gufunc_loop(void **args, int *dimensions, int *steps, void *data)
output = elementwise_function(input) {
char *input_1 = (char*)args[0]; /* these are as previously */
Both output and input can be a single array element only. char *input_2 = (char*)args[1];
generalized ufunc char *output = (char*)args[2];
output and input can be arrays with a fixed number of dimensions 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 */
For example, matrix trace (sum of diag elements): int input_2_strides_n = steps[5]; /* steps */
int input_2_strides_p = steps[6];

(continued from previous page) Check what happens if data is now modified, and img saved again.
int output_strides_n = steps[7];
int output_strides_p = steps[8]; 9.3.3 The old buffer protocol
int m = dimension[1]; /* core dimensions are added after */ """
int n = dimension[2]; /* the main dimension; order as in */ From buffer
int p = dimension[3]; /* signature */ ============
int i; Show how to exchange data between numpy and a library that only knows
the buffer interface.
for (i = 0; i < dimensions[0]; ++i) { """
matmul_for_strided_matrices(input_1, input_2, output,
strides for each array...); import numpy as np
import Image
input_1 += steps[0];
input_2 += steps[1]; # Let's make a sample image, RGBA format
output += steps[2];
} x = np.zeros((200, 200, 4), dtype=np.int8)
}
x[:,:,0] = 254 # red
x[:,:,3] = 255 # opaque
9.3 Interoperability features data = x.view(np.int32) # Check that you understand why this is OK!
9.3.1 Sharing multidimensional, typed data img = Image.frombuffer("RGBA", (200, 200), data)
img.save('test.png')
Suppose you
1. Write a library than handles (multidimensional) binary data, #
# Modify the original data, and save again.
2. Want to make it easy to manipulate the data with NumPy, or whatever other library, #
# It turns out that PIL, which knows next to nothing about Numpy,
3. . . . but would not like to have NumPy as a dependency.
# happily shares the same data.
Currently, 3 solutions: #
1. the “old” buffer interface x[:,:,1] = 254

2. the array interface img.save('test2.png')
3. the “new” buffer interface (PEP 3118)
9.3.2 The old buffer protocol

• Only 1-D buffers
• No data type information
• C-level interface; PyBufferProcs tp_as_buffer in the type object
• But it’s integrated into Python (e.g. strings support it)
Mini-exercise using Pillow (Python Imaging Library):
See also:
pilbuffer.py
>>> from PIL import Image

>>> data = np.zeros((200, 200, 4), dtype=np.int8)
>>> data[:, :] = [255, 0, 0, 255] # Red
>>> # In PIL, RGBA images consist of 32-bit integers whose bytes are [RR,GG,BB,AA]
>>> data = data.view(np.int32).squeeze()
>>> img = Image.frombuffer("RGBA", (200, 200), data, "raw", "RGBA", 0, 1)
>>> img.save('test.png')
Q:
9.3. Interoperability features 304 9.3. Interoperability features 305

9.4 Array siblings: chararray, maskedarray, matrix

9.4.1 chararray: vectorized string operations
>>> x = np.array(['a', ' bbb', ' ccc']).view(np.chararray)
>>> x.lstrip(' ')
chararray(['a', 'bbb', 'ccc'],
dtype='...')
>>> x.upper()
chararray(['A', ' BBB', ' CCC'],
dtype='...')
Note: .view() has a second meaning: it can make an ndarray an instance of a specialized ndarray
subclass
9.4.2 masked_array missing data

9.3.4 Array interface protocol Masked arrays are arrays that may have missing or invalid entries.
• Multidimensional buffers For example, suppose we have an array where the fourth entry is invalid:
• Data type information present
>>> x = np.array([1, 2, 3, -99, 5])
• NumPy-specific approach; slowly deprecated (but not going away)
One way to describe this is to create a masked array:
• Not integrated in Python otherwise
See also: >>> mx = np.ma.masked_array(x, mask=[0, 0, 0, 1, 0])
>>> mx
Documentation: http://numpy.org/doc/stable/reference/arrays.interface.html masked_array(data=[1, 2, 3, --, 5],
mask=[False, False, False, True, False],
>>> x = np.array([[1, 2], [3, 4]]) fill_value=999999)
>>> x.__array_interface__
{'data': (171694552, False), # memory address of data, is readonly?
Masked mean ignores masked data:
'descr': [('', '<i4')], # data type descriptor
'typestr': '<i4', # same, in another form >>> mx.mean()
'strides': None, # strides; or None if in C-order 2.75
'shape': (2, 2), >>> np.mean(mx)
'version': 3, 2.75
}
::
Warning: Not all NumPy functions respect masks, for instance np.dot, so check the return types.
>>> from PIL import Image
>>> img = Image.open('data/test.png')
>>> img.__array_interface__ The masked_array returns a view to the original array:
{'data': ...,
>>> mx[1] = 9
'shape': (200, 200, 4),
>>> x
'typestr': '|u1'}
array([ 1, 9, 3, -99, 5])
>>> x = np.asarray(img)
>>> x.shape
(200, 200, 4)
The mask
You can modify the mask by assigning:
Note: A more C-friendly variant of the array interface is also defined.
>>> mx[1] = np.ma.masked
>>> mx
masked_array(data=[1, --, 3, --, 5],
mask=[False, True, False, True, False],
fill_value=999999)
The mask is cleared on assignment:
9.3. Interoperability features 306 9.4. Array siblings: chararray, maskedarray, matrix 307
>>> mx[1] = 9 Note that Matplotlib knows about masked arrays:

>>> mx >>> plt.plot(year, populations, 'o-')
masked_array(data=[1, 9, 3, --, 5], [<matplotlib.lines.Line2D object at ...>, ...]
mask=[False, False, False, True, False],
fill_value=999999)
The mask is also available directly:

>>> mx.mask
array([False, False, False, True, False])
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], 9.4.3 recarray: purely convenience
fill_value=999999)
>>> arr = np.array([('a', 1), ('b', 2)], dtype=[('x', 'S1'), ('y', int)])
>>> arr2 = arr.view(np.recarray)
Domain-aware functions >>> arr2.x
The masked array package also contains domain-aware functions: chararray(['a', 'b'],
dtype='|S1')
>>> np.ma.log(np.array([1, 2, -1, -2, 3, -5])) >>> arr2.y
masked_array(data=[0.0, 0.6931471805599453, --, --, 1.0986122886681098, --], array([1, 2])
mask=[False, False, True, True, False, True],
fill_value=1e+20)
9.4.4 matrix: convenience?
• always 2-D
Note: Streamlined and more seamless support for dealing with missing data in arrays is making its
way into NumPy 1.7. Stay tuned! • * is the matrix product, not the elementwise one
>>> np.matrix([[1, 0], [0, 1]]) * np.matrix([[1, 2], [3, 4]])

matrix([[1, 2],
Example: Masked statistics [3, 4]])
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.
9.5 Summary
>>> data = np.loadtxt('data/populations.txt') • Anatomy of the ndarray: data, dtype, strides.
>>> populations = np.ma.masked_array(data[:,1:])
>>> year = data[:, 0] • Universal functions: elementwise operations, how to make new ones
• Ndarray subclasses
>>> bad_years = (((year >= 1903) & (year <= 1910))
... | ((year >= 1917) & (year <= 1918))) • Various buffer interfaces for integration with other tools
>>> # '&' means 'and' and '|' means 'or'
>>> populations[bad_years, 0] = np.ma.masked
• Recent additions: PEP 3118, generalized ufuncs
>>> populations[bad_years, 1] = np.ma.masked
>>> populations.mean(axis=0)
9.6 Contributing to NumPy/Scipy
masked_array(data=[40472.72727272727, 18627.272727272728, 42400.0],
mask=[False, False, False], Get this tutorial: http://www.euroscipy.org/talk/882
fill_value=1e+20)
9.6.1 Why
>>> populations.std(axis=0)
masked_array(data=[21087.656489006717, 15625.799814240254, 3322.5062255844787], • “There’s a bug?”
mask=[False, False, False],
fill_value=1e+20)
9.4. Array siblings: chararray, maskedarray, matrix 308 9.5. Summary 309
• “I don’t understand what this is supposed to do?”

>>> print(np.__file__)
• “I have this fancy code. Would you like to have it?” /...
• “I’d like to help! What can I do?” In case you have old/broken NumPy installations lying around.
If unsure, try to remove existing NumPy installations, and reinstall. . .
9.6.2 Reporting bugs
• Bug tracker (prefer this) 9.6.3 Contributing to documentation
– https://github.com/numpy/numpy/issues 1. Documentation editor
– https://github.com/scipy/scipy/issues • http://docs.scipy.org/doc/numpy
– Click the “Register” link to get an account • Registration
• Mailing lists ( scipy.org/Mailing_Lists ) – Register an account
– If you’re unsure – Subscribe to scipy-dev mailing list (subscribers-only)
– No replies in a week or so? Just file a bug ticket. – Problem with mailing lists: you get mail
Good bug report ∗ But: you can turn mail delivery off
∗ “change your subscription options”, at the bottom of
Title: numpy.random.permutations fails for non-integer arguments
http://mail.python.org/mailman/listinfo/scipy-dev
I'm trying to generate random permutations, using numpy.random.permutations
– Send a mail @ scipy-dev mailing list; ask for activation:
When calling numpy.random.permutation with non-integer arguments
To: [email protected]
it fails with a cryptic error message::
Hi,
>>> np.random.permutation(12)
array([11, 5, 8, 4, 6, 1, 9, 3, 7, 2, 10, 0])
I'd like to edit NumPy/Scipy docstrings. My account is XXXXX
>>> np.random.permutation(12.) #doctest: +SKIP
Cheers,
N. N.
File "mtrand.pyx", line 3311, in mtrand.RandomState.permutation
File "mtrand.pyx", line 3254, in mtrand.RandomState.shuffle
TypeError: len() of unsized object • Check the style guide:
– http://docs.scipy.org/doc/numpy/
This also happens with long arguments, and so
np.random.permutation(X.shape[0]) where X is an array fails on 64 – Don’t be intimidated; to fix a small thing, just fix it
bit windows (where shape is a tuple of longs).
• Edit
It would be great if it could cast to integer or at least raise a 2. Edit sources and send patches (as for bugs)
proper error for non-integer types.
3. Complain on the mailing list
I'm using NumPy 1.4.1, built from the official tarball, on Windows
64 with Visual studio 2008, on Python.org 64-bit Python.
9.6.4 Contributing features
0. What are you trying to do? The contribution of features is documented on https://docs.scipy.org/doc/numpy/dev/
1. Small code snippet reproducing the bug (if possible)
9.6.5 How to help, in general
• What actually happens
• Bug fixes always welcome!
• What you’d expect
– What irks you most
2. Platform (Windows / Linux / OSX, 32/64 bits, x86/PPC, . . . )
– Browse the tracker
3. Version of NumPy/Scipy
• Documentation work
>>> print(np.__version__)
1... – API docs: improvements to docstrings
∗ Know some Scipy module well?
Check that the following is what you expect
– User guide
9.6. Contributing to NumPy/Scipy 310 9.6. Contributing to NumPy/Scipy 311

∗ 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:
10
– numpy-discussion list
– scipy-dev list
CHAPTER
Debugging code
Author: Gaël Varoquaux

This section explores tools to understand better your code base: debugging, to find and fix bugs.
It is not specific to the scientific Python community, but the strategies that we will employ are tailored
to its needs.
Prerequisites
• Numpy
• IPython
• nosetests
• pyflakes
• gdb for the C-debugging part.
Chapter contents
• Avoiding bugs
– Coding best practices to avoid getting in trouble
– pyflakes: fast static analysis
• Debugging workflow
• Using the Python debugger
– Invoking the debugger
9.6. Contributing to NumPy/Scipy 312 313

– Debugger commands and interaction kdialog --title "pyflakes %f ilename" --msgbox "$(pyflakes %f ilename)"
• Debugging segmentation faults using gdb
• In TextMate
Menu: TextMate -> Preferences -> Advanced -> Shell variables, add a shell variable:
10.1 Avoiding bugs TM_PYCHECKER = /Library/Frameworks/Python.framework/Versions/Current/bin/pyflakes
10.1.1 Coding best practices to avoid getting in trouble Then Ctrl-Shift-V is binded to a pyflakes report
• In vim In your .vimrc (binds F5 to pyflakes):
Brian Kernighan
autocmd FileType python let &mp = 'echo "*** running % ***" ; pyflakes %'
“Everyone knows that debugging is twice as hard as writing a program in the first place. So if you’re autocmd FileType tex,mp,rst,python imap <Esc>[15~ <C-O>:make!^M
autocmd FileType tex,mp,rst,python map <Esc>[15~ :make!^M
as clever as you can be when you write it, how will you ever debug it?”
autocmd FileType tex,mp,rst,python set autowrite
• We all write buggy code. Accept it. Deal with it. • In emacs In your .emacs (binds F5 to pyflakes):
• Write your code with testing and debugging in mind. (defun pyflakes-thisfile () (interactive)
• Keep It Simple, Stupid (KISS). (compile (format "pyflakes %s " (buffer-file-name)))
)
– What is the simplest thing that could possibly work?
(define-minor-mode pyflakes-mode
• Don’t Repeat Yourself (DRY). "Toggle pyflakes mode.
– Every piece of knowledge must have a single, unambiguous, authoritative representation within With no argument, this command toggles the mode.
a system. Non-null prefix argument turns on the mode.
Null prefix argument turns off the mode."
– Constants, algorithms, etc. . . ;; The initial value.
nil
• Try to limit interdependencies of your code. (Loose Coupling) ;; The indicator for the mode line.
• Give your variables, functions and modules meaningful names (not mathematics names) " Pyflakes"
;; The minor mode bindings.
'( ([f5] . pyflakes-thisfile) )
10.1.2 pyflakes: fast static analysis )
They are several static analysis tools in Python; to name a few:
(add-hook 'python-mode-hook (lambda () (pyflakes-mode t)))
• pylint
• pychecker A type-as-go spell-checker like integration
• pyflakes • In vim
• flake8 – Use the pyflakes.vim plugin:
Here we focus on pyflakes, which is the simplest tool. 1. download the zip file from http://www.vim.org/scripts/script.php?script_id=2441
• Fast, simple 2. extract the files in ~/.vim/ftplugin/python
• Detects syntax errors, missing imports, typos on names. 3. make sure your vimrc has filetype plugin indent on
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 – Alternatively: use the syntastic plugin. This can be configured to use flake8 too and also
handles on-the-fly checking for many other languages.
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:
10.1. Avoiding bugs 314 10.1. Avoiding bugs 315

• Inspect values of variables.

• Modify values of variables.
• Set breakpoints.
print
Yes, print statements do work as a debugging tool. However to inspect runtime, it is often more
efficient to use the debugger.
10.3.1 Invoking the debugger

• In emacs Ways to launch the debugger:
Use the flymake mode with pyflakes, documented on https://www.emacswiki.org/emacs/FlyMake 1. Postmortem, launch debugger after module errors.
and included in Emacs 26 and more recent. To activate it, use M-x (meta-key then x) and enter
flymake-mode at the prompt. To enable it automatically when opening a Python file, add the 2. Launch the module with the debugger.
following line to your .emacs file: 3. Call the debugger inside the module
(add-hook 'python-mode-hook '(lambda () (flymake-mode)))
Postmortem
Situation: You’re working in IPython and you get a traceback.
10.2 Debugging workflow
Here we debug the file index_error.py. When running it, an IndexError is raised. Type %debug and
If you do have a non trivial bug, this is when debugging strategies kick in. There is no silver bullet. Yet, drop into the debugger.
strategies help: In [1]: %run index_error.py
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 IndexError Traceback (most recent call last)
/home/varoquau/dev/scipy-lecture-notes/advanced/debugging/index_error.py in <module>()
code, with short modify-run-fail cycles
6
1. Make it fail reliably. Find a test case that makes the code fail every time. 7 if __name__ == '__main__':
----> 8 index_error()
2. Divide and Conquer. Once you have a failing test case, isolate the failing code. 9
• Which module.
/home/varoquau/dev/scipy-lecture-notes/advanced/debugging/index_error.py in index_error()
• Which function. 3 def index_error():
4 lst = list('foobar')
• Which line of code. ----> 5 print lst[len(lst)]
=> isolate a small reproducible failure: a test case 6
7 if __name__ == '__main__':
3. Change one thing at a time and re-run the failing test case.
IndexError: list index out of range
4. Use the debugger to understand what is going wrong.
5. Take notes and be patient. It may take a while. In [2]: %debug
> /home/varoquau/dev/scipy-lecture-notes/advanced/debugging/index_error.py(5)index_error()
4 lst = list('foobar')
Note: Once you have gone through this process: isolated a tight piece of code reproducing the bug ----> 5 print lst[len(lst)]
and fix the bug using this piece of code, add the corresponding code to your test suite. 6
ipdb> list
1 """Small snippet to raise an IndexError."""
10.3 Using the Python debugger 2
3 def index_error():
The python debugger, pdb: https://docs.python.org/library/pdb.html, allows you to inspect your code 4 lst = list('foobar')
interactively. ----> 5 print lst[len(lst)]
6
Specifically it allows you to: 7 if __name__ == '__main__':
8 index_error()
• View the source code. 9
• Walk up and down the call stack. (continues on next page)
10.2. Debugging workflow 316 10.3. Using the Python debugger 317

3
ipdb> len(lst) 1---> 4 import numpy as np
6 5 import scipy as sp
ipdb> print(lst[len(lst)-1])
r ipdb> b 34
ipdb> quit Breakpoint 2 at /home/varoquau/dev/scipy-lecture-notes/advanced/optimizing/wiener_
˓→filtering.py:34
In [3]:
• Continue execution to next breakpoint with c(ont(inue)):
Post-mortem debugging without IPython ipdb> c

> /home/varoquau/dev/scipy-lecture-notes/advanced/optimizing/wiener_filtering.
˓→py(34)iterated_wiener()
In some situations you cannot use IPython, for instance to debug a script that wants to be called from 33 """
the command line. In this case, you can call the script with python -m pdb script.py: 2--> 34 noisy_img = noisy_img
$ python -m pdb index_error.py 35 denoised_img = local_mean(noisy_img, size=size)
> /home/varoquau/dev/scipy-lecture-notes/advanced/optimizing/index_error.py(1)<module>()
-> """Small snippet to raise an IndexError.""" • Step into code with n(ext) and s(tep): next jumps to the next statement in the current execution
(Pdb) continue context, while step will go across execution contexts, i.e. enable exploring inside function calls:
File "/usr/lib/python2.6/pdb.py", line 1296, in main ipdb> s
pdb._runscript(mainpyfile) > /home/varoquau/dev/scipy-lecture-notes/advanced/optimizing/wiener_filtering.
File "/usr/lib/python2.6/pdb.py", line 1215, in _runscript ˓→py(35)iterated_wiener()
self.run(statement) 2 34 noisy_img = noisy_img

File "/usr/lib/python2.6/bdb.py", line 372, in run ---> 35 denoised_img = local_mean(noisy_img, size=size)
exec cmd in globals, locals 36 l_var = local_var(noisy_img, size=size)
File "<string>", line 1, in <module>
File "index_error.py", line 8, in <module> ipdb> n
index_error() > /home/varoquau/dev/scipy-lecture-notes/advanced/optimizing/wiener_filtering.
File "index_error.py", line 5, in index_error ˓→py(36)iterated_wiener()
print lst[len(lst)] 35 denoised_img = local_mean(noisy_img, size=size)

IndexError: list index out of range ---> 36 l_var = local_var(noisy_img, size=size)
Uncaught exception. Entering post mortem debugging 37 for i in range(3):
Running 'cont' or 'step' will restart the program
> /home/varoquau/dev/scipy-lecture-notes/advanced/optimizing/index_error.py(5)index_error() • Step a few lines and explore the local variables:
-> print(lst[len(lst)])
(Pdb) ipdb> n
> /home/varoquau/dev/scipy-lecture-notes/advanced/optimizing/wiener_filtering.
˓→py(37)iterated_wiener()
36 l_var = local_var(noisy_img, size=size)

Step-by-step execution ---> 37 for i in range(3):
38 res = noisy_img - denoised_img
Situation: You believe a bug exists in a module but are not sure where. ipdb> print(l_var)
For instance we are trying to debug wiener_filtering.py. Indeed the code runs, but the filtering does [[5868 5379 5316 ..., 5071 4799 5149]
not work well. [5013 363 437 ..., 346 262 4355]
[5379 410 344 ..., 392 604 3377]
• Run the script in IPython with the debugger using %run -d wiener_filtering.p : ...,
[ 435 362 308 ..., 275 198 1632]
In [1]: %run -d wiener_filtering.py [ 548 392 290 ..., 248 263 1653]
*** Blank or comment [ 466 789 736 ..., 1835 1725 1940]]
*** Blank or comment ipdb> print(l_var.min())
*** Blank or comment 0
Breakpoint 1 at /home/varoquau/dev/scipy-lecture-notes/advanced/optimizing/wiener_
˓→filtering.py:4
NOTE: Enter 'c' at the ipdb> prompt to start your script.

Oh dear, nothing but integers, and 0 variation. Here is our bug, we are doing integer arithmetic.
> <string>(1)<module>()
Raising exception on numerical errors
• Set a break point at line 34 using b 34:
When we run the wiener_filtering.py file, the following warnings are raised:
ipdb> n
> /home/varoquau/dev/scipy-lecture-notes/advanced/optimizing/wiener_filtering.py(4) In [2]: %run wiener_filtering.py
˓→<module>() wiener_filtering.py:40: RuntimeWarning: divide by zero encountered in divide
(continues on next page) noise_level = (1 - noise/l_var )
10.3. Using the Python debugger 318 10.3. Using the Python debugger 319
We can turn these warnings in exception, which enables us to do post-mortem debugging on them, • The Visual Studio Code integrated development environment includes a debugging mode.
and find our problem more quickly:
• The Mu editor is a simple Python editor that includes a debugging mode.
In [3]: np.seterr(all='raise')
Out[3]: {'divide': 'print', 'invalid': 'print', 'over': 'print', 'under': 'ignore'}
In [4]: %run wiener_filtering.py
---------------------------------------------------------------------------
10.3.2 Debugger commands and interaction
FloatingPointError Traceback (most recent call last)
/home/esc/anaconda/lib/python2.7/site-packages/IPython/utils/py3compat.pyc in execfile(fname, l(list) Lists the code at the current position
˓→ *where) u(p) Walk up the call stack
176 else: d(own) Walk down the call stack
177 filename = fname n(ext) Execute the next line (does not go down in new functions)
--> 178 __builtin__.execfile(filename, *where)
s(tep) Execute the next statement (goes down in new functions)
/home/esc/physique-cuso-python-2013/scipy-lecture-notes/advanced/debugging/wiener_filtering. bt Print the call stack
˓→py in <module>() a Print the local variables
55 pl.matshow(noisy_face[cut], cmap=pl.cm.gray) !command Execute the given Python command (by opposition to pdb commands
56
---> 57 denoised_face = iterated_wiener(noisy_face)
58 pl.matshow(denoised_face[cut], cmap=pl.cm.gray)
59 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
/home/esc/physique-cuso-python-2013/scipy-lecture-notes/advanced/debugging/wiener_filtering.
in the current frame with the same name: use different names than your local variable when
˓→py in iterated_wiener(noisy_img, size)
38 res = noisy_img - denoised_img

typing code in the debugger.
39 noise = (res**2).sum()/res.size
---> 40 noise_level = (1 - noise/l_var )
41 noise_level[noise_level<0] = 0 Getting help when in the debugger
42 denoised_img += noise_level*res
Type h or help to access the interactive help:
FloatingPointError: divide by zero encountered in divide ipdb> help
Documented commands (type help <topic>):

========================================
Other ways of starting a debugger EOF bt cont enable jump pdef r tbreak w
• Raising an exception as a poor man break point a c continue exit l pdoc restart u whatis
alias cl d h list pinfo return unalias where
If you find it tedious to note the line number to set a break point, you can simply raise an exception args clear debug help n pp run unt
at the point that you want to inspect and use IPython’s %debug. Note that in this case you cannot b commands disable ignore next q s until
step or continue the execution. break condition down j p quit step up
• Debugging test failures using nosetests Miscellaneous help topics:

You can run nosetests --pdb to drop in post-mortem debugging on exceptions, and nosetests ==========================
exec pdb
--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 Undocumented commands:
ipdbplugin. You can than pass --ipdb and --ipdb-failure options to nosetests. ======================
retval rv
• Calling the debugger explicitly
Insert the following line where you want to drop in the debugger:
import pdb; pdb.set_trace()
10.4 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
Warning: When running nosetests, the output is captured, and thus it seems that the debugger useless. For this we turn to the gnu debugger, gdb, available on Linux.
does not work. Simply run the nosetests with the -s flag.
Before we start with gdb, let us add a few Python-specific tools to it. For this we add a few macros to
our ~/.gdbinit. The optimal choice of macro depends on your Python version and your gdb version. I
have added a simplified version in gdbinit, but feel free to read DebuggingWithGdb.
Graphical debuggers and alternatives
To debug with gdb the Python script segfault.py, we can run the script in gdb as follows
• pudb is a good semi-graphical debugger with a text user interface in the console.
10.3. Using the Python debugger 320 10.4. Debugging segmentation faults using gdb 321

$ gdb python
... 1630 ../Python/ceval.c: No such file or directory.
(gdb) run segfault.py in ../Python/ceval.c
Starting program: /usr/bin/python segfault.py (gdb) pyframe
[Thread debugging using libthread_db enabled] segfault.py (12): print_big_array
Program received signal SIGSEGV, Segmentation fault. The corresponding code is:
_strided_byte_copy (dst=0x8537478 "\360\343G", outstrides=4, src=
0x86c0690 <Address 0x86c0690 out of bounds>, instrides=32, N=3, def make_big_array(small_array):
elsize=4) big_array = stride_tricks.as_strided(small_array,
at numpy/core/src/multiarray/ctors.c:365 shape=(2e6, 2e6), strides=(32, 32))
365 _FAST_MOVE(Int32); return big_array
(gdb)
def print_big_array(small_array):
We get a segfault, and gdb captures it for post-mortem debugging in the C level stack (not the Python big_array = make_big_array(small_array)
call stack). We can debug the C call stack using gdb’s commands:
Thus the segfault happens when printing big_array[-10:]. The reason is simply that big_array has
(gdb) up been allocated with its end outside the program memory.
#1 0x004af4f5 in _copy_from_same_shape (dest=<value optimized out>,
src=<value optimized out>, myfunc=0x496780 <_strided_byte_copy>,
swap=0) Note: For a list of Python-specific commands defined in the gdbinit, read the source of this file.
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
Wrap up exercise
code that triggers this segfault, so we go up the stack until we hit the Python execution loop:
(gdb) up The following script is well documented and hopefully legible. It seeks to answer a problem of actual
#8 0x080ddd23 in call_function (f= interest for numerical computing, but it does not work. . . Can you debug it?
Frame 0x85371ec, for file /home/varoquau/usr/lib/python2.6/site-packages/numpy/core/
˓→arrayprint.py, line 156, in _leading_trailing (a=<numpy.ndarray at remote 0x85371b0>, _nc=
Python source code: to_debug.py
˓→<module at remote 0xb7f93a64>), throwflag=0)
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.py, line 156, in _leading_trailing (a=<numpy.ndarray at remote 0x85371b0>, _nc=
˓→<module at remote 0xb7f93a64>), throwflag=0)
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 at remote 0x853ecf0>, big_array=<numpy.ndarray at remote 0x853ed20>), throwflag=0)␣
˓→at ../Python/ceval.c:1630
10.4. Debugging segmentation faults using gdb 322 10.4. Debugging segmentation faults using gdb 323
• Writing faster numerical code

– Additional Links
11.1 Optimization workflow
11
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.
CHAPTER 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.
11.2 Profiling Python code
Optimizing code
No optimization without measuring!
• Measure: profiling, timing

• You’ll have surprises: the fastest code is not always what you think
11.2.1 Timeit
In IPython, use timeit (https://docs.python.org/library/timeit.html) to time elementary operations:
Donald Knuth In [1]: import numpy as np
In [2]: a = np.arange(1000)
“Premature optimization is the root of all evil”
In [3]: %timeit a ** 2
Author: Gaël Varoquaux 100000 loops, best of 3: 5.73 us per loop
This chapter deals with strategies to make Python code go faster. In [4]: %timeit a ** 2.1
1000 loops, best of 3: 154 us per loop
Prerequisites
In [5]: %timeit a * a
100000 loops, best of 3: 5.56 us per loop
• line_profiler
Use this to guide your choice between strategies.
Chapters contents Note: For long running calls, using %time instead of %timeit; it is less precise but faster
• Optimization workflow
• Profiling Python code 11.2.2 Profiler
– Timeit Useful when you have a large program to profile, for example the following file:
– Profiler # For this example to run, you also need the 'ica.py' file
– Line-profiler
import numpy as np
• Making code go faster from scipy import linalg
– Algorithmic optimization
from ica import fastica
∗ Example of the SVD
324 11.1. Optimization workflow 325

(continued from previous page) 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
def test(): rest of the code is useless.
data = np.random.random((5000, 100))
u, s, v = linalg.svd(data)
pca = np.dot(u[:, :10].T, data) Profiling outside of IPython, running ‘‘cProfile‘‘
results = fastica(pca.T, whiten=False)
Similar profiling can be done outside of IPython, simply calling the built-in Python profilers cProfile
if __name__ == '__main__': and profile.
test()
$ python -m cProfile -o demo.prof demo.py
Using the -o switch will output the profiler results to the file demo.prof to view with an external tool.
Note: This is a combination of two unsupervised learning techniques, principal component analysis This can be useful if you wish to process the profiler output with a visualization tool.
(PCA) and independent component analysis (ICA). 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
separation technique, for example to unmix multiple signals that have been recorded through multiple 11.2.3 Line-profiler
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. The profiler tells us which function takes most of the time, but not where it is called.
For this, we use the line_profiler: in the source file, we decorate a few functions that we want to inspect
To run it, you also need to download the ica module. In IPython we can time the script: with @profile (no need to import it)
In [1]: %run -t demo.py @profile

def test():
IPython CPU timings (estimated): data = np.random.random((5000, 100))
User : 14.3929 s. u, s, v = linalg.svd(data)
System: 0.256016 s. pca = np.dot(u[:, :10], data)
results = fastica(pca.T, whiten=False)
and profile it:
Then we run the script using the kernprof.py program, with switches -l, --line-by-line and -v,
In [2]: %run -p demo.py --view to use the line-by-line profiler and view the results in addition to saving them:
916 function calls in 14.551 CPU seconds $ kernprof.py -l -v demo.py
Ordered by: internal time Wrote profile results to demo.py.lprof

Timer unit: 1e-06 s
ncalls tottime percall cumtime percall filename:lineno (function)
1 14.457 14.457 14.479 14.479 decomp.py:849 (svd) File: demo.py
1 0.054 0.054 0.054 0.054 {method 'random_sample' of 'mtrand.RandomState'␣ Function: test at line 5
˓→objects} Total time: 14.2793 s
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} Line # Hits Time Per Hit % Time Line Contents
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) 5 @profile
6 0.001 0.000 0.001 0.000 ica.py:192 (g) 6 def test():
14 0.001 0.000 0.001 0.000 {numpy.linalg.lapack_lite.dsyevd} 7 1 19015 19015.0 0.1 data = np.random.random((5000, 100))
19 0.001 0.000 0.001 0.000 twodim_base.py:204 (diag) 8 1 14242163 14242163.0 99.7 u, s, v = linalg.svd(data)
1 0.001 0.001 0.008 0.008 ica.py:69 (_ica_par) 9 1 10282 10282.0 0.1 pca = np.dot(u[:10, :], data)
1 0.001 0.001 14.551 14.551 {execfile} 10 1 7799 7799.0 0.1 results = fastica(pca.T, whiten=False)
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) The SVD is taking all the time. We need to optimise this line.
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>) 11.3 Making code go faster
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__) Once we have identified the bottlenecks, we need to make the corresponding code go faster.
35 0.000 0.000 0.001 0.000 defmatrix.py:43 (asmatrix)
21 0.000 0.000 0.001 0.000 defmatrix.py:287 (__mul__)
41 0.000 0.000 0.000 0.000 {numpy.core.multiarray.zeros} 11.3.1 Algorithmic optimization
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) The first thing to look for is algorithmic optimization: are there ways to compute less, or better?
...
11.2. Profiling Python code 326 11.3. Making code go faster 327
For a high-level view of the problem, a good understanding of the maths behind the algorithm helps. Know your computational linear algebra. When in doubt, explore scipy.linalg, and use %timeit to
However, it is not uncommon to find simple changes, like moving computation or memory allocation try out different alternatives on your data.
outside a for loop, that bring in big gains.
Example of the SVD

11.4 Writing faster numerical code
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. A complete discussion on advanced use of numpy is found in chapter Advanced NumPy, or in the article
However, in both of these example, we are not using all the output of the SVD, but only the first few The NumPy array: a structure for efficient numerical computation by van der Walt et al. Here we discuss
rows of its first return argument. If we use the svd implementation of scipy, we can ask for an incomplete only some commonly encountered tricks to make code faster.
version of the SVD. Note that implementations of linear algebra in scipy are richer then those in numpy • Vectorizing for loops
and should be preferred.
Find tricks to avoid for loops using numpy arrays. For this, masks and indices arrays can be useful.
In [3]: %timeit np.linalg.svd(data)
1 loops, best of 3: 14.5 s per loop • Broadcasting
Use broadcasting to do operations on arrays as small as possible before combining them.
In [4]: from scipy import linalg
• In place operations
In [5]: %timeit linalg.svd(data)
1 loops, best of 3: 14.2 s per loop In [1]: a = np.zeros(1e7)
In [6]: %timeit linalg.svd(data, full_matrices=False) In [2]: %timeit global a ; a = 0*a

1 loops, best of 3: 295 ms per loop 10 loops, best of 3: 111 ms per loop
In [7]: %timeit np.linalg.svd(data, full_matrices=False) In [3]: %timeit global a ; a *= 0

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

10 loops, best of 3: 124 ms per loop
In [2]: %timeit demo.
demo.fastica demo.np demo.prof.pdf demo.py demo.pyc In [3]: %timeit a + 1
demo.linalg demo.prof demo.prof.png demo.py.lprof demo.test 10 loops, best of 3: 112 ms per loop
In [2]: %timeit demo.test() • Beware of cache effects

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 Memory access is cheaper when it is grouped: accessing a big array in a continuous way is much
1 loops, best of 3: 17.5 s per loop faster than random access. This implies amongst other things that smaller strides are faster
(see CPU cache effects):
In [3]: import demo_opt
In [1]: c = np.zeros((1e4, 1e4), order='C')
In [4]: %timeit demo_opt.test()
1 loops, best of 3: 208 ms per loop In [2]: %timeit c.sum(axis=0)
1 loops, best of 3: 3.89 s per loop
Real incomplete SVDs, e.g. computing only the first 10 eigenvectors, can be computed with arpack, In [3]: %timeit c.sum(axis=1)
available in scipy.sparse.linalg.eigsh. 1 loops, best of 3: 188 ms per loop
In [4]: c.strides
Computational linear algebra
Out[4]: (80000, 8)
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 This is the reason why Fortran ordering or C ordering may make a big difference on operations:
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.
11.3. Making code go faster 328 11.4. Writing faster numerical code 329
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)
12
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: CHAPTER
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 Sparse Matrices in SciPy
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. Don’t base your optimization on
theoretical considerations.
Author: Robert Cimrman

11.4.1 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 12.1 Introduction
your repository, you could try: asv
(dense) matrix is:
• If you need some interactive visualization why not try RunSnakeRun
• 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 (*)
12.1.1 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)
11.4. Writing faster numerical code 330 331


>>> plt.xlabel('size n')
Text(...'size n')
>>> plt.ylabel('memory [MB]')
Text(...'memory [MB]')
12.1.2 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
• cons: depends on actual storage scheme, (*) usually does not hold
12.1.3 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
• natural language processing
– nonzero at (i, j) means that the document i contains the word j
• ... 12.2 Storage Schemes
12.1.4 Prerequisites • seven sparse matrix types in scipy.sparse:
• numpy 1. csc_matrix: Compressed Sparse Column format
• scipy 2. csr_matrix: Compressed Sparse Row format
• matplotlib (optional) 3. bsr_matrix: Block Sparse Row format
• ipython (the enhancements come handy) 4. lil_matrix: List of Lists format

5. dok_matrix: Dictionary of Keys format
12.1.5 Sparsity Structure Visualization 6. coo_matrix: COOrdinate format (aka IJV, triplet format)
• spy() from matplotlib 7. dia_matrix: DIAgonal format
• example plots: • 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
12.1. Introduction 332 12.2. Storage Schemes 333

12.2.1 Common Methods – rather specialized

• all scipy.sparse classes are subclasses of spmatrix – solving PDEs by finite differences
– default implementation of arithmetic operations – with an iterative solver
∗ always converts to CSR
Examples
∗ subclasses override for efficiency
• create some DIA matrices:
– shape, data type set/get
>>> data = np.array([[1, 2, 3, 4]]).repeat(3, axis=0)
– nonzero indices
>>> data
– format conversion, interaction with NumPy (toarray(), todense()) array([[1, 2, 3, 4],
[1, 2, 3, 4],
– ... [1, 2, 3, 4]])
• attributes: >>> offsets = np.array([0, -1, 2])
>>> mtx = sparse.dia_matrix((data, offsets), shape=(4, 4))
– mtx.A - same as mtx.toarray() >>> mtx
<4x4 sparse matrix of type '<... 'numpy.int64'>'
– mtx.T - transpose (same as mtx.transpose()) with 9 stored elements (3 diagonals) in DIAgonal format>
– mtx.H - Hermitian (conjugate) transpose >>> mtx.todense()
matrix([[1, 0, 3, 0],
– mtx.real - real part of complex matrix [1, 2, 0, 4],
[0, 2, 3, 0],
– mtx.imag - imaginary part of complex matrix [0, 0, 3, 4]])
– mtx.size - the number of nonzeros (same as self.getnnz())
>>> data = np.arange(12).reshape((3, 4)) + 1
– mtx.shape - the number of rows and columns (tuple) >>> data
array([[ 1, 2, 3, 4],
• data usually stored in NumPy arrays
[ 5, 6, 7, 8],
[ 9, 10, 11, 12]])
12.2.2 Sparse Matrix Classes >>> mtx = sparse.dia_matrix((data, offsets), shape=(4, 4))
>>> mtx.data
Diagonal Format (DIA) array([[ 1, 2, 3, 4],
[ 5, 6, 7, 8],
• very simple scheme
[ 9, 10, 11, 12]])
• diagonals in dense NumPy array of shape (n_diag, length) >>> mtx.offsets
array([ 0, -1, 2], dtype=int32)
– fixed length -> waste space a bit when far from main diagonal >>> print(mtx)
(0, 0) 1
– subclass of _data_matrix (sparse matrix classes with .data attribute)
(1, 1) 2
• offset for each diagonal (2, 2) 3
(3, 3) 4
– 0 is the main diagonal (1, 0) 5
– negative offset = below (2, 1) 6
(3, 2) 7
– positive offset = above (0, 2) 11
(1, 3) 12
• fast matrix * vector (sparsetools) >>> mtx.todense()
• fast and easy item-wise operations matrix([[ 1, 0, 11, 0],
[ 5, 2, 0, 12],
– manipulate data array directly (fast NumPy machinery) [ 0, 6, 3, 0],
[ 0, 0, 7, 4]])
• constructor accepts:
– dense matrix (array) • explanation with a scheme:
– sparse matrix offset: row
– shape tuple (create empty matrix)
2: 9
– (data, offsets) tuple 1: --10------
0: 1 . 11 .
• no slicing, no individual item access -1: 5 2 . 12
• use: -2: . 6 3 .
12.2. Storage Schemes 334 12.2. Storage Schemes 335


-3: . . 7 4 with 5 stored elements in LInked List format>
---------8 >>> print(mtx)
(0, 1) 1.0
• matrix-vector multiplication (0, 2) 1.0
(0, 3) 1.0
>>> vec = np.ones((4, )) (1, 1) 1.0
>>> vec (1, 3) 1.0
array([1., 1., 1., 1.]) >>> mtx.todense()
>>> mtx * vec matrix([[0., 1., 1., 1., 0.],
array([12., 19., 9., 11.]) [0., 1., 0., 1., 0.],
>>> mtx.toarray() * vec [0., 0., 0., 0., 0.],
array([[ 1., 0., 11., 0.], [0., 0., 0., 0., 0.]])
[ 5., 2., 0., 12.], >>> mtx.toarray()
[ 0., 6., 3., 0.], array([[0., 1., 1., 1., 0.],
[ 0., 0., 7., 4.]]) [0., 1., 0., 1., 0.],
[0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0.]])
List of Lists Format (LIL)
• row-based linked list • more slicing and indexing:
– each row is a Python list (sorted) of column indices of non-zero elements >>> mtx = sparse.lil_matrix([[0, 1, 2, 0], [3, 0, 1, 0], [1, 0, 0, 1]])
>>> mtx.todense()
– rows stored in a NumPy array (dtype=np.object) matrix([[0, 1, 2, 0],
[3, 0, 1, 0],
– non-zero values data stored analogously
[1, 0, 0, 1]]...)
• efficient for constructing sparse matrices incrementally >>> print(mtx)
(0, 1) 1
• constructor accepts: (0, 2) 2
(1, 0) 3
– dense matrix (array)
(1, 2) 1
– sparse matrix (2, 0) 1
(2, 3) 1
– shape tuple (create empty matrix) >>> mtx[:2, :]
• flexible slicing, changing sparsity structure is efficient <2x4 sparse matrix of type '<... 'numpy.int64'>'
with 4 stored elements in LInked List format>
• slow arithmetics, slow column slicing due to being row-based >>> mtx[:2, :].todense()
matrix([[0, 1, 2, 0],
• use: [3, 0, 1, 0]]...)
– when sparsity pattern is not known apriori or changes >>> mtx[1:2, [0,2]].todense()
matrix([[3, 1]]...)
– example: reading a sparse matrix from a text file >>> mtx.todense()
matrix([[0, 1, 2, 0],
Examples [3, 0, 1, 0],
[1, 0, 0, 1]]...)
• create an empty LIL matrix:
>>> mtx = sparse.lil_matrix((4, 5)) Dictionary of Keys Format (DOK)
• subclass of Python dict
• prepare random data:
– keys are (row, column) index tuples (no duplicate entries allowed)
>>> from numpy.random import rand
>>> data = np.round(rand(2, 3)) – values are corresponding non-zero values
>>> data
• efficient for constructing sparse matrices incrementally
array([[1., 1., 1.],
[1., 0., 1.]]) • constructor accepts:
• assign the data using fancy indexing: – dense matrix (array)

– sparse matrix
>>> mtx[:2, [1, 2, 3]] = data
>>> mtx – shape tuple (create empty matrix)
<4x5 sparse matrix of type '<... 'numpy.float64'>'
• efficient O(1) access to individual elements

• flexible slicing, changing sparsity structure is efficient • very fast conversion to and from CSR/CSC formats
• can be efficiently converted to a coo_matrix once constructed • fast matrix * vector (sparsetools)
• slow arithmetics (for loops with dict.iteritems()) • fast and easy item-wise operations
• use: – manipulate data array directly (fast NumPy machinery)
– when sparsity pattern is not known apriori or changes • no slicing, no arithmetics (directly)
• use:
Examples
– facilitates fast conversion among sparse formats
• create a DOK matrix element by element:
– when converting to other format (usually CSR or CSC), duplicate entries are summed
>>> mtx = sparse.dok_matrix((5, 5), dtype=np.float64) together
>>> mtx
∗ facilitates efficient construction of finite element matrices
<5x5 sparse matrix of type '<... 'numpy.float64'>'
with 0 stored elements in Dictionary Of Keys format>
>>> for ir in range(5): Examples
... for ic in range(5):
... mtx[ir, ic] = 1.0 * (ir != ic) • create empty COO matrix:
>>> mtx
<5x5 sparse matrix of type '<... 'numpy.float64'>' >>> mtx = sparse.coo_matrix((3, 4), dtype=np.int8)
with 20 stored elements in Dictionary Of Keys format> >>> mtx.todense()
>>> mtx.todense() matrix([[0, 0, 0, 0],
matrix([[0., 1., 1., 1., 1.], [0, 0, 0, 0],
[1., 0., 1., 1., 1.], [0, 0, 0, 0]], dtype=int8)
[1., 1., 0., 1., 1.],
[1., 1., 1., 0., 1.], • create using (data, ij) tuple:
[1., 1., 1., 1., 0.]])
>>> row = np.array([0, 3, 1, 0])
>>> col = np.array([0, 3, 1, 2])
• slicing and indexing:
>>> data = np.array([4, 5, 7, 9])
>>> mtx[1, 1] >>> mtx = sparse.coo_matrix((data, (row, col)), shape=(4, 4))
0.0 >>> mtx
>>> mtx[1, 1:3] <4x4 sparse matrix of type '<... 'numpy.int64'>'
<1x2 sparse matrix of type '<... 'numpy.float64'>' with 4 stored elements in COOrdinate format>
with 1 stored elements in Dictionary Of Keys format> >>> mtx.todense()
>>> mtx[1, 1:3].todense() matrix([[4, 0, 9, 0],
matrix([[0., 1.]]) [0, 7, 0, 0],
>>> mtx[[2,1], 1:3].todense() [0, 0, 0, 0],
matrix([[1., 0.], [0, 0, 0, 5]])
[0., 1.]])
• duplicates entries are summed together:
Coordinate Format (COO) >>> row = np.array([0, 0, 1, 3, 1, 0, 0])

>>> col = np.array([0, 2, 1, 3, 1, 0, 0])
• also known as the ‘ijv’ or ‘triplet’ format >>> data = np.array([1, 1, 1, 1, 1, 1, 1])
>>> mtx = sparse.coo_matrix((data, (row, col)), shape=(4, 4))
– three NumPy arrays: row, col, data >>> mtx.todense()
– data[i] is value at (row[i], col[i]) position matrix([[3, 0, 1, 0],
[0, 2, 0, 0],
– permits duplicate entries [0, 0, 0, 0],
[0, 0, 0, 1]])
– subclass of _data_matrix (sparse matrix classes with .data attribute)
• fast format for constructing sparse matrices • no slicing. . . :
• constructor accepts: >>> mtx[2, 3]
– dense matrix (array) Traceback (most recent call last):
...
– sparse matrix TypeError: 'coo_matrix' object ...
– (data, ij) tuple

Compressed Sparse Row Format (CSR) (continued from previous page)
• row oriented [4, 5, 6]]...)

>>> mtx.data
– three NumPy arrays: indices, indptr, data array([1, 2, 3, 4, 5, 6]...)
>>> mtx.indices
∗ indices is array of column indices array([0, 2, 2, 0, 1, 2], dtype=int32)
∗ data is array of corresponding nonzero values >>> mtx.indptr
∗ indptr points to row starts in indices and data
∗ length is n_row + 1, last item = number of values = length of both indices and • create using (data, indices, indptr) tuple:
data >>> data = np.array([1, 2, 3, 4, 5, 6])
∗ nonzero values of the i-th row are data[indptr[i]:indptr[i+1]] with column indices >>> indices = np.array([0, 2, 2, 0, 1, 2])
indices[indptr[i]:indptr[i+1]] >>> indptr = np.array([0, 2, 3, 6])
>>> mtx = sparse.csr_matrix((data, indices, indptr), shape=(3, 3))
∗ item (i, j) can be accessed as data[indptr[i]+k], where k is position of j in in- >>> mtx.todense()
dices[indptr[i]:indptr[i+1]] matrix([[1, 0, 2],
[0, 0, 3],
– subclass of _cs_matrix (common CSR/CSC functionality) [4, 5, 6]])
∗ subclass of _data_matrix (sparse matrix classes with .data attribute)
• fast matrix vector products and other arithmetics (sparsetools) Compressed Sparse Column Format (CSC)
• constructor accepts: • column oriented
– dense matrix (array) – three NumPy arrays: indices, indptr, data
– sparse matrix ∗ indices is array of row indices
– shape tuple (create empty matrix) ∗ data is array of corresponding nonzero values
– (data, ij) tuple ∗ indptr points to column starts in indices and data
– (data, indices, indptr) tuple ∗ length is n_col + 1, last item = number of values = length of both indices and
data
• efficient row slicing, row-oriented operations
∗ nonzero values of the i-th column are data[indptr[i]:indptr[i+1]] with row indices
• slow column slicing, expensive changes to the sparsity structure
indices[indptr[i]:indptr[i+1]]
• use:
∗ item (i, j) can be accessed as data[indptr[j]+k], where k is position of i in in-
– actual computations (most linear solvers support this format) dices[indptr[j]:indptr[j+1]]
– subclass of _cs_matrix (common CSR/CSC functionality)
Examples
∗ subclass of _data_matrix (sparse matrix classes with .data attribute)
• create empty CSR matrix:
• fast matrix vector products and other arithmetics (sparsetools)
>>> mtx = sparse.csr_matrix((3, 4), dtype=np.int8)
>>> mtx.todense()
matrix([[0, 0, 0, 0], – dense matrix (array)
[0, 0, 0, 0],
[0, 0, 0, 0]], dtype=int8) – sparse matrix
• create using (data, ij) tuple:
– (data, ij) tuple
>>> row = np.array([0, 0, 1, 2, 2, 2])
– (data, indices, indptr) tuple
>>> col = np.array([0, 2, 2, 0, 1, 2])
>>> data = np.array([1, 2, 3, 4, 5, 6]) • efficient column slicing, column-oriented operations
>>> mtx = sparse.csr_matrix((data, (row, col)), shape=(3, 3))
>>> mtx • slow row slicing, expensive changes to the sparsity structure
• use:
with 6 stored elements in Compressed Sparse Row format>
>>> mtx.todense() – actual computations (most linear solvers support this format)
matrix([[1, 0, 2],
[0, 0, 3],

Examples – shape tuple (create empty matrix)
• create empty CSC matrix: – (data, ij) tuple

– (data, indices, indptr) tuple
>>> mtx = sparse.csc_matrix((3, 4), dtype=np.int8)
>>> mtx.todense() • many arithmetic operations considerably more efficient than CSR for sparse matrices with dense
matrix([[0, 0, 0, 0], sub-matrices
[0, 0, 0, 0],
[0, 0, 0, 0]], dtype=int8) • use:
– like CSR
• create using (data, ij) tuple:
– vector-valued finite element discretizations
>>> row = np.array([0, 0, 1, 2, 2, 2])
>>> col = np.array([0, 2, 2, 0, 1, 2]) Examples
>>> data = np.array([1, 2, 3, 4, 5, 6])
>>> mtx = sparse.csc_matrix((data, (row, col)), shape=(3, 3)) • create empty BSR matrix with (1, 1) block size (like CSR. . . ):
>>> mtx
<3x3 sparse matrix of type '<... 'numpy.int64'>' >>> mtx = sparse.bsr_matrix((3, 4), dtype=np.int8)
with 6 stored elements in Compressed Sparse Column format> >>> mtx
>>> mtx.todense() <3x4 sparse matrix of type '<... 'numpy.int8'>'
matrix([[1, 0, 2], with 0 stored elements (blocksize = 1x1) in Block Sparse Row format>
[0, 0, 3], >>> mtx.todense()
[4, 5, 6]]...) matrix([[0, 0, 0, 0],
>>> mtx.data [0, 0, 0, 0],
array([1, 4, 5, 2, 3, 6]...) [0, 0, 0, 0]], dtype=int8)
>>> mtx.indices
array([0, 2, 2, 0, 1, 2], dtype=int32)
• create empty BSR matrix with (3, 2) block size:
>>> mtx.indptr
array([0, 2, 3, 6], dtype=int32) >>> mtx = sparse.bsr_matrix((3, 4), blocksize=(3, 2), dtype=np.int8)
>>> mtx
• create using (data, indices, indptr) tuple: <3x4 sparse matrix of type '<... 'numpy.int8'>'
with 0 stored elements (blocksize = 3x2) in Block Sparse Row format>
>>> data = np.array([1, 4, 5, 2, 3, 6]) >>> mtx.todense()
>>> indices = np.array([0, 2, 2, 0, 1, 2]) matrix([[0, 0, 0, 0],
>>> indptr = np.array([0, 2, 3, 6]) [0, 0, 0, 0],
>>> mtx = sparse.csc_matrix((data, indices, indptr), shape=(3, 3)) [0, 0, 0, 0]], dtype=int8)
>>> mtx.todense()
matrix([[1, 0, 2],
– a bug?
[0, 0, 3],
[4, 5, 6]]) • create using (data, ij) tuple with (1, 1) block size (like CSR. . . ):
>>> row = np.array([0, 0, 1, 2, 2, 2])

Block Compressed Row Format (BSR) >>> col = np.array([0, 2, 2, 0, 1, 2])
>>> data = np.array([1, 2, 3, 4, 5, 6])
• basically a CSR with dense sub-matrices of fixed shape instead of scalar items >>> mtx = sparse.bsr_matrix((data, (row, col)), shape=(3, 3))
– block size (R, C) must evenly divide the shape of the matrix (M, N) >>> mtx
– three NumPy arrays: indices, indptr, data with 6 stored elements (blocksize = 1x1) in Block Sparse Row format>
>>> mtx.todense()
∗ indices is array of column indices for each block matrix([[1, 0, 2],
∗ data is array of corresponding nonzero values of shape (nnz, R, C) [0, 0, 3],
[4, 5, 6]]...)
∗ ... >>> mtx.data
array([[[1]],
– subclass of _cs_matrix (common CSR/CSC functionality)
∗ subclass of _data_matrix (sparse matrix classes with .data attribute) [[2]],
• fast matrix vector products and other arithmetics (sparsetools) [[3]],

[[4]],
– dense matrix (array)
[[5]],
– sparse matrix (continues on next page)

(continued from previous page) 12.3 Linear System Solvers

[[6]]]...) • sparse matrix/eigenvalue problem solvers live in scipy.sparse.linalg
>>> mtx.indices
array([0, 2, 2, 0, 1, 2], dtype=int32) • the submodules:
>>> mtx.indptr – dsolve: direct factorization methods for solving linear systems
– isolve: iterative methods for solving linear systems
• create using (data, indices, indptr) tuple with (2, 2) block size: – eigen: sparse eigenvalue problem solvers
>>> indptr = np.array([0, 2, 3, 6]) • all solvers are accessible from:
>>> indices = np.array([0, 2, 2, 0, 1, 2])
>>> data = np.array([1, 2, 3, 4, 5, 6]).repeat(4).reshape(6, 2, 2) >>> import scipy.sparse.linalg as spla
>>> mtx = sparse.bsr_matrix((data, indices, indptr), shape=(6, 6)) >>> spla.__all__
>>> mtx.todense() ['LinearOperator', 'Tester', 'arpack', 'aslinearoperator', 'bicg',
matrix([[1, 1, 0, 0, 2, 2], 'bicgstab', 'cg', 'cgs', 'csc_matrix', 'csr_matrix', 'dsolve',
[1, 1, 0, 0, 2, 2], 'eigen', 'eigen_symmetric', 'factorized', 'gmres', 'interface',
[0, 0, 0, 0, 3, 3], 'isolve', 'iterative', 'lgmres', 'linsolve', 'lobpcg', 'lsqr',
[0, 0, 0, 0, 3, 3], 'minres', 'np', 'qmr', 'speigs', 'spilu', 'splu', 'spsolve', 'svd',
[4, 4, 5, 5, 6, 6], 'test', 'umfpack', 'use_solver', 'utils', 'warnings']
[4, 4, 5, 5, 6, 6]])
>>> data
array([[[1, 1], 12.3.1 Sparse Direct Solvers
[1, 1]],
• default solver: SuperLU 4.0
[[2, 2],
– included in SciPy
[2, 2]],
– real and complex systems
[[3, 3],
[3, 3]], – both single and double precision
• optional: umfpack
[[4, 4],
[4, 4]], – real and complex systems
[[5, 5], – double precision only

[5, 5]], – recommended for performance
[[6, 6], – wrappers now live in scikits.umfpack
[6, 6]]])
– check-out the new scikits.suitesparse by Nathaniel Smith
12.2.3 Summary Examples

• import the whole module, and see its docstring:
Table 1: Summary of storage schemes.
>>> from scipy.sparse.linalg import dsolve
format matrix get fancy set item fancy solvers note
>>> help(dsolve)
* vector item get set
DIA sparsetools. . . . iterative has data array, specialized
• both superlu and umfpack can be used (if the latter is installed) as follows:
LIL via yes yes yes yes iterative arithmetics via CSR, incre-
CSR mental construction – prepare a linear system:
DOK python yes one yes yes iterative O(1) item access, incremen-
axis tal construction >>> import numpy as np
>>> from scipy import sparse
only
>>> mtx = sparse.spdiags([[1, 2, 3, 4, 5], [6, 5, 8, 9, 10]], [0, 1], 5, 5)
COO sparsetools. . . . iterative has data array, facilitates >>> mtx.todense()
fast conversion matrix([[ 1, 5, 0, 0, 0],
CSR sparsetoolsyes yes slow . any has data array, fast row-wise [ 0, 2, 8, 0, 0],
ops [ 0, 0, 3, 9, 0],
CSC sparsetoolsyes yes slow . any has data array, fast column- [ 0, 0, 0, 4, 10],
wise ops [ 0, 0, 0, 0, 5]])
BSR sparsetools. . . . specializedhas data array, specialized >>> rhs = np.array([1, 2, 3, 4, 5], dtype=np.float32)
– solve as single precision real:
12.2. Storage Schemes 344 12.3. Linear System Solvers 345


>>> mtx1 = mtx.astype(np.float32)
>>> x = dsolve.spsolve(mtx1, rhs, use_umfpack=False) x = linsolve.spsolve(mtx, rhs)
>>> print(x)
[106. -21. 5.5 -1.5 1. ] print('rezidual: %r ' % np.linalg.norm(mtx * x - rhs))
>>> print("Error: %s " % (mtx1 * x - rhs))
Error: [0. 0. 0. 0. 0.] • examples/direct_solve.py
– solve as double precision real: 12.3.2 Iterative Solvers

>>> mtx2 = mtx.astype(np.float64) • the isolve module contains the following solvers:
>>> x = dsolve.spsolve(mtx2, rhs, use_umfpack=True)
>>> print(x) – bicg (BIConjugate Gradient)
[106. -21. 5.5 -1.5 1. ]
– bicgstab (BIConjugate Gradient STABilized)
Error: [0. 0. 0. 0. 0.] – cg (Conjugate Gradient) - symmetric positive definite matrices only
– cgs (Conjugate Gradient Squared)
– solve as single precision complex:
– gmres (Generalized Minimal RESidual)
>>> mtx1 = mtx.astype(np.complex64)
>>> x = dsolve.spsolve(mtx1, rhs, use_umfpack=False) – minres (MINimum RESidual)
>>> print(x)
[106. +0.j -21. +0.j 5.5+0.j -1.5+0.j 1. +0.j] – qmr (Quasi-Minimal Residual)
Error: [0.+0.j 0.+0.j 0.+0.j 0.+0.j 0.+0.j] Common Parameters
• mandatory:
– solve as double precision complex:
A [{sparse matrix, dense matrix, LinearOperator}] The N-by-N matrix of the linear system.
>>> mtx2 = mtx.astype(np.complex128)
>>> x = dsolve.spsolve(mtx2, rhs, use_umfpack=True) b [{array, matrix}] Right hand side of the linear system. Has shape (N,) or (N,1).
>>> print(x)
• optional:
[106. +0.j -21. +0.j 5.5+0.j -1.5+0.j 1. +0.j]
>>> print("Error: %s " % (mtx2 * x - rhs)) x0 [{array, matrix}] Starting guess for the solution.
Error: [0.+0.j 0.+0.j 0.+0.j 0.+0.j 0.+0.j]
tol [float] Relative tolerance to achieve before terminating.
""" maxiter [integer] Maximum number of iterations. Iteration will stop after maxiter steps even if
Solve a linear system the specified tolerance has not been achieved.
=======================
M [{sparse matrix, dense matrix, LinearOperator}] Preconditioner for A. The preconditioner
Construct a 1000x1000 lil_matrix and add some values to it, convert it should approximate the inverse of A. Effective preconditioning dramatically improves the
to CSR format and solve A x = b for x:and solve a linear system with a rate of convergence, which implies that fewer iterations are needed to reach a given error
direct solver. tolerance.
"""
import numpy as np
callback [function] User-supplied function to call after each iteration. It is called as callback(xk),
import scipy.sparse as sps where xk is the current solution vector.
from scipy.sparse.linalg.dsolve import linsolve LinearOperator Class
rand = np.random.rand from scipy.sparse.linalg.interface import LinearOperator
mtx = sps.lil_matrix((1000, 1000), dtype=np.float64) • common interface for performing matrix vector products
mtx[0, :100] = rand(100)
mtx[1, 100:200] = mtx[0, :100] • useful abstraction that enables using dense and sparse matrices within the solvers, as well as
mtx.setdiag(rand(1000)) matrix-free solutions
• has shape and matvec() (+ some optional parameters)
plt.clf()
plt.spy(mtx, marker='.', markersize=2) • example:
plt.show()
mtx = mtx.tocsr() >>> from scipy.sparse.linalg import LinearOperator
rhs = rand(1000) >>> def mv(v):
... return np.array([2*v[0], 3*v[1]])
12.3. Linear System Solvers 346 12.3. Linear System Solvers 347

... plt.figure(figsize=(9,9))
>>> A = LinearOperator((2, 2), matvec=mv)
>>> A for i in range(K):
<2x2 _CustomLinearOperator with dtype=float64> plt.subplot(3, 3, i+1)
>>> A.matvec(np.ones(2)) plt.title('Eigenvector %d ' % i)
array([2., 3.]) plt.pcolor(V[:,i].reshape(N,N))
>>> A * np.ones(2) plt.axis('equal')
array([2., 3.]) plt.axis('off')
plt.show()
A Few Notes on Preconditioning – examples/pyamg_with_lobpcg.py

• problem specific • example by Nils Wagner:
• often hard to develop – examples/lobpcg_sakurai.py
• if not sure, try ILU • output:
– available in dsolve as spilu() $ python examples/lobpcg_sakurai.py
Results by LOBPCG for n=2500
12.3.3 Eigenvalue Problem Solvers
[ 0.06250083 0.06250028 0.06250007]
The eigen module
Exact eigenvalues
• 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 [ 0.06250005 0.0625002 0.06250044]
combination with PyAMG * example by Nathan Bell:
Elapsed time 7.01
"""
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

12.3. Linear System Solvers 348 12.3. Linear System Solvers 349
12.4 Other Interesting Packages

• PyAMG
– algebraic multigrid solvers
13
– https://github.com/pyamg/pyamg
• Pysparse
– own sparse matrix classes
– matrix and eigenvalue problem solvers
– http://pysparse.sourceforge.net/ CHAPTER
Image manipulation and processing

using Numpy and Scipy
Authors: Emmanuelle Gouillart, Gaël Varoquaux

This section addresses basic image manipulation and processing using the core scientific modules NumPy
and SciPy. Some of the operations covered by this tutorial may be useful for other kinds of multidimen-
sional array processing than image processing. In particular, the submodule scipy.ndimage provides
functions operating on n-dimensional NumPy arrays.
See also:
For more advanced image processing and image-specific routines, see the tutorial Scikit-image: image
processing, dedicated to the skimage module.
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 the
documentation:
>>> from scipy import ndimage
Common tasks in image processing:
12.4. Other Interesting Packages 350 351

• 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
• ...
Chapters contents
• Opening and writing to image files

• Displaying images
• Basic manipulations
– Statistical information
– Geometrical transformations
• Image filtering
– Blurring/smoothing
– Sharpening
– Denoising
– Mathematical morphology
Creating a numpy array from an image file:
• Feature extraction
– Edge detection >>> import imageio
>>> face = misc.face()
– Segmentation >>> imageio.imsave('face.png', face) # First we need to create the PNG file
• Measuring objects properties: ndimage.measurements
>>> face = imageio.imread('face.png')
• Full code examples >>> type(face)
<class 'imageio.core.util.Array'>
• Examples for the image processing chapter
>>> face.shape, face.dtype
((768, 1024, 3), dtype('uint8'))
13.1 Opening and writing to image files dtype is uint8 for 8-bit images (0-255)
Opening raw files (camera, 3-D images)
Writing an array to a file:
>>> face.tofile('face.raw') # Create raw file
from scipy import misc >>> face_from_raw = np.fromfile('face.raw', dtype=np.uint8)
import imageio >>> face_from_raw.shape
f = misc.face() (2359296,)
imageio.imsave('face.png', f) # uses the Image module (PIL) >>> face_from_raw.shape = (768, 1024, 3)

Need to know the shape and dtype of the image (how to separate data bytes).
plt.imshow(f)
plt.show() For large data, use np.memmap for memory mapping:
>>> face_memmap = np.memmap('face.raw', dtype=np.uint8, shape=(768, 1024, 3))
(data are read from the file, and not loaded into memory)
Working on a list of image files
13.1. Opening and writing to image files 352 13.1. Opening and writing to image files 353
>>> for i in range(10):

... im = np.random.randint(0, 256, 10000).reshape((100, 100))
... imageio.imsave('random_%02d .png' % i, im)
>>> from glob import glob
>>> filelist = glob('random*.png')
>>> filelist.sort()
13.2 Displaying images

Use matplotlib and imshow to display an image inside a matplotlib figure:
>>> f = misc.face(gray=True) # retrieve a grayscale image

>>> plt.imshow(f, cmap=plt.cm.gray)
Increase contrast by setting min and max values:
>>> plt.imshow(f, cmap=plt.cm.gray, vmin=30, vmax=200)

>>> # Remove axes and ticks • Image plane widgets
>>> plt.axis('off')
(-0.5, 1023.5, 767.5, -0.5) • Isosurfaces
• ...
Draw contour lines:
>>> plt.contour(f, [50, 200])

<matplotlib.contour.QuadContourSet ...>
For smooth intensity variations, use interpolation='bilinear'. For fine inspection of intensity varia-
tions, use interpolation='nearest':
>>> plt.imshow(f[320:340, 510:530], cmap=plt.cm.gray, interpolation='bilinear')

>>> plt.imshow(f[320:340, 510:530], cmap=plt.cm.gray, interpolation='nearest')
13.3 Basic manipulations
See also:
Images are arrays: use the whole numpy machinery.
More interpolation methods are in Matplotlib’s examples.
See also:
3-D visualization: Mayavi
See 3D plotting with Mayavi.
13.2. Displaying images 354 13.3. Basic manipulations 355

13.3.1 Statistical information

>>> face.mean()
113.48026784261067
>>> face.max(), face.min()
(250, 0)
np.histogram
Exercise
• Open as an array the scikit-image logo (http://scikit-image.org/_static/img/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 matplotlib. Change the interpolation method and zoom to see
the difference.
• Transform your image to greyscale
>>> face = misc.face(gray=True) • Increase the contrast of the image by changing its minimum and maximum values. Optional:
>>> face[0, 40] use scipy.stats.scoreatpercentile (read the docstring!) to saturate 5% of the darkest pixels
127 and 5% of the lightest pixels.
>>> # Slicing
>>> face[10:13, 20:23]
• Save the array to two different file formats (png, jpg, tiff)
array([[141, 153, 145],
[133, 134, 125],
[ 96, 92, 94]], dtype=uint8)
>>> face[100:120] = 255
>>>
>>> lx, ly = face.shape
>>> X, Y = np.ogrid[0:lx, 0:ly]
>>> mask = (X - lx / 2) ** 2 + (Y - ly / 2) ** 2 > lx * ly / 4
>>> # Masks
>>> face[mask] = 0
>>> # Fancy indexing
>>> face[range(400), range(400)] = 255
13.3.2 Geometrical transformations
>>> lx, ly = face.shape
>>> # Cropping
>>> crop_face = face[lx // 4: - lx // 4, ly // 4: - ly // 4]
>>> # up <-> down flip
>>> flip_ud_face = np.flipud(face)
>>> # rotation
>>> rotate_face = ndimage.rotate(face, 45)
>>> rotate_face_noreshape = ndimage.rotate(face, 45, reshape=False)
13.3. Basic manipulations 356 13.3. Basic manipulations 357

13.4 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.
13.4.1 Blurring/smoothing
Gaussian filter from scipy.ndimage:
>>> from scipy import misc >>> f = misc.face(gray=True)
>>> face = misc.face(gray=True) >>> f = f[230:290, 220:320]
>>> blurred_face = ndimage.gaussian_filter(face, sigma=3) >>> noisy = f + 0.4 * f.std() * np.random.random(f.shape)
>>> very_blurred = ndimage.gaussian_filter(face, sigma=5)
A Gaussian filter smoothes the noise out. . . and the edges as well:
Uniform filter
>>> gauss_denoised = ndimage.gaussian_filter(noisy, 2)
>>> local_mean = ndimage.uniform_filter(face, size=11)
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)
13.4.2 Sharpening Median filter: better result for straight boundaries (low curvature):
Sharpen a blurred image: >>> im = np.zeros((20, 20))

>>> im[5:-5, 5:-5] = 1
>>> from scipy import misc >>> im = ndimage.distance_transform_bf(im)
>>> face = misc.face(gray=True).astype(float) >>> im_noise = im + 0.2 * np.random.randn(*im.shape)
>>> blurred_f = ndimage.gaussian_filter(face, 3) >>> im_med = ndimage.median_filter(im_noise, 3)
increase the weight of edges by adding an approximation of the Laplacian:
>>> filter_blurred_f = ndimage.gaussian_filter(blurred_f, 1)

>>> alpha = 30
>>> sharpened = blurred_f + alpha * (blurred_f - filter_blurred_f)
13.4.3 Denoising
Noisy face:
13.4. Image filtering 358 13.4. Image filtering 359

Other rank filter: ndimage.maximum_filter, ndimage.percentile_filter (continued from previous page)
Other local non-linear filters: Wiener (scipy.signal.wiener), etc. >>> a

array([[0, 0, 0, 0, 0, 0, 0],
Non-local filters [0, 0, 1, 1, 1, 0, 0],
[0, 0, 1, 1, 1, 0, 0],
[0, 0, 1, 1, 1, 0, 0],
Exercise: denoising [0, 0, 1, 1, 1, 0, 0],
[0, 0, 1, 1, 1, 0, 0],
• Create a binary image (of 0s and 1s) with several objects (circles, ellipses, squares, or random [0, 0, 0, 0, 0, 0, 0]])
shapes). >>> ndimage.binary_erosion(a).astype(a.dtype)
array([[0, 0, 0, 0, 0, 0, 0],
• Add some noise (e.g., 20% of noise) [0, 0, 0, 0, 0, 0, 0],
• Try two different denoising methods for denoising the image: gaussian filtering and median [0, 0, 0, 1, 0, 0, 0],
filtering. [0, 0, 0, 1, 0, 0, 0],
[0, 0, 0, 1, 0, 0, 0],
• Compare the histograms of the two different denoised images. Which one is the closest to the [0, 0, 0, 0, 0, 0, 0],
histogram of the original (noise-free) image? [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)
See also: array([[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0],
More denoising filters are available in skimage.denoising, see the Scikit-image: image processing tuto-
[0, 0, 0, 0, 0, 0, 0],
rial. [0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0],
13.4.4 Mathematical morphology [0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0]])
See wikipedia for a definition of 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]])
>>> el.astype(np.int)
array([[0, 1, 0],
[1, 1, 1],
[0, 1, 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.]])
Erosion = minimum filter. Replace the value of a pixel by the minimal value covered by the structuring
element.: Also works for grey-valued images:
>>> a = np.zeros((7,7), dtype=np.int) >>> np.random.seed(2)

>>> a[1:6, 2:5] = 1 >>> im = np.zeros((64, 64))
13.4. Image filtering 360 13.4. Image filtering 361


>>> 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))) 13.5 Feature extraction
13.5.1 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)
Opening: erosion + dilation: Use a gradient operator (Sobel) to find high intensity variations:
>>> a = np.zeros((5,5), dtype=np.int) >>> sx = ndimage.sobel(im, axis=0, mode='constant')
>>> a[1:4, 1:4] = 1; a[4, 4] = 1 >>> sy = ndimage.sobel(im, axis=1, mode='constant')
>>> a >>> sob = np.hypot(sx, sy)
array([[0, 0, 0, 0, 0],
[0, 1, 1, 1, 0],
[0, 1, 1, 1, 0],
[0, 1, 1, 1, 0],
[0, 0, 0, 0, 1]])
>>> # Opening removes small objects
>>> ndimage.binary_opening(a, structure=np.ones((3,3))).astype(np.int)
array([[0, 0, 0, 0, 0],
[0, 1, 1, 1, 0],
[0, 1, 1, 1, 0],
[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], 13.5.2 Segmentation
[0, 0, 0, 0, 0]]) • Histogram-based segmentation (no spatial information)
Application: remove noise: >>> n = 10

>>> l = 256
>>> square = np.zeros((32, 32)) >>> im = np.zeros((l, l))
>>> square[10:-10, 10:-10] = 1 >>> np.random.seed(1)
>>> np.random.seed(2) >>> points = l*np.random.random((2, n**2))
>>> x, y = (32*np.random.random((2, 20))).astype(np.int) >>> im[(points[0]).astype(np.int), (points[1]).astype(np.int)] = 1
>>> square[x, y] = 1 >>> im = ndimage.gaussian_filter(im, sigma=l/(4.*n))
>>> open_square = ndimage.binary_opening(square) >>> mask = (im > im.mean()).astype(np.float)

>>> mask += 0.1 * im
>>> eroded_square = ndimage.binary_erosion(square) >>> img = mask + 0.2*np.random.randn(*mask.shape)
>>> reconstruction = ndimage.binary_propagation(eroded_square, mask=square)
>>> hist, bin_edges = np.histogram(img, bins=60)
Closing: dilation + erosion >>> bin_centers = 0.5*(bin_edges[:-1] + bin_edges[1:])
Many other mathematical morphology operations: hit and miss transform, tophat, etc. >>> binary_img = img > 0.5
13.4. Image filtering 362 13.5. Feature extraction 363

See also:
Other Scientific Packages provide algorithms that can be useful for image processing. In this example,
we use the spectral clustering function of the scikit-learn in order to segment glued objects.
>>> 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
Use mathematical morphology to clean up the result: >>> circle1 = (x - center1[0])**2 + (y - center1[1])**2 < radius1**2
>>> circle2 = (x - center2[0])**2 + (y - center2[1])**2 < radius2**2
>>> # Remove small white regions >>> circle3 = (x - center3[0])**2 + (y - center3[1])**2 < radius3**2
>>> open_img = ndimage.binary_opening(binary_img) >>> circle4 = (x - center4[0])**2 + (y - center4[1])**2 < radius4**2
>>> # Remove small black hole
>>> close_img = ndimage.binary_closing(open_img) >>> # 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, n_clusters=4, eigen_solver='arpack')

Exercise >>> label_im = -np.ones(mask.shape)
>>> label_im[mask] = labels
Check that reconstruction operations (erosion + propagation) produce a better result than open-
ing/closing:
>>> eroded_img = ndimage.binary_erosion(binary_img)
>>> reconstruct_img = ndimage.binary_propagation(eroded_img, mask=binary_img)
>>> tmp = np.logical_not(reconstruct_img)
>>> eroded_tmp = ndimage.binary_erosion(tmp)
>>> reconstruct_final = np.logical_not(ndimage.binary_propagation(eroded_tmp, mask=tmp))
>>> np.abs(mask - close_img).mean()
0.00727836...
>>> np.abs(mask - reconstruct_final).mean()
0.00059502...
Exercise
Check how a first denoising step (e.g. with a median filter) modifies the histogram, and check that
the resulting histogram-based segmentation is more accurate.
13.6 Measuring objects properties: ndimage.measurements
See also: Synthetic data:
More advanced segmentation algorithms are found in the scikit-image: see Scikit-image: image pro-
cessing.
13.5. Feature extraction 364 13.6. Measuring objects properties: ndimage.measurements 365
>>> 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()
• Analysis of connected components

Label connected components: ndimage.label:
>>> label_im, nb_labels = ndimage.label(mask)

>>> nb_labels # how many regions?
29
>>> 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.
13.6. Measuring objects properties: ndimage.measurements 366 13.6. Measuring objects properties: ndimage.measurements 367
Example: block mean:

>>> f = misc.face(gray=True)
>>> sx, sy = f.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(f, labels=regions, index=np.arange(1,
... regions.max() +1))
>>> block_mean.shape = (sx // 4, sy // 6)

... sizes = range(1, s/2, 2)
... granulo = [ndimage.binary_opening(data, \
... structure=disk_structure(n)).sum() for n in sizes]
... return granulo
...
>>>
>>> np.random.seed(1)
>>> n = 10
>>> l = 256
>>> im = np.zeros((l, l))
When regions are regular blocks, it is more efficient to use stride tricks (Example: fake dimensions with
>>> points = l*np.random.random((2, n**2))
strides). >>> im[(points[0]).astype(np.int), (points[1]).astype(np.int)] = 1
Non-regularly-spaced blocks: radial mean: >>> im = ndimage.gaussian_filter(im, sigma=l/(4.*n))
>>>
>>> sx, sy = f.shape >>> mask = im > im.mean()
>>> X, Y = np.ogrid[0:sx, 0:sy] >>>
>>> r = np.hypot(X - sx/2, Y - sy/2) >>> granulo = granulometry(mask, sizes=np.arange(2, 19, 4))
>>> rbin = (20* r/r.max()).astype(np.int)
>>> radial_mean = ndimage.mean(f, labels=rbin, index=np.arange(1, rbin.max() +1))
• Other measures
Correlation function, Fourier/wavelet spectrum, etc.
One example with mathematical morphology: granulometry
>>> 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 is None:
13.6. Measuring objects properties: ndimage.measurements 368 13.6. Measuring objects properties: ndimage.measurements 369

13.8 Examples for the image processing chapter
13.8.1 Displaying a Racoon Face

Small example to plot a racoon face.
import scipy.misc
f = scipy.misc.face(gray=True)
plt.imshow(f[320:340, 510:530], cmap=plt.cm.gray)
plt.axis('off')
plt.imshow(f[320:340, 510:530], cmap=plt.cm.gray, interpolation='nearest')
plt.axis('off')
plt.subplots_adjust(wspace=0.02, hspace=0.02, top=1, bottom=0, left=0, right=1)

plt.show()
from scipy import misc

import imageio
13.8.3 Image manipulation and numpy arrays
f = misc.face()
imageio.imsave('face.png', f) # uses the Image module (PIL) This example shows how to do image manipulation using common numpy arrays tricks.

plt.imshow(f)
plt.show()
13.8.2 Image interpolation

The example demonstrates image interpolation on a Racoon face.
13.7. Full code examples 370 13.8. Examples for the image processing chapter 371
import numpy as np
import scipy
import scipy.misc
face = scipy.misc.face(gray=True)
face[10:13, 20:23]
face[100:120] = 255
lx, ly = face.shape
X, Y = np.ogrid[0:lx, 0:ly]
mask = (X - lx/2)**2 + (Y - ly/2)**2 > lx*ly/4
face[mask] = 0
face[range(400), range(400)] = 255 import numpy as np
import scipy
plt.figure(figsize=(3, 3)) from scipy import ndimage
plt.axes([0, 0, 1, 1]) import matplotlib.pyplot as plt
plt.imshow(face, cmap=plt.cm.gray)
plt.axis('off') f = scipy.misc.face(gray=True)
sx, sy = f.shape
plt.show() X, Y = np.ogrid[0:sx, 0:sy]

r = np.hypot(X - sx/2, Y - sy/2)
Note: Click here to download the full example code rbin = (20* r/r.max()).astype(np.int)
radial_mean = ndimage.mean(f, labels=rbin, index=np.arange(1, rbin.max() +1))
13.8.4 Radial mean plt.axes([0, 0, 1, 1])
This example shows how to do a radial mean with scikit-image. plt.imshow(rbin, cmap=plt.cm.nipy_spectral)
plt.axis('off')
plt.show()
13.8. Examples for the image processing chapter 372 13.8. Examples for the image processing chapter 373
13.8.5 Plot the block mean of an image 13.8.6 Display a Racoon Face
An example showing how to use broad-casting to plot the mean of blocks of an image. An example that displays a racoon face with matplotlib.
import scipy.misc
plt.subplot(131)
plt.imshow(f, cmap=plt.cm.gray)
plt.subplot(132)
plt.imshow(f, cmap=plt.cm.gray, vmin=30, vmax=200)
plt.axis('off')
plt.subplot(133)
plt.imshow(f, cmap=plt.cm.gray)
plt.contour(f, [50, 200])
plt.axis('off')
import numpy as np
import scipy.misc plt.subplots_adjust(wspace=0, hspace=0., top=0.99, bottom=0.01, left=0.05,
from scipy import ndimage right=0.99)
import matplotlib.pyplot as plt plt.show()
f = scipy.misc.face(gray=True) Total running time of the script: ( 0 minutes 0.383 seconds)

sx, sy = f.shape
X, Y = np.ogrid[0:sx, 0:sy]
regions = sy//6 * (X//4) + Y//6
block_mean = ndimage.mean(f, labels=regions,
index=np.arange(1, regions.max() +1))
block_mean.shape = (sx//4, sy//6)
13.8.7 Image sharpening
This example shows how to sharpen an image in noiseless situation by applying the filter inverse to the
plt.figure(figsize=(5, 5)) blur.
plt.imshow(block_mean, cmap=plt.cm.gray)
plt.axis('off')
plt.show()
import scipy import scipy.misc

from scipy import ndimage from scipy import ndimage
f = scipy.misc.face(gray=True).astype(float) face = scipy.misc.face(gray=True)

blurred_f = ndimage.gaussian_filter(f, 3) blurred_face = ndimage.gaussian_filter(face, sigma=3)
very_blurred = ndimage.gaussian_filter(face, sigma=5)
filter_blurred_f = ndimage.gaussian_filter(blurred_f, 1) local_mean = ndimage.uniform_filter(face, size=11)
alpha = 30 plt.figure(figsize=(9, 3))

sharpened = blurred_f + alpha * (blurred_f - filter_blurred_f) plt.subplot(131)
plt.imshow(blurred_face, cmap=plt.cm.gray)
plt.figure(figsize=(12, 4)) plt.axis('off')
plt.subplot(132)
plt.subplot(131) plt.imshow(very_blurred, cmap=plt.cm.gray)
plt.imshow(f, cmap=plt.cm.gray) plt.axis('off')
plt.subplot(132) plt.imshow(local_mean, cmap=plt.cm.gray)
plt.imshow(blurred_f, cmap=plt.cm.gray) plt.axis('off')
plt.axis('off')
plt.subplot(133) plt.subplots_adjust(wspace=0, hspace=0., top=0.99, bottom=0.01,
plt.imshow(sharpened, cmap=plt.cm.gray) left=0.01, right=0.99)
plt.axis('off')
plt.show()
plt.tight_layout()


13.8.9 Synthetic data
13.8.8 Blurring of images The example generates and displays simple synthetic data.
An example showing various processes that blur an image.

import numpy as np
from scipy import ndimage x, y = (32*np.random.random((2, 20))).astype(np.int)
import matplotlib.pyplot as plt square[x, y] = 1
np.random.seed(1) open_square = ndimage.binary_opening(square)

n = 10
l = 256 eroded_square = ndimage.binary_erosion(square)
im = np.zeros((l, l)) reconstruction = ndimage.binary_propagation(eroded_square, mask=square)
points = l*np.random.random((2, n**2))
im[(points[0]).astype(np.int), (points[1]).astype(np.int)] = 1 plt.figure(figsize=(9.5, 3))
im = ndimage.gaussian_filter(im, sigma=l/(4.*n)) plt.subplot(131)
plt.imshow(square, cmap=plt.cm.gray, interpolation='nearest')
mask = im > im.mean() plt.axis('off')
plt.subplot(132)
label_im, nb_labels = ndimage.label(mask) plt.imshow(open_square, cmap=plt.cm.gray, interpolation='nearest')
plt.axis('off')
plt.figure(figsize=(9,3)) plt.subplot(133)
plt.imshow(reconstruction, cmap=plt.cm.gray, interpolation='nearest')
plt.imshow(im)
plt.axis('off') plt.subplots_adjust(wspace=0, hspace=0.02, top=0.99, bottom=0.01, left=0.01, right=0.99)
plt.subplot(132) plt.show()
plt.imshow(mask, cmap=plt.cm.gray)
plt.axis('off') Total running time of the script: ( 0 minutes 0.030 seconds)
plt.subplot(133)
plt.imshow(label_im, cmap=plt.cm.nipy_spectral)
plt.axis('off') Note: Click here to download the full example code
plt.subplots_adjust(wspace=0.02, hspace=0.02, top=1, bottom=0, left=0, right=1)

plt.show() 13.8.11 Image denoising
Total running time of the script: ( 0 minutes 0.042 seconds) This example demoes image denoising on a Racoon face.
13.8.10 Opening, erosion, and propagation

This example shows simple operations of mathematical morphology.
import numpy as np
import scipy
import scipy.misc
f = f[230:290, 220:320]
noisy = f + 0.4*f.std()*np.random.random(f.shape)
gauss_denoised = ndimage.gaussian_filter(noisy, 2)
import numpy as np med_denoised = ndimage.median_filter(noisy, 3)
plt.figure(figsize=(12,2.8))
square = np.zeros((32, 32))
square[10:-10, 10:-10] = 1 plt.subplot(131)
np.random.seed(2) plt.imshow(noisy, cmap=plt.cm.gray, vmin=40, vmax=220)

plt.axis('off') label_im = np.searchsorted(labels, label_im)
plt.title('noisy', fontsize=20)
plt.subplot(132) # Now that we have only one connected component, extract it's bounding box
plt.imshow(gauss_denoised, cmap=plt.cm.gray, vmin=40, vmax=220) slice_x, slice_y = ndimage.find_objects(label_im==4)[0]
plt.axis('off') roi = im[slice_x, slice_y]
plt.title('Gaussian filter', fontsize=20)
plt.subplot(133) plt.figure(figsize=(4, 2))
plt.imshow(med_denoised, cmap=plt.cm.gray, vmin=40, vmax=220) plt.axes([0, 0, 1, 1])
plt.axis('off') plt.imshow(roi)
plt.title('Median filter', fontsize=20) plt.axis('off')
plt.subplots_adjust(wspace=0.02, hspace=0.02, top=0.9, bottom=0, left=0, plt.show()

right=1)


13.8.13 Measurements from images
13.8.12 Find the bounding box of an object This examples shows how to measure quantities from various images.
This example shows how to extract the bounding box of the largest object
import numpy as np
np.random.seed(1)
import numpy as np
n = 10
l = 256
im = np.zeros((l, l))
np.random.seed(1)
im[(points[0]).astype(np.int), (points[1]).astype(np.int)] = 1
n = 10
im = ndimage.gaussian_filter(im, sigma=l/(4.*n))
l = 256
mask = im > im.mean()
label_im, nb_labels = ndimage.label(mask)
# Find the largest connected component
mask = im > im.mean()
sizes = ndimage.sum(mask, label_im, range(nb_labels + 1))
mask_size = sizes < 1000
label_im, nb_labels = ndimage.label(mask)
remove_pixel = mask_size[label_im]
label_im[remove_pixel] = 0
sizes = ndimage.sum(mask, label_im, range(nb_labels + 1))
labels = np.unique(label_im)

mask_size = sizes < 1000 plt.axis('off')
remove_pixel = mask_size[label_im] plt.subplot(153)
label_im[remove_pixel] = 0 plt.imshow(flip_ud_face, cmap=plt.cm.gray)
labels = np.unique(label_im) plt.axis('off')
label_clean = np.searchsorted(labels, label_im) plt.subplot(154)
plt.imshow(rotate_face, cmap=plt.cm.gray)
plt.axis('off')
plt.figure(figsize=(6 ,3)) plt.subplot(155)
plt.imshow(rotate_face_noreshape, cmap=plt.cm.gray)
plt.imshow(label_im, cmap=plt.cm.nipy_spectral)
plt.axis('off') plt.subplots_adjust(wspace=0.02, hspace=0.3, top=1, bottom=0.1, left=0,
plt.subplot(122) right=1)
plt.imshow(label_clean, vmax=nb_labels, cmap=plt.cm.nipy_spectral)
plt.axis('off') plt.show()
plt.subplots_adjust(wspace=0.01, hspace=0.01, top=1, bottom=0, left=0, right=1) Total running time of the script: ( 0 minutes 0.460 seconds)
plt.show()

13.8.15 Denoising an image with the median filter
This example shows the original image, the noisy image, the denoised one (with the median filter) and
13.8.14 Geometrical transformations the difference between the two.
This examples demos some simple geometrical transformations on a Racoon face.
import numpy as np
import scipy.misc
import numpy as np
face = scipy.misc.face(gray=True)
lx, ly = face.shape
im = np.zeros((20, 20))
# Cropping
im[5:-5, 5:-5] = 1
crop_face = face[lx//4:-lx//4, ly//4:-ly//4]
im = ndimage.distance_transform_bf(im)
# up <-> down flip
im_noise = im + 0.2*np.random.randn(*im.shape)
flip_ud_face = np.flipud(face)
# rotation
im_med = ndimage.median_filter(im_noise, 3)
rotate_face = ndimage.rotate(face, 45)
rotate_face_noreshape = ndimage.rotate(face, 45, reshape=False)
plt.subplot(141)
plt.imshow(im, interpolation='nearest')
plt.axis('off')
plt.subplot(151)
plt.title('Original image', fontsize=20)
plt.imshow(face, cmap=plt.cm.gray)
plt.subplot(142)
plt.axis('off')
plt.imshow(im_noise, interpolation='nearest', vmin=0, vmax=5)
plt.subplot(152)
plt.axis('off')
plt.imshow(crop_face, cmap=plt.cm.gray)

plt.title('Noisy image', fontsize=20) bin_centers = 0.5*(bin_edges[:-1] + bin_edges[1:])
plt.subplot(143)
plt.imshow(im_med, interpolation='nearest', vmin=0, vmax=5) binary_img = img > 0.5
plt.axis('off')
plt.title('Median filter', fontsize=20) plt.figure(figsize=(11,4))
plt.subplot(144)
plt.imshow(np.abs(im - im_med), cmap=plt.cm.hot, interpolation='nearest') plt.subplot(131)
plt.axis('off') plt.imshow(img)
plt.title('Error', fontsize=20) plt.axis('off')
plt.subplot(132)
plt.plot(bin_centers, hist, lw=2)
plt.subplots_adjust(wspace=0.02, hspace=0.02, top=0.9, bottom=0, left=0, plt.axvline(0.5, color='r', ls='--', lw=2)
right=1) plt.text(0.57, 0.8, 'histogram', fontsize=20, transform = plt.gca().transAxes)
plt.yticks([])
plt.show() plt.subplot(133)
plt.imshow(binary_img, cmap=plt.cm.gray, interpolation='nearest')
Total running time of the script: ( 0 minutes 0.046 seconds) plt.axis('off')
plt.subplots_adjust(wspace=0.02, hspace=0.3, top=1, bottom=0.1, left=0, right=1)

Note: Click here to download the full example code plt.show()

13.8.16 Histogram segmentation
This example does simple histogram analysis to perform segmentation. Note: Click here to download the full example code
13.8.17 Finding edges with Sobel filters

The Sobel filter is one of the simplest way of finding edges.
import numpy as np
import numpy as np
np.random.seed(1) from scipy import ndimage
n = 10 import matplotlib.pyplot as plt
l = 256
im = np.zeros((l, l)) im = np.zeros((256, 256))
points = l*np.random.random((2, n**2)) im[64:-64, 64:-64] = 1
im = ndimage.gaussian_filter(im, sigma=l/(4.*n)) im = ndimage.rotate(im, 15, mode='constant')
im = ndimage.gaussian_filter(im, 8)
mask = (im > im.mean()).astype(np.float)
sx = ndimage.sobel(im, axis=0, mode='constant')
mask += 0.1 * im sy = ndimage.sobel(im, axis=1, mode='constant')
sob = np.hypot(sx, sy)
img = mask + 0.2*np.random.randn(*mask.shape)
hist, bin_edges = np.histogram(img, bins=60) plt.subplot(141)

plt.imshow(im, cmap=plt.cm.gray) f = f[230:290, 220:320]
plt.axis('off')
plt.title('square', fontsize=20) noisy = f + 0.4*f.std()*np.random.random(f.shape)
plt.subplot(142)
plt.imshow(sx) tv_denoised = denoise_tv_chambolle(noisy, weight=10)
plt.axis('off')
plt.title('Sobel (x direction)', fontsize=20)
plt.subplot(143) plt.figure(figsize=(12, 2.8))
plt.imshow(sob)
plt.title('Sobel filter', fontsize=20) plt.imshow(noisy, cmap=plt.cm.gray, vmin=40, vmax=220)
plt.axis('off')
im += 0.07*np.random.random(im.shape) plt.title('noisy', fontsize=20)
plt.subplot(132)
sx = ndimage.sobel(im, axis=0, mode='constant') plt.imshow(tv_denoised, cmap=plt.cm.gray, vmin=40, vmax=220)
sy = ndimage.sobel(im, axis=1, mode='constant') plt.axis('off')
sob = np.hypot(sx, sy) plt.title('TV denoising', fontsize=20)
plt.subplot(144) tv_denoised = denoise_tv_chambolle(noisy, weight=50)

plt.imshow(sob) plt.subplot(133)
plt.axis('off') plt.imshow(tv_denoised, cmap=plt.cm.gray, vmin=40, vmax=220)
plt.title('Sobel for noisy image', fontsize=20) plt.axis('off')
plt.title('(more) TV denoising', fontsize=20)
plt.subplots_adjust(wspace=0.02, hspace=0.02, top=0.9, bottom=0, left=0,

plt.subplots_adjust(wspace=0.02, hspace=0.02, top=1, bottom=0, left=0, right=0.9) right=1)
plt.show()
plt.show()
13.8.19 Greyscale dilation

13.8.18 Total Variation denoising
This example illustrates greyscale mathematical morphology.
This example demoes Total-Variation (TV) denoising on a Racoon face.
import numpy as np
import numpy as np from scipy import ndimage
import scipy import matplotlib.pyplot as plt
import scipy.misc
import matplotlib.pyplot as plt im = np.zeros((64, 64))
try: np.random.seed(2)
from skimage.restoration import denoise_tv_chambolle x, y = (63*np.random.random((2, 8))).astype(np.int)
except ImportError: im[x, y] = np.arange(8)
# skimage < 0.12
from skimage.filters import denoise_tv_chambolle bigger_points = ndimage.grey_dilation(im, size=(5, 5), structure=np.ones((5, 5)))
f = scipy.misc.face(gray=True) square = np.zeros((16, 16))


square[4:-4, 4:-4] = 1 binary_img = img > 0.5
dist = ndimage.distance_transform_bf(square)
dilate_dist = ndimage.grey_dilation(dist, size=(3, 3), \ # Remove small white regions
structure=np.ones((3, 3))) open_img = ndimage.binary_opening(binary_img)
# Remove small black hole
plt.figure(figsize=(12.5, 3)) close_img = ndimage.binary_closing(open_img)
plt.subplot(141)
plt.imshow(im, interpolation='nearest', cmap=plt.cm.nipy_spectral) plt.figure(figsize=(12, 3))
plt.axis('off')
plt.subplot(142) l = 128
plt.imshow(bigger_points, interpolation='nearest', cmap=plt.cm.nipy_spectral)
plt.subplot(143) plt.imshow(binary_img[:l, :l], cmap=plt.cm.gray)
plt.imshow(dist, interpolation='nearest', cmap=plt.cm.nipy_spectral) plt.axis('off')
plt.subplot(144) plt.imshow(open_img[:l, :l], cmap=plt.cm.gray)
plt.imshow(dilate_dist, interpolation='nearest', cmap=plt.cm.nipy_spectral) plt.axis('off')
plt.imshow(close_img[:l, :l], cmap=plt.cm.gray)
plt.subplots_adjust(wspace=0, hspace=0.02, top=0.99, bottom=0.01, left=0.01, right=0.99) plt.axis('off')
plt.show() plt.subplot(144)
plt.imshow(mask[:l, :l], cmap=plt.cm.gray)
Total running time of the script: ( 0 minutes 0.045 seconds) plt.contour(close_img[:l, :l], [0.5], linewidths=2, colors='r')
plt.axis('off')
Note: Click here to download the full example code plt.subplots_adjust(wspace=0.02, hspace=0.3, top=1, bottom=0.1, left=0, right=1)
plt.show()
13.8.20 Cleaning segmentation with mathematical morphology
An example showing how to clean segmentation with mathematical morphology: removing small regions
and holes.
13.8.21 Segmentation with Gaussian mixture models

This example performs a Gaussian mixture model analysis of the image histogram to find the right
thresholds for separating foreground from background.
import numpy as np
np.random.seed(1)
n = 10
l = 256
mask = (im > im.mean()).astype(np.float) import numpy as np

img = mask + 0.3*np.random.randn(*mask.shape) from sklearn.mixture import GaussianMixture

np.random.seed(1)
n = 10
l = 256
mask = (im > im.mean()).astype(np.float)
img = mask + 0.3*np.random.randn(*mask.shape)
hist, bin_edges = np.histogram(img, bins=60)

bin_centers = 0.5*(bin_edges[:-1] + bin_edges[1:])
classif = GaussianMixture(n_components=2) import numpy as np

classif.fit(img.reshape((img.size, 1))) from skimage.morphology import watershed
from skimage.feature import peak_local_max
threshold = np.mean(classif.means_) import matplotlib.pyplot as plt
binary_img = img > threshold from scipy import ndimage
# Generate an initial image with two overlapping circles

plt.figure(figsize=(11,4)) x, y = np.indices((80, 80))
x1, y1, x2, y2 = 28, 28, 44, 52
plt.subplot(131) r1, r2 = 16, 20
plt.imshow(img) mask_circle1 = (x - x1) ** 2 + (y - y1) ** 2 < r1 ** 2
plt.axis('off') mask_circle2 = (x - x2) ** 2 + (y - y2) ** 2 < r2 ** 2
plt.subplot(132) image = np.logical_or(mask_circle1, mask_circle2)
plt.plot(bin_centers, hist, lw=2) # Now we want to separate the two objects in image
plt.axvline(0.5, color='r', ls='--', lw=2) # Generate the markers as local maxima of the distance
plt.text(0.57, 0.8, 'histogram', fontsize=20, transform = plt.gca().transAxes) # to the background
plt.yticks([]) distance = ndimage.distance_transform_edt(image)
plt.subplot(133) local_maxi = peak_local_max(
plt.imshow(binary_img, cmap=plt.cm.gray, interpolation='nearest') distance, indices=False, footprint=np.ones((3, 3)), labels=image)
plt.axis('off') markers = ndimage.label(local_maxi)[0]
labels = watershed(-distance, markers, mask=image)
plt.subplots_adjust(wspace=0.02, hspace=0.3, top=1, bottom=0.1, left=0, right=1)
plt.show() plt.figure(figsize=(9, 3.5))
plt.subplot(131)
plt.imshow(image, cmap='gray', interpolation='nearest')
Total running time of the script: ( 0 minutes 0.248 seconds) plt.axis('off')
plt.subplot(132)
Note: Click here to download the full example code plt.imshow(-distance, interpolation='nearest')
plt.axis('off')
plt.subplot(133)
plt.imshow(labels, cmap='nipy_spectral', interpolation='nearest')
13.8.22 Watershed segmentation plt.axis('off')
This example shows how to do segmentation with watershed. plt.subplots_adjust(hspace=0.01, wspace=0.01, top=1, bottom=0, left=0,
right=1)
plt.show()
13.8.23 Granulometry
This example performs a simple granulometry analysis.

plt.show()
13.8.24 Segmentation with spectral clustering

This example uses spectral clustering to do segmentation.
import numpy as np
import numpy as np
from scipy import ndimage from sklearn.feature_extraction import image
import matplotlib.pyplot as plt from sklearn.cluster import spectral_clustering
def disk_structure(n): l = 100

struct = np.zeros((2 * n + 1, 2 * n + 1)) x, y = np.indices((l, l))
x, y = np.indices((2 * n + 1, 2 * n + 1))
mask = (x - n)**2 + (y - n)**2 <= n**2 center1 = (28, 24)
struct[mask] = 1 center2 = (40, 50)
return struct.astype(np.bool) center3 = (67, 58)
center4 = (24, 70)
def granulometry(data, sizes=None): radius1, radius2, radius3, radius4 = 16, 14, 15, 14
s = max(data.shape)
if sizes is None: circle1 = (x - center1[0])**2 + (y - center1[1])**2 < radius1**2
sizes = range(1, s/2, 2) circle2 = (x - center2[0])**2 + (y - center2[1])**2 < radius2**2
granulo = [ndimage.binary_opening(data, \ circle3 = (x - center3[0])**2 + (y - center3[1])**2 < radius3**2
structure=disk_structure(n)).sum() for n in sizes] circle4 = (x - center4[0])**2 + (y - center4[1])**2 < radius4**2
return granulo
4 circles
np.random.seed(1) img = circle1 + circle2 + circle3 + circle4
n = 10 mask = img.astype(bool)
l = 256 img = img.astype(float)
points = l*np.random.random((2, n**2)) img += 1 + 0.2*np.random.randn(*img.shape)
im = ndimage.gaussian_filter(im, sigma=l/(4.*n)) # Convert the image into a graph with the value of the gradient on the
# edges.
mask = im > im.mean() graph = image.img_to_graph(img, mask=mask)
granulo = granulometry(mask, sizes=np.arange(2, 19, 4)) # Take a decreasing function of the gradient: we take it weakly
# dependant from the gradient the segmentation is close to a voronoi
plt.figure(figsize=(6, 2.2)) graph.data = np.exp(-graph.data / graph.data.std())
plt.subplot(121) # Force the solver to be arpack, since amg is numerically

plt.imshow(mask, cmap=plt.cm.gray) # unstable on this example
opened = ndimage.binary_opening(mask, structure=disk_structure(10)) labels = spectral_clustering(graph, n_clusters=4)
opened_more = ndimage.binary_opening(mask, structure=disk_structure(14)) label_im = -np.ones(mask.shape)
plt.contour(opened, [0.5], colors='b', linewidths=2) label_im[mask] = labels
plt.contour(opened_more, [0.5], colors='r', linewidths=2)
plt.axis('off') plt.figure(figsize=(6, 3))
plt.plot(np.arange(2, 19, 4), granulo, 'ok', ms=8) plt.imshow(img, cmap=plt.cm.nipy_spectral, interpolation='nearest')
plt.axis('off')
plt.subplot(122)
plt.subplots_adjust(wspace=0.02, hspace=0.15, top=0.95, bottom=0.15, left=0, right=0.95) plt.imshow(label_im, cmap=plt.cm.nipy_spectral, interpolation='nearest')

plt.axis('off')
plt.subplots_adjust(wspace=0, hspace=0., top=0.99, bottom=0.01, left=0.01, right=0.99)

plt.show()
CHAPTER 14
Mathematical optimization: finding
minima of functions
See also:
Authors: Gaël Varoquaux
More on image-processing:
Mathematical optimization deals with the problem of finding numerically minimums (or maximums or
• The chapter on Scikit-image zeros) of a function. In this context, the function is called cost function, or objective function, or energy.
• Other, more powerful and complete modules: OpenCV (Python bindings), CellProfiler, ITK with Here, we are interested in using scipy.optimize for black-box optimization: we do not rely on the
Python bindings mathematical 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
• Matplotlib
See also:
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 explanations.
13.8. Examples for the image processing chapter 394 395

Chapters contents 14.1.1 Convex versus non-convex optimization
• Knowing your problem

– Convex versus non-convex optimization
– Smooth and non-smooth problems
– Noisy versus exact cost functions
– Constraints
• A review of the different optimizers
– Getting started: 1D optimization
– Gradient based methods
– Newton and quasi-newton methods
• Full code examples
• Examples for the mathematical optimization chapter A convex function: A non-convex function
– Gradient-less methods • f is above all its tangents.
• equivalently, for two point A, B, f(C) lies
– Global optimizers below the segment [f(A), f(B])], if A < C <
• Practical guide to optimization with scipy B
– Choosing a method
– Making your optimizer faster Optimizing convex functions is easy. Optimizing non-convex functions can be very hard.
– Computing gradients
Note: It can be proven that for a convex function a local minimum is also a global minimum. Then,
– Synthetic exercices
in some sense, the minimum is unique.
• Special case: non-linear least-squares
– Minimizing the norm of a vector function
14.1.2 Smooth and non-smooth problems
– Curve fitting
• Optimization with constraints
– Box bounds
– General constraints
• Examples for the mathematical optimization chapter
14.1 Knowing your problem

Not all optimization problems are equal. Knowing your problem enables you to choose the right tool.
Dimensionality of the problem

A smooth function: A non-smooth function
The scale of an optimization problem is pretty much set by the dimensionality of the problem, i.e. the The gradient is defined everywhere, and is a continuous function
number of scalar variables on which the search is performed.
Optimizing smooth functions is easier (true in the context of black-box optimization, otherwise Lin-
ear Programming is an example of methods which deal very efficiently with piece-wise linear functions).
14.1. Knowing your problem 396 14.1. Knowing your problem 397
14.1.3 Noisy versus exact cost functions (continued from previous page)
>>> result.success # check if solver was successful
True
>>> x_min = result.x
>>> x_min
0.699999999...
>>> x_min - 0.7
-2.16...e-10
Brent’s method on a quadratic function: it con-

verges in 3 iterations, as the quadratic approximation is
Noisy (blue) and non-noisy (green) functions then exact.
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 Brent’s method on a non-convex function: note
function is not noisy, a gradient-based optimization may be a noisy optimization. that the fact that the optimizer avoided the local mini-
mum is a matter of luck.
14.1.4 Constraints
Note: You can use different solvers using the parameter method.
Note: scipy.optimize.minimize_scalar() can also be used for optimization constrained to an

interval using the parameter bounds.
14.2.2 Gradient based methods

Some intuitions about gradient descent
Here we focus on intuitions, not code. Code will follow.
Gradient descent basically consists in taking small steps in the direction of the gradient, that is the
Optimizations under constraints direction of the steepest descent.
Here:
−1 < 𝑥1 < 1 Table 1: Fixed step gradient descent
−1 < 𝑥2 < 1
14.2 A review of the different optimizers

A well-conditioned quadratic function.
14.2.1 Getting started: 1D optimization
Let’s get started by finding the minimum of the scalar function 𝑓 (𝑥) = exp[(𝑥−0.7)2 ]. scipy.optimize.
minimize_scalar() uses Brent’s method to find the minimum of a function:
>>> from scipy import optimize An ill-conditioned quadratic function.
>>> def f(x): The core problem of gradient-methods on ill-conditioned prob-
... return -np.exp(-(x - 0.7)**2) lems is that the gradient tends not to point in the direction of
>>> result = optimize.minimize_scalar(f) the minimum.
14.2. A review of the different optimizers 398 14.2. A review of the different optimizers 399
We can see that very anisotropic (ill-conditioned) functions are harder to optimize. Conjugate gradient descent
The gradient descent algorithms above are toys not to be used on real problems.
Take home message: conditioning number and preconditioning
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
If you know natural scaling for your variables, prescale them so that they behave similarly. This is
it cross the valley. The conjugate gradient solves this problem by adding a friction term: each step
related to preconditioning.
depends on the two last values of the gradient and sharp turns are reduced.
Also, it clearly can be advantageous to take bigger steps. This is done in gradient descent code using a Table 3: Conjugate gradient descent
line search.
Table 2: Adaptive step gradient descent
An ill-conditioned non-
quadratic function.
A well-conditioned quadratic
function.
An ill-conditioned very non-

quadratic function.
An ill-conditioned quadratic scipy provides scipy.optimize.minimize() to find the minimum of scalar functions of one or more
function. variables. The simple conjugate gradient method can be used by setting the parameter method to CG
>>> def f(x): # The rosenbrock function
... return .5*(1 - x[0])**2 + (x[1] - x[0]**2)**2
>>> optimize.minimize(f, [2, -1], method="CG")
fun: 1.6...e-11
jac: array([-6.15...e-06, 2.53...e-07])
message: ...'Optimization terminated successfully.'
nfev: 108
nit: 13
An ill-conditioned non- njev: 27
quadratic function. status: 0
success: True
x: array([0.99999..., 0.99998...])
Gradient methods need the Jacobian (gradient) of the function. They can compute it numerically, but
will perform better if you can pass them the gradient:
>>> def jacobian(x):
... return np.array((-2*.5*(1 - x[0]) - 4*x[0]*(x[1] - x[0]**2), 2*(x[1] - x[0]**2)))
>>> optimize.minimize(f, [2, 1], method="CG", jac=jacobian)
An ill-conditioned very non- fun: 2.957...e-14
jac: array([ 7.1825...e-07, -2.9903...e-07])
quadratic function.
message: 'Optimization terminated successfully.'
nfev: 16
The more a function looks like a quadratic function (elliptic iso-curves), the easier it is to optimize. nit: 8
njev: 16
14.2. A review of the different optimizers 400 14.2. A review of the different optimizers 401

>>> def hessian(x): # Computed with sympy
status: 0 ... return np.array(((1 - 4*x[1] + 12*x[0]**2, -4*x[0]), (-4*x[0], 2)))
success: True >>> optimize.minimize(f, [2,-1], method="Newton-CG", jac=jacobian, hess=hessian)
x: array([1.0000..., 1.0000...]) fun: 1.6277...e-15
jac: array([ 1.1104...e-07, -7.7809...e-08])
Note that the function has only been evaluated 27 times, compared to 108 without the gradient. message: ...'Optimization terminated successfully.'
nfev: 11
nhev: 10
14.2.3 Newton and quasi-newton methods nit: 10
njev: 20
Newton methods: using the Hessian (2nd differential)
status: 0
Newton methods use a local quadratic approximation to compute the jump direction. For this purpose, success: True
they rely on the 2 first derivative of the function: the gradient and the Hessian. x: array([0.99999..., 0.99999...])
Note: At very high-dimension, the inversion of the Hessian can be costly and unstable (large scale >
250).
An ill-conditioned quadratic function:

Note that, as the quadratic approximation is exact, the Newton
method is blazing fast Note: Newton optimizers should not to be confused with Newton’s root finding method, based on the
same principles, scipy.optimize.newton().
Quasi-Newton methods: approximating the Hessian on the fly

An ill-conditioned non-quadratic function:
Here we are optimizing a Gaussian, which is always below its BFGS: BFGS (Broyden-Fletcher-Goldfarb-Shanno algorithm) refines at each step an approximation of
quadratic approximation. As a result, the Newton method over- the Hessian.
shoots and leads to oscillations.
14.4 Examples for the mathematical optimization chapter
An ill-conditioned very non-quadratic function:
In scipy, you can use the Newton method by setting method to Newton-CG in scipy.optimize. Note: Click here to download the full example code
minimize(). Here, CG refers to the fact that an internal inversion of the Hessian is performed by
conjugate gradient
14.4.1 Noisy optimization problem
... return .5*(1 - x[0])**2 + (x[1] - x[0]**2)**2 Draws a figure explaining noisy vs non-noisy optimization
>>> optimize.minimize(f, [2,-1], method="Newton-CG", jac=jacobian)
fun: 1.5...e-15
jac: array([ 1.0575...e-07, -7.4832...e-08])
nfev: 11
nhev: 0
nit: 10
njev: 52
status: 0
success: True
x: array([0.99999..., 0.99999...])
Note that compared to a conjugate gradient (above), Newton’s method has required less function evalua-
tions, but more gradient evaluations, as it uses it to approximate the Hessian. Let’s compute the Hessian
and pass it to the algorithm:
14.2. A review of the different optimizers 402 14.3. Full code examples 403
import numpy as np
np.random.seed(0)
x = np.linspace(-5, 5, 101)
x_ = np.linspace(-5, 5, 31)
def f(x):
return -np.exp(-x**2)
# A smooth function
plt.figure(1, figsize=(3, 2.5))
plt.clf()
plt.plot(x_, f(x_) + .2*np.random.normal(size=31), linewidth=2) •

plt.plot(x, f(x), linewidth=2)
import numpy as np
plt.ylim(ymin=-1.3) import matplotlib.pyplot as plt
plt.axis('off')
plt.tight_layout() x = np.linspace(-1.5, 1.5, 101)
plt.show()
# A smooth function
Total running time of the script: ( 0 minutes 0.040 seconds) plt.figure(1, figsize=(3, 2.5))
plt.clf()
Note: Click here to download the full example code plt.plot(x, np.sqrt(.2 + x**2), linewidth=2)
plt.text(-1, 0, '$f$', size=20)
plt.ylim(ymin=-.2)
14.4.2 Smooth vs non-smooth plt.axis('off')
Draws a figure to explain smooth versus non smooth optimization. plt.tight_layout()
# A non-smooth function
plt.clf()
plt.plot(x, np.abs(x), linewidth=2)
plt.text(-1, 0, '$f$', size=20)
plt.ylim(ymin=-.2)
plt.axis('off')
plt.tight_layout()
plt.show()

•
14.4.3 Curve fitting

A curve fitting example
14.4. Examples for the mathematical optimization chapter 404 14.4. Examples for the mathematical optimization chapter 405
14.4.4 Convex function

A figure showing the definition of a convex function
import numpy as np
np.random.seed(0)
# Our test function

def f(t, omega, phi):
return np.cos(omega * t + phi)
•
# Our x and y data
x = np.linspace(0, 3, 50) import numpy as np
y = f(x, 1.5, 1) + .1*np.random.normal(size=50) import matplotlib.pyplot as plt
# Fit the model: the parameters omega and phi can be found in the x = np.linspace(-1, 2)
# `params` vector
params, params_cov = optimize.curve_fit(f, x, y) plt.figure(1, figsize=(3, 2.5))
plt.clf()
# plot the data and the fitted curve
t = np.linspace(0, 3, 1000) # A convex function
plt.plot(x, x**2, linewidth=2)
plt.figure(1) plt.text(-.7, -.6**2, '$f$', size=20)
plt.clf()
plt.plot(x, y, 'bx') # The tangent in one point
plt.plot(t, f(t, *params), 'r-') plt.plot(x, 2*x - 1)
plt.show() plt.plot(1, 1, 'k+')
plt.text(.3, -.75, "Tangent to $f$", size=15)
plt.text(1, 1 - .5, 'C', size=15)
# Convexity as barycenter
plt.plot([.35, 1.85], [.35**2, 1.85**2])

plt.plot([.35, 1.85], [.35**2, 1.85**2], 'k+')
plt.text(.35 - .2, .35**2 + .1, 'A', size=15) plt.figure(1)
plt.text(1.85 - .2, 1.85**2, 'B', size=15) plt.clf()
X, Y = np.mgrid[-1.5:1.5:100j, -1.1:1.1:100j]
plt.ylim(ymin=-1) plt.imshow(f([X, Y]).T, cmap=plt.cm.gray_r, extent=[-1.5, 1.5, -1.1, 1.1],
plt.axis('off') origin='lower')
plt.tight_layout() plt.contour(X, Y, f([X, Y]), cmap=plt.cm.gnuplot)
# Convexity as barycenter # Plot the gradient

plt.figure(2, figsize=(3, 2.5)) dX, dY = g_prime([.1*X[::5, ::5], Y[::5, ::5]])
plt.clf() # Adjust for our preconditioning
plt.plot(x, x**2 + np.exp(-5*(x - .5)**2), linewidth=2) dX *= .1
plt.text(-.7, -.6**2, '$f$', size=20) plt.quiver(X[::5, ::5], Y[::5, ::5], dX, dY, color='.5')
plt.ylim(ymin=-1) # Plot our solution

plt.axis('off') plt.plot(x_min[0], x_min[1], 'r+', markersize=15)
plt.tight_layout()
14.4.5 Finding a minimum in a flat neighborhood

An excercise of finding minimum. This excercise is hard because the function is very flat around the
minimum (all its derivatives are zero). Thus gradient information is unreliable.
The function admits a minimum in [0, 0]. The challenge is to get within 1e-7 of this minimum, starting
at x0 = [1, 1].
The solution that we adopt here is to give up on using gradient or information based on local differences,
and to rely on the Powell algorithm. With 162 function evaluations, we get to 1e-8 of the solution.
import numpy as np
def f(x):
return np.exp(-1/(.01*x[0]**2 + x[1]**2))
# A well-conditionned version of f:
def g(x):
return f([10*x[0], x[1]])
# The gradient of g. We won't use it here for the optimization.

def g_prime(x): •
r = np.sqrt(x[0]**2 + x[1]**2)
return 2/r**3*g(x)*x/r
result = optimize.minimize(g, [1, 1], method="Powell", tol=1e-10)

x_min = result.x
Some pretty plotting
plt.figure(0)
plt.clf()
t = np.linspace(-1.1, 1.1, 100)
plt.plot(t, f([0, t]))

x = x.T
y = y.T

plt.clf()
plt.axes([0, 0, 1, 1])
contours = plt.contour(np.sqrt((x - 3)**2 + (y - 2)**2),

extent=[-2.03, 4.2, -1.6, 3.2],
cmap=plt.cm.gnuplot)
plt.clabel(contours,
inline=1,
fmt='%1.1f ',
fontsize=14)
plt.plot([-1.5, 0, 1.5, 0, -1.5],
[ 0, 1.5, 0, -1.5, 0], 'k', linewidth=2)
plt.fill_between([ -1.5, 0, 1.5],
[ 0, -1.5, 0],
[ 0, 1.5, 0],
color='.8')
plt.axvline(0, color='k')
plt.axhline(0, color='k')
plt.text(-.9, 2.8, '$x_2$', size=20)

plt.text(3.6, -.6, '$x_1$', size=20)
plt.axis('tight')
plt.axis('off')
•
# And now plot the optimization path
Total running time of the script: ( 0 minutes 0.037 seconds) accumulator = list()
Note: Click here to download the full example code def f(x):
# Store the list of function calls
accumulator.append(x)
return np.sqrt((x[0] - 3)**2 + (x[1] - 2)**2)
14.4.6 Optimization with constraints
An example showing how to do optimization with general constraints using SLSQP and cobyla. def constraint(x):
return np.atleast_1d(1.5 - np.sum(np.abs(x)))
optimize.minimize(f, np.array([0, 0]), method="SLSQP",

constraints={"fun": constraint, "type": "ineq"})
accumulated = np.array(accumulator)
plt.plot(accumulated[:, 0], accumulated[:, 1])
plt.show()
14.4.7 Brent’s method

import numpy as np
Illustration of 1D optimization: Brent’s method
x, y = np.mgrid[-2.03:4.2:.04, -1.6:3.2:.04]
•
Out:
Converged at 6
Converged at 23
import numpy as np
• x_0 = np.exp(-1)
def f(x):
return (x - x_0)**2 + epsilon*np.exp(-5*(x - .5 - x_0)**2)
for epsilon in (0, 1):

plt.axes([0, 0, 1, 1])
# A convex function
plt.plot(x, f(x), linewidth=2)
# Apply brent method. To have access to the iteration, do this in an

# artificial way: allow the algorithm to iter only once
all_x = list()
all_y = list()
for iter in range(30):
result = optimize.minimize_scalar(f, bracket=(-5, 2.9, 4.5), method="Brent",
•
options={"maxiter": iter}, tol=np.finfo(1.).eps)
if result.success:
print('Converged at ', iter)
break
this_x = result.x
all_x.append(this_x)
all_y.append(f(this_x))
if iter < 6:

import numpy as np
plt.text(this_x - .05*np.sign(this_x) - .05, import matplotlib.pyplot as plt
f(this_x) + 1.2*(.3 - iter % 2), iter + 1, from scipy import optimize
size=12)
x, y = np.mgrid[-2.9:5.8:.05, -2.5:5:.05]
plt.plot(all_x[:10], all_y[:10], 'k+', markersize=12, markeredgewidth=2) x = x.T
y = y.T
plt.plot(all_x[-1], all_y[-1], 'rx', markersize=12)
plt.axis('off') for i in (1, 2):
plt.ylim(ymin=-1, ymax=8) # Create 2 figure: only the second one will have the optimization
# path
plt.figure(figsize=(4, 3)) plt.figure(i, figsize=(3, 2.5))
plt.semilogy(np.abs(all_y - all_y[-1]), linewidth=2) plt.clf()
plt.ylabel('Error on f(x)') plt.axes([0, 0, 1, 1])
plt.xlabel('Iteration')
plt.tight_layout() contours = plt.contour(np.sqrt((x - 3)**2 + (y - 2)**2),
extent=[-3, 6, -2.5, 5],
plt.show() cmap=plt.cm.gnuplot)
plt.clabel(contours,
Total running time of the script: ( 0 minutes 0.279 seconds) inline=1,
fmt='%1.1f ',
fontsize=14)
Note: Click here to download the full example code plt.plot([-1.5, -1.5, 1.5, 1.5, -1.5],
[-1.5, 1.5, 1.5, -1.5, -1.5], 'k', linewidth=2)
plt.fill_between([ -1.5, 1.5],
[ -1.5, -1.5],
14.4.8 Constraint optimization: visualizing the geometry
[ 1.5, 1.5],
A small figure explaining optimization with constraints color='.8')
plt.axvline(0, color='k')
plt.axhline(0, color='k')
plt.text(-.9, 4.4, '$x_2$', size=20)

plt.text(5.6, -.6, '$x_1$', size=20)
plt.axis('equal')
plt.axis('off')
# And now plot the optimization path

accumulator = list()
def f(x):
# Store the list of function calls
accumulator.append(x)
return np.sqrt((x[0] - 3)**2 + (x[1] - 2)**2)
•
# We don't use the gradient, as with the gradient, L-BFGS is too fast,
# and finds the optimum without showing us a pretty path
def f_prime(x):
r = np.sqrt((x[0] - 3)**2 + (x[0] - 2)**2)
return np.array(((x[0] - 3)/r, (x[0] - 2)/r))
optimize.minimize(f, np.array([0, 0]), method="L-BFGS-B",

bounds=((-1.5, 1.5), (-1.5, 1.5)))
accumulated = np.array(accumulator)
plt.plot(accumulated[:, 0], accumulated[:, 1])
plt.show()

plt.xticks(np.arange(n_methods), method_names, size=11)
plt.xlim(-.2, n_methods - .5)
14.4.9 Plotting the comparison of optimizers plt.legend(loc='best', numpoints=1, handletextpad=0, prop=dict(size=12),
frameon=False)
Plots the results from the comparison of optimizers. plt.ylabel('# function calls (a.u.)')
# Create a second legend for the problem dimensionality

plt.twinx()
for n_dim, color in zip(sorted(results.keys()), colors):

plt.plot([-10, ], [0, ], 'o', color=color,
label='# dim: %i ' % n_dim)
plt.legend(loc=(.47, .07), numpoints=1, handletextpad=0, prop=dict(size=12),
frameon=False, ncol=2)
plt.xlim(-.2, n_methods - .5)
plt.xticks(np.arange(n_methods), method_names)
plt.yticks(())
plt.tight_layout()
plt.show()

import pickle
import sys
import numpy as np
14.4.10 Alternating optimization
results = pickle.load(open(
'helper/compare_optimizers_py%s .pkl' % sys.version_info[0], The challenge here is that Hessian of the problem is a very ill-conditioned matrix. This can easily be
'rb')) seen, as the Hessian of the first term in simply 2*np.dot(K.T, K). Thus the conditioning of the problem
n_methods = len(list(results.values())[0]['Rosenbrock ']) can be judged from looking at the conditioning of K.
n_dims = len(results)
import time
symbols = 'o>*Ds'
import numpy as np
plt.figure(1, figsize=(10, 4)) from scipy import optimize
plt.clf() import matplotlib.pyplot as plt
colors = plt.cm.nipy_spectral(np.linspace(0, 1, n_dims))[:, :3] np.random.seed(0)
method_names = list(list(results.values())[0]['Rosenbrock '].keys()) K = np.random.normal(size=(100, 100))

method_names.sort(key=lambda x: x[::-1], reverse=True)
def f(x):
for n_dim_index, ((n_dim, n_dim_bench), color) in enumerate( return np.sum((np.dot(K, x - 1))**2) + np.sum(x**2)**2
zip(sorted(results.items()), colors)):
for (cost_name, cost_bench), symbol in zip(sorted(n_dim_bench.items()),
symbols): def f_prime(x):
for method_index, method_name, in enumerate(method_names): return 2*np.dot(np.dot(K.T, K), x - 1) + 4*np.sum(x**2)*x
this_bench = cost_bench[method_name]
bench = np.mean(this_bench)
plt.semilogy([method_index + .1*n_dim_index, ], [bench, ], def hessian(x):
marker=symbol, color=color) H = 2*np.dot(K.T, K) + 4*2*x*x[:, np.newaxis]
return H + 4*np.eye(H.shape[0])*np.sum(x**2)
# Create a legend for the problem type
for cost_name, symbol in zip(sorted(n_dim_bench.keys()), Some pretty plotting
symbols):
plt.semilogy([-10, ], [0, ], symbol, color='.5', plt.figure(1)
label=cost_name) plt.clf()

Z = X, Y = np.mgrid[-1.5:1.5:100j, -1.1:1.1:100j]
# Complete in the additional dimensions with zeros
Z = np.reshape(Z, (2, -1)).copy()
Z.resize((100, Z.shape[-1]))
Z = np.apply_along_axis(f, 0, Z)
Z = np.reshape(Z, X.shape)
plt.imshow(Z.T, cmap=plt.cm.gray_r, extent=[-1.5, 1.5, -1.1, 1.1],
origin='lower')
plt.contour(X, Y, Z, cmap=plt.cm.gnuplot)
# A reference but slow solution:

t0 = time.time()
x_ref = optimize.minimize(f, K[0], method="Powell").x
print(' Powell: time %.2f s' % (time.time() - t0))
f_ref = f(x_ref)
# Compare different approaches

t0 = time.time()
x_bfgs = optimize.minimize(f, K[0], method="BFGS").x
print(' BFGS: time %.2f s, x error %.2f , f error %.2f ' % (time.time() - t0,
np.sqrt(np.sum((x_bfgs - x_ref)**2)), f(x_bfgs) - f_ref))
t0 = time.time()
x_l_bfgs = optimize.minimize(f, K[0], method="L-BFGS-B").x
print(' L-BFGS: time %.2f s, x error %.2f , f error %.2f ' % (time.time() - t0,
np.sqrt(np.sum((x_l_bfgs - x_ref)**2)), f(x_l_bfgs) - f_ref))
t0 = time.time()
x_bfgs = optimize.minimize(f, K[0], jac=f_prime, method="BFGS").x
print(" BFGS w f': time %.2f s, x error %.2f , f error %.2f " % ( Out:
time.time() - t0, np.sqrt(np.sum((x_bfgs - x_ref)**2)),
f(x_bfgs) - f_ref)) Powell: time 0.27s
BFGS: time 0.75s, x error 0.02, f error -0.02
t0 = time.time() L-BFGS: time 0.06s, x error 0.02, f error -0.02
x_l_bfgs = optimize.minimize(f, K[0], jac=f_prime, method="L-BFGS-B").x BFGS w f': time 0.08s, x error 0.02, f error -0.02
print("L-BFGS w f': time %.2f s, x error %.2f , f error %.2f " % ( L-BFGS w f': time 0.00s, x error 0.02, f error -0.02
time.time() - t0, np.sqrt(np.sum((x_l_bfgs - x_ref)**2)), Newton: time 0.01s, x error 0.02, f error -0.02
f(x_l_bfgs) - f_ref))
t0 = time.time()
x_newton = optimize.minimize(f, K[0], jac=f_prime, hess=hessian, method="Newton-CG").x
print(" Newton: time %.2f s, x error %.2f , f error %.2f " % ( Note: Click here to download the full example code
time.time() - t0, np.sqrt(np.sum((x_newton - x_ref)**2)),
f(x_newton) - f_ref))
plt.show() 14.4.11 Gradient descent

An example demoing gradient descent by creating figures that trace the evolution of the optimizer.
import numpy as np
import sys, os
sys.path.append(os.path.abspath('helper'))
from cost_functions import mk_quad, mk_gauss, rosenbrock,\
rosenbrock_prime, rosenbrock_hessian, LoggingFunction,\
CountingFunction
x_min, x_max = -1, 2

y_min, y_max = 2.25/3*x_min - .2, 2.25/3*x_max - .2
A formatter to print values on contours (continued from previous page)

optimize.minimize(f, x0, jac=f_prime, method="CG", callback=store, options={"gtol": 1e-12})
def super_fmt(value):
return all_x_i, all_y_i, all_f_i
if value > 1:
if np.abs(int(value) - value) < .1:
out = '$10^{%.1i }$' % value
def newton_cg(x0, f, f_prime, hessian):
else:
all_x_i = [x0[0]]
out = '$10^{%.1f }$' % value
all_y_i = [x0[1]]
else:
all_f_i = [f(x0)]
value = np.exp(value - .01)
def store(X):
if value > .1:
x, y = X
out = '%1.1f ' % value
all_x_i.append(x)
elif value > .01:
all_y_i.append(y)
out = '%.2f ' % value
all_f_i.append(f(X))
else:
optimize.minimize(f, x0, method="Newton-CG", jac=f_prime, hess=hessian, callback=store,␣
out = '%.2e ' % value
˓→options={"xtol": 1e-12})
return out
return all_x_i, all_y_i, all_f_i
A gradient descent algorithm do not use: its a toy, use scipy’s optimize.fmin_cg
def bfgs(x0, f, f_prime, hessian=None):
def gradient_descent(x0, f, f_prime, hessian=None, adaptative=False): all_x_i = [x0[0]]
x_i, y_i = x0 all_y_i = [x0[1]]
all_x_i = list() all_f_i = [f(x0)]
all_y_i = list() def store(X):
all_f_i = list() x, y = X
all_x_i.append(x)
for i in range(1, 100): all_y_i.append(y)
all_x_i.append(x_i) all_f_i.append(f(X))
all_y_i.append(y_i) optimize.minimize(f, x0, method="BFGS", jac=f_prime, callback=store, options={"gtol": 1e-
all_f_i.append(f([x_i, y_i])) ˓→12})
dx_i, dy_i = f_prime(np.asarray([x_i, y_i])) return all_x_i, all_y_i, all_f_i
if adaptative:
# Compute a step size using a line_search to satisfy the Wolf
# conditions def powell(x0, f, f_prime, hessian=None):
step = optimize.line_search(f, f_prime, all_x_i = [x0[0]]
np.r_[x_i, y_i], -np.r_[dx_i, dy_i], all_y_i = [x0[1]]
np.r_[dx_i, dy_i], c2=.05) all_f_i = [f(x0)]
step = step[0] def store(X):
if step is None: x, y = X
step = 0 all_x_i.append(x)
else: all_y_i.append(y)
step = 1 all_f_i.append(f(X))
x_i += - step*dx_i optimize.minimize(f, x0, method="Powell", callback=store, options={"ftol": 1e-12})
y_i += - step*dy_i return all_x_i, all_y_i, all_f_i
if np.abs(all_f_i[-1]) < 1e-16:
break
return all_x_i, all_y_i, all_f_i def nelder_mead(x0, f, f_prime, hessian=None):
all_x_i = [x0[0]]
all_y_i = [x0[1]]
def gradient_descent_adaptative(x0, f, f_prime, hessian=None): all_f_i = [f(x0)]
return gradient_descent(x0, f, f_prime, adaptative=True) def store(X):
x, y = X
all_x_i.append(x)
def conjugate_gradient(x0, f, f_prime, hessian=None): all_y_i.append(y)
all_x_i = [x0[0]] all_f_i.append(f(X))
all_y_i = [x0[1]] optimize.minimize(f, x0, method="Nelder-Mead", callback=store, options={"ftol": 1e-12})
all_f_i = [f(x0)] return all_x_i, all_y_i, all_f_i
def store(X):
x, y = X
all_x_i.append(x) Run different optimizers on these problems
all_y_i.append(y)
levels = dict()
all_f_i.append(f(X))

for index, ((f, f_prime, hessian), optimizer) in enumerate(( cmap=plt.cm.gnuplot, origin='lower')
(mk_quad(.7), gradient_descent), levels[f] = contours.levels
(mk_quad(.7), gradient_descent_adaptative), plt.clabel(contours, inline=1,
(mk_quad(.02), gradient_descent), fmt=super_fmt, fontsize=14)
(mk_quad(.02), gradient_descent_adaptative),
(mk_gauss(.02), gradient_descent_adaptative), plt.plot(all_x_i, all_y_i, 'b-', linewidth=2)
((rosenbrock, rosenbrock_prime, rosenbrock_hessian), plt.plot(all_x_i, all_y_i, 'k+')
gradient_descent_adaptative),
(mk_gauss(.02), conjugate_gradient), plt.plot(logging_f.all_x_i, logging_f.all_y_i, 'k.', markersize=2)
((rosenbrock, rosenbrock_prime, rosenbrock_hessian),
conjugate_gradient), plt.plot([0], [0], 'rx', markersize=12)
(mk_quad(.02), newton_cg),
(mk_gauss(.02), newton_cg),
((rosenbrock, rosenbrock_prime, rosenbrock_hessian), plt.xticks(())
newton_cg), plt.yticks(())
(mk_quad(.02), bfgs), plt.xlim(x_min, x_max)
(mk_gauss(.02), bfgs), plt.ylim(y_min, y_max)
((rosenbrock, rosenbrock_prime, rosenbrock_hessian), plt.draw()
bfgs),
(mk_quad(.02), powell), plt.figure(index + 100, figsize=(4, 3))
(mk_gauss(.02), powell), plt.clf()
((rosenbrock, rosenbrock_prime, rosenbrock_hessian), plt.semilogy(np.maximum(np.abs(all_f_i), 1e-30), linewidth=2,
powell), label='# iterations')
(mk_gauss(.02), nelder_mead), plt.ylabel('Error on f(x)')
((rosenbrock, rosenbrock_prime, rosenbrock_hessian), plt.semilogy(logging_f.counts,
nelder_mead), np.maximum(np.abs(logging_f.all_f_i), 1e-30),
)): linewidth=2, color='g', label='# function calls')
plt.legend(loc='upper right', frameon=True, prop=dict(size=11),
# Compute a gradient-descent borderaxespad=0, handlelength=1.5, handletextpad=.5)
x_i, y_i = 1.6, 1.1 plt.tight_layout()
counting_f_prime = CountingFunction(f_prime) plt.draw()
counting_hessian = CountingFunction(hessian)
logging_f = LoggingFunction(f, counter=counting_f_prime.counter)
all_x_i, all_y_i, all_f_i = optimizer(np.array([x_i, y_i]),
logging_f, counting_f_prime,
hessian=counting_hessian)
# Plot the contour plot

if not max(all_y_i) < y_max:
x_min *= 1.2
x_max *= 1.2
y_min *= 1.2
y_max *= 1.2
x, y = np.mgrid[x_min:x_max:100j, y_min:y_max:100j]
x = x.T
y = y.T
plt.figure(index, figsize=(3, 2.5))

plt.clf() •
plt.axes([0, 0, 1, 1])
X = np.concatenate((x[np.newaxis, ...], y[np.newaxis, ...]), axis=0)

z = np.apply_along_axis(f, 0, X)
log_z = np.log(z + .01)
plt.imshow(log_z,
extent=[x_min, x_max, y_min, y_max],
cmap=plt.cm.gray_r, origin='lower',
vmax=log_z.min() + 1.5*log_z.ptp())
contours = plt.contour(log_z,
levels=levels.get(f, None),
extent=[x_min, x_max, y_min, y_max],
• •
• •
• •
• •
• •
• •
• •
• •
• •
• •
• •
• •
• •
• •
• •
• •
• •
• •
(> 250) the Hessian matrix is too costly to compute and invert. L-BFGS keeps a low-rank version. In
addition, box bounds are also supported by L-BFGS-B:

... return .5*(1 - x[0])**2 + (x[1] - x[0]**2)**2
>>> optimize.minimize(f, [2, 2], method="L-BFGS-B", jac=jacobian)
fun: 1.4417...e-15
hess_inv: <2x2 LbfgsInvHessProduct with dtype=float64>
jac: array([ 1.0233...e-07, -2.5929...e-08])
message: ...'CONVERGENCE: NORM_OF_PROJECTED_GRADIENT_<=_PGTOL'
nfev: 17
nit: 16
status: 0
success: True
x: array([1.0000..., 1.0000...])
• 14.4.12 Gradient-less methods

A shooting method: the Powell algorithm
Almost a gradient approach

On a exactly quadratic function, BFGS is not as fast as New-
ton’s method, but still very fast.
Powell’s method isn’t too sensitive to local ill-
conditionning in low dimensions
An ill-conditioned non-quadratic function:

Here BFGS does better than Newton, as its empirical estimate
of the curvature is better than that given by the Hessian.
Simplex method: the Nelder-Mead

The Nelder-Mead algorithms is a generalization of dichotomy approaches to high-dimensional spaces. The
>>> def f(x): # The rosenbrock function algorithm works by refining a simplex, the generalization of intervals and triangles to high-dimensional
... return .5*(1 - x[0])**2 + (x[1] - x[0]**2)**2 spaces, to bracket the minimum.
>>> def jacobian(x): 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
>>> optimize.minimize(f, [2, -1], method="BFGS", jac=jacobian)
fun: 2.6306...e-16
large-scale bell-shape behavior. However it is slower than gradient-based methods on smooth, non-noisy
hess_inv: array([[0.99986..., 2.0000...], functions.
[2.0000..., 4.498...]])
jac: array([ 6.7089...e-08, -3.2222...e-08])
nfev: 10
nit: 8
njev: 10
status: 0
success: True
x: array([1. , 0.99999...])
L-BFGS: Limited-memory BFGS Sits between BFGS and conjugate gradient: in very high dimensions
14.5 Practical guide to optimization with scipy

14.5.1 Choosing a method
All methods are exposed as the method argument of scipy.optimize.minimize().
An ill-conditioned non-
quadratic function:
An ill-conditioned very non-

quadratic function:
Without knowledge of the gradient
Using the Nelder-Mead solver in scipy.optimize.minimize(): • In general, prefer BFGS or L-BFGS, even if you have to approximate numerically
gradients. These are also the default if you omit the parameter method - depending
if the problem has constraints or bounds
... return .5*(1 - x[0])**2 + (x[1] - x[0]**2)**2
>>> optimize.minimize(f, [2, -1], method="Nelder-Mead") • On well-conditioned problems, Powell and Nelder-Mead, both gradient-free
final_simplex: (array([[1.0000..., 1.0000...], methods, work well in high dimension, but they collapse for ill-conditioned problems.
[0.99998... , 0.99996... ],
[1.0000..., 1.0000... ]]), array([1.1152...e-10, 1.5367...e-10, 4.9883...e-10])) With knowledge of the gradient
fun: 1.1152...e-10
• BFGS or L-BFGS.
nfev: 111 • Computational overhead of BFGS is larger than that L-BFGS, itself larger than
nit: 58 that of conjugate gradient. On the other side, BFGS usually needs less function
status: 0 evaluations than CG. Thus conjugate gradient method is better than BFGS at
success: True optimizing computationally cheap functions.
x: array([1.0000..., 1.0000...])
With the Hessian
• If you can compute the Hessian, prefer the Newton method (Newton-CG or TCG).
14.4.13 Global optimizers
If you have noisy measurements
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 • Use Nelder-Mead or Powell.
may need a global optimizer.
14.5.2 Making your optimizer faster
Brute force: a grid search
• Choose the right method (see above), do compute analytically the gradient and Hessian, if you
scipy.optimize.brute() evaluates the function on a given grid of parameters and returns the param- can.
eters corresponding to the minimum value. The parameters are specified with ranges given to numpy.
mgrid. By default, 20 steps are taken in each direction: • Use preconditionning when possible.
• Choose your initialization points wisely. For instance, if you are running many similar optimiza-
tions, warm-restart one with the results of another.
... return .5*(1 - x[0])**2 + (x[1] - x[0]**2)**2
>>> optimize.brute(f, ((-1, 2), (-1, 2))) • Relax the tolerance if you don’t need precision using the parameter tol.
array([1.0000..., 1.0000...])
14.5.3 Computing gradients

Computing gradients, and even more Hessians, is very tedious but worth the effort. Symbolic computa-
tion with Sympy may come in handy.
14.4. Examples for the mathematical optimization chapter 438 14.5. Practical guide to optimization with scipy 439
Warning: A very common source of optimization not converging well is human error in the com-
putation 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, jacobian, [2, -1])
2.384185791015625e-07
See also scipy.optimize.approx_fprime() to find your errors.
14.5.4 Synthetic exercices
14.6 Special case: non-linear least-squares

14.6.1 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 Levenberg–Marquardt algorithm implemented in scipy.optimize.leastsq().
Lets try to minimize the norm of the following vectorial function:
Exercice: A simple (?) quadratic function >>> def f(x):
... return np.arctan(x) - np.arctan(np.linspace(0, 1, len(x)))
Optimize the following function, using K[0] as a starting point:
np.random.seed(0) >>> x0 = np.zeros(10)
K = np.random.normal(size=(100, 100)) >>> optimize.leastsq(f, x0)
(array([0. , 0.11111111, 0.22222222, 0.33333333, 0.44444444,
def f(x): 0.55555556, 0.66666667, 0.77777778, 0.88888889, 1. ]), 2)
return np.sum((np.dot(K, x - 1))**2) + np.sum(x**2)**2
This took 67 function evaluations (check it with ‘full_output=1’). What if we compute the norm
Time your approach. Find the fastest approach. Why is BFGS not working well? ourselves and use a good generic optimizer (BFGS):
>>> def g(x):

... return np.sum(f(x)**2)
Exercice: A locally flat minimum >>> optimize.minimize(g, x0, method="BFGS")
fun: 2.6940...e-11
Consider the function exp(-1/(.1*x**2 + y**2). This function admits a minimum in (0, 0). Starting hess_inv: array([[...
from an initialization at (1, 1), try to get within 1e-8 of this minimum point. ...
...]])
jac: array([...
...
...])
nfev: 144
nit: 11
njev: 12
status: 0
success: True
x: array([-7.3...e-09, 1.1111...e-01, 2.2222...e-01, 3.3333...e-01,
4.4444...e-01, 5.5555...e-01, 6.6666...e-01, 7.7777...e-01,
8.8889...e-01, 1.0000...e+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,
14.5. Practical guide to optimization with scipy 440 14.6. Special case: non-linear least-squares 441
and larger than the number of parameters to optimize. (continued from previous page)
nfev: 9
nit: 2
status: 0
Warning: If the function is linear, this is a linear-algebra problem, and should be solved with success: True
scipy.linalg.lstsq(). x: array([1.5, 1.5])
14.6.2 Curve fitting
14.7.2 General constraints

Equality and inequality constraints specified as functions: 𝑓 (𝑥) = 0 and 𝑔(𝑥) < 0.
Least square problems occur often when fitting a • scipy.optimize.fmin_slsqp() Sequential least square programming: equality and inequality con-
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([1.5185..., 0.92665...]), array([[ 0.00037..., -0.00056...],
[-0.0005..., 0.00123...]]))
straints:
>>> def f(x):

Exercise ... return np.sqrt((x[0] - 3)**2 + (x[1] - 2)**2)
Do the same with omega = 3. What is the difficulty? >>> def constraint(x):
... return np.atleast_1d(1.5 - np.sum(np.abs(x)))
>>> x0 = np.array([0, 0])

14.7 Optimization with constraints >>> optimize.minimize(f, x0, constraints={"fun": constraint, "type": "ineq"})
fun: 2.4748...
jac: array([-0.70708..., -0.70712...])
14.7.1 Box bounds message: ...'Optimization terminated successfully.'
Box bounds correspond to limiting each of the individual parameters of the optimization. Note that nfev: 20
some problems that are not originally written as box bounds can be rewritten as such via change of vari- nit: 5
ables. Both scipy.optimize.minimize_scalar() and scipy.optimize.minimize() support bound njev: 5
status: 0
constraints with the parameter bounds:
success: True
>>> def f(x): x: array([1.2500..., 0.2499...])
... return np.sqrt((x[0] - 3)**2 + (x[1] - 2)**2)
>>> optimize.minimize(f, np.array([0, 0]), bounds=((-1.5, 1.5), (-1.5, 1.5)))
fun: 1.5811... Warning: The above problem is known as the Lasso problem in statistics, and there exist very
hess_inv: <2x2 LbfgsInvHessProduct with dtype=float64> efficient solvers for it (for instance in scikit-learn). In general do not use generic solvers when specific
jac: array([-0.94868..., -0.31622...])
ones exist.
message: ...'CONVERGENCE: NORM_OF_PROJECTED_GRADIENT_<=_PGTOL'
14.7. Optimization with constraints 442 14.7. Optimization with constraints 443
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.
15
14.9 Examples for the mathematical optimization chapter
CHAPTER
Interfacing with C
Author: Valentin Haenel

This chapter contains an introduction to the many different routes for making your native code (primarily
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
• Introduction
• Python-C-Api
• Ctypes
• SWIG
• Cython
• Summary
• Further Reading and References
• Exercises
15.1 Introduction
This chapter covers the following techniques:
• Python-C-Api
14.8. Full code examples 444 445

• Ctypes • High maintenance cost

• SWIG (Simplified Wrapper and Interface Generator) • No forward compatibility across Python versions as C-API changes
• Cython • Reference count bugs are easy to create and very hard to track down.
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 Note: The Python-C-Api example here serves mainly for didactic reasons. Many of the other techniques
understand the wrapping problem from different angles. Having said that, there are other alternatives actually depend on this, so it is good to have a high-level understanding of how it works. In 99% of the
out there, but having understood the basics of the ones above, you will be in a position to evaluate the use-cases you will be better off, using an alternative technique.
technique of your choice to see if it fits your needs.
The following criteria may be useful when evaluating a technology:
Note: Since reference counting bugs are easy to create and hard to track down, anyone really needing
• Are additional libraries required?
to use the Python C-API should read the section about objects, types and reference counts from the
• Is the code autogenerated? official python documentation. Additionally, there is a tool by the name of cpychecker which can help
discover common errors with reference counting.
• Does it need to be compiled?
• Is there good support for interacting with Numpy arrays?
• Does it support C++? 15.2.1 Example
Before you set out, you should consider your use case. When interfacing with native code, there are The following C-extension module, make the cos function from the standard math library available to
usually two use-cases that come up: Python:
• Existing code in C/C++ that needs to be leveraged, either because it already exists, or because it /* Example of wrapping cos function from math.h with the Python-C-API. */
is faster.
# include <Python.h>
• Python code too slow, push inner loops to native code # include <math.h>
Each technology is demonstrated by wrapping the cos function from math.h. While this is a mostly a
/* wrapped cosine function */
trivial example, it should serve us well to demonstrate the basics of the wrapping solution. Since each
static PyObject* cos_func(PyObject* self, PyObject* args)
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. double value;
Last but not least, two small warnings: double answer;
• All of these techniques may crash (segmentation fault) the Python interpreter, which is (usually) /* parse the input, from python float to c double */
due to bugs in the C code. if (!PyArg_ParseTuple(args, "d", &value))
return NULL;
• All the examples have been done on Linux, they should be possible on other operating systems. /* if the above function returns -1, an appropriate Python exception will
• You will need a C compiler for most of the examples. * have been set, and the function simply returns NULL
*/
15.2 Python-C-Api /* call cos from libm */

answer = cos(value);
The Python-C-API is the backbone of the standard Python interpreter (a.k.a CPython). Using this API
/* construct the output from cos, from c double to python float */
it is possible to write Python extension module in C and C++. Obviously, these extension modules can, return Py_BuildValue("f", answer);
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. /* define functions in module */
static PyMethodDef CosMethods[] =
Advantages {
{"cos_func", cos_func, METH_VARARGS, "evaluate the cosine"},
• Requires no additional libraries {NULL, NULL, 0, NULL}
• Lots of low-level control };
• Entirely usable from C++ # if PY_MAJOR_VERSION >= 3

/* module initialization */
Disadvantages /* Python version 3*/
• May require a substantial amount of effort static struct PyModuleDef cModPyDem =
{
• Much overhead in the code PyModuleDef_HEAD_INIT,
"cos_module", "Some documentation",
• Must be compiled
15.2. Python-C-Api 446 15.2. Python-C-Api 447

(continued from previous page) The file cos_module.so contains the compiled extension, which we can now load in the IPython inter-
-1, preter:
CosMethods
};
Note: In Python 3, the filename for compiled modules includes metadata on the Python interpreter
PyMODINIT_FUNC (see PEP 3149) and is thus longer. The import statement is not affected by this.
PyInit_cos_module(void)
{
In [1]: import cos_module
return PyModule_Create(&cModPyDem);
}
In [2]: cos_module?
Type: module
# else
String Form:<module 'cos_module' from 'cos_module.so'>
File: /home/esc/git-working/scipy-lecture-notes/advanced/interfacing_with_c/python_c_api/
/* module initialization */
˓→cos_module.so
/* Python version 2 */
Docstring: <no docstring>
PyMODINIT_FUNC
initcos_module(void)
In [3]: dir(cos_module)
{
Out[3]: ['__doc__', '__file__', '__name__', '__package__', 'cos_func']
(void) Py_InitModule("cos_module", CosMethods);
}
In [4]: cos_module.cos_func(1.0)
Out[4]: 0.5403023058681398
# endif
As you can see, there is much boilerplate, both to «massage» the arguments and return types into Out[5]: 1.0
place and for the module initialisation. Although some of this is amortised, as the extension grows, the
boilerplate required for each function(s) remains. In [6]: cos_module.cos_func(3.14159265359)
Out[7]: -1.0
The standard python build system distutils supports compiling C-extensions from a setup.py, which
is rather convenient:
Now let’s see how robust this is:
from distutils.core import setup, Extension
In [10]: cos_module.cos_func('foo')
# define the extension module ---------------------------------------------------------------------------
cos_module = Extension('cos_module', sources=['cos_module.c']) TypeError Traceback (most recent call last)
<ipython-input-10-11bee483665d> in <module>()
# run the setup ----> 1 cos_module.cos_func('foo')
setup(ext_modules=[cos_module])
TypeError: a float is required
This can be compiled:
$ cd advanced/interfacing_with_c/python_c_api 15.2.2 Numpy Support

Analog to the Python-C-API, Numpy, which is itself implemented as a C-extension, comes with the
$ ls
Numpy-C-API. This API can be used to create and manipulate Numpy arrays from C, when writing a
cos_module.c setup.py
custom C-extension. See also: Advanced NumPy.
$ python setup.py build_ext --inplace
running build_ext Note: If you do ever need to use the Numpy C-API refer to the documentation about Arrays and
building 'cos_module' extension Iterators.
creating build
creating build/temp.linux-x86_64-2.7
gcc -pthread -fno-strict-aliasing -g -O2 -DNDEBUG -g -fwrapv -O3 -Wall -Wstrict-prototypes - The following example shows how to pass Numpy arrays as arguments to functions and how to iterate
˓→fPIC -I/home/esc/anaconda/include/python2.7 -c cos_module.c -o build/temp.linux-x86_64-2.7/ over Numpy arrays using the (old) Numpy-C-API. It simply takes an array as argument applies the
˓→cos_module.o
cosine function from the math.h and returns a resulting new array.
gcc -pthread -shared build/temp.linux-x86_64-2.7/cos_module.o -L/home/esc/anaconda/lib -
˓→lpython2.7 -o /home/esc/git-working/scipy-lecture-notes/advanced/interfacing_with_c/python_c_ /* Example of wrapping the cos function from math.h using the Numpy-C-API. */
˓→api/cos_module.so
# include <Python.h>
$ ls # include <numpy/arrayobject.h>
build/ cos_module.c cos_module.so setup.py # include <math.h>
• build_ext is to build extension modules

/* wrapped cosine function */
• --inplace will output the compiled extension module into the current directory (continues on next page)


static PyObject* cos_func_np(PyObject* self, PyObject* args) if (iternext == NULL) {
{ NpyIter_Deallocate(iter);
PyArrayObject *arrays[2]; /* holds input and output array */ return NULL;
PyObject *ret; }
NpyIter *iter;
npy_uint32 op_flags[2]; /* Fetch the output array which was allocated by the iterator: */
npy_uint32 iterator_flags; ret = (PyObject *)NpyIter_GetOperandArray(iter)[1];
PyArray_Descr *op_dtypes[2]; Py_INCREF(ret);
NpyIter_IterNextFunc *iternext; if (NpyIter_GetIterSize(iter) == 0) {

/*
/* parse single numpy array argument */ * If there are no elements, the loop cannot be iterated.
if (!PyArg_ParseTuple(args, "O!", &PyArray_Type, &arrays[0])) { * This check is necessary with NPY_ITER_ZEROSIZE_OK.
return NULL; */
} NpyIter_Deallocate(iter);
return ret;
arrays[1] = NULL; /* The result will be allocated by the iterator */ }
/* Set up and create the iterator */ /* The location of the data pointer which the iterator may update */
iterator_flags = (NPY_ITER_ZEROSIZE_OK | char **dataptr = NpyIter_GetDataPtrArray(iter);
/* /* The location of the stride which the iterator may update */
* Enable buffering in case the input is not behaved npy_intp *strideptr = NpyIter_GetInnerStrideArray(iter);
* (native byte order or not aligned), /* The location of the inner loop size which the iterator may update */
* disabling may speed up some cases when it is known to npy_intp *innersizeptr = NpyIter_GetInnerLoopSizePtr(iter);
* be unnecessary.
*/ /* iterate over the arrays */
NPY_ITER_BUFFERED | do {
/* Manually handle innermost iteration for speed: */ npy_intp stride = strideptr[0];
NPY_ITER_EXTERNAL_LOOP | npy_intp count = *innersizeptr;
NPY_ITER_GROWINNER); /* out is always contiguous, so use double */
double *out = (double *)dataptr[1];
op_flags[0] = (NPY_ITER_READONLY | char *in = dataptr[0];
/*
* Required that the arrays are well behaved, since the cos /* The output is allocated and guaranteed contiguous (out++ works): */
* call below requires this. assert(strideptr[1] == sizeof(double));
*/
NPY_ITER_NBO | /*
NPY_ITER_ALIGNED); * For optimization it can make sense to add a check for
* stride == sizeof(double) to allow the compiler to optimize for that.
/* Ask the iterator to allocate an array to write the output to */ */
op_flags[1] = NPY_ITER_WRITEONLY | NPY_ITER_ALLOCATE; while (count--) {
*out = cos(*(double *)in);
/* out++;
* Ensure the iteration has the correct type, could be checked in += stride;
* specifically here. }
*/ } while (iternext(iter));
op_dtypes[0] = PyArray_DescrFromType(NPY_DOUBLE);
op_dtypes[1] = op_dtypes[0]; /* Clean up and return the result */
NpyIter_Deallocate(iter);
/* Create the numpy iterator object: */ return ret;
iter = NpyIter_MultiNew(2, arrays, iterator_flags, }
/* Use input order for output and iteration */
NPY_KEEPORDER,
/* Allow only byte-swapping of input */ /* define functions in module */
NPY_EQUIV_CASTING, op_flags, op_dtypes); static PyMethodDef CosMethods[] =
Py_DECREF(op_dtypes[0]); /* The second one is identical. */ {
{"cos_func_np", cos_func_np, METH_VARARGS,
if (iter == NULL) "evaluate the cosine on a numpy array"},
return NULL; {NULL, NULL, 0, NULL}
};
iternext = NpyIter_GetIterNext(iter, NULL);

# if PY_MAJOR_VERSION >= 3 # The function is OK with `x` not having any elements:
/* module initialization */ x_empty = np.array([], dtype=np.float64)
/* Python version 3*/ y_empty = cos_module_np.cos_func_np(x_empty)
static struct PyModuleDef cModPyDem = { assert np.array_equal(y_empty, np.array([], dtype=np.float64))
PyModuleDef_HEAD_INIT,
"cos_module", "Some documentation", # The function can handle arbitrary dimensions and non-contiguous data.
-1, # `x_2d` contains the same values, but has a different shape.
CosMethods # Note: `x_2d.flags` shows it is not contiguous and `x2.ravel() == x`
}; x_2d = x.repeat(2)[::2].reshape(-1, 3)
PyMODINIT_FUNC PyInit_cos_module_np(void) { y_2d = cos_module_np.cos_func_np(x_2d)
PyObject *module; # When reshaped back, the same result is given:
module = PyModule_Create(&cModPyDem); assert np.array_equal(y_2d.ravel(), y)
if(module==NULL) return NULL;
/* IMPORTANT: this must be called */ # The function handles incorrect byte-order fine:
import_array(); x_not_native_byteorder = x.astype(x.dtype.newbyteorder())
if (PyErr_Occurred()) return NULL; y_not_native_byteorder = cos_module_np.cos_func_np(x_not_native_byteorder)
return module; assert np.array_equal(y_not_native_byteorder, y)
}
# The function fails if the data type is incorrect:
# else x_incorrect_dtype = x.astype(np.float32)
/* module initialization */ try:
/* Python version 2 */ cos_module_np.cos_func_np(x_incorrect_dtype)
PyMODINIT_FUNC initcos_module_np(void) { assert 0, "This cannot be reached."
PyObject *module; except TypeError:
module = Py_InitModule("cos_module_np", CosMethods); # A TypeError will be raised, this can be changed by changing the
if(module==NULL) return; # casting rule.
/* IMPORTANT: this must be called */ pass
import_array();
return; And this should result in the following figure:
}
# endif
To compile this we can use distutils again. However we need to be sure to include the Numpy headers
by using :func: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]) 15.3 Ctypes
To convince ourselves if this does actually works, we run the following test script: 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.
import cos_module_np
import numpy as np Advantages
• Part of the Python standard library
x = np.arange(0, 2 * np.pi, 0.1) • Does not need to be compiled
y = cos_module_np.cos_func_np(x)
plt.plot(x, y) • Wrapping code entirely in Python
plt.show()
Disadvantages
• Requires code to be wrapped to be available as a shared library (roughly speaking *.dll in Windows
# Below are more specific tests for less common usage *.so in Linux and *.dylib in Mac OSX.)
# ---------------------------------------------------
(continues on next page) • No good support for C++
15.2. Python-C-Api 452 15.3. Ctypes 453

15.3.1 Example (continued from previous page)
As advertised, the wrapper code is in pure Python. Out[5]: 1.0
""" Example of wrapping cos function from math.h using ctypes. """ In [6]: cos_module.cos_func(3.14159265359)
Out[6]: -1.0
import ctypes
As with the previous example, this code is somewhat robust, although the error message is not quite as
# find and load the library helpful, since it does not tell us what the type should be.
# OSX or linux In [7]: cos_module.cos_func('foo')

from ctypes.util import find_library ---------------------------------------------------------------------------
libm = ctypes.cdll.LoadLibrary(find_library('m')) ArgumentError Traceback (most recent call last)
# Windows ----> 1 cos_module.cos_func('foo')
# from ctypes import windll
# libm = cdll.msvcrt /home/esc/git-working/scipy-lecture-notes/advanced/interfacing_with_c/ctypes/cos_module.py in␣
˓→cos_func(arg)
12 def cos_func(arg):
# set the argument type 13 ''' Wrapper for cos from math.h '''
libm.cos.argtypes = [ctypes.c_double] ---> 14 return libm.cos(arg)
# set the return type
libm.cos.restype = ctypes.c_double ArgumentError: argument 1: <type 'exceptions.TypeError'>: wrong type
def cos_func(arg): 15.3.2 Numpy Support

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

Out[4]: 0.5403023058681398 gcc -shared -Wl,-soname,libcos_doubles.so -o libcos_doubles.so cos_doubles.o
In [5]: cos_module.cos_func(0.0) cos_doubles.o : cos_doubles.c

15.3. Ctypes 454 15.3. Ctypes 455


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 15.4 SWIG
cos_doubles.h cos_doubles.py makefile
SWIG, the Simplified Wrapper Interface Generator, is a software development tool that connects pro-
Now we can proceed to wrap this library via ctypes with direct support for (certain kinds of) Numpy grams written in C and C++ with a variety of high-level programming languages, including Python.
arrays: 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
""" Example of wrapping a C library function that accepts a C double array as large and may not be too human readable and the multiple levels of indirection which are a result of the
input using the numpy.ctypeslib. """ wrapping process, may be a bit tricky to understand.
import numpy as np
import numpy.ctypeslib as npct Note: The autogenerated C code uses the Python-C-Api.
from ctypes import c_int
# input type for the cos_doubles function Advantages

# must be a double array, with single dimension that is contiguous
• Can automatically wrap entire libraries given the headers
array_1d_double = npct.ndpointer(dtype=np.double, ndim=1, flags='CONTIGUOUS')
• Works nicely with C++
# load the library, using numpy mechanisms
libcd = npct.load_library("libcos_doubles", ".") Disadvantages
• Autogenerates enormous files
# setup the return types and argument types
libcd.cos_doubles.restype = None • Hard to debug if something goes wrong
libcd.cos_doubles.argtypes = [array_1d_double, array_1d_double, c_int]
• Steep learning curve
def cos_doubles_func(in_array, out_array): 15.4.1 Example

return libcd.cos_doubles(in_array, out_array, len(in_array))
Let’s imagine that our cos function lives in a cos_module which has been written in c and consists of
• Note the inherent limitation of contiguous single dimensional Numpy arrays, since the C functions the source file cos_module.c:
requires this kind of buffer. # include <math.h>
• Also note that the output array must be preallocated, for example with numpy.zeros() and the
function will write into it’s buffer. double cos_func(double arg){
return cos(arg);
• 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: and the header file cos_module.h:
double cos_func(double arg);
import numpy as np
import cos_doubles 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.
x = np.arange(0, 2 * np.pi, 0.1)
y = np.empty_like(x) /* Example of wrapping cos function from math.h using SWIG. */
cos_doubles.cos_doubles_func(x, y) %module cos_module

plt.plot(x, y) %{
plt.show() /* the resulting C file should be built as a python extension */
15.3. Ctypes 456 15.4. SWIG 457


# define SWIG_FILE_WITH_INIT Docstring: <no docstring>
/* Includes the header in the wrapper code */
# include "cos_module.h" In [3]: dir(cos_module)
%} Out[3]:
/* Parse the header file to generate wrappers */ ['__builtins__',
%include "cos_module.h" '__doc__',
'__file__',
As you can see, not too much code is needed here. For this simple example it is enough to simply include '__name__',
the header file in the interface file, to expose the function to Python. However, SWIG does allow for '__package__',
'_cos_module',
more fine grained inclusion/exclusion of functions found in header files, check the documentation for
'_newclass',
details.
'_object',
Generating the compiled wrappers is a two stage process: '_swig_getattr',
'_swig_property',
1. Run the swig executable on the interface file to generate the files cos_module_wrap.c, which '_swig_repr',
is the source file for the autogenerated Python C-extension and cos_module.py, which is the '_swig_setattr',
autogenerated pure python module. '_swig_setattr_nondynamic',
'cos_func']
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: In [4]: cos_module.cos_func(1.0)
Out[4]: 0.5403023058681398
setup(ext_modules=[Extension("_cos_module",
Out[5]: 1.0
sources=["cos_module.c", "cos_module.i"])])
$ cd advanced/interfacing_with_c/swig Out[6]: -1.0
$ ls Again we test for robustness, and we see that we get a better error message (although, strictly speaking
cos_module.c cos_module.h cos_module.i setup.py in Python there is no double type):
$ python setup.py build_ext --inplace In [7]: cos_module.cos_func('foo')
running build_ext ---------------------------------------------------------------------------
building '_cos_module' extension TypeError Traceback (most recent call last)
swigging cos_module.i to cos_module_wrap.c <ipython-input-7-11bee483665d> in <module>()
swig -python -o cos_module_wrap.c cos_module.i ----> 1 cos_module.cos_func('foo')
creating build
creating build/temp.linux-x86_64-2.7 TypeError: in method 'cos_func', argument 1 of type 'double'
gcc -pthread -fno-strict-aliasing -g -O2 -DNDEBUG -g -fwrapv -O3 -Wall -Wstrict-prototypes -
˓→fPIC -I/home/esc/anaconda/include/python2.7 -c cos_module.c -o build/temp.linux-x86_64-2.7/
˓→cos_module.o
15.4.2 Numpy Support
˓→fPIC -I/home/esc/anaconda/include/python2.7 -c cos_module_wrap.c -o build/temp.linux-x86_64- Numpy provides support for SWIG with the numpy.i file. This interface file defines various so-called
˓→2.7/cos_module_wrap.o typemaps which support conversion between Numpy arrays and C-Arrays. In the following example we
gcc -pthread -shared build/temp.linux-x86_64-2.7/cos_module.o build/temp.linux-x86_64-2.7/cos_ will take a quick look at how such typemaps work in practice.
˓→module_wrap.o -L/home/esc/anaconda/lib -lpython2.7 -o /home/esc/git-working/scipy-lecture-
˓→notes/advanced/interfacing_with_c/swig/_cos_module.so We have the same cos_doubles function as in the ctypes example:
$ ls void cos_doubles(double * in_array, double * out_array, int size);

build/ cos_module.c cos_module.h cos_module.i cos_module.py _cos_module.so* cos_module_
˓→wrap.c setup.py # include <math.h>
We can now load and execute the cos_module as we have done in the previous examples: /* Compute the cosine of each element in in_array, storing the result in
* out_array. */
In [1]: import cos_module void cos_doubles(double * in_array, double * out_array, int size){
int i;
In [2]: cos_module? for(i=0;i<size;i++){
Type: module out_array[i] = cos(in_array[i]);
String Form:<module 'cos_module' from 'cos_module.py'> }
File: /home/esc/git-working/scipy-lecture-notes/advanced/interfacing_with_c/swig/cos_ }
˓→module.py
(continues on next page) This is wrapped as cos_doubles_func using the following SWIG interface file:
15.4. SWIG 458 15.4. SWIG 459


/* Example of wrapping a C function that takes a C double array as input using
* numpy typemaps for SWIG. */ creating build
%module cos_doubles gcc -pthread -fno-strict-aliasing -g -O2 -DNDEBUG -g -fwrapv -O3 -Wall -Wstrict-prototypes -
%{ ˓→fPIC -I/home/esc/anaconda/lib/python2.7/site-packages/numpy/core/include -I/home/esc/
/* the resulting C file should be built as a python extension */ ˓→anaconda/include/python2.7 -c cos_doubles.c -o build/temp.linux-x86_64-2.7/cos_doubles.o
# define SWIG_FILE_WITH_INIT gcc -pthread -fno-strict-aliasing -g -O2 -DNDEBUG -g -fwrapv -O3 -Wall -Wstrict-prototypes -
/* Includes the header in the wrapper code */ ˓→fPIC -I/home/esc/anaconda/lib/python2.7/site-packages/numpy/core/include -I/home/esc/
# include "cos_doubles.h" ˓→anaconda/include/python2.7 -c cos_doubles_wrap.c -o build/temp.linux-x86_64-2.7/cos_doubles_
%} ˓→wrap.o
In file included from /home/esc/anaconda/lib/python2.7/site-packages/numpy/core/include/numpy/

/* include the numpy typemaps */ ˓→ndarraytypes.h:1722,
%include "numpy.i" from /home/esc/anaconda/lib/python2.7/site-packages/numpy/core/include/numpy/

/* need this for correct module initialization */ ˓→ndarrayobject.h:17,
%init %{ from /home/esc/anaconda/lib/python2.7/site-packages/numpy/core/include/numpy/

import_array(); ˓→arrayobject.h:15,
%} from cos_doubles_wrap.c:2706:
/home/esc/anaconda/lib/python2.7/site-packages/numpy/core/include/numpy/npy_deprecated_api.
/* typemaps for the two arrays, the second will be modified in-place */ ˓→h:11:2: warning: #warning "Using deprecated NumPy API, disable it by #defining NPY_NO_
%apply (double* IN_ARRAY1, int DIM1) {(double * in_array, int size_in)} ˓→DEPRECATED_API NPY_1_7_API_VERSION"
%apply (double* INPLACE_ARRAY1, int DIM1) {(double * out_array, int size_out)} gcc -pthread -shared build/temp.linux-x86_64-2.7/cos_doubles.o build/temp.linux-x86_64-2.7/cos_
˓→doubles_wrap.o -L/home/esc/anaconda/lib -lpython2.7 -o /home/esc/git-working/scipy-lecture-
/* Wrapper for cos_doubles that massages the types */ ˓→notes/advanced/interfacing_with_c/swig_numpy/_cos_doubles.so
%inline %{ $ ls
/* takes as input two numpy arrays */ build/ cos_doubles.h cos_doubles.py cos_doubles_wrap.c setup.py
void cos_doubles_func(double * in_array, int size_in, double * out_array, int size_out) { cos_doubles.c cos_doubles.i _cos_doubles.so* numpy.i test_cos_doubles.py
/* calls the original funcion, providing only the size of the first */
cos_doubles(in_array, out_array, size_in); And, as before, we convince ourselves that it worked:
}
%} import numpy as np
• To use the Numpy typemaps, we need include the numpy.i file. import cos_doubles
• Observe the call to import_array() which we encountered already in the Numpy-C-API example. x = np.arange(0, 2 * np.pi, 0.1)
y = np.empty_like(x)
• 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. cos_doubles.cos_doubles_func(x, y)
• As opposed to the simple SWIG example, we don’t include the cos_doubles.h header, There plt.plot(x, y)
plt.show()
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:

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
15.5 Cython
building '_cos_doubles' extension
swigging cos_doubles.i to cos_doubles_wrap.c
Cython is both a Python-like language for writing C-extensions and an advanced compiler for this
swig -python -o cos_doubles_wrap.c cos_doubles.i language. The Cython language is a superset of Python, which comes with additional constructs that
cos_doubles.i:24: Warning(490): Fragment 'NumPy_Backward_Compatibility' not found. allow you call C functions and annotate variables and class attributes with c types. In this sense one
cos_doubles.i:24: Warning(490): Fragment 'NumPy_Backward_Compatibility' not found. could also call it a Python with types.
cos_doubles.i:24: Warning(490): Fragment 'NumPy_Backward_Compatibility' not found.
In addition to the basic use case of wrapping native code, Cython supports an additional use-case,
15.4. SWIG 460 15.5. Cython 461

namely interactive optimization. Basically, one starts out with a pure-Python script and incrementally (continued from previous page)
adds Cython types to the bottleneck code to optimize only those code paths that really matter. creating build
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.
˓→fPIC -I/home/esc/anaconda/include/python2.7 -c cos_module.c -o build/temp.linux-x86_64-2.7/
While others solutions that autogenerate code can be quite difficult to debug (for example SWIG) Cython ˓→cos_module.o
comes with an extension to the GNU debugger that helps debug Python, Cython and C code. gcc -pthread -shared build/temp.linux-x86_64-2.7/cos_module.o -L/home/esc/anaconda/lib -
˓→lpython2.7 -o /home/esc/git-working/scipy-lecture-notes/advanced/interfacing_with_c/cython/
˓→cos_module.so
Note: The autogenerated C code uses the Python-C-Api. $ ls

build/ cos_module.c cos_module.pyx cos_module.so* setup.py
Advantages
And running it:
• Python like language for writing C-extensions
• Autogenerated code
In [2]: cos_module?
• Supports incremental optimization Type: module
• Includes a GNU debugger extension String Form:<module 'cos_module' from 'cos_module.so'>
File: /home/esc/git-working/scipy-lecture-notes/advanced/interfacing_with_c/cython/cos_
• Support for C++ (Since version 0.13) ˓→module.so
Docstring: <no docstring>

Disadvantages
• Must be compiled In [3]: dir(cos_module)
Out[3]:
• Requires an additional library ( but only at build time, at this problem can be overcome by shipping ['__builtins__',
the generated C files) '__doc__',
'__file__',
'__name__',
15.5.1 Example
'__package__',
The main Cython code for our cos_module is contained in the file cos_module.pyx: '__test__',
'cos_func']
""" Example of wrapping cos function from math.h using Cython. """
cdef extern from "math.h": Out[4]: 0.5403023058681398
double cos(double arg)
def cos_func(arg): Out[5]: 1.0
return cos(arg)
Note the additional keywords such as cdef and extern. Also the cos_func is then pure Python. Out[6]: -1.0
Again we can use the standard distutils module, but this time we need some additional pieces from And, testing a little for robustness, we can see that we get good error messages:
the Cython.Distutils:
In [7]: cos_module.cos_func('foo')
from distutils.core import setup, Extension ---------------------------------------------------------------------------
from Cython.Distutils import build_ext TypeError Traceback (most recent call last)
setup( ----> 1 cos_module.cos_func('foo')
cmdclass={'build_ext': build_ext},
ext_modules=[Extension("cos_module", ["cos_module.pyx"])] /home/esc/git-working/scipy-lecture-notes/advanced/interfacing_with_c/cython/cos_module.so in␣
) ˓→cos_module.cos_func (cos_module.c:506)()
Compiling this: TypeError: a float is required
$ cd advanced/interfacing_with_c/cython
Additionally, it is worth noting that Cython ships with complete declarations for the C math library,
$ ls
cos_module.pyx setup.py
which simplifies the code above to become:
$ python setup.py build_ext --inplace """ Simpler example of wrapping cos function from math.h using Cython. """
running build_ext
cythoning cos_module.pyx to cos_module.c from libc.math cimport cos
building 'cos_module' extension
15.5. Cython 462 15.5. Cython 463


def cos_func(arg): ext_modules=[Extension("cos_doubles",
return cos(arg) sources=["_cos_doubles.pyx", "cos_doubles.c"],
include_dirs=[numpy.get_include()])],
In this case the cimport statement is used to import the cos function. )
• As with the previous compiled Numpy examples, we need the include_dirs option.
15.5.2 Numpy Support
Cython has support for Numpy via the numpy.pyx file which allows you to add the Numpy array type $ ls
cos_doubles.c cos_doubles.h _cos_doubles.pyx setup.py test_cos_doubles.py
to your Cython code. I.e. like specifying that variable i is of type int, you can specify that variable
$ python setup.py build_ext -i
a is of type numpy.ndarray with a given dtype. Also, certain optimizations such as bounds checking running build_ext
are supported. Look at the corresponding section in the Cython documentation. In case you want to cythoning _cos_doubles.pyx to _cos_doubles.c
pass Numpy arrays as C arrays to your Cython wrapped C functions, there is a section about this in the building 'cos_doubles' extension
Cython documentation. creating build
In the following example, we will show how to wrap the familiar cos_doubles function using Cython.
˓→fPIC -I/home/esc/anaconda/lib/python2.7/site-packages/numpy/core/include -I/home/esc/
void cos_doubles(double * in_array, double * out_array, int size);
˓→anaconda/include/python2.7 -c _cos_doubles.c -o build/temp.linux-x86_64-2.7/_cos_doubles.o
In file included from /home/esc/anaconda/lib/python2.7/site-packages/numpy/core/include/numpy/

# include <math.h> ˓→ndarraytypes.h:1722,
from /home/esc/anaconda/lib/python2.7/site-packages/numpy/core/include/numpy/
/* Compute the cosine of each element in in_array, storing the result in ˓→ndarrayobject.h:17,
* out_array. */ from /home/esc/anaconda/lib/python2.7/site-packages/numpy/core/include/numpy/
void cos_doubles(double * in_array, double * out_array, int size){ ˓→arrayobject.h:15,
int i; from _cos_doubles.c:253:
for(i=0;i<size;i++){ /home/esc/anaconda/lib/python2.7/site-packages/numpy/core/include/numpy/npy_deprecated_api.
out_array[i] = cos(in_array[i]); ˓→h:11:2: warning: #warning "Using deprecated NumPy API, disable it by #defining NPY_NO_
} ˓→DEPRECATED_API NPY_1_7_API_VERSION"
} /home/esc/anaconda/lib/python2.7/site-packages/numpy/core/include/numpy/__ufunc_api.h:236:␣
˓→warning: ‘_import_umath’ defined but not used
This is wrapped as cos_doubles_func using the following Cython code: gcc -pthread -fno-strict-aliasing -g -O2 -DNDEBUG -g -fwrapv -O3 -Wall -Wstrict-prototypes -
˓→fPIC -I/home/esc/anaconda/lib/python2.7/site-packages/numpy/core/include -I/home/esc/
""" Example of wrapping a C function that takes C double arrays as input using ˓→anaconda/include/python2.7 -c cos_doubles.c -o build/temp.linux-x86_64-2.7/cos_doubles.o
the Numpy declarations from Cython """ gcc -pthread -shared build/temp.linux-x86_64-2.7/_cos_doubles.o build/temp.linux-x86_64-2.7/
˓→cos_doubles.o -L/home/esc/anaconda/lib -lpython2.7 -o /home/esc/git-working/scipy-lecture-
# cimport the Cython declarations for numpy ˓→notes/advanced/interfacing_with_c/cython_numpy/cos_doubles.so
cimport numpy as np $ ls
build/ _cos_doubles.c cos_doubles.c cos_doubles.h _cos_doubles.pyx cos_doubles.so* setup.
# if you want to use the Numpy-C-API from Cython ˓→py test_cos_doubles.py
# (not strictly necessary for this example, but good practice)
np.import_array()
And, as before, we convince ourselves that it worked:
# cdefine the signature of our c function import numpy as np
cdef extern from "cos_doubles.h": import matplotlib.pyplot as plt
void cos_doubles (double * in_array, double * out_array, int size) import cos_doubles
# create the wrapper code, with numpy type annotations x = np.arange(0, 2 * np.pi, 0.1)
def cos_doubles_func(np.ndarray[double, ndim=1, mode="c"] in_array not None, y = np.empty_like(x)
np.ndarray[double, ndim=1, mode="c"] out_array not None):
cos_doubles(<double*> np.PyArray_DATA(in_array), cos_doubles.cos_doubles_func(x, y)
<double*> np.PyArray_DATA(out_array), plt.plot(x, y)
in_array.shape[0]) plt.show()
And can be compiled using distutils:

import numpy
from Cython.Distutils import build_ext
setup(
cmdclass={'build_ext': build_ext},
15.5. Cython 464 15.5. Cython 465

15.8.1 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
15.6 Summary have the correct type, are one dimensional and contiguous in memory.
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. 15.8.2 Ctypes
1. Modify the Numpy example such that cos_doubles_func handles the preallocation for you, thus
x Part of CPython Compiled Autogenerated Numpy Support making it more like the Numpy-C-API example.
Python-C-API True True False True
Ctypes True False False True 15.8.3 SWIG
Swig False True True True
Cython False True True True 1. Look at the code that SWIG autogenerates, how much of it do you understand?
2. Modify the Numpy example such that cos_doubles_func handles the preallocation for you, thus
Of all three presented techniques, Cython is the most modern and advanced. In particular, the ability making it more like the Numpy-C-API example.
to optimize code incrementally by adding types to your Python code is unique.
3. Modify the cos_doubles C function so that it returns an allocated array. Can you wrap this using
SWIG typemaps? If not, why not? Is there a workaround for this specific situation? (Hint: you
15.7 Further Reading and References know the size of the output array, so it may be possible to construct a Numpy array from the
returned double *.)
• Gaël Varoquaux’s 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 15.8.4 Cython
come back to for some inspiration.
1. Look at the code that Cython autogenerates. Take a closer look at some of the comments that
Cython inserts. What do you see?
15.8 Exercises 2. Look at the section Working with Numpy from the Cython documentation to learn how to incre-
mentally optimize a pure python script that uses Numpy.
Since this is a brand new section, the exercises are considered more as pointers as to what to look at
next, so pick the ones that you find more interesting. If you have good ideas for exercises, please let us 3. Modify the Numpy example such that cos_doubles_func handles the preallocation for you, thus
know! making it more like the Numpy-C-API example.
1. Download the source code for each example and compile and run them on your machine.
2. Make trivial changes to each example and convince yourself that this works. ( E.g. change cos for
sin.)
3. Most of the examples, especially the ones involving Numpy may still be fragile and respond badly
to input errors. Look for ways to crash the examples, figure what the problem is and devise a
potential solution. Here are some ideas:
(a) Numerical overflow.
(b) Input and output arrays that have different lengths.
(c) Multidimensional array.
(d) Empty array
(e) Arrays with non-double types
4. Use the %timeit IPython magic to measure the execution time of the various solutions
15.6. Summary 466 15.8. Exercises 467

This part of the Scipy lecture notes is dedicated to various scientific packages useful for extended needs.
Part III
Packages and applications
468 469
R is a language dedicated to statistics. Python is a general-purpose language with statistics modules. R

has more statistical analysis features than Python, and specialized syntaxes. However, when it comes to
building complex analysis pipelines that mix statistics with e.g. image analysis, text mining, or control
of a physical experiment, the richness of Python is an invaluable asset.
16
Contents
• Data representation and interaction

– Data as a table
CHAPTER – The pandas data-frame
• Hypothesis testing: comparing two groups
– Student’s t-test: the simplest statistical test
– Paired tests: repeated measurements on the same individuals
• Linear models, multiple factors, and analysis of variance
– “formulas” to specify statistical models in Python
Statistics in Python – Multiple Regression: including multiple factors

– Post-hoc hypothesis testing: analysis of variance (ANOVA)
• More visualization: seaborn for statistical exploration
– Pairplot: scatter matrices
– lmplot: plotting a univariate regression
• Testing for interactions
• Full code for the figures
Requirements • Solutions to this chapter’s exercises
• Standard scientific Python environment (numpy, scipy, matplotlib)

Tip: In this document, the Python inputs are represented with the sign “>>>”.
• Pandas
• Statsmodels
• Seaborn
To install Python and these dependencies, we recommend that you download Anaconda Python or
Enthought Canopy, or preferably use the package manager if you are under Ubuntu or other linux. Disclaimer: Gender questions
Some of the examples of this tutorial are chosen around gender questions. The reason is that on such
See also: questions controlling the truth of a claim actually matters to many people.
• Bayesian statistics in Python: This chapter does not cover tools for Bayesian statistics. Of
particular interest for Bayesian modelling is PyMC, which implements a probabilistic programming
language in Python. 16.1 Data representation and interaction
• Read a statistics book: The Think stats book is available as free PDF or in print and is a great
introduction to statistics. 16.1.1 Data as a table
The setting that we consider for statistical analysis is that of multiple observations or samples described
by a set of different attributes or features. The data can than be seen as a 2D table, or matrix, with
columns giving the different attributes of the data, and rows the observations. For instance, the data
contained in examples/brain_size.csv:
"";"Gender";"FSIQ";"VIQ";"PIQ";"Weight";"Height";"MRI_Count"
Tip: Why Python for statistics? "1";"Female";133;132;124;"118";"64.5";816932
470 16.1. Data representation and interaction 471


"2";"Male";140;150;124;".";"72.5";1001121 0 -6.000000 0.279415 0.960170
"3";"Male";139;123;150;"143";"73.3";1038437 1 -5.368421 0.792419 0.609977
"4";"Male";133;129;128;"172";"68.8";965353 2 -4.736842 0.999701 0.024451
"5";"Female";137;132;134;"147";"65.0";951545 3 -4.105263 0.821291 -0.570509
4 -3.473684 0.326021 -0.945363
5 -2.842105 -0.295030 -0.955488
16.1.2 The pandas data-frame 6 -2.210526 -0.802257 -0.596979
7 -1.578947 -0.999967 -0.008151
8 -0.947368 -0.811882 0.583822
Tip: We will store and manipulate this data in a pandas.DataFrame, from the pandas module. It is ...
the Python equivalent of the spreadsheet table. It is different from a 2D numpy array as it has named
columns, can contain a mixture of different data types by column, and has elaborate selection and pivotal
mechanisms.
Creating dataframes: reading data files or converting arrays

Other inputs: pandas can input data from SQL, excel files, or other formats. See the pandas docu-
mentation.
Separator
It is a CSV file, but the separator is “;”
Reading from a CSV file: Using the above CSV file that gives observations of brain size and weight
Manipulating data
and IQ (Willerman et al. 1991), the data are a mixture of numerical and categorical values:
data is a pandas.DataFrame, that resembles R’s dataframe:
>>> import pandas
>>> data = pandas.read_csv('examples/brain_size.csv', sep=';', na_values=".") >>> data.shape # 40 rows and 8 columns
>>> data (40, 8)
Unnamed: 0 Gender FSIQ VIQ PIQ Weight Height MRI_Count
0 1 Female 133 132 124 118.0 64.5 816932 >>> data.columns # It has columns
1 2 Male 140 150 124 NaN 72.5 1001121 Index([u'Unnamed: 0', u'Gender', u'FSIQ', u'VIQ', u'PIQ', u'Weight', u'Height', u'MRI_Count'],␣
2 3 Male 139 123 150 143.0 73.3 1038437 ˓→dtype='object')
3 4 Male 133 129 128 172.0 68.8 965353

4 5 Female 137 132 134 147.0 65.0 951545 >>> print(data['Gender']) # Columns can be addressed by name
... 0 Female
1 Male
2 Male
3 Male
Warning: Missing values
4 Female
The weight of the second individual is missing in the CSV file. If we don’t specify the missing value ...
(NA = not available) marker, we will not be able to do statistical analysis.
>>> # Simpler selector
>>> data[data['Gender'] == 'Female']['VIQ'].mean()
109.45
Note: For a quick view on a large dataframe, use its describe method: pandas.DataFrame.describe().
Creating from arrays: A pandas.DataFrame can also be seen as a dictionary of 1D ‘series’, eg arrays
or lists. If we have 3 numpy arrays:

>>> t = np.linspace(-6, 6, 20)
>>> sin_t = np.sin(t)
>>> cos_t = np.cos(t)
groupby: splitting a dataframe on values of categorical variables:
We can expose them as a pandas.DataFrame: >>> groupby_gender = data.groupby('Gender')
>>> for gender, value in groupby_gender['VIQ']:
>>> pandas.DataFrame({'t': t, 'sin': sin_t, 'cos': cos_t}) ... print((gender, value.mean()))
t sin cos (continues on next page)
16.1. Data representation and interaction 472 16.1. Data representation and interaction 473

>>> from pandas.tools import plotting
('Female', 109.45) >>> plotting.scatter_matrix(data[['Weight', 'Height', 'MRI_Count']])
('Male', 115.25)
groupby_gender is a powerful object that exposes many operations on the resulting group of dataframes:
>>> groupby_gender.mean()
Unnamed: 0 FSIQ VIQ PIQ Weight Height MRI_Count
Gender
Female 19.65 111.9 109.45 110.45 137.200000 65.765000 862654.6
Male 21.35 115.0 115.25 111.60 166.444444 71.431579 954855.4
Tip: Use tab-completion on groupby_gender to find more. Other common grouping functions are
median, count (useful for checking to see the amount of missing values in different subsets) or sum.
Groupby evaluation is lazy, no work is done until an aggregation function is applied.
>>> plotting.scatter_matrix(data[['PIQ', 'VIQ', 'FSIQ']])
Two populations
The IQ metrics are bimodal, as if there are 2 sub-populations.
Exercise
• What is the mean value for VIQ for the full population?
• How many males/females were included in this study?
Hint use ‘tab completion’ to find out the methods that can be called, instead of ‘mean’ in the
above example.
• What is the average value of MRI counts expressed in log units, for males and females?
Note: groupby_gender.boxplot is used for the plots above (see this example).
Plotting data
Exercise
Pandas comes with some plotting tools (pandas.tools.plotting, using matplotlib behind the scene)
to display statistics of the data in dataframes: Plot the scatter matrix for males only, and for females only. Do you think that the 2 sub-populations
correspond to gender?
Scatter matrices:
16.1. Data representation and interaction 474 16.1. Data representation and interaction 475
16.2 Hypothesis testing: comparing two groups 16.2.2 Paired tests: repeated measurements on the same individuals
For simple statistical tests, we will use the scipy.stats sub-module of scipy:
>>> from scipy import stats
See also:
Scipy is a vast library. For a quick summary to the whole library, see the scipy chapter.
16.2.1 Student’s t-test: the simplest statistical test

1-sample t-test: testing the value of a population mean
PIQ, VIQ, and FSIQ give 3 measures of IQ. Let us test

if FISQ and PIQ are significantly different. We can use a 2 sample test:
>>> stats.ttest_ind(data['FSIQ'], data['PIQ'])

Ttest_indResult(statistic=0.46563759638..., pvalue=0.64277250...)
The problem with this approach is that it forgets that there are links between observations: FSIQ
and PIQ are measured on the same individuals. Thus the variance due to inter-subject variability is
confounding, and can be removed, using a “paired test”, or “repeated measures test”:
>>> stats.ttest_rel(data['FSIQ'], data['PIQ'])

Ttest_relResult(statistic=1.784201940..., pvalue=0.082172638183...)
scipy.stats.ttest_1samp() tests if the population mean of data is likely to be equal to a given value
(technically if observations are drawn from a Gaussian distributions of given population mean). It returns
the T statistic, and the p-value (see the function’s help):
>>> stats.ttest_1samp(data['VIQ'], 0) This is equivalent to a 1-sample test on the difference:

Ttest_1sampResult(statistic=30.088099970..., pvalue=1.32891964...e-28)
>>> stats.ttest_1samp(data['FSIQ'] - data['PIQ'], 0)
Ttest_1sampResult(statistic=1.784201940..., pvalue=0.082172638...)
Tip: With a p-value of 10^-28 we can claim that the population mean for the IQ (VIQ measure) is not
0.
2-sample t-test: testing for difference across populations

T-tests assume Gaussian errors. We can use a Wilcoxon signed-rank test, that relaxes this assumption:
We have seen above that the mean VIQ in the male and female populations were different. To test if
this is significant, we do a 2-sample t-test with scipy.stats.ttest_ind(): >>> stats.wilcoxon(data['FSIQ'], data['PIQ'])
WilcoxonResult(statistic=274.5, pvalue=0.106594927...)
>>> female_viq = data[data['Gender'] == 'Female']['VIQ']
>>> male_viq = data[data['Gender'] == 'Male']['VIQ']
>>> stats.ttest_ind(female_viq, male_viq) Note: The corresponding test in the non paired case is the Mann–Whitney U test, scipy.stats.
Ttest_indResult(statistic=-0.77261617232..., pvalue=0.4445287677858...) mannwhitneyu().
Exercise
16.2. Hypothesis testing: comparing two groups 476 16.2. Hypothesis testing: comparing two groups 477
• Test the difference between weights in males and females.

• Use non parametric statistics to test the difference between VIQ in males and females.
Conclusion: we find that the data does not support the hypothesis that males and females have Then we specify an OLS model and fit it:
different VIQ.
>>> from statsmodels.formula.api import ols
>>> model = ols("y ~ x", data).fit()
We can inspect the various statistics derived from the fit:
>>> print(model.summary())
OLS Regression Results
16.3 Linear models, multiple factors, and analysis of variance ==========================...
Dep. Variable: y R-squared: 0.804
Model: OLS Adj. R-squared: 0.794
16.3.1 “formulas” to specify statistical models in Python Method: Least Squares F-statistic: 74.03
Date: ... Prob (F-statistic): 8.56e-08
A simple linear regression Time: ... Log-Likelihood: -57.988
No. Observations: 20 AIC: 120.0
Df Residuals: 18 BIC: 122.0
Df Model: 1
Covariance Type: nonrobust
==========================...
coef std err t P>|t| [0.025 0.975]
------------------------------------------...
Intercept -5.5335 1.036 -5.342 0.000 -7.710 -3.357
x 2.9369 0.341 8.604 0.000 2.220 3.654
==========================...
Omnibus: 0.100 Durbin-Watson: 2.956
Prob(Omnibus): 0.951 Jarque-Bera (JB): 0.322
Skew: -0.058 Prob(JB): 0.851
Kurtosis: 2.390 Cond. No. 3.03
==========================...
Given two set of observations, x and y, we want to
Warnings:
test the hypothesis that y is a linear function of x. In other terms: [1] Standard Errors assume that the covariance matrix of the errors is correctly specified.
𝑦 = 𝑥 * coef + intercept + 𝑒
where e is observation noise. We will use the statsmodels module to:
Terminology:
1. Fit a linear model. We will use the simplest strategy, ordinary least squares (OLS).
Statsmodels uses a statistical terminology: the y variable in statsmodels is called ‘endogenous’ while
2. Test that coef is non zero.
the x variable is called exogenous. This is discussed in more detail here.
To simplify, y (endogenous) is the value you are trying to predict, while x (exogenous) represents the
features you are using to make the prediction.
First, we generate simulated data according to the model:

Exercise
>>> x = np.linspace(-5, 5, 20) Retrieve the estimated parameters from the model above. Hint: use tab-completion to find the
>>> np.random.seed(1) relevent attribute.
>>> # normal distributed noise
>>> y = -5 + 3*x + 4 * np.random.normal(size=x.shape)
>>> # Create a data frame containing all the relevant variables
>>> data = pandas.DataFrame({'x': x, 'y': y})
“formulas” for statistics in Python Categorical variables: comparing groups or multiple categories
Let us go back the data on brain size:
See the statsmodels documentation
16.3. Linear models, multiple factors, and analysis of variance 478 16.3. Linear models, multiple factors, and analysis of variance 479
>>> data = pandas.read_csv('examples/brain_size.csv', sep=';', na_values=".") >>> data_fisq = pandas.DataFrame({'iq': data['FSIQ'], 'type': 'fsiq'})
>>> data_piq = pandas.DataFrame({'iq': data['PIQ'], 'type': 'piq'})
We can write a comparison between IQ of male and female using a linear model: >>> data_long = pandas.concat((data_fisq, data_piq))
>>> print(data_long)
>>> model = ols("VIQ ~ Gender + 1", data).fit() iq type
>>> print(model.summary()) 0 133 fsiq
OLS Regression Results 1 140 fsiq
==========================... 2 139 fsiq
Dep. Variable: VIQ R-squared: 0.015 ...
Model: OLS Adj. R-squared: -0.010 31 137 piq
Method: Least Squares F-statistic: 0.5969 32 110 piq
Date: ... Prob (F-statistic): 0.445 33 86 piq
Time: ... Log-Likelihood: -182.42 ...
Df Residuals: 38 BIC: 372.2 >>> model = ols("iq ~ type", data_long).fit()
Df Model: 1 >>> print(model.summary())
Covariance Type: nonrobust OLS Regression Results
==========================... ...
coef std err t P>|t| [0.025 0.975] ==========================...
-----------------------------------------------------------------------... coef std err t P>|t| [0.025 0.975]
Intercept 109.4500 5.308 20.619 0.000 98.704 120.196 ------------------------------------------...
Gender[T.Male] 5.8000 7.507 0.773 0.445 -9.397 20.997 Intercept 113.4500 3.683 30.807 0.000 106.119 120.781
==========================... type[T.piq] -2.4250 5.208 -0.466 0.643 -12.793 7.943
Omnibus: 26.188 Durbin-Watson: 1.709 ...
Skew: 0.010 Prob(JB): 0.157 We can see that we retrieve the same values for t-test and corresponding p-values for the effect of the
Kurtosis: 1.510 Cond. No. 2.62 type of iq than the previous t-test:
==========================...
>>> stats.ttest_ind(data['FSIQ'], data['PIQ'])
Ttest_indResult(statistic=0.46563759638..., pvalue=0.64277250...)
Warnings:
[1] Standard Errors assume that the covariance matrix of the errors is correctly specified.
16.3.2 Multiple Regression: including multiple factors

Tips on specifying model
Forcing categorical: the ‘Gender’ is automatically detected as a categorical variable, and thus each
of its different values are treated as different entities.
An integer column can be forced to be treated as categorical using:
>>> model = ols('VIQ ~ C(Gender)', data).fit()
Intercept: We can remove the intercept using - 1 in the formula, or force the use of an intercept
using + 1.
Tip: By default, statsmodels treats a categorical variable with K possible values as K-1 ‘dummy’
boolean variables (the last level being absorbed into the intercept term). This is almost always a
good default choice - however, it is possible to specify different encodings for categorical variables
(http://statsmodels.sourceforge.net/devel/contrasts.html).
Consider a linear model explaining a variable z (the dependent variable) with 2 variables x and y:
𝑧 = 𝑥 𝑐1 + 𝑦 𝑐2 + 𝑖 + 𝑒
Such a model can be seen in 3D as fitting a plane to a cloud of (x, y, z) points.
Link to t-tests between different FSIQ and PIQ
To compare different types of IQ, we need to create a “long-form” table, listing IQs, where the type
of IQ is indicated by a categorical variable:
Example: the iris data (examples/iris.csv)
16.3. Linear models, multiple factors, and analysis of variance 480 16.3. Linear models, multiple factors, and analysis of variance 481
Tip: Sepal and petal size tend to be related: bigger flowers are bigger! But is there in addition a
systematic effect of species?
16.3.3 Post-hoc hypothesis testing: analysis of variance (ANOVA)
In the above iris example, we wish to test if the petal length is different between versicolor and virginica,
after removing the effect of sepal width. This can be formulated as testing the difference between the
coefficient associated to versicolor and virginica in the linear model estimated above (it is an Analysis of
Variance, ANOVA). For this, we write a vector of ‘contrast’ on the parameters estimated: we want
to test "name[T.versicolor] - name[T.virginica]", with an F-test:
>>> print(model.f_test([0, 1, -1, 0]))

<F test: F=array([[3.24533535]]), p=0.07369..., df_denom=146, df_num=1>
Is this difference significant?
Exercise
Going back to the brain size + IQ data, test if the VIQ of male and female are different after removing
the effect of brain size, height and weight.
>>> data = pandas.read_csv('examples/iris.csv')

>>> model = ols('sepal_width ~ name + petal_length', data).fit()
>>> print(model.summary()) 16.4 More visualization: seaborn for statistical exploration
==========================... Seaborn combines simple statistical fits with plotting on pandas dataframes.
Dep. Variable: sepal_width R-squared: 0.478
Model: OLS Adj. R-squared: 0.468 Let us consider a data giving wages and many other personal information on 500 individuals (Berndt,
Method: Least Squares F-statistic: 44.63 ER. The Practice of Econometrics. 1991. NY: Addison-Wesley).
Date: ... Prob (F-statistic): 1.58e-20
Time: ... Log-Likelihood: -38.185
No. Observations: 150 AIC: 84.37 Tip: The full code loading and plotting of the wages data is found in corresponding example.
Df Model: 3
Covariance Type: nonrobust >>> print(data)
==========================... EDUCATION SOUTH SEX EXPERIENCE UNION WAGE AGE RACE \
coef std err t P>|t| [0.025 0.975] 0 8 0 1 21 0 0.707570 35 2
------------------------------------------... 1 9 0 1 42 0 0.694605 57 3
Intercept 2.9813 0.099 29.989 0.000 2.785 3.178 2 12 0 0 1 0 0.824126 19 3
name[T.versicolor] -1.4821 0.181 -8.190 0.000 -1.840 -1.124 3 12 0 0 4 0 0.602060 22 3
name[T.virginica] -1.6635 0.256 -6.502 0.000 -2.169 -1.158 ...
petal_length 0.2983 0.061 4.920 0.000 0.178 0.418
==========================...
Omnibus: 2.868 Durbin-Watson: 1.753 16.4.1 Pairplot: scatter matrices
We can easily have an intuition on the interactions between continuous variables using seaborn.
Skew: -0.082 Prob(JB): 0.236
pairplot() to display a scatter matrix:
==========================... >>> import seaborn
>>> seaborn.pairplot(data, vars=['WAGE', 'AGE', 'EDUCATION'],
Warnings: ... kind='reg')
16.3. Linear models, multiple factors, and analysis of variance 482 16.4. More visualization: seaborn for statistical exploration 483
Categorical variables can be

plotted as the hue:
>>> seaborn.pairplot(data, vars=['WAGE', 'AGE', 'EDUCATION'], Look and feel and matplotlib settings
... kind='reg', hue='SEX')
Seaborn changes the default of matplotlib figures to achieve a more “modern”, “excel-like” look. It
does that upon import. You can reset the default using:
>>> from matplotlib import pyplot as plt
>>> plt.rcdefaults()
Tip: To switch back to seaborn settings, or understand better styling in seaborn, see the relevent
section of the seaborn documentation.
16.4. More visualization: seaborn for statistical exploration 484 16.4. More visualization: seaborn for statistical exploration 485
16.4.2 lmplot: plotting a univariate regression 16.5 Testing for interactions
A regression capturing the relation between one vari-

able and another, eg wage and eduction, can be plotted using seaborn.lmplot():
>>> seaborn.lmplot(y='WAGE', x='EDUCATION', data=data)

Do wages increase more with edu-
cation for males than females?
Robust regression
Tip: The plot above is made of two different fits. We need to formulate a single model that tests for a
variance of slope across the two populations. This is done via an “interaction”.
Tip: Given that, in the above plot, there seems to be a couple of data points that are outside of the
main cloud to the right, they might be outliers, not representative of the population, but driving the
regression. >>> result = sm.ols(formula='wage ~ education + gender + education * gender',
... data=data).fit()
>>> print(result.summary())
To compute a regression that is less sentive to outliers, one must use a robust model. This is done ...
in seaborn using robust=True in the plotting functions, or in statsmodels by replacing the use of the coef std err t P>|t| [0.025 0.975]
OLS by a “Robust Linear Model”, statsmodels.formula.api.rlm(). ------------------------------------------------------------------------------
Intercept 0.2998 0.072 4.173 0.000 0.159 0.441
gender[T.male] 0.2750 0.093 2.972 0.003 0.093 0.457
education 0.0415 0.005 7.647 0.000 0.031 0.052
education:gender[T.male] -0.0134 0.007 -1.919 0.056 -0.027 0.000
==========================...
...
Can we conclude that education benefits males more than females?
Take home messages
• Hypothesis testing and p-values give you the significance of an effect / difference.
• Formulas (with categorical variables) enable you to express rich links in your data.
• Visualizing your data and fitting simple models give insight into the data.
• Conditionning (adding factors that can explain all or part of the variation) is an important
modeling aspect that changes the interpretation.
16.4. More visualization: seaborn for statistical exploration 486 16.5. Testing for interactions 487

16.6 Full code for the figures data = pandas.read_csv('brain_size.csv', sep=';', na_values='.')
# Box plot of FSIQ and PIQ (different measures od IQ)

Code examples for the statistics chapter. plt.figure(figsize=(4, 3))
data.boxplot(column=['FSIQ', 'PIQ'])
# Boxplot of the difference
plt.boxplot(data['FSIQ'] - data['PIQ'])
16.6.1 Boxplots and paired differences plt.xticks((1, ), ('FSIQ - PIQ', ))
Plot boxplots for FSIQ, PIQ, and the paired difference between the two: while the spread (error bars) plt.show()
for FSIQ and PIQ are very large, there is a systematic (common) effect due to the subjects. This effect
is cancelled out in the difference and the spread of the difference (“paired” by subject) is much smaller Total running time of the script: ( 0 minutes 0.053 seconds)
than the spread of the individual measures.
16.6.2 Plotting simple quantities of a pandas dataframe

This example loads from a CSV file data with mixed numerical and categorical entries, and plots a few
quantities, separately for females and males, thanks to the pandas integrated plotting tool (that uses
matplotlib behind the scene).
See http://pandas.pydata.org/pandas-docs/stable/visualization.html
• •
import pandas
16.6. Full code for the figures 488 16.6. Full code for the figures 489

data = pandas.read_csv('brain_size.csv', sep=';', na_values='.')
# Box plots of different columns for each gender

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

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

plt.show()
16.6.3 Analysis of Iris petal and sepal sizes

Ilustrate an analysis on a real dataset:
• Visualizing the data to formulate intuitions
• • Fitting of a linear model

• Hypothesis test of the effect of a categorical variable in the presence of a continuous confound
import pandas
from statsmodels.formula.api import ols
# Load the data

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

categories = pandas.Categorical(data['name'])
# The parameter 'c' is passed to plt.scatter and will control the color
plotting.scatter_matrix(data, c=categories.codes, marker='o')
fig = plt.gcf()
fig.suptitle("blue: setosa, green: versicolor, red: virginica", size=13)
import pandas

Intercept 2.9813 0.099 29.989 0.000 2.785 3.178
name[T.versicolor] -1.4821 0.181 -8.190 0.000 -1.840 -1.124
name[T.virginica] -1.6635 0.256 -6.502 0.000 -2.169 -1.158
petal_length 0.2983 0.061 4.920 0.000 0.178 0.418
==============================================================================
Skew: -0.082 Prob(JB): 0.236
==============================================================================
Warnings:
Testing the difference between effect of versicolor and virginica
<F test: F=array([[3.24533535]]), p=0.07369058781700064, df_denom=146, df_num=1>
16.6.4 Simple Regression

Fit a simple linear regression using ‘statsmodels’, compute corresponding p-values.
# Original author: Thomas Haslwanter
import numpy as np
Statistical analysis import pandas
# Let us try to explain the sepal length as a function of the petal # For statistics. Requires statsmodels 5.0 or more
# width and the category of iris from statsmodels.formula.api import ols
# Analysis of Variance (ANOVA) on linear models
model = ols('sepal_width ~ name + petal_length', data).fit() from statsmodels.stats.anova import anova_lm
print(model.summary())
Generate and show the data
# Now formulate a "contrast", to test if the offset for versicolor and
# virginica are identical x = np.linspace(-5, 5, 20)
print('Testing the difference between effect of versicolor and virginica') # To get reproducable values, provide a seed value
print(model.f_test([0, 1, -1, 0])) np.random.seed(1)
plt.show()
y = -5 + 3*x + 4 * np.random.normal(size=x.shape)
Out:
# Plot the data
OLS Regression Results plt.figure(figsize=(5, 4))
============================================================================== plt.plot(x, y, 'o')
Dep. Variable: sepal_width R-squared: 0.478
Method: Least Squares F-statistic: 44.63
Date: Thu, 18 Aug 2022 Prob (F-statistic): 1.58e-20
Time: 10:40:00 Log-Likelihood: -38.185
Df Model: 3
======================================================================================
coef std err t P>|t| [0.025 0.975]
--------------------------------------------------------------------------------------

x 2.9369 0.341 8.604 0.000 2.220 3.654
==============================================================================
Skew: -0.058 Prob(JB): 0.851
==============================================================================
Warnings:
ANOVA results
df sum_sq mean_sq F PR(>F)
x 1.0 1588.873443 1588.873443 74.029383 8.560649e-08
Residual 18.0 386.329330 21.462741 NaN NaN
Plot the fitted model
# Retrieve the parameter estimates

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

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

# Peform analysis of variance on fitted linear model

anova_results = anova_lm(model)
print('\nANOVA results')
print(anova_results)
Out:

==============================================================================
Dep. Variable: y R-squared: 0.804
Date: Thu, 18 Aug 2022 Prob (F-statistic): 8.56e-08
Df Model: 1
==============================================================================
coef std err t P>|t| [0.025 0.975] Total running time of the script: ( 0 minutes 0.051 seconds)
------------------------------------------------------------------------------
Intercept -5.5335 1.036 -5.342 0.000 -7.710 -3.357
16.6.5 Multiple Regression

Calculate using ‘statsmodels’ just the best fit, or all the corresponding statistical parameters.
Also shows how to make 3d plots.
# Original author: Thomas Haslwanter
import numpy as np
import pandas
# For 3d plots. This import is necessary to have 3D plotting below

# For statistics. Requires statsmodels 5.0 or more

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

np.random.seed(1)
Multilinear regression model, calculating fit, P-values, confidence intervals etc.
# Z is the elevation of this 2D grid
# Convert the data into a Pandas DataFrame to use the formulas framework
Z = -5 + 3*X - 0.5*Y + 8 * np.random.normal(size=X.shape)
# in statsmodels
# Plot the data
# First we need to flatten the data: it's 2D layout is not relevent.
fig = plt.figure()
X = X.flatten()
ax = fig.gca(projection='3d')
Y = Y.flatten()
surf = ax.plot_surface(X, Y, Z, cmap=plt.cm.coolwarm,
Z = Z.flatten()
rstride=1, cstride=1)
ax.view_init(20, -120)
data = pandas.DataFrame({'x': X, 'y': Y, 'z': Z})
ax.set_xlabel('X')
ax.set_ylabel('Y')
# Fit the model
ax.set_zlabel('Z')
model = ols("z ~ x + y", data).fit()
# Print the summary

print("\nRetrieving manually the parameter estimates:")

print(model._results.params)
# should be array([-4.99754526, 3.00250049, -0.50514907])
# Peform analysis of variance on fitted linear model

anova_results = anova_lm(model)
print('\nANOVA results')
print(anova_results)
plt.show()
Out:

==============================================================================
Dep. Variable: z R-squared: 0.594 # EDUCATION: Number of years of education
Model: OLS Adj. R-squared: 0.592 # SEX: 1=Female, 0=Male
Method: Least Squares F-statistic: 320.4 # WAGE: Wage (dollars per hour)
Date: Thu, 18 Aug 2022 Prob (F-statistic): 1.89e-86 data = pandas.read_csv('wages.txt', skiprows=27, skipfooter=6, sep=None,
Time: 10:40:00 Log-Likelihood: -1537.7 header=None, names=['education', 'gender', 'wage'],
No. Observations: 441 AIC: 3081. usecols=[0, 2, 5],
Df Residuals: 438 BIC: 3094. )
Df Model: 2
Covariance Type: nonrobust # Convert genders to strings (this is particulary useful so that the
============================================================================== # statsmodels formulas detects that gender is a categorical variable)
coef std err t P>|t| [0.025 0.975] import numpy as np
------------------------------------------------------------------------------ data['gender'] = np.choose(data.gender, ['male', 'female'])
Intercept -4.5052 0.378 -11.924 0.000 -5.248 -3.763
x 3.1173 0.125 24.979 0.000 2.872 3.363 # Log-transform the wages, because they typically are increased with
y -0.5109 0.125 -4.094 0.000 -0.756 -0.266 # multiplicative factors
============================================================================== data['wage'] = np.log10(data['wage'])
Prob(Omnibus): 0.878 Jarque-Bera (JB): 0.204 simple plotting
Skew: -0.052 Prob(JB): 0.903
Kurtosis: 3.015 Cond. No. 3.03 import seaborn
==============================================================================
# Plot 2 linear fits for male and female.
Warnings: seaborn.lmplot(y='wage', x='education', hue='gender', data=data)
Retrieving manually the parameter estimates:

[-4.50523303 3.11734237 -0.51091248]
ANOVA results
df sum_sq mean_sq F PR(>F)
x 1.0 39284.301219 39284.301219 623.962799 2.888238e-86
y 1.0 1055.220089 1055.220089 16.760336 5.050899e-05
Residual 438.0 27576.201607 62.959364 NaN NaN
16.6.6 Test for an education/gender interaction in wages

Wages depend mostly on education. Here we investigate how this dependence is related to gender: not
only does gender create an offset in wages, it also seems that wages increase more with education for
males than females.
Does our data support this last hypothesis? We will test this using statsmodels’ formulas (http://
statsmodels.sourceforge.net/stable/example_formulas.html).
Load and massage the data
import pandas
import urllib
import os
if not os.path.exists('wages.txt'):
# Download the file if it is not present
urllib.urlretrieve('http://lib.stat.cmu.edu/datasets/CPS_85_Wages', statistical analysis
'wages.txt')

import statsmodels.formula.api as sm
gender[T.male] 0.2750 0.093 2.972 0.003 0.093 0.457
# Note that this model is not the plot displayed above: it is one education 0.0415 0.005 7.647 0.000 0.031 0.052
# joined model for male and female, not separate models for male and education:gender[T.male] -0.0134 0.007 -1.919 0.056 -0.027 0.000
# female. The reason is that a single model enables statistical testing ==============================================================================
result = sm.ols(formula='wage ~ education + gender', data=data).fit() Omnibus: 4.838 Durbin-Watson: 1.825
print(result.summary()) Prob(Omnibus): 0.089 Jarque-Bera (JB): 5.000
Skew: -0.156 Prob(JB): 0.0821
Kurtosis: 3.356 Cond. No. 194.
Out:
==============================================================================
============================================================================== Warnings:
Dep. Variable: wage R-squared: 0.193 [1] Standard Errors assume that the covariance matrix of the errors is correctly specified.
Method: Least Squares F-statistic: 63.42 Looking at the p-value of the interaction of gender and education, the data does not support the hy-
Date: Thu, 18 Aug 2022 Prob (F-statistic): 2.01e-25 pothesis that education benefits males more than female (p-value > 0.05).
Time: 10:40:01 Log-Likelihood: 86.654
No. Observations: 534 AIC: -167.3 import matplotlib.pyplot as plt
Df Residuals: 531 BIC: -154.5 plt.show()
Df Model: 2
Covariance Type: nonrobust Total running time of the script: ( 0 minutes 0.513 seconds)
==================================================================================
coef std err t P>|t| [0.025 0.975]
---------------------------------------------------------------------------------- Note: Click here to download the full example code
Intercept 0.4053 0.046 8.732 0.000 0.314 0.496
gender[T.male] 0.1008 0.018 5.625 0.000 0.066 0.136
education 0.0334 0.003 9.768 0.000 0.027 0.040
==============================================================================
16.6.7 Visualizing factors influencing wages
Omnibus: 4.675 Durbin-Watson: 1.792 This example uses seaborn to quickly plot various factors relating wages, experience and eduction.
Skew: -0.147 Prob(JB): 0.0873 Seaborn (https://seaborn.pydata.org) is a library that combines visualization and statistical fits to show
Kurtosis: 3.365 Cond. No. 69.7 trends in data.
==============================================================================
Note that importing seaborn changes the matplotlib style to have an “excel-like” feeling. This changes
Warnings: affect other matplotlib figures. To restore defaults once this example is run, we would need to call
[1] Standard Errors assume that the covariance matrix of the errors is correctly specified. plt.rcdefaults().
# Standard library imports

The plots above highlight that there is not only a different offset in wage but also a different slope import urllib
We need to model this using an interaction import os
result = sm.ols(formula='wage ~ education + gender + education * gender', import matplotlib.pyplot as plt

data=data).fit()
print(result.summary()) Load the data
Out: import pandas
OLS Regression Results if not os.path.exists('wages.txt'):

============================================================================== # Download the file if it is not present
Dep. Variable: wage R-squared: 0.198 urllib.urlretrieve('http://lib.stat.cmu.edu/datasets/CPS_85_Wages',
Model: OLS Adj. R-squared: 0.194 'wages.txt')
Date: Thu, 18 Aug 2022 Prob (F-statistic): 2.94e-25 # Give names to the columns
Time: 10:40:01 Log-Likelihood: 88.503 names = [
No. Observations: 534 AIC: -169.0 'EDUCATION: Number of years of education',
Df Residuals: 530 BIC: -151.9 'SOUTH: 1=Person lives in South, 0=Person lives elsewhere',
Df Model: 3 'SEX: 1=Female, 0=Male',
Covariance Type: nonrobust 'EXPERIENCE: Number of years of work experience',
============================================================================================ 'UNION: 1=Union member, 0=Not union member',
coef std err t P>|t| [0.025 0.975] 'WAGE: Wage (dollars per hour)',
-------------------------------------------------------------------------------------------- 'AGE: years',
Intercept 0.2998 0.072 4.173 0.000 0.159 0.441 'RACE: 1=Other, 2=Hispanic, 3=White',

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

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

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

kind='reg', hue='SEX')
plt.suptitle('Effect of gender: 1=Female, 0=Male')

kind='reg', hue='RACE')
plt.suptitle('Effect of race: 1=Other, 2=Hispanic, 3=White')

kind='reg', hue='UNION')
plt.suptitle('Effect of union: 1=Union member, 0=Not union member')
• •

•
Plot a simple regression
seaborn.lmplot(y='WAGE', x='EDUCATION', data=data)

16.6.8 Air fares before and after 9/11
plt.show()
This is a business-intelligence (BI) like application.
What is interesting here is that we may want to study fares as a function of the year, paired accordingly
to the trips, or forgetting the year, only as a function of the trip endpoints.
Using statsmodels’ linear models, we find that both with an OLS (ordinary least square) and a robust
fit, the intercept and the slope are significantly non-zero: the air fares have decreased between 2000 and
2001, and their dependence on distance travelled has also decreased
# Standard library imports

import urllib
import os
Load the data
import pandas
if not os.path.exists('airfares.txt'):
# Download the file if it is not present
urllib.urlretrieve(
'http://www.stat.ufl.edu/~winner/data/airq4.dat',

'airfares.txt')
# As a seperator, ' +' is a regular expression that means 'one of more

# space'
data = pandas.read_csv('airfares.txt', sep=' +', header=0,
names=['city1', 'city2', 'pop1', 'pop2',
'dist', 'fare_2000', 'nb_passengers_2000',
'fare_2001', 'nb_passengers_2001'])
# we log-transform the number of passengers

import numpy as np
data['nb_passengers_2000'] = np.log10(data['nb_passengers_2000'])
data['nb_passengers_2001'] = np.log10(data['nb_passengers_2001'])
Make a dataframe whith the year as an attribute, instead of separate columns
# This involves a small danse in which we separate the dataframes in 2,

# one for year 2000, and one for 2001, before concatenating again.
# Make an index of each flight

data_flat = data.reset_index()
data_2000 = data_flat[['city1', 'city2', 'pop1', 'pop2',

'dist', 'fare_2000', 'nb_passengers_2000']]
# Rename the columns
data_2000.columns = ['city1', 'city2', 'pop1', 'pop2', 'dist', 'fare',
'nb_passengers']
# Add a column with the year
data_2000['year'] = 2000
data_2001 = data_flat[['city1', 'city2', 'pop1', 'pop2',

'dist', 'fare_2001', 'nb_passengers_2001']]
# Rename the columns
data_2001.columns = ['city1', 'city2', 'pop1', 'pop2', 'dist', 'fare',
'nb_passengers']
# Add a column with the year
data_2001['year'] = 2001
•
data_flat = pandas.concat([data_2000, data_2001])
Plot scatter matrices highlighting different aspects
import seaborn
seaborn.pairplot(data_flat, vars=['fare', 'dist', 'nb_passengers'],
kind='reg', markers='.')
# A second plot, to show the effect of the year (ie the 9/11 effect)
seaborn.pairplot(data_flat, vars=['fare', 'dist', 'nb_passengers'],
kind='reg', hue='year', markers='.')
•
Statistical testing: dependence of fare on distance and number of passengers
import statsmodels.formula.api as sm
result = sm.ols(formula='fare ~ 1 + dist + nb_passengers', data=data_flat).fit()

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

result = sm.rlm(formula='fare ~ 1 + dist + nb_passengers', data=data_flat).fit()
print(result.summary())
Out:
==============================================================================
Dep. Variable: fare R-squared: 0.275
Method: Least Squares F-statistic: 1585.
Date: Thu, 18 Aug 2022 Prob (F-statistic): 0.00
Time: 10:40:20 Log-Likelihood: -45532.
No. Observations: 8352 AIC: 9.107e+04
•
Df Residuals: 8349 BIC: 9.109e+04
Plot the difference in fare Df Model: 2
import matplotlib.pyplot as plt =================================================================================
coef std err t P>|t| [0.025 0.975]
plt.figure(figsize=(5, 2)) ---------------------------------------------------------------------------------
seaborn.boxplot(data.fare_2001 - data.fare_2000) Intercept 211.2428 2.466 85.669 0.000 206.409 216.076
plt.title('Fare: 2001 - 2000') dist 0.0484 0.001 48.149 0.000 0.046 0.050
plt.subplots_adjust() nb_passengers -32.8925 1.127 -29.191 0.000 -35.101 -30.684
==============================================================================
plt.figure(figsize=(5, 2)) Omnibus: 604.051 Durbin-Watson: 1.446
seaborn.boxplot(data.nb_passengers_2001 - data.nb_passengers_2000) Prob(Omnibus): 0.000 Jarque-Bera (JB): 740.733
plt.title('NB passengers: 2001 - 2000') Skew: 0.710 Prob(JB): 1.42e-161
plt.subplots_adjust() Kurtosis: 3.338 Cond. No. 5.23e+03
==============================================================================
Warnings:
[2] The condition number is large, 5.23e+03. This might indicate that there are
strong multicollinearity or other numerical problems.
Robust linear Model Regression Results
==============================================================================
Dep. Variable: fare No. Observations: 8352
Model: RLM Df Residuals: 8349
Method: IRLS Df Model: 2
Norm: HuberT
Scale Est.: mad
• Cov Type: H1

Date: Thu, 18 Aug 2022 ==============================================================================
Time: 10:40:20 Dep. Variable: fare_2001 R-squared: 0.159
No. Iterations: 12 Model: OLS Adj. R-squared: 0.159
================================================================================= Method: Least Squares F-statistic: 791.7
coef std err z P>|z| [0.025 0.975] Date: Thu, 18 Aug 2022 Prob (F-statistic): 1.20e-159
--------------------------------------------------------------------------------- Time: 10:40:20 Log-Likelihood: -22640.
Intercept 215.0848 2.448 87.856 0.000 210.287 219.883 No. Observations: 4176 AIC: 4.528e+04
dist 0.0460 0.001 46.166 0.000 0.044 0.048 Df Residuals: 4174 BIC: 4.530e+04
nb_passengers -35.2686 1.119 -31.526 0.000 -37.461 -33.076 Df Model: 1
================================================================================= Covariance Type: nonrobust
==============================================================================
If the model instance has been used for another fit with different fit coef std err t P>|t| [0.025 0.975]
parameters, then the fit options might not be the correct ones anymore . ------------------------------------------------------------------------------
Intercept 148.0279 1.673 88.480 0.000 144.748 151.308
Statistical testing: regression of fare on distance: 2001/2000 difference dist 0.0388 0.001 28.136 0.000 0.036 0.041
==============================================================================
result = sm.ols(formula='fare_2001 - fare_2000 ~ 1 + dist', data=data).fit() Omnibus: 136.558 Durbin-Watson: 1.544
print(result.summary()) Prob(Omnibus): 0.000 Jarque-Bera (JB): 149.624
Skew: 0.462 Prob(JB): 3.23e-33
# Plot the corresponding regression Kurtosis: 2.920 Cond. No. 2.40e+03
data['fare_difference'] = data['fare_2001'] - data['fare_2000'] ==============================================================================
seaborn.lmplot(x='dist', y='fare_difference', data=data)
Warnings:
plt.show() [1] Standard Errors assume that the covariance matrix of the errors is correctly specified.
16.7 Solutions to this chapter’s exercises
16.7.1 Relating Gender and IQ

Going back to the brain size + IQ data, test if the VIQ of male and female are different after removing
the effect of brain size, height and weight.
Notice that here ‘Gender’ is a categorical value. As it is a non-float data type, statsmodels is able to
automatically infer this.
import pandas
data = pandas.read_csv('../brain_size.csv', sep=';', na_values='.')
model = ols('VIQ ~ Gender + MRI_Count + Height', data).fit()

# Here, we don't need to define a contrast, as we are testing a single

# coefficient of our model, and not a combination of coefficients.
# However, defining a contrast, which would then be a 'unit contrast',
# will give us the same results
print(model.f_test([0, 1, 0, 0]))
Out: Out:
16.6. Full code for the figures 512 16.7. Solutions to this chapter’s exercises 513

==============================================================================
Dep. Variable: VIQ R-squared: 0.246
Date: Thu, 18 Aug 2022 Prob (F-statistic): 0.0184
Df Model: 3
==================================================================================
coef std err t P>|t| [0.025 0.975]
----------------------------------------------------------------------------------
Intercept 166.6258 88.824 1.876 0.069 -13.696 346.948
Gender[T.Male] 8.8524 10.710 0.827 0.414 -12.890 30.595
MRI_Count 0.0002 6.46e-05 2.615 0.013 3.78e-05 0.000
Height -3.0837 1.276 -2.417 0.021 -5.674 -0.494
==============================================================================
Skew: 0.005 Prob(JB): 0.324
Kurtosis: 1.823 Cond. No. 2.40e+07
==============================================================================
Warnings:
<F test: F=array([[0.68319608]]), p=0.4140878441244722, df_denom=35, df_num=1>
Here we plot a scatter matrix to get intuitions on our results. This goes beyond what was asked in the Total running time of the script: ( 0 minutes 0.328 seconds)
exercise
# This plotting is useful to get an intuitions on the relationships between

# our different variables

# Fill in the missing values for Height for plotting

data['Height'].fillna(method='pad', inplace=True)
# The parameter 'c' is passed to plt.scatter and will control the color
# The same holds for parameters 'marker', 'alpha' and 'cmap', that
# control respectively the type of marker used, their transparency and
# the colormap
plotting.scatter_matrix(data[['VIQ', 'MRI_Count', 'Height']],
c=(data['Gender'] == 'Female'), marker='o',
alpha=1, cmap='winter')
fig = plt.gcf()
fig.suptitle("blue: male, green: female", size=13)
plt.show()
16.7. Solutions to this chapter’s exercises 514 16.7. Solutions to this chapter’s exercises 515
• Algebraic manipulations
– Expand
– Simplify
• Calculus
17
– Limits
– Differentiation
– Series expansion
– Integration
CHAPTER
• Equation solving
• Linear Algebra
– Matrices
– Differential Equations
Sympy : Symbolic Mathematics in 17.1 First Steps with SymPy
Python 17.1.1 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:
>>> import sympy as sym

>>> a = sym.Rational(1, 2)
>>> a
Author: Fabian Pedregosa
1/2
Objectives >>> a*2

1
1. Evaluate expressions with arbitrary precision.
SymPy uses mpmath in the background, which makes it possible to perform computations using
2. Perform algebraic manipulations on symbolic expressions. arbitrary-precision arithmetic. That way, some special constants, like 𝑒, 𝑝𝑖, 𝑜𝑜 (Infinity), are treated
3. Perform basic calculus tasks (limits, differentiation and integration) with symbolic ex- as symbols and can be evaluated with arbitrary precision:
pressions.
>>> sym.pi**2
4. Solve polynomial and transcendental equations. pi**2
5. Solve some differential equations. >>> sym.pi.evalf()

3.14159265358979
What is SymPy? SymPy is a Python library for symbolic mathematics. It aims to be an alternative to
>>> (sym.pi + sym.exp(1)).evalf()
systems such as Mathematica or Maple while keeping the code as simple as possible and easily extensible.
5.85987448204884
SymPy is written entirely in Python and does not require any external libraries.
Sympy documentation and packages for installation can be found on http://www.sympy.org/ as you see, evalf evaluates the expression to a floating-point number.
There is also a class representing mathematical infinity, called oo:
Chapters contents
>>> sym.oo > 99999
True
• First Steps with SymPy
>>> sym.oo + 1
– Using SymPy as a calculator oo
– Symbols
516 17.1. First Steps with SymPy 517

Exercises (continued from previous page)
√ >>> sym.cos(x) * sym.cos(y) - sym.sin(x) * sym.sin(y)

1. Calculate 2 with 100 decimals. -sin(x)*sin(y) + cos(x)*cos(y)
2. Calculate 1/2 + 1/3 in rational arithmetic.

17.2.2 Simplify
Use simplify if you would like to transform an expression into a simpler form:
17.1.2 Symbols
In contrast to other Computer Algebra Systems, in SymPy you have to declare symbolic variables >>> sym.simplify((x + x * y) / x)
y + 1
explicitly:
>>> x = sym.Symbol('x') Simplification is a somewhat vague term, and more precises alternatives to simplify exists: powsimp
>>> y = sym.Symbol('y') (simplification of exponents), trigsimp (for trigonometric expressions) , logcombine, radsimp, together.
Then you can manipulate them:

Exercises
>>> x + y + x - y
2*x 1. Calculate the expanded form of (𝑥 + 𝑦)6 .
2. Simplify the trigonometric expression sin(𝑥)/ cos(𝑥)
>>> (x + y) ** 2
(x + y)**2
Symbols can now be manipulated using some of python operators: +, -`, ``*, ** (arithmetic), &, |, ~ 17.3 Calculus
, >>, << (boolean).
17.3.1 Limits
Printing
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):
Sympy allows for control of the display of the output. From here we use the following setting for
printing: >>> sym.limit(sym.sin(x) / x, x, 0)
>>> sym.init_printing(use_unicode=False, wrap_line=True) 1
you can also calculate the limit at infinity:
>>> sym.limit(x, x, sym.oo)

17.2 Algebraic manipulations oo
SymPy is capable of performing powerful algebraic manipulations. We’ll take a look into some of the >>> sym.limit(1 / x, x, sym.oo)
most frequently used: expand and simplify. 0
>>> sym.limit(x ** x, x, 0)
17.2.1 Expand 1
Use this to expand an algebraic expression. It will try to denest powers and multiplications:
>>> sym.expand((x + y) ** 3) 17.3.2 Differentiation

3 2 2 3
x + 3*x *y + 3*x*y + y You can differentiate any SymPy expression using diff(func, var). Examples:
>>> 3 * x * y ** 2 + 3 * y * x ** 2 + x ** 3 + y ** 3
3 2 2 3 >>> sym.diff(sym.sin(x), x)
x + 3*x *y + 3*x*y + y cos(x)
>>> sym.diff(sym.sin(2 * x), x)
2*cos(2*x)
Further options can be given in form on keywords:
>>> sym.expand(x + y, complex=True) >>> sym.diff(sym.tan(x), x)

re(x) + re(y) + I*im(x) + I*im(y) 2
>>> sym.I * sym.im(x) + sym.I * sym.im(y) + sym.re(x) + sym.re(y) tan (x) + 1
re(x) + re(y) + I*im(x) + I*im(y)
You can check, that it is correct by:
>>> sym.expand(sym.cos(x + y), trig=True)
-sin(x)*sin(y) + cos(x)*cos(y)
17.2. Algebraic manipulations 518 17.3. Calculus 519


>>> sym.limit((sym.tan(x + y) - sym.tan(x)) / y, y, 0)
2 --------------
tan (x) + 1 4
Higher derivatives can be calculated using the diff(func, var, n) method: It is possible to compute definite integral:
>>> sym.diff(sym.sin(2 * x), x, 1) >>> sym.integrate(x**3, (x, -1, 1))

2*cos(2*x) 0
>>> sym.integrate(sym.sin(x), (x, 0, sym.pi / 2))
>>> sym.diff(sym.sin(2 * x), x, 2) 1
-4*sin(2*x) >>> sym.integrate(sym.cos(x), (x, -sym.pi / 2, sym.pi / 2))
2
>>> sym.diff(sym.sin(2 * x), x, 3)
-8*cos(2*x) Also improper integrals are supported as well:
>>> sym.integrate(sym.exp(-x), (x, 0, sym.oo))

17.3.3 Series expansion 1
>>> sym.integrate(sym.exp(-x ** 2), (x, -sym.oo, sym.oo))
SymPy also knows how to compute the Taylor series of an expression at a point. Use series(expr, ____
var): \/ pi
>>> sym.series(sym.cos(x), x)
2 4
x x / 6\ 17.4 Equation solving
1 - -- + -- + O\x /
2 24 SymPy is able to solve algebraic equations, in one and several variables using solveset():
>>> sym.series(1/sym.cos(x), x)
2 4 >>> sym.solveset(x ** 4 - 1, x)
x 5*x / 6\ {-1, 1, -I, I}
1 + -- + ---- + O\x /
2 24 As you can see it takes as first argument an expression that is supposed to be equaled to 0. It also has
(limited) support for transcendental equations:
Exercises >>> sym.solveset(sym.exp(x) + 1, x)

{I*(2*n*pi + pi) | n in Integers}
1. Calculate lim𝑥→0 sin(𝑥)/𝑥
2. Calculate the derivative of 𝑙𝑜𝑔(𝑥) for 𝑥. Systems of linear equations
Sympy is able to solve a large part of polynomial equations, and is also capable of solving multiple
17.3.4 Integration equations with respect to multiple variables giving a tuple as second argument. To do this you use
the solve() command:
SymPy has support for indefinite and definite integration of transcendental elementary and special func-
tions via integrate() facility, which uses the powerful extended Risch-Norman algorithm and some >>> solution = sym.solve((x + 5 * y - 2, -3 * x + 6 * y - 15), (x, y))
heuristics and pattern matching. You can integrate elementary functions: >>> solution[x], solution[y]
(-3, 1)
>>> sym.integrate(6 * x ** 5, x)
6
x Another alternative in the case of polynomial equations is factor. factor returns the polynomial factorized
>>> sym.integrate(sym.sin(x), x) into irreducible terms, and is capable of computing the factorization over various domains:
-cos(x)
>>> sym.integrate(sym.log(x), x) >>> f = x ** 4 - 3 * x ** 2 + 1
x*log(x) - x >>> sym.factor(f)
>>> sym.integrate(2 * x + sym.sinh(x), x) / 2 \ / 2 \
2 \x - x - 1/*\x + x - 1/
x + cosh(x)
>>> sym.factor(f, modulus=5)
Also special functions are handled easily: 2 2
(x - 2) *(x + 2)
>>> sym.integrate(sym.exp(-x ** 2) * sym.erf(x), x)
____ 2 SymPy is also able to solve boolean equations, that is, to decide if a certain boolean expression is
\/ pi *erf (x) satisfiable or not. For this, we use the function satisfiable:
17.3. Calculus 520 17.4. Equation solving 521


>>> sym.satisfiable(x & y)
{x: True, y: True}
>>> sym.dsolve(f(x).diff(x, x) + f(x), f(x))
f(x) = C1*sin(x) + C2*cos(x)
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:
Keyword arguments can be given to this function in order to help if find the best possible resolution sys-
>>> sym.satisfiable(x & ~x) tem. For example, if you know that it is a separable equations, you can use keyword hint='separable'
False to force dsolve to resolve it as a separable equation:
>>> sym.dsolve(sym.sin(x) * sym.cos(f(x)) + sym.cos(x) * sym.sin(f(x)) * f(x).diff(x), f(x),␣

˓→hint='separable')
Exercises
/ C1 \ / C1 \
[f(x) = - acos|------| + 2*pi, f(x) = acos|------|]
1. Solve the system of equations 𝑥 + 𝑦 = 2, 2 · 𝑥 + 𝑦 = 0 \cos(x)/ \cos(x)/
2. Are there boolean values x, y that make (~x | y) & (~y | x) true?
Exercises
17.5 Linear Algebra 1. Solve the Bernoulli differential equation
17.5.1 Matrices
𝑑𝑓 (𝑥)
Matrices are created as instances from the Matrix class: 𝑥 + 𝑓 (𝑥) − 𝑓 (𝑥)2 = 0
𝑥
>>> sym.Matrix([[1, 0], [0, 1]])
[1 0]
2. Solve the same equation using hint='Bernoulli'. What do you observe ?
[ ]
[0 1]
unlike a NumPy array, you can also put Symbols in it:

>>> x, y = sym.symbols('x, y')
>>> A = sym.Matrix([[1, x], [y, 1]])
>>> A
[1 x]
[ ]
[y 1]
>>> A**2
[x*y + 1 2*x ]
[ ]
[ 2*y x*y + 1]
17.5.2 Differential Equations

SymPy is capable of solving (some) Ordinary Differential. To solve differential equations, use dsolve.
First, create an undefined function by passing cls=Function to the symbols function:
>>> f, g = sym.symbols('f g', cls=sym.Function)
f and g are now undefined functions. We can call f(x), and it will represent an unknown function:
>>> f(x)
f(x)
>>> f(x).diff(x, x) + f(x)

2
d
f(x) + ---(f(x))
2
dx
17.5. Linear Algebra 522 17.5. Linear Algebra 523

– Mathematical morphology
• Image segmentation
– Binary segmentation: foreground + background
– Marker based methods
18
• Measuring regions’ properties
• Data visualization and interaction
• Feature extraction for computer vision
CHAPTER
• Examples for the scikit-image chapter
18.1 Introduction and concepts

Images are NumPy’s arrays np.ndarray
image np.ndarray
Scikit-image: image processing pixels array values: a[2, 3]

channels array dimensions
image encoding dtype (np.uint8, np.uint16, np.float)
filters functions (numpy, skimage, scipy)

>>> check = np.zeros((8, 8))
Author: Emmanuelle Gouillart >>> check[::2, 1::2] = 1
>>> check[1::2, ::2] = 1
scikit-image is a Python package dedicated to image processing, and using natively NumPy arrays as >>> import matplotlib.pyplot as plt
image objects. This chapter describes how to use scikit-image on various image processing tasks, and >>> plt.imshow(check, cmap='gray', interpolation='nearest')
insists on the link with other scientific Python modules such as NumPy and SciPy.
See also:
For basic image manipulation, such as image cropping or simple filtering, a large number of simple
operations can be realized with NumPy and SciPy only. See Image manipulation and processing using
Numpy and Scipy.
Note that you should be familiar with the content of the previous chapter before reading the current
one, as basic operations such as masking and labeling are a prerequisite.
Chapters contents
• Introduction and concepts

– scikit-image and the SciPy ecosystem
– What’s to be found in scikit-image
• Input/output, data types and colorspaces
– Data types
– Colorspaces
18.1.1 scikit-image and the SciPy ecosystem
• Image preprocessing / enhancement
Recent versions of scikit-image is packaged in most Scientific Python distributions, such as Anaconda
– Local filters or Enthought Canopy. It is also packaged for Ubuntu/Debian.
– Non-local filters
524 18.1. Introduction and concepts 525

>>> import skimage

>>> from skimage import data # most functions are in subpackages
Most scikit-image functions take NumPy ndarrays as arguments
>>> camera = data.camera()

>>> camera.dtype
dtype('uint8')
>>> camera.shape
(512, 512)
>>> from skimage import filters
>>> filtered_camera = filters.gaussian(camera, 1)
>>> type(filtered_camera)
<type 'numpy.ndarray'>
Other Python packages are available for image processing and work with NumPy arrays:
• scipy.ndimage : for nd-arrays. Basic filtering, mathematical morphology, regions properties
• Mahotas
Works with all data formats supported by the
Also, powerful image processing libraries have Python bindings: Python Imaging Library (or any other I/O plugin provided to imread with the plugin keyword ar-
gument).
• OpenCV (computer vision)
Also works with URL image paths:
• ITK (3D images and registration)
• and many others >>> logo = io.imread('http://scikit-image.org/_static/img/logo.png')
(but they are less Pythonic and NumPy friendly, to a variable extent). Saving to files:
18.1.2 What’s to be found in scikit-image >>> io.imsave('local_logo.png', logo)
• Website: http://scikit-image.org/ (imsave also uses an external plugin such as PIL)

• Gallery of examples: http://scikit-image.org/docs/stable/auto_examples/
Different kinds of functions, from boilerplate utility functions to high-level recent algorithms.
18.2.1 Data types
• Filters: functions transforming images into other images.
– NumPy machinery
– Common filtering algorithms
• Data reduction functions: computation of image histogram, position of local maxima, of corners,
etc.
• Other actions: I/O, visualization, etc.
Image ndarrays can be represented either by integers
18.2 Input/output, data types and colorspaces (signed or unsigned) or floats.
Careful with overflows with integer data types
I/O: skimage.io
>>> camera = data.camera()
>>> from skimage import io >>> camera.dtype
dtype('uint8')
Reading from files: skimage.io.imread() >>> camera_multiply = 3 * camera
>>> import os Different integer sizes are possible: 8-, 16- or 32-bytes, signed or unsigned.
>>> filename = os.path.join(skimage.data_dir, 'camera.png')
>>> camera = io.imread(filename)
Warning: An important (if questionable) skimage convention: float images are supposed to lie
in [-1, 1] (in order to have comparable contrast for all float images)
18.2. Input/output, data types and colorspaces 526 18.2. Input/output, data types and colorspaces 527
>>> from skimage import img_as_float

>>> camera_float = img_as_float(camera)
>>> camera.max(), camera_float.max()
(255, 1.0)
Some image processing routines need to work with float arrays, and may hence output an array with a
different type and the data range from the input array
>>> from skimage import filters

>>> camera_sobel = filters.sobel(camera) Example : horizontal Sobel filter
>>> camera_sobel.max()
0.591502... >>> text = data.text()
>>> hsobel_text = filters.sobel_h(text)
Utility functions are provided in skimage to convert both the dtype and the data range, following
skimage’s conventions: util.img_as_float, util.img_as_ubyte, etc. Uses the following linear kernel for computing horizontal gradients:
See the user guide for more details. 1 2 1
0 0 0
-1 -2 -1
18.2.2 Colorspaces
Color images are of shape (N, M, 3) or (N, M, 4) (when an alpha channel encodes transparency)
>>> face = scipy.misc.face()

>>> face.shape
(768, 1024, 3)
Routines converting between different colorspaces (RGB, HSV, LAB etc.) are available in skimage.
color : color.rgb2hsv, color.lab2rgb, etc. Check the docstring for the expected dtype (and data
range) of input images.
18.3.2 Non-local filters
Non-local filters use a large region of the image (or all the image) to transform the value of one pixel:
3D images
>>> from skimage import exposure
Most functions of skimage can take 3D images as input arguments. Check the docstring to know if a >>> camera = data.camera()
function can be used on 3D images (for example MRI or CT images). >>> camera_equalized = exposure.equalize_hist(camera)
Exercise
Open a color image on your disk as a NumPy array.

Find a skimage function computing the histogram of an image and plot the histogram of
each color channel
Convert the image to grayscale and plot its histogram.
Enhances contrast in large almost uniform regions.

18.3 Image preprocessing / enhancement
Goals: denoising, feature (edges) extraction, . . . 18.3.3 Mathematical morphology
See wikipedia for an introduction on mathematical morphology.
18.3.1 Local filters Probe an image with a simple shape (a structuring element), and modify this image according to how
Local filters replace the value of pixels by a function of the values of neighboring pixels. The function the shape locally fits or misses the image.
can be linear or non-linear. Default structuring element: 4-connectivity of a pixel
Neighbourhood: square (choose size), disk, or more complicated structuring element.
18.3. Image preprocessing / enhancement 528 18.3. Image preprocessing / enhancement 529

>>> from skimage import morphology
>>> morphology.diamond(1) array([[0, 0, 0, 0, 0],
array([[0, 1, 0], [0, 0, 1, 0, 0],
[1, 1, 1], [0, 1, 1, 1, 0],
[0, 1, 0]], dtype=uint8) [0, 0, 1, 0, 0],
[0, 0, 0, 0, 0]], dtype=uint8)
Opening: erosion + dilation:
>>> a = np.zeros((5,5), dtype=np.int)

>>> 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]])
>>> morphology.binary_opening(a, morphology.diamond(1)).astype(np.uint8)
array([[0, 0, 0, 0, 0],
[0, 0, 1, 0, 0],
Erosion = minimum filter. Replace the value of a pixel by the minimal value covered by the structuring [0, 1, 1, 1, 0],
element.: [0, 0, 1, 0, 0],
[0, 0, 0, 0, 0]], dtype=uint8)
>>> a = np.zeros((7,7), dtype=np.uint8)
>>> a[1:6, 2:5] = 1 Opening removes small objects and smoothes corners.
>>> a
array([[0, 0, 0, 0, 0, 0, 0],
[0, 0, 1, 1, 1, 0, 0], Grayscale mathematical morphology
[0, 0, 1, 1, 1, 0, 0],
[0, 0, 1, 1, 1, 0, 0], Mathematical morphology operations are also available for (non-binary) grayscale images (int or float
[0, 0, 1, 1, 1, 0, 0], type). Erosion and dilation correspond to minimum (resp. maximum) filters.
[0, 0, 1, 1, 1, 0, 0],
[0, 0, 0, 0, 0, 0, 0]], dtype=uint8)
>>> morphology.binary_erosion(a, morphology.diamond(1)).astype(np.uint8) Higher-level mathematical morphology are available: tophat, skeletonization, etc.
array([[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0], See also:
[0, 0, 0, 1, 0, 0, 0], Basic mathematical morphology is also implemented in scipy.ndimage.morphology. The scipy.
[0, 0, 0, 1, 0, 0, 0], ndimage implementation works on arbitrary-dimensional arrays.
[0, 0, 0, 1, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0]], dtype=uint8)
>>> #Erosion removes objects smaller than the structure
Example of filters comparison: image denoising
>>> morphology.binary_erosion(a, morphology.diamond(2)).astype(np.uint8)
array([[0, 0, 0, 0, 0, 0, 0],
>>> from skimage.morphology import disk
[0, 0, 0, 0, 0, 0, 0],
>>> coins = data.coins()
[0, 0, 0, 0, 0, 0, 0],
>>> coins_zoom = coins[10:80, 300:370]
[0, 0, 0, 0, 0, 0, 0],
>>> median_coins = filters.median(coins_zoom, disk(1))
[0, 0, 0, 0, 0, 0, 0],
>>> from skimage import restoration
[0, 0, 0, 0, 0, 0, 0],
>>> tv_coins = restoration.denoise_tv_chambolle(coins_zoom, weight=0.1)
[0, 0, 0, 0, 0, 0, 0]], dtype=uint8)
>>> gaussian_coins = filters.gaussian(coins, sigma=2)
Dilation: maximum filter:

>>> a = np.zeros((5, 5))
>>> a[2, 2] = 1
>>> a
array([[0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0.],
[0., 0., 1., 0., 0.],
[0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0.]])
>>> morphology.binary_dilation(a, morphology.diamond(1)).astype(np.uint8)
18.3. Image preprocessing / enhancement 530 18.3. Image preprocessing / enhancement 531
18.4 Image segmentation >>> blobs_labels = measure.label(blobs, background=0)
Image segmentation is the attribution of different labels to different regions of the image, for example in
order to extract the pixels of an object of interest.
18.4.1 Binary segmentation: foreground + background

Histogram-based method: Otsu thresholding
Tip: The Otsu method is a simple heuristic to find a threshold to separate the foreground from the
background.
Earlier scikit-image versions
skimage.filters is called skimage.filter in earlier versions of scikit-image See also:

scipy.ndimage.find_objects() is useful to return slices on object in an image.
from skimage import data
from skimage import filters 18.4.2 Marker based methods
camera = data.camera()
val = filters.threshold_otsu(camera) If you have markers inside a set of regions, you can use these to segment the regions.
mask = camera < val
Watershed segmentation
The Watershed (skimage.morphology.watershed()) is a region-growing approach that fills “basins” in
the image
>>> from skimage.morphology import watershed

>>> from skimage.feature import peak_local_max
>>>
>>> # Generate an initial image with two overlapping circles
>>> x, y = np.indices((80, 80))
>>> x1, y1, x2, y2 = 28, 28, 44, 52
>>> r1, r2 = 16, 20
>>> mask_circle1 = (x - x1) ** 2 + (y - y1) ** 2 < r1 ** 2
>>> mask_circle2 = (x - x2) ** 2 + (y - y2) ** 2 < r2 ** 2
>>> image = np.logical_or(mask_circle1, mask_circle2)
>>> # Now we want to separate the two objects in image
Labeling connected components of a discrete image
>>> # Generate the markers as local maxima of the distance
>>> # to the background
Tip: Once you have separated foreground objects, it is use to separate them from each other. For this, >>> from scipy import ndimage
we can assign a different integer labels to each one. >>> distance = ndimage.distance_transform_edt(image)
>>> local_maxi = peak_local_max(distance, indices=False, footprint=np.ones((3, 3)),␣
˓→labels=image)
Synthetic data: >>> markers = morphology.label(local_maxi)

>>> labels_ws = watershed(-distance, markers, mask=image)
>>> n = 20
>>> l = 256
>>> im = np.zeros((l, l)) Random walker segmentation
>>> points = l * np.random.random((2, n ** 2))
>>> im[(points[0]).astype(np.int), (points[1]).astype(np.int)] = 1 The random walker algorithm (skimage.segmentation.random_walker()) is similar to the Watershed,
>>> im = filters.gaussian(im, sigma=l / (4. * n)) but with a more “probabilistic” approach. It is based on the idea of the diffusion of labels in the image:
>>> blobs = im > im.mean()
>>> from skimage import segmentation
>>> # Transform markers image so that 0-valued pixels are to
Label all connected components:
>>> # be labelled, and -1-valued pixels represent background
>>> from skimage import measure >>> markers[~image] = -1
>>> all_labels = measure.label(blobs) >>> labels_rw = segmentation.random_walker(image, markers)
Label only foreground connected components:
18.4. Image segmentation 532 18.4. Image segmentation 533

Some image processing operations:
>>> coins = data.coins()

>>> mask = coins > filters.threshold_otsu(coins)
>>> clean_border = segmentation.clear_border(mask)
Visualize binary result:
>>> plt.figure()
<Figure size ... with 0 Axes>
>>> plt.imshow(clean_border, cmap='gray')
Postprocessing label images

Visualize contour
skimage provides several utility functions that can be used on label images (ie images where >>> plt.figure()
different discrete values identify different regions). Functions names are often self-explaining: <Figure size ... with 0 Axes>
skimage.segmentation.clear_border(), skimage.segmentation.relabel_from_one(), skimage. >>> plt.imshow(coins, cmap='gray')
morphology.remove_small_objects(), etc. <matplotlib.image.AxesImage object at 0x...>
>>> plt.contour(clean_border, [0.5])
<matplotlib.contour.QuadContourSet ...>
Exercise Use skimage dedicated utility function:
• Load the coins image from the data submodule. >>> coins_edges = segmentation.mark_boundaries(coins, clean_border.astype(np.int))
• Separate the coins from the background by testing several segmentation methods: Otsu thresh-
olding, adaptive thresholding, and watershed or random walker segmentation.
• If necessary, use a postprocessing function to improve the coins / background segmentation.
18.5 Measuring regions’ properties

>>> from skimage import measure
Example: compute the size and perimeter of the two segmented regions:
>>> properties = measure.regionprops(labels_rw)

>>> [prop.area for prop in properties]
[770, 1168]
>>> [prop.perimeter for prop in properties] The (ex-
[100.91..., 126.81...] perimental) scikit-image viewer
skimage.viewer = matplotlib-based canvas for displaying images + experimental Qt-based GUI-toolkit
See also:
>>> from skimage import viewer
for some properties, functions are available as well in scipy.ndimage.measurements with a different >>> new_viewer = viewer.ImageViewer(coins)
API (a list is returned). >>> new_viewer.show()
Exercise (continued) Useful for displaying pixel values.

For more interaction, plugins can be added to the viewer:
• Use the binary image of the coins and background from the previous exercise.
>>> new_viewer = viewer.ImageViewer(coins)
• Compute an image of labels for the different coins. >>> from skimage.viewer.plugins import lineprofile
• Compute the size and eccentricity of all coins. >>> new_viewer += lineprofile.LineProfile()
>>> new_viewer.show()
18.6 Data visualization and interaction

Meaningful visualizations are useful when testing a given processing pipeline.
18.5. Measuring regions’ properties 534 18.6. Data visualization and interaction 535
>>> from skimage import feature
Example: detecting corners using Harris detector
from skimage.feature import corner_harris, corner_subpix, corner_peaks

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

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

coords_subpix = corner_subpix(image, coords, window_size=13)
(this ex-
ample is taken from the plot_corner example in scikit-image)
Points of interest such as corners can then be used to match objects in different images, as described in
the plot_matching example of scikit-image.

18.9 Examples for the scikit-image chapter
18.7 Feature extraction for computer vision
Geometric or textural descriptor can be extracted from images in order to
• classify parts of the image (e.g. sky vs. buildings)
• match parts of different images (e.g. for object detection) 18.9.1 Creating an image
• and many other applications of Computer Vision How to create an image with basic NumPy commands : np.zeros, slicing. . .
This examples show how to create a simple checkerboard.
18.7. Feature extraction for computer vision 536 18.8. Full code examples 537

import numpy as np
import matplotlib.pyplot as plt plt.figure(figsize=(4, 4))
plt.imshow(camera, cmap='gray', interpolation='nearest')
check = np.zeros((8, 8)) plt.axis('off')
check[::2, 1::2] = 1
check[1::2, ::2] = 1 plt.tight_layout()
plt.matshow(check, cmap='gray') plt.show()
plt.show()
18.9.3 Integers can overflow

18.9.2 Displaying a simple image An illustration of overflow problem arising when working with integers
Load and display an image
18.9. Examples for the scikit-image chapter 538 18.9. Examples for the scikit-image chapter 539
from skimage import data, exposure

import matplotlib.pyplot as plt camera = data.camera()

from skimage import data camera_equalized = exposure.equalize_hist(camera)
camera = data.camera() plt.figure(figsize=(7, 3))

camera_multiply = 3 * camera
plt.subplot(121)
plt.figure(figsize=(8, 4)) plt.imshow(camera, cmap='gray', interpolation='nearest')
plt.imshow(camera, cmap='gray', interpolation='nearest') plt.subplot(122)
plt.axis('off') plt.imshow(camera_equalized, cmap='gray', interpolation='nearest')
plt.imshow(camera_multiply, cmap='gray', interpolation='nearest') plt.tight_layout()
plt.axis('off') plt.show()
plt.tight_layout() Total running time of the script: ( 0 minutes 0.074 seconds)

plt.show()

18.9.5 Computing horizontal gradients with the Sobel filter
This example illustrates the use of the horizontal Sobel filter, to compute horizontal gradients.
18.9.4 Equalizing the histogram of an image
Histogram equalizing makes images have a uniform histogram.

from skimage import filters
text = data.text()
hsobel_text = filters.sobel_h(text)
(continued from previous page) Total running time of the script: ( 0 minutes 0.083 seconds)
plt.subplot(121)
plt.imshow(text, cmap='gray', interpolation='nearest')
plt.axis('off') 18.9.7 Otsu thresholding
plt.subplot(122)
plt.imshow(hsobel_text, cmap='nipy_spectral', interpolation='nearest') This example illustrates automatic Otsu thresholding.
plt.axis('off')
plt.tight_layout()
plt.show()
18.9.6 Segmentation contours

Visualize segmentation contours on original grayscale image.

from skimage import exposure
val = filters.threshold_otsu(camera)
hist, bins_center = exposure.histogram(camera)
plt.subplot(131)
plt.imshow(camera, cmap='gray', interpolation='nearest')
from skimage import data, segmentation plt.axis('off')
from skimage import filters plt.subplot(132)
import matplotlib.pyplot as plt plt.imshow(camera < val, cmap='gray', interpolation='nearest')
import numpy as np plt.axis('off')
plt.subplot(133)
coins = data.coins() plt.plot(bins_center, hist, lw=2)
mask = coins > filters.threshold_otsu(coins) plt.axvline(val, color='k', ls='--')
clean_border = segmentation.clear_border(mask).astype(np.int)
plt.tight_layout()
coins_edges = segmentation.mark_boundaries(coins, clean_border) plt.show()
plt.figure(figsize=(8, 3.5)) Total running time of the script: ( 0 minutes 0.104 seconds)
plt.subplot(121)
plt.imshow(clean_border, cmap='gray')
plt.axis('off') Note: Click here to download the full example code
plt.subplot(122)
plt.imshow(coins_edges)
plt.axis('off')
18.9.8 Labelling connected components of an image
plt.tight_layout() This example shows how to label connected components of a binary image, using the dedicated skim-
plt.show() age.measure.label function.
from skimage import measure

import numpy as np
n = 12
l = 256
np.random.seed(1)
points = l * np.random.random((2, n ** 2))
im = filters.gaussian(im, sigma= l / (4. * n))
blobs = im > 0.7 * im.mean()
all_labels = measure.label(blobs)
blobs_labels = measure.label(blobs, background=0) from matplotlib import pyplot as plt
plt.figure(figsize=(9, 3.5)) from skimage import data

plt.subplot(131) from skimage.feature import corner_harris, corner_subpix, corner_peaks
plt.imshow(blobs, cmap='gray') from skimage.transform import warp, AffineTransform
plt.axis('off')
plt.subplot(132)
plt.imshow(all_labels, cmap='nipy_spectral') tform = AffineTransform(scale=(1.3, 1.1), rotation=1, shear=0.7,
plt.axis('off') translation=(210, 50))
plt.subplot(133) image = warp(data.checkerboard(), tform.inverse, output_shape=(350, 350))
plt.imshow(blobs_labels, cmap='nipy_spectral')
plt.axis('off') coords = corner_peaks(corner_harris(image), min_distance=5)
coords_subpix = corner_subpix(image, coords, window_size=13)
plt.tight_layout()
plt.show() plt.gray()
plt.imshow(image, interpolation='nearest')
Total running time of the script: ( 0 minutes 0.092 seconds) plt.plot(coords_subpix[:, 1], coords_subpix[:, 0], '+r', markersize=15, mew=5)
plt.plot(coords[:, 1], coords[:, 0], '.b', markersize=7)
plt.axis('off')
Note: Click here to download the full example code plt.show()

18.9.9 Affine transform
Warping and affine transforms of images. Note: Click here to download the full example code
18.9.10 Various denoising filters

This example compares several denoising filters available in scikit-image: a Gaussian filter, a median
filter, and total variation denoising.
import numpy as np
from skimage import data import numpy as np
from skimage import filters from skimage.morphology import watershed
from skimage import restoration from skimage.feature import peak_local_max
from skimage import measure
coins = data.coins() from skimage.segmentation import random_walker
gaussian_filter_coins = filters.gaussian(coins, sigma=2) import matplotlib.pyplot as plt
med_filter_coins = filters.median(coins, np.ones((3, 3))) from scipy import ndimage
tv_filter_coins = restoration.denoise_tv_chambolle(coins, weight=0.1)
# Generate an initial image with two overlapping circles
plt.figure(figsize=(16, 4)) x, y = np.indices((80, 80))
plt.subplot(141) x1, y1, x2, y2 = 28, 28, 44, 52
plt.imshow(coins[10:80, 300:370], cmap='gray', interpolation='nearest') r1, r2 = 16, 20
plt.axis('off') mask_circle1 = (x - x1) ** 2 + (y - y1) ** 2 < r1 ** 2
plt.title('Image') mask_circle2 = (x - x2) ** 2 + (y - y2) ** 2 < r2 ** 2
plt.subplot(142) image = np.logical_or(mask_circle1, mask_circle2)
plt.imshow(gaussian_filter_coins[10:80, 300:370], cmap='gray', # Now we want to separate the two objects in image
interpolation='nearest') # Generate the markers as local maxima of the distance
plt.axis('off') # to the background
plt.title('Gaussian filter') distance = ndimage.distance_transform_edt(image)
plt.subplot(143) local_maxi = peak_local_max(
plt.imshow(med_filter_coins[10:80, 300:370], cmap='gray', distance, indices=False, footprint=np.ones((3, 3)), labels=image)
interpolation='nearest') markers = measure.label(local_maxi)
plt.axis('off') labels_ws = watershed(-distance, markers, mask=image)
plt.title('Median filter')
plt.subplot(144) markers[~image] = -1
plt.imshow(tv_filter_coins[10:80, 300:370], cmap='gray', labels_rw = random_walker(image, markers)
interpolation='nearest')
plt.axis('off') plt.figure(figsize=(12, 3.5))
plt.title('TV filter') plt.subplot(141)
plt.show() plt.imshow(image, cmap='gray', interpolation='nearest')
plt.axis('off')
plt.title('image')
plt.subplot(142)
plt.imshow(-distance, interpolation='nearest')
Note: Click here to download the full example code plt.axis('off')
plt.title('distance map')
plt.subplot(143)
plt.imshow(labels_ws, cmap='nipy_spectral', interpolation='nearest')
18.9.11 Watershed and random walker for segmentation plt.axis('off')
plt.title('watershed segmentation')
This example compares two segmentation methods in order to separate two connected disks: the water- plt.subplot(144)
shed algorithm, and the random walker algorithm. plt.imshow(labels_rw, cmap='nipy_spectral', interpolation='nearest')
Both segmentation methods require seeds, that are pixels belonging unambigusouly to a reagion. Here, plt.axis('off')
local maxima of the distance map to the background are used as seeds. plt.title('random walker segmentation')
plt.tight_layout()
plt.show()
Tutorial content
• Introduction
• Example
19
• What are Traits
– Initialisation
– Validation
– Documentation
CHAPTER – Visualization: opening a dialog
– Deferral
– Notification
– Some more advanced traits
19.1 Introduction
Traits: building interactive dialogs Tip: 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.
Author: Didrik Pinte

The Traits project allows you to simply add validation, initialization, delegation, notification and a
graphical user interface to Python object attributes.
Tip: 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
• Either wxPython, PyQt or PySide

• Numpy and Scipy
• Enthought Tool Suite
• All required software can be obtained by installing the EPD Free The main packages of the Enthought Tool Suite are:
• Traits - component based approach to build our applications.
• Kiva - 2D primitives supporting path based rendering, affine transforms, alpha blending and more.
• Enable - object based 2D drawing canvas.
548 19.1. Introduction 549

• Chaco - plotting toolkit for building complex interactive 2D plots. – Deferral

• Mayavi - 3D visualization of scientific data based on VTK. • Notification
• Envisage - application plugin framework for building scriptable and extensible applications • Visualization
In this tutorial, we will focus on Traits. • Documentation
A class can freely mix trait-based attributes with normal Python attributes, or can opt to allow the use
19.2 Example 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.
Throughout this tutorial, we will use an example based on a water resource management simple case. We The common way of creating a traits class is by extending from the HasTraits base class and defining
will try to model a dam and reservoir system. The reservoir and the dams do have a set of parameters : class traits :
• Name
from traits.api import HasTraits, Str, Float
• Minimal and maximal capacity of the reservoir [hm3]
class Reservoir(HasTraits):
• Height and length of the dam [m]
• Catchment area [km2] name = Str
max_storage = Float
• Hydraulic head [m]
• Power of the turbines [MW]
Warning: For Traits 3.x users
• Minimal and maximal release [m3/s]
If using Traits 3.x, you need to adapt the namespace of the traits packages:
• Efficiency of the turbines
• traits.api should be enthought.traits.api
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 • traitsui.api should be enthought.traits.ui.api
𝑃 = 𝜌ℎ𝑟𝑔𝑘, where:
• 𝑃 is Power in watts, Using a traits class like that is as simple as any other Python class. Note that the trait value are passed
using keyword arguments:
• 𝜌 is the density of water (~1000 kg/m3),
• ℎ is height in meters, reservoir = Reservoir(name='Lac de Vouglans', max_storage=605)
• 𝑟 is flow rate in cubic meters per second,

• 𝑔 is acceleration due to gravity of 9.8 m/s2,
19.3.1 Initialisation
All the traits do have a default value that initialise the variables. For example, the basic python types
• 𝑘 is a coefficient of efficiency ranging from 0 to 1.
do have the following trait equivalents:
Tip: Annual electric energy production depends on the available water supply. In some installations Trait Python Type Built-in Default Value
the water flow rate can vary by a factor of 10:1 over the course of a year. Bool Boolean False
Complex Complex number 0+0j
The second part of the behaviour is the state of the storage that depends on controlled and uncontrolled Float Floating point number 0.0
parameters : Int Plain integer 0
Long Long integer 0L
𝑠𝑡𝑜𝑟𝑎𝑔𝑒𝑡+1 = 𝑠𝑡𝑜𝑟𝑎𝑔𝑒𝑡 + 𝑖𝑛𝑓 𝑙𝑜𝑤𝑠 − 𝑟𝑒𝑙𝑒𝑎𝑠𝑒 − 𝑠𝑝𝑖𝑙𝑙𝑎𝑔𝑒 − 𝑖𝑟𝑟𝑖𝑔𝑎𝑡𝑖𝑜𝑛
Str String ‘’
Unicode Unicode u”
Warning: The data used in this tutorial are not real and might even not have sense in the reality.
A number of other predefined trait type do exist : Array, Enum, Range, Event, Dict, List, Color, Set,
Expression, Code, Callable, Type, Tuple, etc.
19.3 What are Traits Custom default values can be defined in the code:
A trait is a type definition that can be used for normal Python object attributes, giving the attributes from traits.api import HasTraits, Str, Float
some additional characteristics:
• Standardization:
name = Str
– Initialization
max_storage = Float(100)
– Validation (continues on next page)
19.2. Example 550 19.3. What are Traits 551


from traits.api import HasTraits, Str, Float, Range
reservoir = Reservoir(name='Lac de Vouglans') class Reservoir(HasTraits):

name = Str
max_storage = Float(1e6, desc='Maximal storage [hm3]')
Complex initialisation max_release = Float(10, desc='Maximal release [m3/s]')
head = Float(10, desc='Hydraulic head [m]')
When a complex initialisation is required for a trait, a _XXX_default magic method can be imple- efficiency = Range(0, 1.)
mented. It will be lazily called when trying to access the XXX trait. For example:
def energy_production(self, release):
def _name_default(self): ''' Returns the energy production [Wh] for the given release [m3/s]
""" Complex initialisation of the reservoir name. """ '''
power = 1000 * 9.81 * self.head * release * self.efficiency
return 'Undefined' return power * 3600
if __name__ == '__main__':
19.3.2 Validation reservoir = Reservoir(
name = 'Project A',
Every trait does validation when the user tries to set its content: max_storage = 30,
max_release = 100.0,
reservoir = Reservoir(name='Lac de Vouglans', max_storage=605)
head = 60,
efficiency = 0.8
reservoir.max_storage = '230'
)
---------------------------------------------------------------------------
TraitError Traceback (most recent call last)
release = 80
.../scipy-lecture-notes/advanced/traits/<ipython-input-7-979bdff9974a> in <module>()
print('Releasing {} m3/s produces {} kWh'.format(
----> 1 reservoir.max_storage = '230'
release, reservoir.energy_production(release)
))
.../traits/trait_handlers.pyc in error(self, object, name, value)
166 """
167 raise TraitError( object, name, self.full_info( object, name, value ),
--> 168 value ) 19.3.4 Visualization: opening a dialog
169 The Traits library is also aware of user interfaces and can pop up a default view for the Reservoir class:
170 def arg_error ( self, method, arg_num, object, name, value ):
reservoir1 = Reservoir()
TraitError: The 'max_storage' trait of a Reservoir instance must be a float, but a value of '23 reservoir1.edit_traits()
˓→' <type 'str'> was specified.
19.3.3 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 :
TraitsUI simplifies the way user interfaces are created. Every trait on a HasTraits class has a default
name = Str
editor that will manage the way the trait is rendered to the screen (e.g. the Range trait is displayed as
max_storage = Float(100, desc='Maximal storage [hm3]')
a slider, etc.).
Let’s now define the complete reservoir class: In the very same vein as the Traits declarative way of creating classes, TraitsUI provides a declarative
interface to build user interfaces code:
19.3. What are Traits 552 19.3. What are Traits 553

from traits.api import HasTraits, Str, Float, Range
from traitsui.api import View max_storage = DelegatesTo('reservoir')
min_release = Float
class Reservoir(HasTraits): max_release = DelegatesTo('reservoir')
name = Str
max_storage = Float(1e6, desc='Maximal storage [hm3]') # state attributes
max_release = Float(10, desc='Maximal release [m3/s]') storage = Range(low='min_storage', high='max_storage')
head = Float(10, desc='Hydraulic head [m]')
efficiency = Range(0, 1.) # control attributes
inflows = Float(desc='Inflows [hm3]')
traits_view = View( release = Range(low='min_release', high='max_release')
'name', 'max_storage', 'max_release', 'head', 'efficiency', spillage = Float(desc='Spillage [hm3]')
title = 'Reservoir',
resizable = True, def print_state(self):
) print('Storage\tRelease\tInflows\tSpillage')
str_format = '\t'.join(['{:7.2f} 'for i in range(4)])
def energy_production(self, release): print(str_format.format(self.storage, self.release, self.inflows,
''' Returns the energy production [Wh] for the given release [m3/s] self.spillage))
''' print('-' * 79)
power = 1000 * 9.81 * self.head * release * self.efficiency
return power * 3600
if __name__ == '__main__':
projectA = Reservoir(
if __name__ == '__main__': name = 'Project A',
reservoir = Reservoir( max_storage = 30,
name = 'Project A', max_release = 100.0,
max_storage = 30, hydraulic_head = 60,
max_release = 100.0, efficiency = 0.8
head = 60, )
efficiency = 0.8
) state = ReservoirState(reservoir=projectA, storage=10)
state.release = 90
reservoir.configure_traits() 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
from reservoir import Reservoir
class ReservoirState(HasTraits):
"""Keeps track of the reservoir state given the initial storage.
For the simplicity of the example, the release is considered in

hm3/timestep and not in m3/s.
"""
19.3.5 Deferral reservoir = Instance(Reservoir, ())
min_storage = Float
Being able to defer the definition of a trait and its value to another object is a powerful feature of Traits. max_storage = DelegatesTo('reservoir')
min_release = Float
from traits.api import HasTraits, Instance, DelegatesTo, Float, Range
max_release = DelegatesTo('reservoir')
# state attributes
storage = Range(low='min_storage', high='max_storage')
# control attributes
"""
reservoir = Instance(Reservoir, ())
release = Range(low='min_release', high='max_release')
min_storage = Float
spillage = Float(desc='Spillage [hm3]')
update_storage = Event(desc='Updates the storage to the next time step') # state attributes
storage = Property(depends_on='inflows, release')
def _update_storage_fired(self):
# update storage state # control attributes
new_storage = self.storage - self.release + self.inflows inflows = Float(desc='Inflows [hm3]')
self.storage = min(new_storage, self.max_storage) release = Range(low='min_release', high='max_release')
overflow = new_storage - self.max_storage spillage = Property(
self.spillage = max(overflow, 0) desc='Spillage [hm3]', depends_on=['storage', 'inflows', 'release']
)
def print_state(self):
print('Storage\tRelease\tInflows\tSpillage') ### Private traits.
str_format = '\t'.join(['{:7.2f} 'for i in range(4)]) _storage = Float
print(str_format.format(self.storage, self.release, self.inflows,
self.spillage)) ### Traits property implementation.
print('-' * 79) def _get_storage(self):
new_storage = self._storage - self.release + self.inflows
return min(new_storage, self.max_storage)
if __name__ == '__main__':
projectA = Reservoir( def _set_storage(self, storage_value):
name = 'Project A', self._storage = storage_value
max_storage = 30,
max_release = 5.0, def _get_spillage(self):
hydraulic_head = 60, new_storage = self._storage - self.release + self.inflows
efficiency = 0.8 overflow = new_storage - self.max_storage
) return max(overflow, 0)
state = ReservoirState(reservoir=projectA, storage=15) def print_state(self):

state.release = 5 print('Storage\tRelease\tInflows\tSpillage')
state.inflows = 0 str_format = '\t'.join(['{:7.2f} 'for i in range(4)])
# release the maximum amount of water during 3 time steps self.spillage))
state.update_storage = True print('-' * 79)
state.print_state()
state.update_storage = True if __name__ == '__main__':
state.print_state() projectA = Reservoir(
state.update_storage = True name = 'Project A',
state.print_state() max_storage = 30,
max_release = 5,
Dependency between objects can be made automatic using the trait Property. The depends_on hydraulic_head = 60,
attribute expresses the dependency between the property and other traits. When the other traits gets efficiency = 0.8
)
changed, the property is invalidated. Again, Traits uses magic method names for the property :
• _get_XXX for the getter of the XXX Property trait state = ReservoirState(reservoir=projectA, storage=25)
state.release = 4
• _set_XXX for the setter of the XXX Property trait state.inflows = 0
from traits.api import HasTraits, Instance, DelegatesTo, Float, Range
state.print_state()
from traits.api import Property

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
For the simplicity of the example, the release is considered in recompute them once invalidated.
"""
Let’s extend the TraitsUI introduction with the ReservoirState example:
max_storage = DelegatesTo('reservoir') from traits.api import HasTraits, Instance, DelegatesTo, Float, Range, Property
min_release = Float from traitsui.api import View, Item, Group, VGroup
max_release = DelegatesTo('reservoir') (continues on next page)

max_storage = 30,
from reservoir import Reservoir max_release = 5,
hydraulic_head = 60,
class ReservoirState(HasTraits): efficiency = 0.8
"""Keeps track of the reservoir state given the initial storage. )
For the simplicity of the example, the release is considered in state = ReservoirState(reservoir=projectA, storage=25)
hm3/timestep and not in m3/s. state.release = 4
""" state.inflows = 0
name = DelegatesTo('reservoir') state.print_state()
max_storage = DelegatesTo('reservoir') state.configure_traits()
min_release = Float
# state attributes
# control attributes
release = Range(low='min_release', high='max_release')
spillage = Property(
desc='Spillage [hm3]', depends_on=['storage', 'inflows', 'release']
)
### Traits view

traits_view = View(
Group(
VGroup(Item('name'), Item('storage'), Item('spillage'),
label = 'State', style = 'readonly'
),
VGroup(Item('inflows'), Item('release'), label='Control'),
)
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.
### Private traits. from traits.api import HasTraits, Str, Float, Range, PrototypedFrom, Instance
_storage = Float
class Turbine(HasTraits):
### Traits property implementation. turbine_type = Str
def _get_storage(self): power = Float(1.0, desc='Maximal power delivered by the turbine [Mw]')
return min(new_storage, self.max_storage)
def _set_storage(self, storage_value): name = Str
self._storage = storage_value max_storage = Float(1e6, desc='Maximal storage [hm3]')
max_release = Float(10, desc='Maximal release [m3/s]')
def _get_spillage(self): head = Float(10, desc='Hydraulic head [m]')
new_storage = self._storage - self.release + self.inflows efficiency = Range(0, 1.)
overflow = new_storage - self.max_storage
return max(overflow, 0) turbine = Instance(Turbine)
installed_capacity = PrototypedFrom('turbine', 'power')
print('Storage\tRelease\tInflows\tSpillage')
str_format = '\t'.join(['{:7.2f} 'for i in range(4)]) if __name__ == '__main__':
print(str_format.format(self.storage, self.release, self.inflows, turbine = Turbine(turbine_type='type1', power=5.0)
self.spillage))
print('-' * 79) reservoir = Reservoir(
name = 'Project A',
if __name__ == '__main__': max_storage = 30,
projectA = Reservoir( max_release = 100.0,
name = 'Project A', head = 60,

efficiency = 0.8, print('Warning, we are releasing {} hm3 of water'.format(new))
turbine = turbine,
)
if __name__ == '__main__':
print('installed capacity is initialised with turbine.power') projectA = Reservoir(
print(reservoir.installed_capacity) name = 'Project A',
max_storage = 30,
print('-' * 15) max_release = 100.0,
print('updating the turbine power updates the installed capacity') hydraulic_head = 60,
turbine.power = 10 efficiency = 0.8
print(reservoir.installed_capacity) )
print('-' * 15) state = ReservoirState(reservoir=projectA, storage=10)

print('setting the installed capacity breaks the link between turbine.power') state.release = 90
print('and the installed_capacity trait') state.inflows = 0
state.print_state()
reservoir.installed_capacity = 8
print(turbine.power, reservoir.installed_capacity) The static trait notification signatures can be:
• def _release_changed(self): pass
19.3.6 Notification • def _release_changed(self, new): pass
Traits implements a Listener pattern. For each trait a list of static and dynamic listeners can be fed • def _release_changed(self, old, new): pass
with callbacks. When the trait does change, all the listeners are called.
• def _release_changed(self, name, old, new pass
Static listeners are defined using the _XXX_changed magic methods:
from traits.api import HasTraits, Instance, DelegatesTo, Float, Range 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
reservoir = Instance(Reservoir, ()) the ability to register listeners on the fly with the dynamic listeners
min_storage = Float
max_storage = DelegatesTo('reservoir') from reservoir import Reservoir
min_release = Float from reservoir_state_property import ReservoirState
def wake_up_watchman_if_spillage(new_value):
# state attributes if new_value > 0:
storage = Range(low='min_storage', high='max_storage') print('Wake up watchman! Spilling {} hm3'.format(new_value))
# control attributes if __name__ == '__main__':

inflows = Float(desc='Inflows [hm3]') projectA = Reservoir(
release = Range(low='min_release', high='max_release') name = 'Project A',
spillage = Float(desc='Spillage [hm3]') max_storage = 30,
max_release = 100.0,
def print_state(self): hydraulic_head = 60,
print('Storage\tRelease\tInflows\tSpillage') efficiency = 0.8
str_format = '\t'.join(['{:7.2f} 'for i in range(4)]) )
self.spillage)) state = ReservoirState(reservoir=projectA, storage=10)
print('-' * 79)
#register the dynamic listener
### Traits listeners ########### state.on_trait_change(wake_up_watchman_if_spillage, name='spillage')
def _release_changed(self, new):
"""When the release is higher than zero, warn all the inhabitants of state.release = 90
the valley. state.inflows = 0
""" state.print_state()
if new > 0: print('Forcing spillage')


state.inflows = 100 def _get_spillage(self):
state.release = 0 new_storage = self._storage - self.release + self.inflows
overflow = new_storage - self.max_storage
print('Why do we have two executions of the callback ?') return max(overflow, 0)
The dynamic trait notification signatures are not the same as the static ones : @on_trait_change('storage')
• def wake_up_watchman(): pass print('Storage\tRelease\tInflows\tSpillage')
str_format = '\t'.join(['{:7.2f} 'for i in range(4)])
• def wake_up_watchman(new): pass
• def wake_up_watchman(name, new): pass self.spillage))
print('-' * 79)
• def wake_up_watchman(object, name, new): pass
if __name__ == '__main__':
• def wake_up_watchman(object, name, old, new): pass
projectA = Reservoir(
Removing a dynamic listener can be done by: name = 'Project A',
max_storage = 30,
• calling the remove_trait_listener method on the trait with the listener method as argument, max_release = 5,
• calling the on_trait_change method with listener method and the keyword remove=True, hydraulic_head = 60,
efficiency = 0.8
• deleting the instance that holds the listener. )
Listeners can also be added to classes using the on_trait_change decorator: state = ReservoirState(reservoir=projectA, storage=25)
from traits.api import HasTraits, Instance, DelegatesTo, Float, Range state.release = 4
from traits.api import Property, on_trait_change state.inflows = 0
from reservoir import Reservoir 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.
19.3.7 Some more advanced traits
For the simplicity of the example, the release is considered in The following example demonstrate the usage of the Enum and List traits :
""" from traits.api import HasTraits, Str, Float, Range, Enum, List
reservoir = Instance(Reservoir, ()) from traitsui.api import View, Item
max_storage = DelegatesTo('reservoir')
min_release = Float class IrrigationArea(HasTraits):
max_release = DelegatesTo('reservoir') name = Str
surface = Float(desc='Surface [ha]')
# state attributes crop = Enum('Alfalfa', 'Wheat', 'Cotton')
# control attributes class Reservoir(HasTraits):

inflows = Float(desc='Inflows [hm3]') name = Str
release = Range(low='min_release', high='max_release') max_storage = Float(1e6, desc='Maximal storage [hm3]')
spillage = Property( max_release = Float(10, desc='Maximal release [m3/s]')
desc='Spillage [hm3]', depends_on=['storage', 'inflows', 'release'] head = Float(10, desc='Hydraulic head [m]')
) efficiency = Range(0, 1.)
irrigated_areas = List(IrrigationArea)
### Private traits. ########## def energy_production(self, release):

_storage = Float ''' Returns the energy production [Wh] for the given release [m3/s]
'''
### Traits property implementation. power = 1000 * 9.81 * self.head * release * self.efficiency
def _get_storage(self): return power * 3600
return min(new_storage, self.max_storage) traits_view = View(
Item('name'),
def _set_storage(self, storage_value): Item('max_storage'),
self._storage = storage_value Item('max_release'),
Item('head'),

Item('efficiency'), Item('irrigated_areas'),
Item('irrigated_areas'), Item('total_crop_surface'),
resizable = True resizable = True
) )
if __name__ == '__main__': if __name__ == '__main__':

upper_block = IrrigationArea(name='Section C', surface=2000, crop='Wheat') upper_block = IrrigationArea(name='Section C', surface=2000, crop='Wheat')
reservoir = Reservoir( reservoir = Reservoir(

name='Project A', name='Project A',
max_storage=30, max_storage=30,
max_release=100.0, max_release=100.0,
head=60, head=60,
efficiency=0.8, efficiency=0.8,
irrigated_areas=[upper_block] irrigated_areas=[upper_block],
) )
release = 80 release = 80
print('Releasing {} m3/s produces {} kWh'.format( print('Releasing {} m3/s produces {} kWh'.format(
release, reservoir.energy_production(release) release, reservoir.energy_production(release)
)) ))
Trait listeners can be used to listen to changes in the content of the list to e.g. keep track of the total The next example shows how the Array trait can be used to feed a specialised TraitsUI Item, the
crop surface on linked to a given reservoir. ChacoPlotItem:
from traits.api import HasTraits, Str, Float, Range, Enum, List, Property import numpy as np
from traitsui.api import View, Item
from traits.api import HasTraits, Array, Instance, Float, Property
class IrrigationArea(HasTraits): from traits.api import DelegatesTo
name = Str from traitsui.api import View, Item, Group
surface = Float(desc='Surface [ha]') from chaco.chaco_plot_editor import ChacoPlotItem
crop = Enum('Alfalfa', 'Wheat', 'Cotton')
name = Str class ReservoirEvolution(HasTraits):
max_storage = Float(1e6, desc='Maximal storage [hm3]') reservoir = Instance(Reservoir)
max_release = Float(10, desc='Maximal release [m3/s]')
head = Float(10, desc='Hydraulic head [m]') name = DelegatesTo('reservoir')
efficiency = Range(0, 1.)
inflows = Array(dtype=np.float64, shape=(None))
irrigated_areas = List(IrrigationArea) releass = Array(dtype=np.float64, shape=(None))
initial_stock = Float
total_crop_surface = Property(depends_on='irrigated_areas.surface') stock = Property(depends_on='inflows, releases, initial_stock')
def _get_total_crop_surface(self): month = Property(depends_on='stock')

return sum([iarea.surface for iarea in self.irrigated_areas])
### Traits view ###########
def energy_production(self, release): traits_view = View(
''' Returns the energy production [Wh] for the given release [m3/s] Item('name'),
''' Group(
power = 1000 * 9.81 * self.head * release * self.efficiency ChacoPlotItem('month', 'stock', show_label=False),
return power * 3600 ),
width = 500,
traits_view = View( resizable = True
Item('name'), )
Item('max_storage'),
Item('max_release'), ### Traits properties #####
Item('head'), def _get_stock(self):
Item('efficiency'), """

fixme: should handle cases where we go over the max storage
"""
return self.initial_stock + (self.inflows - self.releases).cumsum()
20
def _get_month(self):
return np.arange(self.stock.size)
if __name__ == '__main__':
reservoir = Reservoir(
name = 'Project A',
max_storage = 30,
max_release = 100.0, CHAPTER
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])
3D plotting with Mayavi

view = ReservoirEvolution(
reservoir = reservoir,
inflows = inflows_ts,
releases = releases_ts
)
view.configure_traits()
Tip: Mayavi is an interactive 3D plotting package. matplotlib can also do simple 3D plotting, but
Mayavi relies on a more powerful engine ( VTK ) and is more suited to displaying large or complex data.
Chapters contents
• Mlab: the scripting interface

– 3D plotting functions
∗ Points
∗ Lines
∗ Elevation surface
See also:
∗ Arbitrary regular mesh
References
∗ Volumetric data
• ETS repositories
– Figures and decorations
• Traits manual
∗ Figure management
• Traits UI manual
∗ Changing plot properties
• Mailing list : [email protected]
∗ Decorations
• Interactive work
– The “pipeline dialog”
19.3. What are Traits 566 567

– The script recording button Lines
• Slicing and dicing data: sources, modules and filters

– An example: inspecting magnetic fields
– Different views on data: sources and modules
∗ Different sources: scatters and fields
∗ Transforming data: filters
∗ mlab.pipeline: the scripting layer
• Animating the data
• Making interactive dialogs
– A simple dialog
Hint: A line connecting points in 3D, with optional thickness and varying color.
– Making it interactive
• Putting it together
mlab.clf() # Clear the figure
t = np.linspace(0, 20, 200)
mlab.plot3d(np.sin(t), np.cos(t), 0.1*t, t)
20.1 Mlab: the scripting interface
The mayavi.mlab module provides simple plotting functions to apply to numpy arrays, similar to mat- Elevation surface
plotlib or matlab’s plotting interface. Try using them in IPython, by starting IPython with the switch
--gui=wx.
20.1.1 3D plotting functions

Points
Hint: A surface given by its elevation, coded as a 2D array
mlab.clf()
x, y = np.mgrid[-10:10:100j, -10:10:100j]
r = np.sqrt(x**2 + y**2)
z = np.sin(r)/r
Hint: Points in 3D, represented with markers (or “glyphs”) and optionaly different sizes. mlab.surf(z, warp_scale='auto')
x, y, z, value = np.random.random((4, 40))

mlab.points3d(x, y, z, value)
20.1. Mlab: the scripting interface 568 20.1. Mlab: the scripting interface 569
Arbitrary regular mesh
This function works with a regular orthogonal grid: the value array is a 3D array that gives the
Hint: A surface mesh given by x, y, z positions of its node points shape of the grid.
mlab.clf() 20.1.2 Figures and decorations

phi, theta = np.mgrid[0:np.pi:11j, 0:2*np.pi:11j]
x = np.sin(phi) * np.cos(theta) Figure management
y = np.sin(phi) * np.sin(theta)
z = np.cos(phi)
mlab.mesh(x, y, z)
Tip: Here is a list of functions useful to control the current figure
mlab.mesh(x, y, z, representation='wireframe', color=(0, 0, 0))
Get the current figure: mlab.gcf()

Note: A surface is defined by points connected to form triangles or polygones. In mayavi.mlab. Clear the current figure: mlab.clf()
surf() and mayavi.mlab.mesh(), the connectivity is implicity given by the layout of the arrays. See Set the current figure: mlab.figure(1, bgcolor=(1, 1, 1), fgcolor=(0.5, 0.5, 0.5)
also mayavi.mlab.triangular_mesh(). Save figure to image file: mlab.savefig(‘foo.png’, size=(300, 300))
Change the view: mlab.view(azimuth=45, elevation=54, distance=1.)
Our data is often more than points and values: it needs some connectivity information
Changing plot properties
Volumetric data
Tip: In general, many properties of the various objects on the figure can be changed. If these visual-
ization 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
Hint: If your data is dense in 3D, it is more difficult to display. One option is to take iso-contours of connectivity between these points is implied by the connectivity on the arrays.
the data. For simple structures (such as orthogonal grids) prefer the surf function, as it will create more efficient
data structures.
mlab.clf() Keyword arguments:
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 color the color of the vtk object. Overides the colormap, if any, when specified.
mlab.contour3d(values) 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 dis-
played. 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
Example:
scalars optional scalar data.
In [1]: import numpy as np
scale_factor scale factor of the glyphs used to represent the vertices, in
fancy_mesh mode. Must be a float. Default: 0.05 In [2]: r, theta = np.mgrid[0:10, -np.pi:np.pi:10j]
scale_mode the scaling mode for the glyphs (‘vector’, ‘scalar’, or ‘none’).
In [3]: x = r * np.cos(theta)
transparent make the opacity of the actor depend on the scalar.
In [4]: y = r * np.sin(theta)
tube_radius radius of the tubes used to represent the lines, in mesh mode. If
None, simple lines are used. In [5]: z = np.sin(r)/r
tube_sides number of sides of the tubes used to represent the lines. Must be
In [6]: from mayavi import mlab
an integer. Default: 6
vmax vmax is used to scale the colormap If None, the max of the data will be In [7]: mlab.mesh(x, y, z, colormap='gist_earth', extent=[0, 1, 0, 1, 0, 1])
used Out[7]: <mayavi.modules.surface.Surface object at 0xde6f08c>
vmin vmin is used to scale the colormap If None, the min of the data will be In [8]: mlab.mesh(x, y, z, extent=[0, 1, 0, 1, 0, 1],
used ...: representation='wireframe', line_width=1, color=(0.5, 0.5, 0.5))
Out[8]: <mayavi.modules.surface.Surface object at 0xdd6a71c>
Decorations
Tip: 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>
• 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
20.2.2 The script recording button

To find out what code can be used to program these changes, click on the red button as you modify
Warning: extent: If we specified extents for a plotting object, mlab.outline’ and ‘mlab.axes don’t those properties, and it will generate the corresponding lines of code.
get them by default.
20.3 Slicing and dicing data: sources, modules and filters

20.2 Interactive work 20.3.1 An example: inspecting magnetic fields
Suppose we are simulating the magnetic field generated by Helmholtz coils. The examples/
Tip: The quickest way to create beautiful visualization with Mayavi is probably to interactively tweak compute_field.py script does this computation and gives you a B array, that is (3 x n), where the
the various settings. first axis is the direction of the field (Bx, By, Bz), and the second axis the index number of the point.
Arrays X, Y and Z give the positions of these data points.
20.2.1 The “pipeline dialog” Exercise

Click on the ‘Mayavi’ button in the scene, and you can control properties of objects with dialogs.
Visualize this field. Your goal is to make sure that the simulation code is correct.
20.2. Interactive work 574 20.3. Slicing and dicing data: sources, modules and filters 575
such data a field, borrowing from terminology used in physics, as it is continuously defined in
space.
• A set of data points measured at random positions in a random order gives rise to much more
difficult and ill-posed interpolation problems: the data structure itself does not tell us what are
the neighbors of a data point. We call such data a scatter.
Unstructured and unconnected data: a scatter Structured and connected data: a field
mlab.points3d, mlab.quiver3d mlab.contour3d
Data sources corresponding to scatters can be created with mayavi.mlab.pipeline.scalar_scatter()

or mayavi.mlab.pipeline.vector_scatter(); field data sources can be created with mlab.pipeline.
scalar_field() or mlab.pipeline.vector_field().
Suggestions
• If you compute the norm of the vector field, you can apply an isosurface to it. Exercice:
• using mayavi.mlab.quiver3d() you can plot vectors. You can also use the ‘masking’ options
1. Create a contour (for instance of the magnetic field norm) by using one of those functions and
(in the GUI) to make the plot a bit less dense.
adding the right module by clicking on the GUI dialog.
2. Create the right source to apply a ‘vector_cut_plane’ and reproduce the picture of the magnetic
20.3.2 Different views on data: sources and modules field shown previously.
Note that one of the difficulties is providing the data in the right form (number of arrays, shape) to
Tip: As we see above, it may be desirable to look at the same data in different ways. the functions. This is often the case with real-life data.
See also:
Mayavi visualization are created by loading the data in a data source and then displayed on the screen
using modules. Sources are described in details in the Mayavi manual.
This can be seen by looking at the “pipeline” view. By right-clicking on the nodes of the pipeline, you
Transforming data: filters
can add new modules.
If you create a vector field, you may want to visualize the iso-contours of its magnitude. But the isosurface
module can only be applied to scalar data, and not vector data. We can use a filter, ExtractVectorNorm
Quiz
to add this scalar value to the vector field.
Why is it not possible to add a VectorCutPlane to the vectors created by mayavi.mlab.quiver3d()? Filters apply a transformation to data, and can be added between sources and modules
Excercice
Different sources: scatters and fields
Using the GUI, add the ExtractVectorNorm filter to display iso-contours of the field magnitude.
Tip: Data comes in different descriptions.
• A 3D block of regularly-spaced value is structured: it is easy to know how one measurement is
related to another neighboring and how to continuously interpolate between these. We can call
20.3. Slicing and dicing data: sources, modules and filters 576 20.3. Slicing and dicing data: sources, modules and filters 577
mlab.pipeline: the scripting layer (continued from previous page)
The mlab scripting layer builds pipelines for you. You can reproduce these pipelines programmati- iso = mlab.contour3d(scalars, transparent=True, contours=[0.5])
cally with the mlab.pipeline interface: each step has a corresponding mlab.pipeline function (sim-
scalars = np.sin(i * x * y * z) /(x * y * z)
ply convert the name of the step to lower-case underscore-separated: ExtractVectorNorm gives ex- iso.mlab_source.scalars = scalars
tract_vector_norm). This function takes as an argument the node that it applies to, as well as optional
parameters, and returns the new node.
For example, iso-contours of the magnitude are coded as: See also:
mlab.pipeline.iso_surface(mlab.pipeline.extract_vector_norm(field), More details in the Mayavi documentation

contours=[0.1*Bmax, 0.4*Bmax],
opacity=0.5)
Event loops
For the interaction with the user (for instance changing the view with the mouse), Mayavi needs some
time to process these events. The for loop above prevents this. The Mayavi documentation details a
workaround
20.5 Making interactive dialogs

It is very simple to make interactive dialogs with Mayavi using the Traits library (see the dedicated
chapter Traits: building interactive dialogs).
20.5.1 A simple dialog

from traits.api import HasTraits, Instance
from traitsui.api import View, Item, HGroup
from mayavi.core.ui.api import SceneEditor, MlabSceneModel
def curve(n_turns):
"The function creating the x, y, z coordinates needed to plot"
phi = np.linspace(0, 2*np.pi, 2000)
return [np.cos(phi) * (1 + 0.5*np.cos(n_turns*phi)),
np.sin(phi) * (1 + 0.5*np.cos(n_turns*phi)),
0.5*np.sin(n_turns*phi)]
Excercice class Visualization(HasTraits):

"The class that contains the dialog"
scene = Instance(MlabSceneModel, ())
Using the mlab.pipeline interface, generate a complete visualization, with iso-contours of the field
magnitude, and a vector cut plane. def __init__(self):
(click on the figure for a solution) HasTraits.__init__(self)
x, y, z = curve(n_turns=2)
# Populating our plot
self.plot = self.scene.mlab.plot3d(x, y, z)
20.4 Animating the data
# Describe the dialog
view = View(Item('scene', height=300, show_label=False,
editor=SceneEditor()),
Tip: To make movies, or interactive application, you may want to change the data represented on a
HGroup('n_turns'), resizable=True)
given visualization.
# Fire up the dialog
If you have built a visualization, using the mlab plotting functions, or the mlab.pipeline function, we Visualization().configure_traits()
can update the data by assigning new values to the mlab_source attributes
x , y , z = np.ogrid[-5:5:100j ,-5:5:100j, -5:5:100j] Tip: Let us read a bit the code above (examples/mlab_dialog.py).
scalars = np.sin(x * y * z) / (x * y * z)
First, the curve function is used to compute the coordinate of the curve we want to plot.
20.4. Animating the data 578 20.5. Making interactive dialogs 579
Second, the dialog is defined by an object inheriting from HasTraits, as it is done with Traits. The (continued from previous page)
important point here is that a Mayavi scene is added as a specific Traits attribute (Instance). This is self.plot.mlab_source.set(x=x, y=y, z=z)
important for embedding it in the dialog. The view of this dialog is defined by the view attribute of the
object. In the init of this object, we populate the 3D scene with a curve. view = View(Item('scene', height=300, show_label=False,
editor=SceneEditor()),
Finally, the configure_traits method creates the dialog and starts the event loop. HGroup('n_turns'), resizable=True)
See also: # Fire up the dialog

Visualization().configure_traits()
There are a few things to be aware of when doing dialogs with Mayavi. Please read the Mayavi docu-
mentation
Tip: Full code of the example: examples/mlab_dialog.py.
20.5.2 Making it interactive
We can combine the Traits events handler with the mlab_source to modify the visualization with the
dialog. 20.6 Putting it together
We will enable the user to vary the n_turns parameter in the definition of the curve. For this, we need:
• to define an n_turns attribute on our visualization object, so that it can appear in the dialog. We Exercise
use a Range type.
Using the code from the magnetic field simulation, create a dialog that enable to move the 2 coils:
• to wire modification of this attribute to a recomputation of the curve. For this, we use the change their parameters.
on_traits_change decorator.
Hint: to define a dialog entry for a vector of dimension 3
direction = Array(float, value=(0, 0, 1), cols=3, shape=(3,))
You can look at the example coil application to see a full-blown application for coil design in 270 lines
of code.
from traits.api import Range, on_trait_change
class Visualization(HasTraits):
n_turns = Range(0, 30, 11)
scene = Instance(MlabSceneModel, ())
def __init__(self):
HasTraits.__init__(self)
x, y, z = curve(self.n_turns)
self.plot = self.scene.mlab.plot3d(x, y, z)
@on_trait_change('n_turns')
def update_plot(self):
x, y, z = curve(self.n_turns)
20.5. Making interactive dialogs 580 20.6. Putting it together 581

Chapters contents
• Introduction: problem settings

• Basic principles of machine learning with scikit-learn
21
• Supervised Learning: Classification of Handwritten Digits
• Supervised Learning: Regression of Housing Data
• Measuring prediction performance
• Unsupervised Learning: Dimensionality Reduction and Visualization
CHAPTER
• The eigenfaces example: chaining PCA and SVMs
• The eigenfaces example: chaining PCA and SVMs
• Parameter selection, Validation, and Testing
• Examples for the scikit-learn chapter
scikit-learn: machine learning in Python 21.1 Introduction: problem settings

21.1.1 What is machine learning?
Tip: Machine Learning is about building programs with tunable parameters that are adjusted
automatically so as to improve their behavior by adapting to previously seen data.
Machine Learning can be considered a subfield of Artificial Intelligence since those algorithms can be
seen as building blocks to make computers learn to behave more intelligently by somehow generalizing
Authors: Gael Varoquaux
rather that just storing and retrieving data items like a database system would do.
Prerequisites
• numpy
• scipy
• matplotlib (optional)
• ipython (the enhancements come handy)
Acknowledgements
This chapter is adapted from a tutorial given by Gaël Varoquaux, Jake Vanderplas, Olivier Grisel. Fig. 1: A classification problem
See also:
We’ll take a look at two very simple machine learning tasks here. The first is a classification task:
Data science in Python the figure shows a collection of two-dimensional data, colored according to two different class labels. A
classification algorithm may be used to draw a dividing boundary between the two clusters of points:
• The Statistics in Python chapter may also be of interest for readers looking into machine learning.
By drawing this separating line, we have learned a model which can generalize to new data: if you were
• The documentation of scikit-learn is very complete and didactic.
to drop another point onto the plane which is unlabeled, this algorithm could now predict whether it’s
582 21.1. Introduction: problem settings 583

a blue or a red point.
Setosa Iris Versicolor Iris Virginica Iris
Quick Question:
If we want to design an algorithm to recognize iris species, what might the

data be?
Remember: we need a 2D array of size [n_samples x n_features].
Fig. 2: A regression problem
• What would the n_samples refer to?
• What might the n_features refer to?
Remember that there must be a fixed number of features for each sample, and feature number i must
The next simple task we’ll look at is a regression task: a simple best-fit line to a set of data. be a similar kind of quantity for each sample.
Again, this is an example of fitting a model to data, but our focus here is that the model can make Loading the Iris Data with Scikit-learn
generalizations about new data. The model has been learned from the training data, and can be used
to predict the result of test data: here, we might be given an x-value, and the model would allow us to Scikit-learn has a very straightforward set of data on these iris species. The data consist of the following:
predict the y value.
• Features in the Iris dataset:
21.1.2 Data in scikit-learn – sepal length (cm)
The data matrix – sepal width (cm)
Machine learning algorithms implemented in scikit-learn expect data to be stored in a two-dimensional – petal length (cm)
array or matrix. The arrays can be either numpy arrays, or in some cases scipy.sparse matrices. The – petal width (cm)
size of the array is expected to be [n_samples, n_features]
• Target classes to predict:
• n_samples: The number of samples: each sample is an item to process (e.g. classify). A sample
can be a document, a picture, a sound, a video, an astronomical object, a row in database or CSV – Setosa
file, or whatever you can describe with a fixed set of quantitative traits. – Versicolour
• n_features: The number of features or distinct traits that can be used to describe each item in – Virginica
a quantitative manner. Features are generally real-valued, but may be boolean or discrete-valued
in some cases. scikit-learn embeds a copy of the iris CSV file along with a function to load it into numpy arrays:
>>> from sklearn.datasets import load_iris

Tip: The number of features must be fixed in advance. However it can be very high dimensional (e.g. >>> iris = load_iris()
millions of features) with most of them being zeros for a given sample. This is a case where scipy.sparse
matrices can be useful, in that they are much more memory-efficient than numpy arrays.
Note: Import sklearn Note that scikit-learn is imported as sklearn
A Simple Example: the Iris Dataset The features of each sample flower are stored in the data attribute of the dataset:
The application problem >>> print(iris.data.shape)
(150, 4)
As an example of a simple dataset, let us a look at the iris data stored by scikit-learn. Suppose we want >>> n_samples, n_features = iris.data.shape
to recognize species of irises. The data consists of measurements of three different species of irises: (continues on next page)
21.1. Introduction: problem settings 584 21.1. Introduction: problem settings 585
(continued from previous page) 21.2 Basic principles of machine learning with scikit-learn
>>> print(n_samples)
150 21.2.1 Introducing the scikit-learn estimator object
>>> print(n_features)
4 Every algorithm is exposed in scikit-learn via an ‘’Estimator” object. For instance a linear regression is:
>>> print(iris.data[0]) sklearn.linear_model.LinearRegression
[5.1 3.5 1.4 0.2]
>>> from sklearn.linear_model import LinearRegression
The information about the class of each sample is stored in the target attribute of the dataset:
Estimator parameters: All the parameters of an estimator can be set when it is instantiated:
>>> print(iris.target.shape)
(150,) >>> model = LinearRegression(n_jobs=1, normalize=True)
>>> print(iris.target) >>> print(model.normalize)
[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 True
0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 >>> print(model)
1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 2 LinearRegression(n_jobs=1, normalize=True)
2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2
2 2]
Fitting on data
The names of the classes are stored in the last attribute, namely target_names: Let’s create some simple data with numpy:
>>> print(iris.target_names) >>> import numpy as np
['setosa' 'versicolor' 'virginica'] >>> x = np.array([0, 1, 2])
>>> y = np.array([0, 1, 2])
This data is four-dimensional, but we can visualize two of the dimensions at a time using a scatter plot:
>>> X = x[:, np.newaxis] # The input data for sklearn is 2D: (samples == 3 x features == 1)
>>> X
array([[0],
[1],
[2]])
>>> model.fit(X, y)
LinearRegression(n_jobs=1, normalize=True)
Estimated parameters: When data is fitted with an estimator, parameters are estimated from the data
at hand. All the estimated parameters are attributes of the estimator object ending by an underscore:
>>> model.coef_
array([1.])
21.2.2 Supervised Learning: Classification and regression

In Supervised Learning, we have a dataset consisting of both features and labels. The task is to
construct an estimator which is able to predict the label of an object given the set of features. A
relatively simple example is predicting the species of iris given a set of measurements of its flower. This
is a relatively simple task. Some more complicated examples are:
• given a multicolor image of an object through a telescope, determine whether that object is a star,
a quasar, or a galaxy.
• given a photograph of a person, identify the person in the photo.
• given a list of movies a person has watched and their personal rating of the movie, recommend a
Exercise: list of movies they would like (So-called recommender systems: a famous example is the Netflix
Prize).
Can you choose 2 features to find a plot where it is easier to seperate the different classes of irises?
Hint: click on the figure above to see the code that generates it, and modify this code. Tip: What these tasks have in common is that there is one or more unknown quantities associated
with the object which needs to be determined from other observed quantities.
Supervised learning is further broken down into two categories, classification and regression. In
classification, the label is discrete, while in regression, the label is continuous. For example, in astronomy,
21.1. Introduction: problem settings 586 21.2. Basic principles of machine learning with scikit-learn 587
the task of determining whether an object is a star, a galaxy, or a quasar is a classification problem: the (continued from previous page)
label is from three distinct categories. On the other hand, we might wish to estimate the age of an object model = LinearRegression()
based on such observations: this would be a regression problem, because the label (age) is a continuous model.fit(x, y)
quantity.
# predict y from the data
Classification: K nearest neighbors (kNN) is one of the simplest learning strategies: given a new, x_new = np.linspace(0, 30, 100)
unknown observation, look up in your reference database which ones have the closest features and assign y_new = model.predict(x_new[:, np.newaxis])
the predominant class. Let’s try it out on our iris classification problem:
from sklearn import neighbors, datasets
iris = datasets.load_iris()
X, y = iris.data, iris.target
knn = neighbors.KNeighborsClassifier(n_neighbors=1)
knn.fit(X, y)
# What kind of iris has 3cm x 5cm sepal and 4cm x 2cm petal?
print(iris.target_names[knn.predict([[3, 5, 4, 2]])])
Fig. 4: A plot of a simple linear regression.
21.2.3 A recap on Scikit-learn’s estimator interface

Scikit-learn strives to have a uniform interface across all methods, and we’ll see examples of these below.
Given a scikit-learn estimator object named model, the following methods are available:
In all Estimators
• model.fit() : fit training data. For supervised learning applications, this accepts
two arguments: the data X and the labels y (e.g. model.fit(X, y)). For unsu-
pervised learning applications, this accepts only a single argument, the data X (e.g.
model.fit(X)).
In supervised estimators
• model.predict() : given a trained model, predict the label of a new set of
data. This method accepts one argument, the new data X_new (e.g. model.
predict(X_new)), and returns the learned label for each object in the array.
Fig. 3: A plot of the sepal space and the prediction of the KNN
• model.predict_proba() : For classification problems, some estimators also pro-
Regression: The simplest possible regression setting is the linear regression one: vide this method, which returns the probability that a new observation has each
categorical label. In this case, the label with the highest probability is returned by
from sklearn.linear_model import LinearRegression model.predict().
# x from 0 to 30 • model.score() : for classification or regression problems, most (all?) estimators

x = 30 * np.random.random((20, 1)) implement a score method. Scores are between 0 and 1, with a larger score indicating
a better fit.
# y = a*x + b with noise
In unsupervised estimators
y = 0.5 * x + 1.0 + np.random.normal(size=x.shape)
• model.transform() : given an unsupervised model, transform new data into the
# create a linear regression model new basis. This also accepts one argument X_new, and returns the new representa-
(continues on next page) tion of the data based on the unsupervised model.
21.2. Basic principles of machine learning with scikit-learn 588 21.2. Basic principles of machine learning with scikit-learn 589
• model.fit_transform() : some estimators implement this method, which more

efficiently performs a fit and a transform on the same input data.
21.2.4 Regularization: what it is and why it is necessary

Prefering simpler models
Train errors Suppose you are using a 1-nearest neighbor estimator. How many errors do you expect
on your train set?
• Train set error is not a good measurement of prediction performance. You need to leave out a test
set.
• In general, we should accept errors on the train set.
An example of regularization The core idea behind regularization is that we are going to prefer
models that are simpler, for a certain definition of ‘’simpler”, even if they lead to more errors on the
train set.
As an example, let’s generate with a 9th order polynomial, with noise:
And now, let’s fit a 4th order and a 9th order polynomial to the data.
With your naked eyes, which model do you prefer, the 4th order one, or the 9th order one?
Let’s look at the ground truth:
Tip: Regularization is ubiquitous in machine learning. Most scikit-learn estimators have a parameter
to tune the amount of regularization. For instance, with k-NN, it is ‘k’, the number of nearest neighbors
used to make the decision. k=1 amounts to no regularization: 0 error on the training set, whereas large
k will push toward smoother decision boundaries in the feature space.
21.2. Basic principles of machine learning with scikit-learn 590 21.2. Basic principles of machine learning with scikit-learn 591
Simple versus complex models for classification
A linear separation A non-linear separation
Tip: For classification models, the decision boundary, that separates the class expresses the complexity
of the model. For instance, a linear model, that makes a decision based on a linear combination of
features, is more complex than a non-linear one.
21.3 Supervised Learning: Classification of Handwritten Digits

21.3.1 The nature of the data
Code and notebook
Python code and Jupyter notebook for this section are found here Let
us visualize the data and remind us what we’re looking at (click on the figure for the full code):
In this section we’ll apply scikit-learn to the classification of handwritten digits. This will go a bit beyond # plot the digits: each image is 8x8 pixels
the iris classification we saw before: we’ll discuss some of the metrics which can be used in evaluating for i in range(64):
the effectiveness of a classification model. ax = fig.add_subplot(8, 8, i + 1, xticks=[], yticks=[])
ax.imshow(digits.images[i], cmap=plt.cm.binary, interpolation='nearest')
>>> from sklearn.datasets import load_digits
>>> digits = load_digits()
21.3.2 Visualizing the Data on its principal components
A good first-step for many problems is to visualize the data using a Dimensionality Reduction technique.
We’ll start with the most straightforward one, Principal Component Analysis (PCA).
PCA seeks orthogonal linear combinations of the features which show the greatest variance, and as such,
can help give you a good idea of the structure of the data set.
>>> from sklearn.decomposition import PCA

>>> pca = PCA(n_components=2)
>>> proj = pca.fit_transform(digits.data)
>>> plt.scatter(proj[:, 0], proj[:, 1], c=digits.target)
<matplotlib.collections.PathCollection object at ...>
>>> plt.colorbar()
<matplotlib.colorbar.Colorbar object at ...>
21.3. Supervised Learning: Classification of Handwritten Digits 592 21.3. Supervised Learning: Classification of Handwritten Digits 593
>>> # split the data into training and validation sets

>>> X_train, X_test, y_train, y_test = train_test_split(digits.data, digits.target)
>>> # train the model

>>> clf = GaussianNB()
>>> clf.fit(X_train, y_train)
GaussianNB()
>>> # use the model to predict the labels of the test data
>>> predicted = clf.predict(X_test)
>>> expected = y_test
>>> print(predicted)
[1 7 7 7 8 2 8 0 4 8 7 7 0 8 2 3 5 8 5 3 7 9 6 2 8 2 2 7 3 5...]
>>> print(expected)
[1 0 4 7 8 2 2 0 4 3 7 7 0 8 2 3 4 8 5 3 7 9 6 3 8 2 2 9 3 5...]
As above, we plot the digits with the predicted labels to get an idea of how well the classification is work-
Question
Given these projections of the data, which numbers do you think a classifier might have trouble
distinguishing?
21.3.3 Gaussian Naive Bayes Classification

For most classification problems, it’s nice to have a simple, fast method to provide a quick baseline
classification. If the simple and fast method is sufficient, then we don’t have to waste CPU cycles on
more complex models. If not, we can use the results of the simple method to give us clues about our
data.
One good method to keep in mind is Gaussian Naive Bayes (sklearn.naive_bayes.GaussianNB).
Old scikit-learn versions
train_test_split() is imported from sklearn.cross_validation
Tip: Gaussian Naive Bayes fits a Gaussian distribution to each training label independantly on each
feature, and uses this to quickly give a rough classification. It is generally not sufficiently accurate for
real-world data, but can perform surprisingly well, for instance on text data.
ing.
>>> from sklearn.naive_bayes import GaussianNB
>>> from sklearn.model_selection import train_test_split
21.3. Supervised Learning: Classification of Handwritten Digits 594 21.3. Supervised Learning: Classification of Handwritten Digits 595
Question 21.4 Supervised Learning: Regression of Housing Data

Why did we split the data into training and validation sets? Here we’ll do a short example of a regression problem: learning a continuous value from a set of features.
21.4.1 A quick look at the data

21.3.4 Quantitative Measurement of Performance
We’d like to measure the performance of our estimator without having to resort to plotting examples. A Code and notebook
simple method might be to simply compare the number of matches:
Python code and Jupyter notebook for this section are found here
>>> matches = (predicted == expected)
>>> print(matches.sum())
367 We’ll use the California house prices set, available in scikit-learn. This records measurements of 8
>>> print(len(matches)) attributes of housing markets in California, as well as the median price. The question is: can you predict
450 the price of a new market given its attributes?:
>>> matches.sum() / float(len(matches))
0.81555555555555559 >>> from sklearn.datasets import fetch_california_housing
>>> data = fetch_california_housing(as_frame=True)
We see that more than 80% of the 450 predictions match the input. But there are other more sophisticated >>> print(data.data.shape)
metrics that can be used to judge the performance of a classifier: several are available in the sklearn. (20640, 8)
>>> print(data.target.shape)
metrics submodule.
(20640,)
One of the most useful metrics is the classification_report, which combines several measures and
prints a table with the results: We can see that there are just over 20000 data points.
>>> from sklearn import metrics The DESCR variable has a long description of the dataset:
>>> print(metrics.classification_report(expected, predicted))
precision recall f1-score support >>> print(data.DESCR)
.. _california_housing_dataset:
0 1.00 0.91 0.95 46
1 0.76 0.64 0.69 44 California Housing dataset
2 0.85 0.62 0.72 47 --------------------------
3 0.98 0.82 0.89 49
4 0.89 0.86 0.88 37 **Data Set Characteristics:**
5 0.97 0.93 0.95 41
6 1.00 0.98 0.99 44 :Number of Instances: 20640
7 0.73 1.00 0.84 45
8 0.50 0.90 0.64 49 :Number of Attributes: 8 numeric, predictive attributes and the target
9 0.93 0.54 0.68 48
:Attribute Information:
accuracy 0.82 450 - MedInc median income in block
macro avg 0.86 0.82 0.82 450 - HouseAge median house age in block
weighted avg 0.86 0.82 0.82 450 - AveRooms average number of rooms
- AveBedrms average number of bedrooms
- Population block population
Another enlightening metric for this sort of multi-label classification is a confusion matrix: it helps us
- AveOccup average house occupancy
visualize which labels are being interchanged in the classification errors: - Latitude house block latitude
- Longitude house block longitude
>>> print(metrics.confusion_matrix(expected, predicted))
[[42 0 0 0 3 0 0 1 0 0]
:Missing Attribute Values: None
[ 0 28 0 0 0 0 0 1 13 2]
[ 0 3 29 0 0 0 0 0 15 0]
This dataset was obtained from the StatLib repository.
[ 0 0 2 40 0 0 0 2 5 0]
http://lib.stat.cmu.edu/datasets/
[ 0 0 1 0 32 1 0 3 0 0]
[ 0 0 0 0 0 38 0 2 1 0]
The target variable is the median house value for California districts.
[ 0 0 1 0 0 0 43 0 0 0]
[ 0 0 0 0 0 0 0 45 0 0]
This dataset was derived from the 1990 U.S. census, using one row per census
[ 0 3 1 0 0 0 0 1 44 0]
block group. A block group is the smallest geographical unit for which the U.S.
[ 0 3 0 1 1 0 0 7 10 26]]
Census Bureau publishes sample data (a block group typically has a population
of 600 to 3,000 people).
We see here that in particular, the numbers 1, 2, 3, and 9 are often being labeled 8.
It can be downloaded/loaded using the
21.3. Supervised Learning: Classification of Handwritten Digits 596 21.4. Supervised Learning: Regression of Housing Data 597

:func:`sklearn.datasets.fetch_california_housing` function. >>> clf = LinearRegression()
>>> clf.fit(X_train, y_train)
.. topic:: References LinearRegression()
>>> predicted = clf.predict(X_test)
- Pace, R. Kelley and Ronald Barry, Sparse Spatial Autoregressions, >>> expected = y_test
Statistics and Probability Letters, 33 (1997) 291-297 >>> print("RMS: %s " % np.sqrt(np.mean((predicted - expected) ** 2)))
RMS: 0.7...
It often helps to quickly visualize pieces of the data using histograms, scatter plots, or other plot types.
With matplotlib, let us show a histogram of the target values: the median price in each neighborhood:
>>> plt.hist(data.target)
(array([...
Let’s have a quick look to see if some features are more

relevant than others for our problem: We can plot the error: expected as a
function of predicted:
>>> for index, feature_name in enumerate(data.feature_names):
... plt.figure() >>> plt.scatter(expected, predicted)
... plt.scatter(data.data[feature_name], data.target) <matplotlib.collections.PathCollection object at ...>
<Figure size...
Tip: The prediction at least correlates with the true price, though there are clearly some biases. We
could imagine evaluating the performance of the regressor by, say, computing the RMS residuals between
the true and predicted price. There are some subtleties in this, however, which we’ll cover in a later
section.
Exercise: Gradient Boosting Tree Regression
There are many other types of regressors available in scikit-learn: we’ll try a more powerful one here.
This is a manual version of a technique called feature selection.
Use the GradientBoostingRegressor class to fit the housing data.
Tip: Sometimes, in Machine Learning it is useful to use feature selection to decide which features are hint You can copy and paste some of the above code, replacing LinearRegression with
the most useful for a particular problem. Automated methods exist which quantify this sort of exercise GradientBoostingRegressor:
of choosing the most informative features. from sklearn.ensemble import GradientBoostingRegressor
# Instantiate the model, fit the results, and scatter in vs. out
21.4.2 Predicting Home Prices: a Simple Linear Regression Solution The solution is found in the code of this chapter
Now we’ll use scikit-learn to perform a simple linear regression on the housing data. There are many
possibilities of regressors to use. A particularly simple one is LinearRegression: this is basically a
wrapper around an ordinary least squares calculation. 21.5 Measuring prediction performance
>>> from sklearn.model_selection import train_test_split
21.5.1 A quick test on the K-neighbors classifier
>>> X_train, X_test, y_train, y_test = train_test_split(data.data, data.target)
>>> from sklearn.linear_model import LinearRegression Here we’ll continue to look at the digits data, but we’ll switch to the K-Neighbors classifier. The K-
(continues on next page) neighbors classifier is an instance-based classifier. The K-neighbors classifier predicts the label of an
21.4. Supervised Learning: Regression of Housing Data 598 21.5. Measuring prediction performance 599
unknown point based on the labels of the K nearest points in the parameter space.
>>> # Get the data

>>> X = digits.data
>>> y = digits.target
>>> # Instantiate and train the classifier

>>> from sklearn.neighbors import KNeighborsClassifier
>>> clf = KNeighborsClassifier(n_neighbors=1)
>>> clf.fit(X, y)
KNeighborsClassifier(...)
>>> # Check the results using metrics

>>> from sklearn import metrics
>>> y_pred = clf.predict(X)
>>> print(metrics.confusion_matrix(y_pred, y))

[[178 0 0 0 0 0 0 0 0 0]
[ 0 182 0 0 0 0 0 0 0 0] Here again the predictions are seemingly perfect as the model was able to perfectly memorize the training
[ 0 0 177 0 0 0 0 0 0 0] set.
[ 0 0 0 183 0 0 0 0 0 0]
[ 0 0 0 0 181 0 0 0 0 0]
[ 0 0 0 0 0 182 0 0 0 0] Warning: Performance on test set
[ 0 0 0 0 0 0 181 0 0 0]
[ 0 0 0 0 0 0 0 179 0 0] Performance on test set does not measure overfit (as described above)
[ 0 0 0 0 0 0 0 0 174 0]
[ 0 0 0 0 0 0 0 0 0 180]]
21.5.2 A correct approach: Using a validation set
Apparently, we’ve found a perfect classifier! But this is misleading for the reasons we saw before: the
Learning the parameters of a prediction function and testing it on the same data is a methodological
classifier essentially “memorizes” all the samples it has already seen. To really test how well this algorithm
mistake: a model that would just repeat the labels of the samples that it has just seen would have a
does, we need to try some samples it hasn’t yet seen.
perfect score but would fail to predict anything useful on yet-unseen data.
This problem also occurs with regression models. In the following we fit an other instance-based model
To avoid over-fitting, we have to define two different sets:
named “decision tree” to the California Housing price dataset we introduced previously:
• a training set X_train, y_train which is used for learning the parameters of a predictive model
>>> from sklearn.datasets import fetch_california_housing
>>> from sklearn.tree import DecisionTreeRegressor • a testing set X_test, y_test which is used for evaluating the fitted predictive model
In scikit-learn such a random split can be quickly computed with the train_test_split() function:
>>> data = fetch_california_housing(as_frame=True)
>>> clf = DecisionTreeRegressor().fit(data.data, data.target) >>> from sklearn import model_selection
>>> predicted = clf.predict(data.data) >>> X = digits.data
>>> expected = data.target >>> y = digits.target
>>> plt.scatter(expected, predicted) >>> X_train, X_test, y_train, y_test = model_selection.train_test_split(X, y,

<matplotlib.collections.PathCollection object at ...> ... test_size=0.25, random_state=0)
>>> plt.plot([0, 50], [0, 50], '--k')
[<matplotlib.lines.Line2D object at ...] >>> print("%r , %r , %r " % (X.shape, X_train.shape, X_test.shape))
(1797, 64), (1347, 64), (450, 64)
Now we train on the training data, and test on the testing data:
>>> clf = KNeighborsClassifier(n_neighbors=1).fit(X_train, y_train)

>>> y_pred = clf.predict(X_test)
>>> print(metrics.confusion_matrix(y_test, y_pred))

[[37 0 0 0 0 0 0 0 0 0]
[ 0 43 0 0 0 0 0 0 0 0]
[ 0 0 43 1 0 0 0 0 0 0]
[ 0 0 0 45 0 0 0 0 0 0]
[ 0 0 0 0 38 0 0 0 0 0]
21.5. Measuring prediction performance 600 21.5. Measuring prediction performance 601

[ 0 0 0 0 0 47 0 0 0 1]
[ 0 0 0 0 0 0 52 0 0 0] >>> for Model in [GaussianNB, KNeighborsClassifier, LinearSVC]:
[ 0 0 0 0 0 0 0 48 0 0] ... clf = Model().fit(X_train, y_train)
[ 0 0 0 0 0 0 0 0 48 0] ... y_pred = clf.predict(X_test)
[ 0 0 0 1 0 1 0 0 0 45]] ... print('%s : %s ' %
>>> print(metrics.classification_report(y_test, y_pred)) ... (Model.__name__, metrics.f1_score(y_test, y_pred, average="macro")))
precision recall f1-score support GaussianNB: 0.8...
KNeighborsClassifier: 0.9...
0 1.00 1.00 1.00 37 LinearSVC: 0.9...
1 1.00 1.00 1.00 43
2 1.00 0.98 0.99 44 • For each classifier, which value for the hyperparameters gives the best results for the digits data?
3 0.96 1.00 0.98 45 For LinearSVC, use loss='l2' and loss='l1'. For KNeighborsClassifier we use n_neighbors
4 1.00 1.00 1.00 38
between 1 and 10. Note that GaussianNB does not have any adjustable hyperparameters.
5 0.98 0.98 0.98 48
6 1.00 1.00 1.00 52 LinearSVC(loss='l1'): 0.930570687535
7 1.00 1.00 1.00 48 LinearSVC(loss='l2'): 0.933068826918
8 1.00 1.00 1.00 48 -------------------
9 0.98 0.96 0.97 47 KNeighbors(n_neighbors=1): 0.991367521884
KNeighbors(n_neighbors=2): 0.984844206884
accuracy 0.99 450 KNeighbors(n_neighbors=3): 0.986775344954
macro avg 0.99 0.99 0.99 450 KNeighbors(n_neighbors=4): 0.980371905382
weighted avg 0.99 0.99 0.99 450 KNeighbors(n_neighbors=5): 0.980456280495
The averaged f1-score is often used as a convenient measure of the overall performance of an algorithm. KNeighbors(n_neighbors=7): 0.978064579214
It appears in the bottom row of the classification report; it can also be accessed directly: KNeighbors(n_neighbors=8): 0.978064579214
>>> metrics.f1_score(y_test, y_pred, average="macro") KNeighbors(n_neighbors=10): 0.975555089773
0.991367...
Solution: code source
The over-fitting we saw previously can be quantified by computing the f1-score on the training data
itself:
21.5.4 Cross-validation
>>> metrics.f1_score(y_train, clf.predict(X_train), average="macro") Cross-validation consists in repetively splitting the data in pairs of train and test sets, called ‘folds’.
1.0
Scikit-learn comes with a function to automatically compute score on all these folds. Here we do KFold
with k=5.
Note: Regression metrics In the case of regression models, we need to use different metrics, such as >>> clf = KNeighborsClassifier()
explained variance. >>> from sklearn.model_selection import cross_val_score
>>> cross_val_score(clf, X, y, cv=5)
array([0.947..., 0.955..., 0.966..., 0.980..., 0.963... ])
21.5.3 Model Selection via Validation

We can use different splitting strategies, such as random splitting:
Tip: We have applied Gaussian Naives, support vectors machines, and K-nearest neighbors classifiers >>> from sklearn.model_selection import ShuffleSplit
>>> cv = ShuffleSplit(n_splits=5)
to the digits dataset. Now that we have these validation tools in place, we can ask quantitatively which
>>> cross_val_score(clf, X, y, cv=cv)
of the three estimators works best for this dataset.
array([...])
• With the default hyper-parameters for each estimator, which gives the best f1 score on the valida-
tion set? Recall that hyperparameters are the parameters set when you instantiate the classifier: Tip: There exists many different cross-validation strategies in scikit-learn. They are often useful to
for example, the n_neighbors in clf = KNeighborsClassifier(n_neighbors=1) take in account non iid datasets.
>>> from sklearn.naive_bayes import GaussianNB

>>> from sklearn.neighbors import KNeighborsClassifier
>>> from sklearn.svm import LinearSVC 21.5.5 Hyperparameter optimization with cross-validation
Consider regularized linear models, such as Ridge Regression, which uses l2 regularlization, and Lasso
>>> X = digits.data
>>> y = digits.target
Regression, which uses l1 regularization. Choosing their regularization parameter is important.
>>> X_train, X_test, y_train, y_test = model_selection.train_test_split(X, y, Let us set these parameters on the Diabetes dataset, a simple regression problem. The diabetes data
... test_size=0.25, random_state=0) consists of 10 physiological variables (age, sex, weight, blood pressure) measure on 442 patients, and an
(continues on next page) indication of disease progression after one year:
21.5. Measuring prediction performance 602 21.5. Measuring prediction performance 603
Built-in Hyperparameter Search

>>> from sklearn.datasets import load_diabetes
>>> data = load_diabetes() For some models within scikit-learn, cross-validation can be performed more efficiently on large datasets.
>>> X, y = data.data, data.target In this case, a cross-validated version of the particular model is included. The cross-validated versions
>>> print(X.shape) of Ridge and Lasso are RidgeCV and LassoCV, respectively. Parameter search on these estimators can
(442, 10)
be performed as follows:
With the default hyper-parameters: we compute the cross-validation score: >>> from sklearn.linear_model import RidgeCV, LassoCV
>>> for Model in [RidgeCV, LassoCV]:
>>> from sklearn.linear_model import Ridge, Lasso ... model = Model(alphas=alphas, cv=3).fit(X, y)
... print('%s : %s ' % (Model.__name__, model.alpha_))
>>> for Model in [Ridge, Lasso]: RidgeCV: 0.0621016941892
... model = Model() LassoCV: 0.0126896100317
... print('%s : %s ' % (Model.__name__, cross_val_score(model, X, y).mean()))
Ridge: 0.4...
We see that the results match those returned by GridSearchCV
Lasso: 0.3...
Nested cross-validation
Basic Hyperparameter Optimization
How do we measure the performance of these estimators? We have used data to set the hyperparameters,
We compute the cross-validation score as a function of alpha, the strength of the regularization for Lasso so we need to test on actually new data. We can do this by running cross_val_score() on our CV
and Ridge. We choose 20 values of alpha between 0.0001 and 1: objects. Here there are 2 cross-validation loops going on, this is called ‘nested cross validation’:
>>> alphas = np.logspace(-3, -1, 30) for Model in [RidgeCV, LassoCV]:

scores = cross_val_score(Model(alphas=alphas, cv=3), X, y, cv=3)
>>> for Model in [Lasso, Ridge]: print(Model.__name__, np.mean(scores))
... scores = [cross_val_score(Model(alpha), X, y, cv=3).mean()
... for alpha in alphas]
... plt.plot(alphas, scores, label=Model.__name__) Note: Note that these results do not match the best results of our curves above, and LassoCV seems
[<matplotlib.lines.Line2D object at ... to under-perform RidgeCV. The reason is that setting the hyper-parameter is harder for Lasso, thus the
estimation error on this hyper-parameter is larger.
21.6 Unsupervised Learning: Dimensionality Reduction and Visual-

ization
Unsupervised learning is applied on X without y: data without labels. A typical use case is to find
hidden structure in the data.
21.6.1 Dimensionality Reduction: PCA

Dimensionality reduction derives a set of new artificial features smaller than the original feature set.
Here we’ll use Principal Component Analysis (PCA), a dimensionality reduction that strives to retain
most of the variance of the original data. We’ll use sklearn.decomposition.PCA on the iris dataset:
Question
>>> X = iris.data
Can we trust our results to be actually useful? >>> y = iris.target
Automatically Performing Grid Search Tip: PCA computes linear combinations of the original features using a truncated Singular Value
Decomposition of the matrix X, to project the data onto a base of the top singular vectors.
sklearn.grid_search.GridSearchCV is constructed with an estimator, as well as a dictionary of pa-
rameter values to be searched. We can find the optimal parameters this way:
>>> from sklearn.decomposition import PCA
>>> from sklearn.model_selection import GridSearchCV >>> pca = PCA(n_components=2, whiten=True)
>>> for Model in [Ridge, Lasso]: >>> pca.fit(X)
... gscv = GridSearchCV(Model(), dict(alpha=alphas), cv=3).fit(X, y) PCA(n_components=2, ...)
... print('%s : %s ' % (Model.__name__, gscv.best_params_))
Ridge: {'alpha': 0.062101694189156162} Once fitted, PCA exposes the singular vectors in the components_ attribute:
Lasso: {'alpha': 0.01268961003167922}
21.5. Measuring prediction performance 604 21.6. Unsupervised Learning: Dimensionality Reduction and Visualization 605
by the colors): this is the sense in which the learning is unsupervised. Nevertheless, we see that the
>>> pca.components_
array([[ 0.3..., -0.08..., 0.85..., 0.3...], projection gives us insight into the distribution of the different flowers in parameter space: notably, iris
[ 0.6..., 0.7..., -0.1..., -0.07...]]) setosa is much more distinct than the other two species.
Other attributes are available as well:

21.6.2 Visualization with a non-linear embedding: tSNE
>>> pca.explained_variance_ratio_
array([0.92..., 0.053...]) For visualization, more complex embeddings can be useful (for statistical analysis, they are harder to
control). sklearn.manifold.TSNE is such a powerful manifold learning method. We apply it to the digits
Let us project the iris dataset along those first two dimensions:: dataset, as the digits are vectors of dimension 8*8 = 64. Embedding them in 2D enables visualization:
>>> X_pca = pca.transform(X) >>> # Take the first 500 data points: it's hard to see 1500 points
>>> X_pca.shape >>> X = digits.data[:500]
(150, 2) >>> y = digits.target[:500]
PCA normalizes and whitens the data, which means that the data is now centered on both components >>> # Fit and transform with a TSNE
with unit variance: >>> from sklearn.manifold import TSNE
>>> tsne = TSNE(n_components=2, random_state=0)
>>> X_pca.mean(axis=0) >>> X_2d = tsne.fit_transform(X)
array([...e-15, ...e-15])
>>> X_pca.std(axis=0, ddof=1) >>> # Visualize the data
array([1., 1.]) >>> plt.scatter(X_2d[:, 0], X_2d[:, 1], c=y)
<matplotlib.collections.PathCollection object at ...>
Furthermore, the samples components do no longer carry any linear correlation:
>>> np.corrcoef(X_pca.T)
array([[1.00000000e+00, 0.0],
[0.0, 1.00000000e+00]])
With a number of retained components 2 or 3, PCA is useful to visualize the dataset:

>>> target_ids = range(len(iris.target_names))
>>> for i, c, label in zip(target_ids, 'rgbcmykw', iris.target_names):
... plt.scatter(X_pca[y == i, 0], X_pca[y == i, 1],
... c=c, label=label)
<matplotlib.collections.PathCollection ...
fit_transform
As TSNE cannot be applied to new data, we need to use its fit_transform method.
sklearn.manifold.TSNE separates quite well the different classes of digits eventhough it had no access
to the class information.
Tip: Note that this projection was determined without any information about the labels (represented
21.6. Unsupervised Learning: Dimensionality Reduction and Visualization 606 21.6. Unsupervised Learning: Dimensionality Reduction and Visualization 607
Exercise: Other dimension reduction of digits
sklearn.manifold has many other non-linear embeddings. Try them out on the digits dataset. Could
you judge their quality without knowing the labels y?
>>> # ...
21.7 The eigenfaces example: chaining PCA and SVMs
Code and notebook
21.8 The eigenfaces example: chaining PCA and SVMs

The goal of this example is to show how an unsupervised method and a supervised one can be chained
for better prediction. It starts with a didactic but lengthy way of doing things, and finishes with the
idiomatic approach to pipelining in scikit-learn.
Here we’ll take a look at a simple facial recognition example. Ideally, we would use a dataset con-
sisting of a subset of the Labeled Faces in the Wild data that is available with sklearn.datasets.
fetch_lfw_people(). However, this is a relatively large download (~200MB) so we will do the tutorial
on a simpler, less rich dataset. Feel free to explore the LFW dataset.
from sklearn import datasets

faces = datasets.fetch_olivetti_faces() Tip: Note is that these faces have already been localized and scaled to a common size. This is an
faces.data.shape important preprocessing piece for facial recognition, and is a process that can require a large collection
of training data. This can be done in scikit-learn, but the challenge is gathering a sufficient amount of
Let’s visualize these faces to see what we’re working with training data for the algorithm to work. Fortunately, this piece is common enough that it has been done.
One good resource is OpenCV, the Open Computer Vision Library.
# plot several images We’ll perform a Support Vector classification of the images. We’ll do a typical train-test split on the
for i in range(15): images:
ax = fig.add_subplot(3, 5, i + 1, xticks=[], yticks=[])
ax.imshow(faces.images[i], cmap=plt.cm.bone) from sklearn.model_selection import train_test_split
X_train, X_test, y_train, y_test = train_test_split(faces.data,
faces.target, random_state=0)
print(X_train.shape, X_test.shape)
Out:
(300, 4096) (100, 4096)
21.8.1 Preprocessing: Principal Component Analysis

1850 dimensions is a lot for SVM. We can use PCA to reduce these 1850 features to a manageable size,
while maintaining most of the information in the dataset.
from sklearn import decomposition

pca = decomposition.PCA(n_components=150, whiten=True)
pca.fit(X_train)
One interesting part of PCA is that it computes the “mean” face, which can be interesting to examine:
21.7. The eigenfaces example: chaining PCA and SVMs 608 21.8. The eigenfaces example: chaining PCA and SVMs 609
plt.imshow(pca.mean_.reshape(faces.images[0].shape),
cmap=plt.cm.bone)
The components (“eigenfaces”) are ordered by their importance from top-left to bottom-right. We
see that the first few components seem to primarily take care of lighting conditions; the remaining
components pull out certain identifying features: the nose, eyes, eyebrows, etc.
With this projection computed, we can now project our original training and test data onto the PCA
basis:
X_train_pca = pca.transform(X_train)
X_test_pca = pca.transform(X_test)
print(X_train_pca.shape)
Out:
(300, 150)
print(X_test_pca.shape)
Out:
The principal components measure deviations about this mean along orthogonal axes. (100, 150)
print(pca.components_.shape)
These projected components correspond to factors in a linear combination of component images such
that the combination approaches the original face.
Out:
(150, 4096) 21.8.2 Doing the Learning: Support Vector Machines

It is also interesting to visualize these principal components: Now we’ll perform support-vector-machine classification on this reduced dataset:
fig = plt.figure(figsize=(16, 6)) from sklearn import svm

for i in range(30): clf = svm.SVC(C=5., gamma=0.001)
ax = fig.add_subplot(3, 10, i + 1, xticks=[], yticks=[]) clf.fit(X_train_pca, y_train)
ax.imshow(pca.components_[i].reshape(faces.images[0].shape),
cmap=plt.cm.bone) Finally, we can evaluate how well this classification did. First, we might plot a few of the test-cases with
the labels learned from the training set:
import numpy as np
for i in range(15):
ax.imshow(X_test[i].reshape(faces.images[0].shape),
cmap=plt.cm.bone)
y_pred = clf.predict(X_test_pca[i, np.newaxis])[0]
color = ('black' if y_pred == y_test[i] else 'red')
ax.set_title(y_pred, fontsize='small', color=color)

17 1.00 1.00 1.00 6
19 1.00 1.00 1.00 4
20 1.00 1.00 1.00 1
21 1.00 1.00 1.00 1
22 1.00 1.00 1.00 2
23 1.00 1.00 1.00 1
24 1.00 1.00 1.00 2
25 1.00 0.50 0.67 2
26 1.00 0.75 0.86 4
27 1.00 1.00 1.00 1
28 0.67 1.00 0.80 2
29 1.00 1.00 1.00 3
30 1.00 1.00 1.00 4
31 1.00 1.00 1.00 3
32 1.00 1.00 1.00 3
33 1.00 1.00 1.00 2
34 1.00 1.00 1.00 3
35 1.00 1.00 1.00 1
36 1.00 1.00 1.00 3
37 1.00 1.00 1.00 3
38 1.00 1.00 1.00 1
39 1.00 1.00 1.00 3
accuracy 0.94 100

macro avg 0.95 0.96 0.94 100
weighted avg 0.97 0.94 0.94 100
Another interesting metric is the confusion matrix, which indicates how often any two items are mixed-
up. The confusion matrix of a perfect classifier would only have nonzero entries on the diagonal, with
zeros on the off-diagonal:
The classifier is correct on an impressive number of images given the simplicity of its learning model!
Using a linear classifier on 150 features derived from the pixel-level data, the algorithm correctly identifies print(metrics.confusion_matrix(y_test, y_pred))
a large number of the people in the images.
Out:
Again, we can quantify this effectiveness using one of several measures from sklearn.metrics. First we
can do the classification report, which shows the precision, recall and other measures of the “goodness” [[3 0 0 ... 0 0 0]
of the classification: [0 4 0 ... 0 0 0]
[0 0 2 ... 0 0 0]
from sklearn import metrics ...
y_pred = clf.predict(X_test_pca) [0 0 0 ... 3 0 0]
print(metrics.classification_report(y_test, y_pred)) [0 0 0 ... 0 1 0]
[0 0 0 ... 0 0 3]]
Out:
precision recall f1-score support 21.8.3 Pipelining
0 1.00 0.50 0.67 6 Above we used PCA as a pre-processing step before applying our support vector machine classifier.
1 1.00 1.00 1.00 4 Plugging the output of one estimator directly into the input of a second estimator is a commonly used
2 0.50 1.00 0.67 2 pattern; for this reason scikit-learn provides a Pipeline object which automates this process. The above
3 1.00 1.00 1.00 1 problem can be re-expressed as a pipeline as follows:
4 0.33 1.00 0.50 1
5 1.00 1.00 1.00 5 from sklearn.pipeline import Pipeline
6 1.00 1.00 1.00 4 clf = Pipeline([('pca', decomposition.PCA(n_components=150, whiten=True)),
7 1.00 0.67 0.80 3 ('svm', svm.LinearSVC(C=1.0))])
9 1.00 1.00 1.00 1
10 1.00 1.00 1.00 4 clf.fit(X_train, y_train)
11 1.00 1.00 1.00 1
12 0.67 1.00 0.80 2 y_pred = clf.predict(X_test)
13 1.00 1.00 1.00 3 print(metrics.confusion_matrix(y_pred, y_test))
14 1.00 1.00 1.00 5
15 1.00 1.00 1.00 3
21.9 Parameter selection, Validation, and Testing np.random.seed(0)

for _ in range(6):
21.9.1 Hyperparameters, Over-fitting, and Under-fitting noisy_X = X + np.random.normal(loc=0, scale=.1, size=X.shape)
plt.plot(noisy_X, y, 'o')
See also: regr.fit(noisy_X, y)
plt.plot(X_test, regr.predict(X_test))
This section is adapted from Andrew Ng’s excellent Coursera course
The issues associated with validation and cross-validation are some of the most important aspects of the
practice of machine learning. Selecting the optimal model for your data is vital, and is a piece of the
problem that is not often appreciated by machine learning practitioners.
The central question is: If our estimator is underperforming, how should we move forward?
• Use simpler or more complicated model?
• Add more features to each observed data point?
• Add more training samples?
The answer is often counter-intuitive. In particular, Sometimes using a more complicated model
will give worse results. Also, Sometimes adding training data will not improve your results.
The ability to determine what steps will improve your model is what separates the successful machine
learning practitioners from the unsuccessful. As we can see, our linear model captures and amplifies the noise in the data. It displays a lot of variance.
Bias-variance trade-off: illustration on a simple regression problem We can use another linear estimator that uses regularization, the Ridge estimator. This estimator
regularizes the coefficients by shrinking them to zero, under the assumption that very high correlations
are often spurious. The alpha parameter controls the amount of shrinkage used.
Code and notebook
regr = linear_model.Ridge(alpha=.1)
np.random.seed(0)
for _ in range(6):
noisy_X = X + np.random.normal(loc=0, scale=.1, size=X.shape)
Let us start with a simple 1D regression problem. This will help us to easily visualize the data and the plt.plot(noisy_X, y, 'o')
model, and the results generalize easily to higher-dimensional datasets. We’ll explore a simple linear regr.fit(noisy_X, y)
regression problem, with sklearn.linear_model. plt.plot(X_test, regr.predict(X_test))
X = np.c_[ .5, 1].T plt.show()

y = [.5, 1]
X_test = np.c_[ 0, 2].T
Without noise, as linear regression fits the data perfectly
from sklearn import linear_model

regr = linear_model.LinearRegression()
regr.fit(X, y)
plt.plot(X, y, 'o')
As we can see, the estimator displays much less variance. However it systematically under-estimates the
coefficient. It displays a biased behavior.
This is a typical example of bias/variance tradeof : non-regularized estimator are not biased, but they
can display a lot of variance. Highly-regularized models have little variance, but high bias. This bias is
not necessarily a bad thing: what matters is choosing the tradeoff between bias and variance that leads
to the best prediction performance. For a specific dataset there is a sweet spot corresponding to the
highest complexity that the data can support, depending on the amount of noise and of observations
available.
In real life situation, we have noise (e.g. measurement noise) in our data:
21.9.2 Visualizing the Bias/Variance Tradeoff
21.9. Parameter selection, Validation, and Testing 614 21.9. Parameter selection, Validation, and Testing 615

Tip: Given a particular dataset and a model (e.g. a polynomial), we’d like to understand whether bias >>> x = np.random.random(size=200)
(underfit) or variance limits prediction, and how to tune the hyperparameter (here d, the degree of the >>> y = generating_func(x, err=1.)
polynomial) to give the best fit.
On a given data, let us fit a simple polynomial regression model with varying degrees:
Central to quantify bias and variance of a

model is to apply it on test data, sampled from the same distribution as the train, but that will capture
independent noise:
>>> xtrain, xtest, ytrain, ytest = train_test_split(x, y, test_size=0.4)
Tip: In the above figure, we see fits for three different values of d. For d = 1, the data is under-fit.
Validation curve A validation curve consists in varying a model parameter that controls its complexity
This means that the model is too simplistic: no straight line will ever be a good fit to this data. In this
(here the degree of the polynomial) and measures both error of the model on training data, and on test
case, we say that the model suffers from high bias. The model itself is biased, and this will be reflected
data (eg with cross-validation). The model parameter is then adjusted so that the test error is minimized:
in the fact that the data is poorly fit. At the other extreme, for d = 6 the data is over-fit. This means
that the model has too many free parameters (6 in this case) which can be adjusted to perfectly fit the We use sklearn.model_selection.validation_curve() to compute train and test error, and plot it:
training data. If we add a new point to this plot, though, chances are it will be very far from the curve
representing the degree-6 fit. In this case, we say that the model suffers from high variance. The reason >>> from sklearn.model_selection import validation_curve
for the term “high variance” is that if any of the input points are varied slightly, it could result in a very
>>> degrees = np.arange(1, 21)
different model.
In the middle, for d = 2, we have found a good mid-point. It fits the data fairly well, and does not suffer >>> model = make_pipeline(PolynomialFeatures(), LinearRegression())
from the bias and variance problems seen in the figures on either side. What we would like is a way to
quantitatively identify bias and variance, and optimize the metaparameters (in this case, the polynomial >>> # Vary the "degrees" on the pipeline step "polynomialfeatures"
>>> train_scores, validation_scores = validation_curve(
degree d) in order to determine the best algorithm.
... model, x[:, np.newaxis], y,
... param_name='polynomialfeatures__degree',
... param_range=degrees)
Polynomial regression with scikit-learn
>>> # Plot the mean train score and validation score across folds
>>> plt.plot(degrees, validation_scores.mean(axis=1), label='cross-validation')
A polynomial regression is built by pipelining PolynomialFeatures and a LinearRegression:
>>> from sklearn.pipeline import make_pipeline >>> plt.plot(degrees, train_scores.mean(axis=1), label='training')
>>> from sklearn.preprocessing import PolynomialFeatures [<matplotlib.lines.Line2D object at ...>]
>>> from sklearn.linear_model import LinearRegression >>> plt.legend(loc='best')
>>> model = make_pipeline(PolynomialFeatures(degree=2), LinearRegression()) <matplotlib.legend.Legend object at ...>
Validation Curves
Let us create a dataset like in the example above:
>>> def generating_func(x, err=0.5):

... return np.random.normal(10 - 1. / (x + 0.1), err)
>>> # randomly sample more data

>>> np.random.seed(1)
This figure shows why validation is impor-

tant. On the left side of the plot, we have very low-degree polynomial, which under-fit the data. This
leads to a low explained variance for both the training set and the validation set. On the far right side Fig. 5: For a degree=1 model
of the plot, we have a very high degree polynomial, which over-fits the data. This can be seen in the
fact that the training explained variance is very high, while on the validation set, it is low. Choosing d
Note that the validation score generally increases with a growing training set, while the training score
around 4 or 5 gets us the best tradeoff.
generally decreases with a growing training set. As the training size increases, they will converge to a
single value.
Tip: The astute reader will realize that something is amiss here: in the above plot, d = 4 gives the
From the above discussion, we know that d = 1 is a high-bias estimator which under-fits the data. This
best results. But in the previous plot, we found that d = 6 vastly over-fits the data. What’s going on
is indicated by the fact that both the training and validation scores are low. When confronted with this
here? The difference is the number of training points used. In the previous example, there were only
type of learning curve, we can expect that adding more training data will not help: both lines converge
eight training points. In this example, we have 100. As a general rule of thumb, the more training points
to a relatively low score.
used, the more complicated model can be used. But how can you determine for a given model whether
more training points will be helpful? A useful diagnostic for this are learning curves. When the learning curves have converged to a low score, we have a high bias model.
A high-bias model can be improved by:
Learning Curves • Using a more sophisticated model (i.e. in this case, increase d)
A learning curve shows the training and validation score as a function of the number of training points. • Gather more features for each sample.
Note that when we train on a subset of the training data, the training score is computed using this • Decrease regularization in a regularized model.
subset, not the full training set. This curve gives a quantitative view into how beneficial it will be to
add training samples. Increasing the number of samples, however, does not improve a high-bias model.
Now let’s look at a high-variance (i.e. over-fit) model:
Questions:
• As the number of training samples are increased, what do you expect to see for the training
score? For the validation score?
• Would you expect the training score to be higher or lower than the validation score? Would you
ever expect this to change?
scikit-learn provides sklearn.model_selection.learning_curve():
>>> from sklearn.model_selection import learning_curve

>>> train_sizes, train_scores, validation_scores = learning_curve(
... model, x[:, np.newaxis], y, train_sizes=np.logspace(-1, 0, 20))
>>> # Plot the mean train score and validation score across folds
>>> plt.plot(train_sizes, validation_scores.mean(axis=1), label='cross-validation')
>>> plt.plot(train_sizes, train_scores.mean(axis=1), label='training') Fig. 6: For a degree=15 model
Here we show the learning curve for d = 15. From the above discussion, we know that d = 15 is a
high-variance estimator which over-fits the data. This is indicated by the fact that the training score
is much higher than the validation score. As we add more samples to this training set, the training score
will continue to decrease, while the cross-validation error will continue to increase, until they meet in 21.9.4 A last word of caution: separate validation and test set
the middle.
Using validation schemes to determine hyper-parameters means that we are fitting the hyper-parameters
Learning curves that have not yet converged with the full training set indicate a high- to the particular validation set. In the same way that parameters can be over-fit to the training set,
variance, over-fit model. hyperparameters can be over-fit to the validation set. Because of this, the validation error tends to
under-predict the classification error of new data.
A high-variance model can be improved by:
For this reason, it is recommended to split the data into three sets:
• Gathering more training samples.
• The training set, used to train the model (usually ~60% of the data)
• Using a less-sophisticated model (i.e. in this case, make d smaller)
• The validation set, used to validate the model (usually ~20% of the data)
• Increasing regularization.
• The test set, used to evaluate the expected error of the validated model (usually ~20% of the
In particular, gathering more features for each sample will not help the results.
data)
21.9.3 Summary on model selection Many machine learning practitioners do not separate test set and validation set. But if your goal is to
gauge the error of a model on unknown data, using an independent test set is vital.
We’ve seen above that an under-performing algorithm can be due to two possible situations: high
bias (under-fitting) and high variance (over-fitting). In order to evaluate our algorithm, we set aside
a portion of our training data for cross-validation. Using the technique of learning curves, we can
train on progressively larger subsets of the data, evaluating the training error and cross-validation error
to determine whether our algorithm has high variance or high bias. But what do we do with this
information?
21.10 Examples for the scikit-learn chapter
High Bias
If a model shows high bias, the following actions might help: Note: Click here to download the full example code
• Add more features. In our example of predicting home prices, it may be helpful to make use of
information such as the neighborhood the house is in, the year the house was built, the size of the
lot, etc. Adding these features to the training and test sets can improve a high-bias estimator 21.10.1 Measuring Decision Tree performance
• Use a more sophisticated model. Adding complexity to the model can help improve on bias. Demonstrates overfit when testing on train set.
For a polynomial fit, this can be accomplished by increasing the degree d. Each learning technique Get the data
has its own methods of adding complexity.
from sklearn.datasets import load_boston
• Use fewer samples. Though this will not improve the classification, a high-bias algorithm can data = load_boston()
attain nearly the same error with a smaller training sample. For algorithms which are compu-
tationally expensive, reducing the training sample size can lead to very large improvements in Train and test a model
speed.
from sklearn.tree import DecisionTreeRegressor
• Decrease regularization. Regularization is a technique used to impose simplicity in some ma- clf = DecisionTreeRegressor().fit(data.data, data.target)
chine learning models, by adding a penalty term that depends on the characteristics of the param-
eters. If a model has high bias, decreasing the effect of regularization can lead to better results. predicted = clf.predict(data.data)
expected = data.target
High Variance
Plot predicted as a function of expected
If a model shows high variance, the following actions might help:
• Use fewer features. Using a feature selection technique may be useful, and decrease the over- from matplotlib import pyplot as plt
fitting of the estimator.
plt.scatter(expected, predicted)
• Use a simpler model. Model complexity and over-fitting go hand-in-hand. plt.plot([0, 50], [0, 50], '--k')
plt.axis('tight')
• Use more training samples. Adding training samples can reduce the effect of over-fitting, and plt.xlabel('True price ($1000s)')
lead to improvements in a high variance estimator. plt.ylabel('Predicted price ($1000s)')
plt.tight_layout()
• Increase Regularization. Regularization is designed to prevent over-fitting. In a high-variance
model, increasing regularization can lead to better results.
These choices become very important in real-world situations. For example, due to limited telescope
time, astronomers must seek a balance between observing a large number of objects, and observing a
large number of features for each object. Determining which is more important for a particular learning
task can inform the observing strategy that the astronomer employs.
21.9. Parameter selection, Validation, and Testing 620 21.10. Examples for the scikit-learn chapter 621
Pretty much no errors!

This is too good to be true: we are testing the model on the train data, which is not a mesure of
generalization.
The results are not valid
21.10.2 Demo PCA in 2D

Load the iris data
from sklearn import datasets Note: Click here to download the full example code
X = iris.data
y = iris.target
21.10.3 A simple linear regression
Fit a PCA
from sklearn.decomposition import PCA

pca = PCA(n_components=2, whiten=True)
pca.fit(X)
Project the data in 2D
X_pca = pca.transform(X)
Visualize the data
target_ids = range(len(iris.target_names))

for i, c, label in zip(target_ids, 'rgbcmykw', iris.target_names):
plt.scatter(X_pca[y == i, 0], X_pca[y == i, 1],
c=c, label=label)
plt.legend()
plt.show()
21.10. Examples for the scikit-learn chapter 622 21.10. Examples for the scikit-learn chapter 623
import numpy as np
from sklearn.linear_model import LinearRegression
# x from 0 to 30
x = 30 * np.random.random((20, 1))
# y = a*x + b with noise

y = 0.5 * x + 1.0 + np.random.normal(size=x.shape)
# create a linear regression model

model = LinearRegression()
model.fit(x, y)
# predict y from the data

x_new = np.linspace(0, 30, 100)
y_new = model.predict(x_new[:, np.newaxis])
# plot the results

ax = plt.axes()
ax.scatter(x, y)
ax.plot(x_new, y_new)
ax.set_xlabel('x')
ax.set_ylabel('y')
ax.axis('tight') # Load the data

from sklearn.datasets import load_iris
iris = load_iris()
plt.show()
# The indices of the features that we are plotting
x_index = 0
Note: Click here to download the full example code y_index = 1
# this formatter will label the colorbar with the correct target names
formatter = plt.FuncFormatter(lambda i, *args: iris.target_names[int(i)])
21.10.4 Plot 2D views of the iris dataset
Plot a simple scatter plot of 2 features of the iris dataset. plt.scatter(iris.data[:, x_index], iris.data[:, y_index], c=iris.target)
Note that more elaborate visualization of this dataset is detailed in the Statistics in Python chapter. plt.colorbar(ticks=[0, 1, 2], format=formatter)
plt.xlabel(iris.feature_names[x_index])
plt.ylabel(iris.feature_names[y_index])
plt.tight_layout()
plt.show()
21.10.5 tSNE to visualize digits

Here we use sklearn.manifold.TSNE to visualize the digits datasets. Indeed, the digits are vectors in
a 8*8 = 64 dimensional space. We want to project them in 2D for visualization. tSNE is often a good
solution, as it groups and separates data points based on their local relationship.
Load the iris data
from sklearn import datasets

digits = datasets.load_digits() Note: Click here to download the full example code
# Take the first 500 data points: it's hard to see 1500 points
X = digits.data[:500]
y = digits.target[:500] 21.10.6 Use the RidgeCV and LassoCV to set the regularization parameter
Fit and transform with a TSNE Load the diabetes dataset
from sklearn.manifold import TSNE from sklearn.datasets import load_diabetes

tsne = TSNE(n_components=2, random_state=0) data = load_diabetes()
X, y = data.data, data.target
print(X.shape)
Project the data in 2D
X_2d = tsne.fit_transform(X) Out:
(442, 10)
Visualize the data
target_ids = range(len(digits.target_names)) Compute the cross-validation score with the default hyper-parameters
from matplotlib import pyplot as plt from sklearn.model_selection import cross_val_score

plt.figure(figsize=(6, 5)) from sklearn.linear_model import Ridge, Lasso
colors = 'r', 'g', 'b', 'c', 'm', 'y', 'k', 'w', 'orange', 'purple'
for i, c, label in zip(target_ids, colors, digits.target_names): for Model in [Ridge, Lasso]:
plt.scatter(X_2d[y == i, 0], X_2d[y == i, 1], c=c, label=label) model = Model()
plt.legend() print('%s : %s ' % (Model.__name__,
plt.show() cross_val_score(model, X, y).mean()))
Out:
Ridge: 0.4101758336587286
Lasso: 0.3375597834274947
We compute the cross-validation score as a function of alpha, the strength of the regularization for Lasso
and Ridge
import numpy as np
alphas = np.logspace(-3, -1, 30)
for Model in [Lasso, Ridge]:

scores = [cross_val_score(Model(alpha), X, y, cv=3).mean()
for alpha in alphas]
plt.plot(alphas, scores, label=Model.__name__)
plt.legend(loc='lower left')
plt.xlabel('alpha')
plt.ylabel('cross validation score')
plt.tight_layout()
plt.show()
np.random.seed(0)
for _ in range(6):
regr.fit(noisy_X, y)

As we can see, our linear model captures and amplifies the noise in the data. It displays a lot of variance.
Note: Click here to download the full example code We can use another linear estimator that uses regularization, the Ridge estimator. This estimator
regularizes the coefficients by shrinking them to zero, under the assumption that very high correlations
are often spurious. The alpha parameter controls the amount of shrinkage used.
21.10.7 Plot variance and regularization in linear models regr = linear_model.Ridge(alpha=.1)
np.random.seed(0)
import numpy as np
for _ in range(6):
# Smaller figures
regr.fit(noisy_X, y)
plt.rcParams['figure.figsize'] = (3, 2)
We consider the situation where we have only 2 data point plt.show()

X = np.c_[ .5, 1].T
y = [.5, 1]
X_test = np.c_[ 0, 2].T
Without noise, as linear regression fits the data perfectly

from sklearn import linear_model
regr = linear_model.LinearRegression()
regr.fit(X, y)
plt.plot(X, y, 'o')
21.10.8 Simple picture of the formal problem of machine learning

This example generates simple synthetic data ploints and shows a separating hyperplane on them.
In real life situation, we have noise (e.g. measurement noise) in our data:
21.10.9 Compare classifiers on the digits data

Compare the performance of a variety of classifiers on a test set for the digits data.
Out:
LinearSVC: 0.9341800269333108
GaussianNB: 0.8332741681010101
KNeighborsClassifier: 0.9804562804949924
------------------
LinearSVC(loss='hinge'): 0.9294570108037394
LinearSVC(loss='squared_hinge'): 0.9341371852581549
-------------------
import numpy as np
from sklearn.linear_model import SGDClassifier
from sklearn.datasets.samples_generator import make_blobs
# we create 50 separable synthetic points

X, Y = make_blobs(n_samples=50, centers=2,
random_state=0, cluster_std=0.60)
from sklearn import model_selection, datasets, metrics
# fit the model from sklearn.svm import LinearSVC
clf = SGDClassifier(loss="hinge", alpha=0.01, from sklearn.naive_bayes import GaussianNB
fit_intercept=True) from sklearn.neighbors import KNeighborsClassifier
clf.fit(X, Y)
digits = datasets.load_digits()
# plot the line, the points, and the nearest vectors to the plane X = digits.data
xx = np.linspace(-1, 5, 10) y = digits.target
yy = np.linspace(-1, 5, 10) X_train, X_test, y_train, y_test = model_selection.train_test_split(X, y,
test_size=0.25, random_state=0)
X1, X2 = np.meshgrid(xx, yy)
Z = np.empty(X1.shape) for Model in [LinearSVC, GaussianNB, KNeighborsClassifier]:
for (i, j), val in np.ndenumerate(X1): clf = Model().fit(X_train, y_train)
x1 = val y_pred = clf.predict(X_test)
x2 = X2[i, j] print('%s : %s ' %
p = clf.decision_function([[x1, x2]]) (Model.__name__, metrics.f1_score(y_test, y_pred, average="macro")))
Z[i, j] = p[0]
print('------------------')
ax = plt.axes() # test SVC loss
ax.contour(X1, X2, Z, [-1.0, 0.0, 1.0], colors='k', for loss in ['hinge', 'squared_hinge']:
linestyles=['dashed', 'solid', 'dashed']) clf = LinearSVC(loss=loss).fit(X_train, y_train)
ax.scatter(X[:, 0], X[:, 1], c=Y, cmap=plt.cm.Paired) y_pred = clf.predict(X_test)
print("LinearSVC(loss='{0} '): {1} ".format(loss,
ax.axis('tight') metrics.f1_score(y_test, y_pred, average="macro")))
print('-------------------')
plt.show()
# test the number of neighbors
Total running time of the script: ( 0 minutes 0.026 seconds) for n_neighbors in range(1, 11):
clf = KNeighborsClassifier(n_neighbors=n_neighbors).fit(X_train, y_train)
y_pred = clf.predict(X_test)
Note: Click here to download the full example code print("KNeighbors(n_neighbors={0} ): {1} ".format(n_neighbors,
metrics.f1_score(y_test, y_pred, average="macro")))
Total running time of the script: ( 0 minutes 1.025 seconds) For this we need to engineer features: the n_th powers of x:
Note: Click here to download the full example code plt.scatter(x, y, s=4)
X = np.array([x**i for i in range(5)]).T

X_test = np.array([x_test**i for i in range(5)]).T
21.10.10 Plot fitting a 9th order polynomial regr = linear_model.LinearRegression()
Fits data generated from a 9th order polynomial with model of 4th order and 9th order polynomials, to regr.fit(X, y)
demonstrate that often simpler models are to be prefered plt.plot(x_test, regr.predict(X_test), label='4th order')
import numpy as np X = np.array([x**i for i in range(10)]).T

from matplotlib import pyplot as plt X_test = np.array([x_test**i for i in range(10)]).T
from matplotlib.colors import ListedColormap regr = linear_model.LinearRegression()
regr.fit(X, y)
from sklearn import linear_model plt.plot(x_test, regr.predict(X_test), label='9th order')
# Create color maps for 3-class classification problem, as with iris plt.legend(loc='best')
cmap_light = ListedColormap(['#FFAAAA', '#AAFFAA', '#AAAAFF']) plt.axis('tight')
cmap_bold = ListedColormap(['#FF0000', '#00FF00', '#0000FF']) plt.title('Fitting a 4th and a 9th order polynomial')
rng = np.random.RandomState(0)
x = 2*rng.rand(100) - 1
f = lambda t: 1.2 * t**2 + .1 * t**3 - .4 * t **5 - .5 * t ** 9

y = f(x) + .4 * rng.normal(size=100)
x_test = np.linspace(-1, 1, 100)
The data
plt.scatter(x, y, s=4)
Ground truth
plt.scatter(x, y, s=4)
plt.plot(x_test, f(x_test), label="truth")
plt.axis('tight')
plt.title('Ground truth (9th order polynomial)')
plt.show()
Fitting 4th and 9th order polynomials
Print the join histogram for each feature
for index, feature_name in enumerate(data.feature_names):

plt.scatter(data.data[feature_name], data.target)
plt.ylabel('Price', size=15)
plt.xlabel(feature_name, size=15)
Total running time of the script: ( 0 minutes 0.047 seconds) plt.tight_layout()
21.10.11 A simple regression analysis on the California housing data

Here we perform a simple regression analysis on the California housing data, exploring two types of
regressors.
from sklearn.datasets import fetch_california_housing

data = fetch_california_housing(as_frame=True)
Print a histogram of the quantity to predict: price

plt.hist(data.target)
plt.xlabel('price ($1000s)')
plt.ylabel('count')
plt.tight_layout() •
• •
• •
• •
from sklearn.ensemble import GradientBoostingRegressor
clf = GradientBoostingRegressor()
clf.fit(X_train, y_train)
predicted = clf.predict(X_test)
expected = y_test
plt.plot([0, 50], [0, 50], '--k')
plt.axis('tight')
plt.xlabel('True price ($1000s)')
plt.ylabel('Predicted price ($1000s)')
plt.tight_layout()
•
Simple prediction
from sklearn.model_selection import train_test_split

X_train, X_test, y_train, y_test = train_test_split(data.data, data.target)

clf = LinearRegression()
clf.fit(X_train, y_train)
predicted = clf.predict(X_test)
expected = y_test
plt.plot([0, 50], [0, 50], '--k')
plt.axis('tight')
plt.xlabel('True price ($1000s)')
plt.ylabel('Predicted price ($1000s)')
plt.tight_layout() Print the error rate
import numpy as np
print("RMS: %r " % np.sqrt(np.mean((predicted - expected) ** 2)))
plt.show()
Out:
RMS: 0.5314909993118918
21.10.12 Nearest-neighbor prediction on iris

Plot the decision boundary of nearest neighbor decision on iris, first with a single nearest neighbor, and
then using 3 nearest neighbors.
import numpy as np
Prediction with gradient boosted tree from matplotlib import pyplot as plt
from sklearn import neighbors, datasets

from matplotlib.colors import ListedColormap
# Create color maps for 3-class classification problem, as with iris

cmap_light = ListedColormap(['#FFAAAA', '#AAFFAA', '#AAAAFF'])
cmap_bold = ListedColormap(['#FF0000', '#00FF00', '#0000FF'])
X = iris.data[:, :2] # we only take the first two features. We could
# avoid this ugly slicing by using a two-dim dataset
y = iris.target
knn.fit(X, y)
x_min, x_max = X[:, 0].min() - .1, X[:, 0].max() + .1

y_min, y_max = X[:, 1].min() - .1, X[:, 1].max() + .1
xx, yy = np.meshgrid(np.linspace(x_min, x_max, 100),
np.linspace(y_min, y_max, 100))
Z = knn.predict(np.c_[xx.ravel(), yy.ravel()])
Put the result into a color plot
Z = Z.reshape(xx.shape)
plt.figure()
plt.pcolormesh(xx, yy, Z, cmap=cmap_light)
# Plot also the training points

plt.scatter(X[:, 0], X[:, 1], c=y, cmap=cmap_bold)
plt.xlabel('sepal length (cm)')
plt.ylabel('sepal width (cm)')
plt.axis('tight') And now, redo the analysis with 3 neighbors
knn.fit(X, y)
Z = knn.predict(np.c_[xx.ravel(), yy.ravel()])
# Put the result into a color plot

Z = Z.reshape(xx.shape)
plt.figure()
plt.pcolormesh(xx, yy, Z, cmap=cmap_light)
# Plot also the training points

plt.scatter(X[:, 0], X[:, 1], c=y, cmap=cmap_bold)
plt.xlabel('sepal length (cm)')
plt.ylabel('sepal width (cm)')
plt.axis('tight')
plt.show()
21.10.13 Simple visualization and classification of the digits dataset

Plot the first few samples of the digits dataset and a 2D representation built using PCA, then do a simple
classification Plot a projection on the 2 first principal axis
from sklearn.datasets import load_digits plt.figure()
digits = load_digits()
from sklearn.decomposition import PCA
pca = PCA(n_components=2)
Plot the data: images of digits proj = pca.fit_transform(digits.data)
plt.scatter(proj[:, 0], proj[:, 1], c=digits.target, cmap="Paired")
Each data in a 8x8 image
plt.colorbar()
fig = plt.figure(figsize=(6, 6)) # figure size in inches
fig.subplots_adjust(left=0, right=1, bottom=0, top=1, hspace=0.05, wspace=0.05)
for i in range(64):
ax.imshow(digits.images[i], cmap=plt.cm.binary, interpolation='nearest')
# label the image with the target value
ax.text(0, 7, str(digits.target[i]))
Classify with Gaussian naive Bayes
from sklearn.naive_bayes import GaussianNB

# split the data into training and validation sets

X_train, X_test, y_train, y_test = train_test_split(digits.data, digits.target)
# train the model

clf = GaussianNB() Quantify the performance
clf.fit(X_train, y_train) First print the number of correct matches
# use the model to predict the labels of the test data matches = (predicted == expected)
predicted = clf.predict(X_test) print(matches.sum())
expected = y_test
Out:
# Plot the prediction
fig = plt.figure(figsize=(6, 6)) # figure size in inches 388
fig.subplots_adjust(left=0, right=1, bottom=0, top=1, hspace=0.05, wspace=0.05)
The total number of data points
# plot the digits: each image is 8x8 pixels
for i in range(64): print(len(matches))
ax.imshow(X_test.reshape(-1, 8, 8)[i], cmap=plt.cm.binary,
interpolation='nearest') Out:
450
# label the image with the target value
if predicted[i] == expected[i]:
ax.text(0, 7, str(predicted[i]), color='green') And now, the ration of correct predictions
else:
ax.text(0, 7, str(predicted[i]), color='red')
matches.sum() / float(len(matches)) from sklearn import datasets

faces = datasets.fetch_olivetti_faces()
Print the classification report faces.data.shape
from sklearn import metrics Let’s visualize these faces to see what we’re working with
print(metrics.classification_report(expected, predicted))
Out: fig = plt.figure(figsize=(8, 6))
# plot several images
precision recall f1-score support for i in range(15):
0 1.00 1.00 1.00 51 ax.imshow(faces.images[i], cmap=plt.cm.bone)
1 0.62 0.93 0.75 41
2 0.94 0.70 0.80 46
3 0.93 0.87 0.90 47
4 1.00 0.84 0.91 43
5 0.86 0.93 0.89 40
6 0.98 0.98 0.98 45
7 0.86 0.96 0.91 52
8 0.65 0.69 0.67 49
9 0.96 0.69 0.81 36
accuracy 0.86 450

macro avg 0.88 0.86 0.86 450
weighted avg 0.88 0.86 0.86 450
Print the confusion matrix
print(metrics.confusion_matrix(expected, predicted))
plt.show()
Out:
[[51 0 0 0 0 0 0 0 0 0]
[ 0 38 0 0 0 0 0 0 3 0]
[ 0 5 32 0 0 0 0 0 9 0]
[ 0 1 0 41 0 2 0 0 2 1]
[ 0 2 1 0 36 0 1 2 1 0]
[ 0 1 0 0 0 37 0 1 1 0]
[ 0 0 1 0 0 0 44 0 0 0]
[ 0 0 0 0 0 1 0 50 1 0]
[ 0 12 0 0 0 1 0 2 34 0]
[ 0 2 0 3 0 2 0 3 1 25]]
Total running time of the script: ( 0 minutes 1.639 seconds) Tip: Note is that these faces have already been localized and scaled to a common size. This is an
important preprocessing piece for facial recognition, and is a process that can require a large collection
Note: Click here to download the full example code of training data. This can be done in scikit-learn, but the challenge is gathering a sufficient amount of
training data for the algorithm to work. Fortunately, this piece is common enough that it has been done.
One good resource is OpenCV, the Open Computer Vision Library.
21.10.14 The eigenfaces example: chaining PCA and SVMs
We’ll perform a Support Vector classification of the images. We’ll do a typical train-test split on the
The goal of this example is to show how an unsupervised method and a supervised one can be chained
images:
for better prediction. It starts with a didactic but lengthy way of doing things, and finishes with the
idiomatic approach to pipelining in scikit-learn. from sklearn.model_selection import train_test_split
X_train, X_test, y_train, y_test = train_test_split(faces.data,
Here we’ll take a look at a simple facial recognition example. Ideally, we would use a dataset con-
faces.target, random_state=0)
sisting of a subset of the Labeled Faces in the Wild data that is available with sklearn.datasets.
fetch_lfw_people(). However, this is a relatively large download (~200MB) so we will do the tutorial print(X_train.shape, X_test.shape)
on a simpler, less rich dataset. Feel free to explore the LFW dataset.
Out:

(300, 4096) (100, 4096)
ax.imshow(pca.components_[i].reshape(faces.images[0].shape),
cmap=plt.cm.bone)
Preprocessing: Principal Component Analysis
1850 dimensions is a lot for SVM. We can use PCA to reduce these 1850 features to a manageable size,
while maintaining most of the information in the dataset.
from sklearn import decomposition

pca = decomposition.PCA(n_components=150, whiten=True)
pca.fit(X_train)
One interesting part of PCA is that it computes the “mean” face, which can be interesting to examine:
plt.imshow(pca.mean_.reshape(faces.images[0].shape),
cmap=plt.cm.bone)
The components (“eigenfaces”) are ordered by their importance from top-left to bottom-right. We
see that the first few components seem to primarily take care of lighting conditions; the remaining
components pull out certain identifying features: the nose, eyes, eyebrows, etc.
With this projection computed, we can now project our original training and test data onto the PCA
basis:
X_train_pca = pca.transform(X_train)
X_test_pca = pca.transform(X_test)
print(X_train_pca.shape)
Out:
(300, 150)
print(X_test_pca.shape)
Out:
(100, 150)
These projected components correspond to factors in a linear combination of component images such
that the combination approaches the original face.
Doing the Learning: Support Vector Machines

Now we’ll perform support-vector-machine classification on this reduced dataset:
The principal components measure deviations about this mean along orthogonal axes. from sklearn import svm
clf = svm.SVC(C=5., gamma=0.001)
print(pca.components_.shape)
clf.fit(X_train_pca, y_train)
Out: Finally, we can evaluate how well this classification did. First, we might plot a few of the test-cases with
(150, 4096) the labels learned from the training set:
import numpy as np
It is also interesting to visualize these principal components: fig = plt.figure(figsize=(8, 6))
for i in range(15):
for i in range(30):
ax.imshow(X_test[i].reshape(faces.images[0].shape),
cmap=plt.cm.bone)

y_pred = clf.predict(X_test_pca[i, np.newaxis])[0] 10 1.00 1.00 1.00 4
color = ('black' if y_pred == y_test[i] else 'red') 11 1.00 1.00 1.00 1
ax.set_title(y_pred, fontsize='small', color=color) 12 0.67 1.00 0.80 2
13 1.00 1.00 1.00 3
14 1.00 1.00 1.00 5
15 1.00 1.00 1.00 3
17 1.00 1.00 1.00 6
19 1.00 1.00 1.00 4
20 1.00 1.00 1.00 1
21 1.00 1.00 1.00 1
22 1.00 1.00 1.00 2
23 1.00 1.00 1.00 1
24 1.00 1.00 1.00 2
25 1.00 0.50 0.67 2
26 1.00 0.75 0.86 4
27 1.00 1.00 1.00 1
28 0.67 1.00 0.80 2
29 1.00 1.00 1.00 3
30 1.00 1.00 1.00 4
31 1.00 1.00 1.00 3
32 1.00 1.00 1.00 3
33 1.00 1.00 1.00 2
34 1.00 1.00 1.00 3
35 1.00 1.00 1.00 1
36 1.00 1.00 1.00 3
37 1.00 1.00 1.00 3
38 1.00 1.00 1.00 1
39 1.00 1.00 1.00 3
accuracy 0.94 100

macro avg 0.95 0.96 0.94 100
weighted avg 0.97 0.94 0.94 100
Another interesting metric is the confusion matrix, which indicates how often any two items are mixed-
up. The confusion matrix of a perfect classifier would only have nonzero entries on the diagonal, with
The classifier is correct on an impressive number of images given the simplicity of its learning model! zeros on the off-diagonal:
Using a linear classifier on 150 features derived from the pixel-level data, the algorithm correctly identifies print(metrics.confusion_matrix(y_test, y_pred))
a large number of the people in the images.
Again, we can quantify this effectiveness using one of several measures from sklearn.metrics. First we Out:
can do the classification report, which shows the precision, recall and other measures of the “goodness”
[[3 0 0 ... 0 0 0]
of the classification:
[0 4 0 ... 0 0 0]
from sklearn import metrics [0 0 2 ... 0 0 0]
y_pred = clf.predict(X_test_pca) ...
print(metrics.classification_report(y_test, y_pred)) [0 0 0 ... 3 0 0]
[0 0 0 ... 0 1 0]
[0 0 0 ... 0 0 3]]
Out:
precision recall f1-score support

Pipelining
0 1.00 0.50 0.67 6 Above we used PCA as a pre-processing step before applying our support vector machine classifier.
1 1.00 1.00 1.00 4 Plugging the output of one estimator directly into the input of a second estimator is a commonly used
2 0.50 1.00 0.67 2
pattern; for this reason scikit-learn provides a Pipeline object which automates this process. The above
3 1.00 1.00 1.00 1
4 0.33 1.00 0.50 1
problem can be re-expressed as a pipeline as follows:
5 1.00 1.00 1.00 5
from sklearn.pipeline import Pipeline
6 1.00 1.00 1.00 4
clf = Pipeline([('pca', decomposition.PCA(n_components=150, whiten=True)),
7 1.00 0.67 0.80 3
('svm', svm.LinearSVC(C=1.0))])
9 1.00 1.00 1.00 1

clf.fit(X_train, y_train) ax = plt.subplot(111, xticks=[], yticks=[])
ax.scatter(X[:, 0], X[:, 1], c=y, cmap=plt.cm.bone)
y_pred = clf.predict(X_test)
print(metrics.confusion_matrix(y_pred, y_test)) ax.scatter(clf.support_vectors_[:, 0],
plt.show() clf.support_vectors_[:, 1],
s=80, edgecolors="k", facecolors="none")
Out:
delta = 1
[[5 0 0 ... 0 0 0] y_min, y_max = -50, 50
[0 4 0 ... 0 0 0] x_min, x_max = -50, 50
[0 0 1 ... 0 0 0] x = np.arange(x_min, x_max + delta, delta)
... y = np.arange(y_min, y_max + delta, delta)
[0 0 0 ... 3 0 0] X1, X2 = np.meshgrid(x, y)
[0 0 0 ... 0 1 0] Z = clf.decision_function(np.c_[X1.ravel(), X2.ravel()])
[0 0 0 ... 0 0 3]] Z = Z.reshape(X1.shape)
ax.contour(X1, X2, Z, [-1.0, 0.0, 1.0], colors='k',

A Note on Facial Recognition linestyles=['dashed', 'solid', 'dashed'])
Here we have used PCA “eigenfaces” as a pre-processing step for facial recognition. The reason we
chose this is because PCA is a broadly-applicable technique, which can be useful for a wide array of
data types. Research in the field of facial recognition in particular, however, has shown that other more
specific feature extraction methods are can be much more effective.
21.10.15 Example of linear and non-linear models

This is an example plot from the tutorial which accompanies an explanation of the support vector
machine GUI.
import numpy as np
from sklearn import svm
data that is linearly separable
def linear_model(rseed=42, n_samples=30):

" Generate data according to a linear model"
np.random.seed(rseed)
data = np.random.normal(0, 10, (n_samples, 2))

data with a non-linear separation
data[:n_samples // 2] -= 15
data[n_samples // 2:] += 15 def nonlinear_model(rseed=42, n_samples=30):
radius = 40 * np.random.random(n_samples)
labels = np.ones(n_samples) far_pts = radius > 20
labels[:n_samples // 2] = -1 radius[far_pts] *= 1.2
radius[~far_pts] *= 1.1
return data, labels
theta = np.random.random(n_samples) * np.pi * 2
X, y = linear_model() data = np.empty((n_samples, 2))

clf = svm.SVC(kernel='linear') data[:, 0] = radius * np.cos(theta)
clf.fit(X, y) data[:, 1] = radius * np.sin(theta)
plt.figure(figsize=(6, 4)) labels = np.ones(n_samples)

(continued from previous page) 21.10.16 Bias and variance of polynomial fit
labels[far_pts] = -1 Demo overfitting, underfitting, and validation and learning curves with polynomial regression.
return data, labels Fit polynomes of different degrees to a dataset: for too small a degree, the model underfits, while for too
large a degree, it overfits.
X, y = nonlinear_model() import numpy as np

clf = svm.SVC(kernel='rbf', gamma=0.001, coef0=0, degree=3) import matplotlib.pyplot as plt
clf.fit(X, y)
plt.figure(figsize=(6, 4)) def generating_func(x, err=0.5):

ax = plt.subplot(1, 1, 1, xticks=[], yticks=[]) return np.random.normal(10 - 1. / (x + 0.1), err)
ax.scatter(X[:, 0], X[:, 1], c=y, cmap=plt.cm.bone, zorder=2)
A polynomial regression
ax.scatter(clf.support_vectors_[:, 0], clf.support_vectors_[:, 1],
s=80, edgecolors="k", facecolors="none") from sklearn.pipeline import make_pipeline
delta = 1 from sklearn.preprocessing import PolynomialFeatures
y_min, y_max = -50, 50
x_min, x_max = -50, 50 A simple figure to illustrate the problem
x = np.arange(x_min, x_max + delta, delta)
y = np.arange(y_min, y_max + delta, delta) n_samples = 8
X1, X2 = np.meshgrid(x, y)
Z = clf.decision_function(np.c_[X1.ravel(), X2.ravel()]) np.random.seed(0)
Z = Z.reshape(X1.shape) x = 10 ** np.linspace(-2, 0, n_samples)
y = generating_func(x)
ax.contour(X1, X2, Z, [-1.0, 0.0, 1.0], colors='k',
linestyles=['dashed', 'solid', 'dashed'], zorder=1) x_test = np.linspace(-0.2, 1.2, 1000)
plt.show() titles = ['d = 1 (under-fit; high bias)',

'd = 2',
'd = 6 (over-fit; high variance)']
degrees = [1, 2, 6]
fig = plt.figure(figsize=(9, 3.5))

fig.subplots_adjust(left=0.06, right=0.98, bottom=0.15, top=0.85, wspace=0.05)
for i, d in enumerate(degrees):
ax = fig.add_subplot(131 + i, xticks=[], yticks=[])
ax.scatter(x, y, marker='x', c='k', s=50)
model = make_pipeline(PolynomialFeatures(d), LinearRegression())

model.fit(x[:, np.newaxis], y)
ax.plot(x_test, model.predict(x_test[:, np.newaxis]), '-b')
ax.set_xlim(-0.2, 1.2)
ax.set_ylim(0, 12)
ax.set_xlabel('house size')
if i == 0:
ax.set_ylabel('price')
ax.set_title(titles[i])
Generate a larger dataset
n_samples = 200
test_size = 0.4
error = 1.0
# randomly sample the data

np.random.seed(1)
x = np.random.random(n_samples)
Plot a validation curve
y = generating_func(x, error)
from sklearn.model_selection import validation_curve
# split into training, validation, and testing sets.
x_train, x_test, y_train, y_test = train_test_split(x, y, test_size=test_size) degrees = np.arange(1, 21)
# show the training and validation sets model = make_pipeline(PolynomialFeatures(), LinearRegression())

plt.scatter(x_train, y_train, color='red', label='Training set') # The parameter to vary is the "degrees" on the pipeline step
plt.scatter(x_test, y_test, color='blue', label='Test set') # "polynomialfeatures"
plt.title('The data') train_scores, validation_scores = validation_curve(
plt.legend(loc='best') model, x[:, np.newaxis], y,
param_name='polynomialfeatures__degree',
param_range=degrees)
# Plot the mean train error and validation error across folds
plt.plot(degrees, validation_scores.mean(axis=1), lw=2,
label='cross-validation')
plt.plot(degrees, train_scores.mean(axis=1), lw=2, label='training')
plt.xlabel('degree of fit')
plt.ylabel('explained variance')
plt.title('Validation curve')
plt.tight_layout()
Learning curves
Plot train and test error with an increasing number of samples
# A learning curve for d=1, 5, 15

for d in [1, 5, 15]:
model = make_pipeline(PolynomialFeatures(degree=d), LinearRegression())
from sklearn.model_selection import learning_curve

train_sizes, train_scores, validation_scores = learning_curve(
model, x[:, np.newaxis], y,
train_sizes=np.logspace(-1, 0, 20))
# Plot the mean train error and validation error across folds
plt.plot(train_sizes, validation_scores.mean(axis=1),
lw=2, label='cross-validation')
plt.plot(train_sizes, train_scores.mean(axis=1),
lw=2, label='training')
plt.ylim(ymin=-.1, ymax=1)
plt.xlabel('number of train samples')
plt.ylabel('explained variance')
plt.title('Learning curve (degree=%i )' % d) •
plt.tight_layout()
plt.show()
• •
21.10.17 Tutorial Diagrams

This script plots the flow-charts used in the scikit-learn tutorials.
import numpy as np
from matplotlib.patches import Circle, Rectangle, Polygon, Arrow, FancyArrow
def create_base(box_bg = '#CCCCCC',

arrow1 = '#88CCFF',
arrow2 = '#88FF88',
supervised=True):
• fig = plt.figure(figsize=(9, 6), facecolor='w')

ax = plt.axes((0, 0, 1, 1), plt.text(5.5, 3.5, "Machine\nLearning\nAlgorithm",
xticks=[], yticks=[], frameon=False) ha='center', va='center', fontsize=14)
ax.set_xlim(0, 9)
ax.set_ylim(0, 6) plt.text(1.05, 1.1, "New Text,\nDocument,\nImage,\netc.",
ha='center', va='center', fontsize=14)
patches = [Rectangle((0.3, 3.6), 1.5, 1.8, zorder=1, fc=box_bg),
Rectangle((0.5, 3.8), 1.5, 1.8, zorder=2, fc=box_bg), plt.text(3.3, 1.7, "Feature\nVector",
Rectangle((0.7, 4.0), 1.5, 1.8, zorder=3, fc=box_bg), ha='left', va='center', fontsize=14)
Rectangle((2.9, 3.6), 0.2, 1.8, fc=box_bg), plt.text(5.5, 1.1, "Predictive\nModel",

Rectangle((3.1, 3.8), 0.2, 1.8, fc=box_bg), ha='center', va='center', fontsize=12)
Rectangle((3.3, 4.0), 0.2, 1.8, fc=box_bg),
if supervised:
Rectangle((0.3, 0.2), 1.5, 1.8, fc=box_bg), plt.text(1.45, 3.05, "Labels",
Rectangle((2.9, 0.2), 0.2, 1.8, fc=box_bg),
plt.text(8.05, 1.1, "Expected\nLabel",
Circle((5.5, 3.5), 1.0, fc=box_bg), ha='center', va='center', fontsize=14)
plt.text(8.8, 5.8, "Supervised Learning Model",
Polygon([[5.5, 1.7], ha='right', va='top', fontsize=18)
[6.1, 1.1],
[5.5, 0.5], else:
[4.9, 1.1]], fc=box_bg), plt.text(8.05, 1.1,
"Likelihood\nor Cluster ID\nor Better\nRepresentation",
FancyArrow(2.3, 4.6, 0.35, 0, fc=arrow1, ha='center', va='center', fontsize=12)
width=0.25, head_width=0.5, head_length=0.2), plt.text(8.8, 5.8, "Unsupervised Learning Model",
ha='right', va='top', fontsize=18)
FancyArrow(3.75, 4.2, 0.5, -0.2, fc=arrow1,
width=0.25, head_width=0.5, head_length=0.2),
FancyArrow(5.5, 2.4, 0, -0.4, fc=arrow1, def plot_supervised_chart(annotate=False):

width=0.25, head_width=0.5, head_length=0.2), create_base(supervised=True)
if annotate:
FancyArrow(2.0, 1.1, 0.5, 0, fc=arrow2, fontdict = dict(color='r', weight='bold', size=14)
width=0.25, head_width=0.5, head_length=0.2), plt.text(1.9, 4.55, 'X = vec.fit_transform(input)',
fontdict=fontdict,
FancyArrow(3.3, 1.1, 1.3, 0, fc=arrow2, rotation=20, ha='left', va='bottom')
width=0.25, head_width=0.5, head_length=0.2), plt.text(3.7, 3.2, 'clf.fit(X, y)',
fontdict=fontdict,
FancyArrow(6.2, 1.1, 0.8, 0, fc=arrow2, rotation=20, ha='left', va='bottom')
width=0.25, head_width=0.5, head_length=0.2)] plt.text(1.7, 1.5, 'X_new = vec.transform(input)',
fontdict=fontdict,
if supervised: rotation=20, ha='left', va='bottom')
patches += [Rectangle((0.3, 2.4), 1.5, 0.5, zorder=1, fc=box_bg), plt.text(6.1, 1.5, 'y_new = clf.predict(X_new)',
Rectangle((0.5, 2.6), 1.5, 0.5, zorder=2, fc=box_bg), fontdict=fontdict,
Rectangle((0.7, 2.8), 1.5, 0.5, zorder=3, fc=box_bg), rotation=20, ha='left', va='bottom')
FancyArrow(2.3, 2.9, 2.0, 0, fc=arrow1,
width=0.25, head_width=0.5, head_length=0.2), def plot_unsupervised_chart():
Rectangle((7.3, 0.85), 1.5, 0.5, fc=box_bg)] create_base(supervised=False)
else:
patches += [Rectangle((7.3, 0.2), 1.5, 1.8, fc=box_bg)]
if __name__ == '__main__':
for p in patches: plot_supervised_chart(False)
ax.add_patch(p) plot_supervised_chart(True)
plot_unsupervised_chart()
plt.text(1.45, 4.9, "Training\nText,\nDocuments,\nImages,\netc.", plt.show()
plt.text(3.6, 4.9, "Feature\nVectors",
ha='left', va='center', fontsize=14) See also:
Going further
• The documentation of scikit-learn is very complete and didactic.

• Introduction to Machine Learning with Python, by Sarah Guido, Andreas Müller (notebooks avail-
able here).
Part IV
About the SciPy lecture notes
21.10. Examples for the scikit-learn chapter 664 665

• Matthieu Boileau • Tim Hoffmann

• Joris Van den Bossche • 2. Hohl
• Michael Boyle • Tarek Hoteit
• Matthew Brett • Gert-Ludwig Ingold
22
• BSGalvan • Zbigniew Jędrzejewski-Szmek
• Lars Buitinck • Thouis (Ray) Jones
• Pierre de Buyl • jorgeprietoarranz
• Ozan Çağlayan • josephsalmon
CHAPTER
• Lawrence Chan • Greg Kiar
• Adrien Chauve • kikocorreoso
• Robert Cimrman • Vince Knight
• Christophe Combelles • LFP6
• David Cournapeau • Manuel López-Ibáñez
About the SciPy lecture notes

• Dave • Marco Mangan
• dogacan dugmeci • Nicola Masarone
• Török Edwin • John McLaughlin
• egens • mhemantha
• Andre Espaze • michelemaroni89
• André Espaze • Mohammad
Release: 2022.1 • Loïc Estève • Zachary Moon
The lecture notes are archived on zenodo: http://dx.doi.org/10.5281/zenodo.594102 • Corey Farwell • Mike Mueller
All code and material is licensed under a Creative Commons Attribution 4.0 International License (CC- • Tim Gates • negm
by) http://creativecommons.org/licenses/by/4.0/ • Stuart Geiger • John B Nelson
22.1 Authors • Mike Müller • Olivier Georg • nicoguaro

• Fabian Pedregosa • Daniel Gerigk • Sergio Oller
22.1.1 Editors
• Didrik Pinte • Robert Gieseke • Theofilos Papapanagiotou
• Gaël Varoquaux
• Nicolas Rougier • Philip Gillißen • patniharshit
• Emmanuelle Gouillart • Ralf Gommers • Fabian Pedregosa
• Gaël Varoquaux
• Olav Vahtras • Emmanuelle Gouillart • Philippe Pepiot
• Pauli Virtanen
• Pierre de Buyl • Julia Gustavsen • Tiago M. D. Pereira
• Zbigniew Jędrzejewski-Szmek
• Valentin Haenel (editor from 2011 to 2015) • Omar Gutiérrez • Nicolas Pettiaux
22.1.2 Chapter authors
• Valentin Haenel • Didrik Pinte
Listed by alphabetical order. 22.1.3 Additional Contributions • Pierre Haessig • Evgeny Pogrebnyak
• Christopher Burns Listed by alphabetical order • Bruno Hanzen • reverland
• Adrian Chauve • Osayd Abdu • Michael Hartmann • Maximilien Riehl
• Robert Cimrman • arunpersaud • Jonathan Helmus • Kristian Rother
• Christophe Combelles • Ross Barnowski • Andreas Hilboll • Nicolas P. Rougier
• André Espaze • Sebastian Berg • Himanshu • Nicolas Rougier
• Emmanuelle Gouillart • Lilian Besson • Julian Hofer • Rutzmoser
666 22.1. Authors 667

• Sander • Wes Turner

• João Felipe Santos • Akihiro Uchida
• Mark Setchell • Utkarsh Upadhyay
• Helen Sherwood-Taylor • Olav Vahtras
• Shoeboxam
• Gael Varoquaux Index
• Simon
• Nelle Varoquaux
• solarjoe
• Olivier Verdier
• ssmiller
• VirgileFritsch
• Scott Staniewicz
• strpeter • Pauli Virtanen
• surfer190 • Yosh Wakeham
• Bartosz Telenczuk • Stefan van der Walt D

• tommyod • yasutomo57jp diff, 519, 522
differentiation, 519
dsolve, 522
E
equations
algebraic, 521
differential, 522
I
integration, 520
M
Matrix, 522
P
Python Enhancement Proposals
PEP 255, 268
PEP 3118, 304
PEP 3129, 277
PEP 318, 270, 277
PEP 342, 268
PEP 343, 278
PEP 380, 269
PEP 380#id13, 269
PEP 8, 272
S
solve, 521
22.1. Authors 668 669

Scip y Lectures

Uploaded by

Copyright:

Available Formats

Scip y Lectures

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Scip y Lectures

Uploaded by

Copyright:

Available Formats

Contents

SciKits Numpy Matplotlib I Getting started with Python for science 2

2020 1.4 The workflow: interactive environments and text editors . . . . . . . . . . . . . . . . . . 8

3 Python 2 and Python 3 47

4 NumPy: creating and manipulating numerical data 50

and man y ot hers . ..

20.1 Mlab: the scripting interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 568

21 scikit-learn: machine learning in Python 582

IV About the SciPy lecture notes 665

Getting started with Python for

1.1.3 How does Python compare to other solutions?

CHAPTER • Painful usage: no interactivity during development, mandatory compilation steps,

Matlab scripting language

Python scientific computing ecosystem guage.

1.1 Why Python? Pros

Other scripting languages: Scilab, Octave, R, IDL, etc.

4 1.1. Why Python? 5

1.2 The Scientific Python ecosystem

Python, a generic and modern computing language

Core numeric libraries

• sympy for symbolic computing

To execute code, press “shift enter” From a script to functions

>>> print("Hello, world!")

To get yourself started, type the following stack of instructions

12 2.1. First steps 13

(continued from previous page)

2.2 Basic types Warning: Integer division

Future behavior: to always get the behavior of Python3

2.2. Basic types 14 2.2. Basic types 15

Tip: All slicing parameters are optional: Reverse:

>>> colors >>> rcolors = colors[::-1]

>>> colors[0] = 'yellow' >>> rcolors + colors

>>> sorted(rcolors) # new object

2.2. Basic types 16 2.2. Basic types 17

(continued from previous page) (continued from previous page)

Methods and Object-Oriented Programming

In [53]: a = "hello, world!"

rules. 'An integer: 1; a float: 0.100000; another string: string'

2.2. Basic types 18 2.2. Basic types 19

• to change a list in place, use indexing/slices:

Sets: unordered, unique items:

Tip: Python library reference says:

2.2. Basic types 20 2.3. Control Flow 21

But most often, it is more readable to iterate over values: >>> 1 is 1.

a in b For any collection b: b contains a

More advanced features 2.3.5 Advanced iteration

2.3. Control Flow 22 2.3. Control Flow 23

Looping over a dictionary

>>> d = {'a': 1, 'b':1.2, 'c':1j}

2.3.6 List Comprehensions

2.3. Control Flow 24 2.4. Defining functions 25

In [81]: def double_it(x):

In [84]: def double_it(x=2): In [103]: slicer(rhyme)

The order of the keyword arguments does not matter:

In [126]: bigx = 1e9 # Now really big

In [4]: add_to_dict() >>> def try_to_modify(x, y, z):

2.4. Defining functions 26 2.4. Defining functions 27

(continued from previous page) (continued from previous page)

In [115]: def addx(y): In [67]: def funcname(params):

2.4. Defining functions 28 2.4. Defining functions 29