SciKits Numpy Matplotlib

2024
SciPy Python EDITION

IP[y]:

Cython IPython
Edited by

Scien�fic Python Gaël Varoquaux

Emmanuelle Gouillart
Olaf Vahtras

Lectures
Pierre de Buyl
K. Jarrod Millman
Stéfan van der Walt

lectures.scientific-python.org

Gaël Varoquaux • Emmanuelle Gouillart • Olav Vahtras

Pierre de Buyl • K. Jarrod Millman • Stéfan van der Walt
Christopher Burns • Adrian Chauve • Robert Cimrman • Christophe Combelles
Ralf Gommers • André Espaze • Zbigniew Jędrzejewski-Szmek
Valen� n Haenel • Michael Hartmann • Gert-Ludwig Ingold • Fabian Pedregosa
Didrik Pinte • Nicolas P. Rougier • Joris Van den Bossche • Pauli Virtanen
and many others...
 Contents

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 . . . . . . . . . . . . . . . . . . . . . . 7
1.4 The workflow: interactive environments and text editors . . . . . . . . . . . . . . . . . . 8

2 The Python language 12

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 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.6 Input and Output . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
2.7 Standard Library . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.8 Exception handling in Python . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
2.9 Object-oriented programming (OOP) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

3 NumPy: creating and manipulating numerical data 53

3.1 The NumPy array object . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.2 Numerical operations on arrays . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
3.3 More elaborate arrays . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
3.4 Advanced operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
3.5 Some exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
3.6 Full code examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94

4 Matplotlib: plotting 106

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
4.2 Simple plot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
4.3 Figures, Subplots, Axes and Ticks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
4.4 Other Types of Plots: examples and exercises . . . . . . . . . . . . . . . . . . . . . . . . . 118
4.5 Beyond this tutorial . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
4.6 Quick references . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
4.7 Full code examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132

5 SciPy : high-level scientific computing 210

5.1 File input/output: scipy.io . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
5.2 Special functions: scipy.special . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212

i
 5.3 Linear algebra operations: scipy.linalg . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
5.4 Interpolation: scipy.interpolate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
5.5 Optimization and fit: scipy.optimize . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
5.6 Statistics and random numbers: scipy.stats . . . . . . . . . . . . . . . . . . . . . . . . 222
5.7 Numerical integration: scipy.integrate . . . . . . . . . . . . . . . . . . . . . . . . . . . 224
5.8 Fast Fourier transforms: scipy.fft . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 226
5.9 Signal processing: scipy.signal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
5.10 Image manipulation: scipy.ndimage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
5.11 Summary exercises on scientific computing . . . . . . . . . . . . . . . . . . . . . . . . . . 236
5.12 Full code examples for the SciPy chapter . . . . . . . . . . . . . . . . . . . . . . . . . . . 249

6 Getting help and finding documentation 292

II Advanced topics 295

7 Advanced Python Constructs 297
7.1 Iterators, generator expressions and generators . . . . . . . . . . . . . . . . . . . . . . . . 298
7.2 Decorators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 303
7.3 Context managers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311

8 Advanced NumPy 315

8.1 Life of ndarray . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 316
8.2 Universal functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 330
8.3 Interoperability features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 340
8.4 Array siblings: chararray, maskedarray . . . . . . . . . . . . . . . . . . . . . . . . . . . 343
8.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 346
8.6 Contributing to NumPy/SciPy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 346

9 Debugging code 349

9.1 Avoiding bugs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 350
9.2 Debugging workflow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 352
9.3 Using the Python debugger . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 353
9.4 Debugging segmentation faults using gdb . . . . . . . . . . . . . . . . . . . . . . . . . . . 358

10 Optimizing code 361

10.1 Optimization workflow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 362
10.2 Profiling Python code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 362
10.3 Making code go faster . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 365
10.4 Writing faster numerical code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 366

11 Sparse Arrays in SciPy 369

11.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 369
11.2 Storage Schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 372
11.3 Linear System Solvers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 384
11.4 Other Interesting Packages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 390

12 Image manipulation and processing using NumPy and SciPy 391

12.1 Opening and writing to image files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 392
12.2 Displaying images . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 394
12.3 Basic manipulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 395
12.4 Image filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 397
12.5 Feature extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 403
12.6 Measuring objects properties: scipy.ndimage.measurements . . . . . . . . . . . . . . . . 406
12.7 Full code examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 409
12.8 Examples for the image processing chapter . . . . . . . . . . . . . . . . . . . . . . . . . . 409

13 Mathematical optimization: finding minima of functions 436

13.1 Knowing your problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 437
13.2 A review of the different optimizers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 440

ii
 13.3 Full code examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 446
13.4 Examples for the mathematical optimization chapter . . . . . . . . . . . . . . . . . . . . 446
13.5 Practical guide to optimization with SciPy . . . . . . . . . . . . . . . . . . . . . . . . . . 486
13.6 Special case: non-linear least-squares . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 488
13.7 Optimization with constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 490
13.8 Full code examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 491
13.9 Examples for the mathematical optimization chapter . . . . . . . . . . . . . . . . . . . . 491

14 Interfacing with C 492

14.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 493
14.2 Python-C-Api . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 493
14.3 Ctypes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 501
14.4 SWIG . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 505
14.5 Cython . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 510
14.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 515
14.7 Further Reading and References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 515
14.8 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 515

III Packages and applications 517

15 Statistics in Python 519
15.1 Data representation and interaction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 520
15.2 Hypothesis testing: comparing two groups . . . . . . . . . . . . . . . . . . . . . . . . . . 525
15.3 Linear models, multiple factors, and analysis of variance . . . . . . . . . . . . . . . . . . . 528
15.4 More visualization: seaborn for statistical exploration . . . . . . . . . . . . . . . . . . . . 534
15.5 Testing for interactions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 537
15.6 Full code for the figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 538
15.7 Solutions to this chapter’s exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 566

16 Sympy : Symbolic Mathematics in Python 569

16.1 First Steps with SymPy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 570
16.2 Algebraic manipulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 571
16.3 Calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 572
16.4 Equation solving . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 575
16.5 Linear Algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 576

17 scikit-image: image processing 578

17.1 Introduction and concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 579
17.2 Importing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 581
17.3 Example data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 582
17.4 Input/output, data types and colorspaces . . . . . . . . . . . . . . . . . . . . . . . . . . . 582
17.5 Image preprocessing / enhancement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 584
17.6 Image segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 588
17.7 Measuring regions’ properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 591
17.8 Data visualization and interaction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 591
17.9 Feature extraction for computer vision . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 592
17.10 Full code examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 593
17.11 Examples for the scikit-image chapter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 593

18 scikit-learn: machine learning in Python 605

18.1 Introduction: problem settings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 606
18.2 Basic principles of machine learning with scikit-learn . . . . . . . . . . . . . . . . . . . . 610
18.3 Supervised Learning: Classification of Handwritten Digits . . . . . . . . . . . . . . . . . . 616
18.4 Supervised Learning: Regression of Housing Data . . . . . . . . . . . . . . . . . . . . . . 620
18.5 Measuring prediction performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 625
18.6 Unsupervised Learning: Dimensionality Reduction and Visualization . . . . . . . . . . . . 631
18.7 Parameter selection, Validation, and Testing . . . . . . . . . . . . . . . . . . . . . . . . . 634
18.8 Examples for the scikit-learn chapter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 642

iii
IV About the Scientific Python Lectures 690
19 About the Scientific Python Lectures 691
19.1 Authors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 691

Index 694

iv
 Scientific Python Lectures, Edition 2024.2rc0.dev0

Contents 1
 Part I

Getting started with Python for

science

2
 Scientific Python Lectures, Edition 2024.2rc0.dev0

This part of the Scientific Python Lectures is a self-contained introduction to everything that is needed
to use Python for science, from the language itself, to numerical computing or plotting.

3
 CHAPTER 1
Python scientific computing ecosystem

Authors: Fernando Perez, Emmanuelle Gouillart, Gaël Varoquaux, Valentin Haenel

1.1 Why Python?

1.1.1 The scientist’s needs

• Get data (simulation, experiment control),

• Manipulate and process data,
• Visualize results, quickly to understand, but also with high quality figures, for reports or publica-
tions.

1.1.2 Python’s strengths

• 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
transform or a fitting algorithm. Don’t reinvent the wheel!
• Easy to learn Most scientists are not paid as programmers, neither have they been trained so.
They need to be able to draw a curve, smooth a signal, do a Fourier transform in a few minutes.
• 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
strange symbols or lengthy routine specifications that would divert the reader from mathematical
or scientific understanding of the code.

4
 Scientific Python Lectures, Edition 2024.2rc0.dev0

• 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.
• 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?

Compiled languages: C, C++, Fortran. . .

Pros
• Very fast. For heavy computations, it’s difficult to outperform these languages.
Cons
• 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-
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 and not everything is open sourced.

Julia

Pros
• Fast code, yet interactive and simple.
• Easily connects to Python or C.
Cons
• Ecosystem limited to numerical computing.
• Still young.

1.1. Why Python? 5

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Other scripting languages: Scilab, Octave, R, IDL, etc.

Pros
• Open-source, free, or at least cheaper than Matlab.
• Some features can be very advanced (statistics in R, etc.)
Cons
• Fewer available algorithms than in Matlab, and the language is not more advanced.
• Some software are dedicated to one domain. Ex: Gnuplot to draw curves. These
programs are very powerful, but they are restricted to a single type of usage, such
as plotting.

Python

Pros
• Very rich scientific computing libraries
• Well thought out language, allowing to write very readable and well structured code:
we “code what we think”.
• Many libraries beyond scientific computing (web server, serial port access, etc.)
• Free and open-source software, widely spread, with a vibrant community.
• 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.
• Modules of the standard library: string processing, file management, simple network protocols.
• A large number of specialized modules or applications written in Python: web framework, etc. . . .
and scientific computing.
• Development tools (automatic testing, documentation generation)
See also:
chapter on Python language
Core numeric libraries

1.2. The scientific Python ecosystem 6

 Scientific Python Lectures, Edition 2024.2rc0.dev0

• NumPy: numerical computing with powerful numerical arrays objects, and routines to manip-
ulate them. https://numpy.org/
See also:
chapter on numpy
• SciPy : high-level numerical routines. Optimization, regression, interpolation, etc https://scipy.
org/
See also:
chapter on SciPy
• Matplotlib : 2-D visualization, “publication-ready” plots https://matplotlib.org/
See also:
chapter on matplotlib
Advanced interactive environments:
• IPython, an advanced Python console https://ipython.org/
• Jupyter, notebooks in the browser https://jupyter.org/
Domain-specific packages,
• pandas, statsmodels, seaborn for statistics
• sympy for symbolic computing
• scikit-image for image processing
• scikit-learn for machine learning
and many more packages not documented in the Scientific Python Lectures.
See also:
chapters on advanced topics
chapters on packages and applications

1.3 Before starting: Installing a working environment

Python comes in many flavors, and there are many ways to install it. However, we recommend to install
a scientific-computing distribution, that comes readily with optimized versions of scientific modules.
Under Linux
If you have a recent distribution, most of the tools are probably packaged, and it is recommended to use
your package manager.
Other systems
There are several fully-featured scientific Python distributions:
• Anaconda
• WinPython

1.3. Before starting: Installing a working environment 7

 Scientific Python Lectures, Edition 2024.2rc0.dev0

1.4 The workflow: interactive environments and text editors

Interactive work to test and understand algorithms: In this section, we describe a workflow
combining interactive work and consolidation.
Python is a general-purpose language. As such, there is not one blessed environment to work in, and
not only one way of using it. Although this makes it harder for beginners to find their way, it makes it
possible for Python to be used for programs, in web servers, or embedded devices.

1.4.1 Interactive work

We recommend an interactive work with the IPython console, or its offspring, the Jupyter notebook.
They are handy to explore and understand algorithms.

Under the notebook

To execute code, press “shift enter”

Start ipython:

In [1]: print('Hello world')

Hello world

Getting help by using the ? operator after an object:

In [2]: print?
Signature: print(*args, sep=' ', end='\n', file=None, flush=False)
Docstring:
Prints the values to a stream, or to sys.stdout by default.

sep
string inserted between values, default a space.
end
string appended after the last value, default a newline.
file
a file-like object (stream); defaults to the current sys.stdout.
flush
whether to forcibly flush the stream.
Type: builtin_function_or_method

See also:
• IPython user manual: https://ipython.readthedocs.io/en/stable/
• Jupyter Notebook QuickStart: https://docs.jupyter.org/en/latest/start/index.html

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
Python files. For this, a powerful code editor will get you far. Here are several good easy-to-use editors:
• Spyder: integrates an IPython console, a debugger, a profiler. . .
• PyCharm: integrates an IPython console, notebooks, a debugger. . . (freely available, but commer-
cial)
• Visual Studio Code: integrates a Python console, notebooks, a debugger, . . .

1.4. The workflow: interactive environments and text editors 8

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Some of these are shipped by the various scientific Python distributions, and you can find them in the
menus.
As an exercise, create a file my_file.py in a code editor, and add the following lines:

s = 'Hello world'
print(s)

Now, you can run it in IPython console or a notebook and explore the resulting variables:

In [3]: %run my_file.py

Hello world

In [4]: s
Out[4]: 'Hello world'

In [5]: %whos
Variable Type Data/Info
----------------------------
s str Hello world

From a script to functions

While it is tempting to work only with scripts, that is a file full of instructions following each other,
do plan to progressively evolve the script to a set of functions:
• A script is not reusable, functions are.
• Thinking in terms of functions helps breaking the problem in small blocks.

1.4.3 IPython and Jupyter Tips and Tricks

The user manuals contain a wealth of information. Here we give a quick introduction to four useful
features: history, tab completion, magic functions, and aliases.

Command history Like a UNIX shell, the IPython console supports command history. Type up and
down to navigate previously typed commands:

In [6]: x = 10

In [7]: <UP>

In [8]: x = 10

Tab completion Tab completion, is a convenient way to explore the structure of any object you’re
dealing with. Simply type object_name.<TAB> to view the object’s attributes. Besides Python objects
and keywords, tab completion also works on file and directory names.*

1.4. The workflow: interactive environments and text editors 9

 Scientific Python Lectures, Edition 2024.2rc0.dev0

In [9]: x = 10

In [10]: x.<TAB>
as_integer_ratio() conjugate() imag to_bytes()
bit_count() denominator numerator
bit_length() from_bytes() real

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.
Other useful magic functions are:
• %cd to change the current directory.

In [11]: cd /tmp
/tmp

• %cpaste allows you to paste code, especially code from websites which has been prefixed with the
standard Python prompt (e.g. >>>) or with an ipython prompt, (e.g. in [3]):

In [12]: %cpaste

• %timeit allows you to time the execution of short snippets using the timeit module from the
standard library:

In [12]: %timeit x = 10
11.6 ns +- 0.818 ns per loop (mean +- std. dev. of 7 runs, 100,000,000 loops␣
˓→each)

See also:
Chapter on optimizing code
• %debug allows you to enter post-mortem debugging. That is to say, if the code you try to execute,
raises an exception, using %debug will enter the debugger at the point where the exception was
thrown.

In [13]: x === 10
Cell In[13], line 1
x === 10
^
SyntaxError: invalid syntax

In [14]: %debug
> /home/jarrod/.venv/lectures/lib64/python3.11/site-packages/IPython/core/
˓→compilerop.py(86)ast_parse()

84 Arguments are exactly the same as ast.parse (in the standard␣

˓→library),

85 and are passed to the built-in compile function."""

---> 86 return compile(source, filename, symbol, self.flags | PyCF_ONLY_
˓→AST, 1)

87
(continues on next page)

1.4. The workflow: interactive environments and text editors 10

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

88 def reset_compiler_flags(self):
ipdb> locals()
{'self': <IPython.core.compilerop.CachingCompiler object at 0x7f30d02efc10>,
˓→'source': 'x === 10\n', 'filename': '<ipython-input-1-8e8bc565444b>', 'symbol

˓→': 'exec'}

ipdb>

See also:
Chapter on debugging

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
typing alias).

Getting help

• The built-in cheat-sheet is accessible via the %quickref magic function.

• A list of all available magic functions is shown when typing %magic.

1.4. The workflow: interactive environments and text editors 11

 CHAPTER 2
The Python language

Authors: Chris Burns, Christophe Combelles, Emmanuelle Gouillart, Gaël Varoquaux

Python for scientific computing

We introduce here the Python language. Only the bare minimum necessary for getting started with
NumPy and SciPy is addressed here. To learn more about the language, consider going through the
excellent tutorial https://docs.python.org/3/tutorial. Dedicated books are also available, such as Dive
into Python 3.

Tip: Python is a programming language, as are C, Fortran, BASIC, PHP, etc. Some specific
features of Python are as follows:
• an interpreted (as opposed to compiled) language. Contrary to e.g. C or Fortran, one does not
compile Python code before executing it. In addition, Python can be used interactively: many
Python interpreters are available, from which commands and scripts can be executed.
• a free software released under an open-source license: Python can be used and distributed free
of charge, even for building commercial software.
• multi-platform: Python is available for all major operating systems, Windows, Linux/Unix,
MacOS X, most likely your mobile phone OS, etc.
• a very readable language with clear non-verbose syntax

12
 Scientific Python Lectures, Edition 2024.2rc0.dev0

• a language for which a large variety of high-quality packages are available for various applications,
from web frameworks to scientific computing.
• a language very easy to interface with other languages, in particular C and C++.
• Some other features of the language are illustrated just below. For example, Python is an object-
oriented language, with dynamic typing (the same variable can contain objects of different types
during the course of a program).
See https://www.python.org/about/ for more information about distinguishing features of Python.

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. the Anaconda Navigator, the Python(x,y) 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.

Once you have started the interpreter, type

>>> print("Hello, world!")

Hello, world!

Tip: The message “Hello, world!” is then displayed. You just executed your first Python instruction,
congratulations!

To get yourself started, type the following stack of instructions

>>> a = 3
>>> b = 2*a
>>> type(b)
<class 'int'>
>>> print(b)
6
>>> a*b
18
>>> b = 'hello'
>>> type(b)
<class 'str'>
>>> b + b
'hellohello'
>>> 2*b
'hellohello'

Tip: Two variables a and b have been defined above. Note that one does not declare the type of a
variable before assigning its value. In C, conversely, one should write:

2.1. First steps 13

 Scientific Python Lectures, Edition 2024.2rc0.dev0

int a = 3;

In addition, the type of a variable may change, in the sense that at one point in time it can be equal
to a value of a certain type, and a second point in time, it can be equal to a value of a different type.
b was first equal to an integer, but it became equal to a string when it was assigned the value ‘hello’.
Operations on integers (b=2*a) are coded natively in Python, and so are some operations on strings such
as additions and multiplications, which amount respectively to concatenation and repetition.

2.2 Basic types

2.2.1 Numerical types

Tip: Python supports the following numerical, scalar types:

Integer

>>> 1 + 1
2
>>> a = 4
>>> type(a)
<class 'int'>

Floats

>>> c = 2.1
>>> type(c)
<class 'float'>

Complex

>>> a = 1.5 + 0.5j

>>> a.real
1.5
>>> a.imag
0.5
>>> type(1. + 0j)
<class 'complex'>

Booleans

>>> 3 > 4
False
>>> test = (3 > 4)
>>> test
False
>>> type(test)
<class 'bool'>

Tip: A Python shell can therefore replace your pocket calculator, with the basic arithmetic operations
+, -, *, /, % (modulo) natively implemented

2.2. Basic types 14

 Scientific Python Lectures, Edition 2024.2rc0.dev0

>>> 7 * 3.
21.0
>>> 2**10
1024
>>> 8 % 3
2

Type conversion (casting):

>>> float(1)
1.0

2.2.2 Containers

Tip: Python provides many efficient types of containers, in which collections of objects can be stored.

Lists

Tip: A list is an ordered collection of objects, that may have different types. For example:

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

>>> type(colors)
<class 'list'>

Indexing: accessing individual objects contained in the list:

>>> colors[2]
'green'

Counting from the end with negative indices:

>>> colors[-1]
'white'
>>> colors[-2]
'black'

Warning: Indexing starts at 0 (as in C), not at 1 (as in Fortran or Matlab)!

Slicing: obtaining sublists of regularly-spaced elements:

>>> colors
['red', 'blue', 'green', 'black', 'white']
>>> colors[2:4]
['green', 'black']

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

2.2. Basic types 15

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Slicing syntax: colors[start:stop:stride]

Tip: All slicing parameters are optional:

>>> colors
['red', 'blue', 'green', 'black', 'white']
>>> colors[3:]
['black', 'white']
>>> colors[:3]
['red', 'blue', 'green']
>>> colors[::2]
['red', 'green', 'white']

Lists are mutable objects and can be modified:

>>> colors[0] = 'yellow'
>>> colors
['yellow', 'blue', 'green', 'black', 'white']
>>> colors[2:4] = ['gray', 'purple']
>>> colors
['yellow', 'blue', 'gray', 'purple', 'white']

Note: The elements of a list may have different types:

>>> colors = [3, -200, 'hello']
>>> colors
[3, -200, 'hello']
>>> colors[1], colors[2]
(-200, 'hello')

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
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.

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/3/tutorial/datastructures.html#more-on-lists

Add and remove elements:

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

>>> colors.append('pink')
>>> colors
['red', 'blue', 'green', 'black', 'white', 'pink']
>>> colors.pop() # removes and returns the last item
'pink'
>>> colors
['red', 'blue', 'green', 'black', 'white']
>>> colors.extend(['pink', 'purple']) # extend colors, in-place
>>> colors
['red', 'blue', 'green', 'black', 'white', 'pink', 'purple']
>>> colors = colors[:-2]
(continues on next page)

2.2. Basic types 16

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

>>> colors
['red', 'blue', 'green', 'black', 'white']

Reverse:

>>> rcolors = colors[::-1]

>>> rcolors
['white', 'black', 'green', 'blue', 'red']
>>> rcolors2 = list(colors) # new object that is a copy of colors in a different␣
˓→memory area

>>> rcolors2
['red', 'blue', 'green', 'black', 'white']
>>> rcolors2.reverse() # in-place; reversing rcolors2 does not affect colors
>>> rcolors2
['white', 'black', 'green', 'blue', 'red']

Concatenate and repeat lists:

>>> rcolors + colors

['white', 'black', 'green', 'blue', 'red', 'red', 'blue', 'green', 'black', 'white']
>>> rcolors * 2
['white', 'black', 'green', 'blue', 'red', 'white', 'black', 'green', 'blue', 'red']

Tip: Sort:

>>> sorted(rcolors) # new object

['black', 'blue', 'green', 'red', 'white']
>>> rcolors
['white', 'black', 'green', 'blue', 'red']
>>> rcolors.sort() # in-place
>>> rcolors
['black', 'blue', 'green', 'red', 'white']

Methods and Object-Oriented Programming

The notation rcolors.method() (e.g. rcolors.append(3) and colors.pop()) is our first example
of object-oriented programming (OOP). Being a list, the object rcolors owns the method function
that is called using the notation .. No further knowledge of OOP than understanding the notation .
is necessary for going through this tutorial.

Discovering methods:

Reminder: in Ipython: tab-completion (press tab)

In [1]: rcolors.<TAB>
append() count() insert() reverse()
clear() extend() pop() sort()
copy() index() remove()

2.2. Basic types 17

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Strings

Different string syntaxes (simple, double or triple quotes):

s = 'Hello, how are you?'

s = "Hi, what's up"
s = '''Hello,
how are you''' # tripling the quotes allows the
# string to span more than one line
s = """Hi,
what's up?"""

In [2]: 'Hi, what's up?'

Cell In[2], line 1
'Hi, what's up?'
^
SyntaxError: unterminated string literal (detected at line 1)

This syntax error can be avoided by enclosing the string in double quotes instead of single quotes.
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.

Tip: Strings are collections like lists. Hence they can be indexed and sliced, using the same syntax and
rules.

Indexing:

>>> a = "hello"
>>> a[0]
'h'
>>> a[1]
'e'
>>> a[-1]
'o'

Tip: (Remember that negative indices correspond to counting from the right end.)

Slicing:

>>> a = "hello, world!"

>>> a[3:6] # 3rd to 6th (excluded) elements: elements 3, 4, 5
'lo,'
>>> a[2:10:2] # Syntax: a[start:stop:step]
'lo o'
>>> a[::3] # every three characters, from beginning to end
'hl r!'

Tip: Accents and special characters can also be handled as in Python 3 strings consist of Unicode
characters.

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.

2.2. Basic types 18

 Scientific Python Lectures, Edition 2024.2rc0.dev0

In [3]: a = "hello, world!"

In [4]: a.replace('l', 'z', 1)

Out[4]: 'hezlo, world!'

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.

See also:
Python offers advanced possibilities for manipulating strings, looking for patterns or formatting. The
interested reader is referred to https://docs.python.org/3/library/stdtypes.html#string-methods and
https://docs.python.org/3/library/string.html#format-string-syntax
String formatting:

>>> 'An integer: %i ; a float: %f ; another string: %s ' % (1, 0.1, 'string') # with␣
˓→more values use tuple after %

'An integer: 1; a float: 0.100000; another string: string'

>>> i = 102
>>> filename = 'processing_of_dataset_%d .txt' % i # no need for tuples with just␣
˓→one value after %

>>> filename
'processing_of_dataset_102.txt'

Dictionaries

Tip: A dictionary is basically an efficient table that maps keys to values.

>>> tel = {'emmanuelle': 5752, 'sebastian': 5578}

>>> tel['francis'] = 5915
>>> tel
{'emmanuelle': 5752, 'sebastian': 5578, 'francis': 5915}
>>> tel['sebastian']
5578
>>> tel.keys()
dict_keys(['emmanuelle', 'sebastian', 'francis'])
>>> tel.values()
dict_values([5752, 5578, 5915])
>>> 'francis' in tel
True

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/3/tutorial/datastructures.html#dictionaries for more
information.
A dictionary can have keys (resp. values) with different types:

>>> d = {'a':1, 'b':2, 3:'hello'}

>>> d
{'a': 1, 'b': 2, 3: 'hello'}

2.2. Basic types 19

 Scientific Python Lectures, Edition 2024.2rc0.dev0

More container types

Tuples
Tuples are basically immutable lists. The elements of a tuple are written between parentheses, or just
separated by commas:

>>> t = 12345, 54321, 'hello!'

>>> t[0]
12345
>>> t
(12345, 54321, 'hello!')
>>> u = (0, 2)

Sets: unordered, unique items:

>>> s = set(('a', 'b', 'c', 'a'))

>>> s
{'a', 'b', 'c'}
>>> s.difference(('a', 'b'))
{'c'}

2.2.3 Assignment operator

Tip: Python library reference says:

Assignment statements are used to (re)bind names to values and to modify attributes or
items of mutable objects.
In short, it works as follows (simple assignment):
1. an expression on the right hand side is evaluated, the corresponding object is created/obtained
2. a name on the left hand side is assigned, or bound, to the r.h.s. object

Things to note:
• a single object can have several names bound to it:

In [5]: a = [1, 2, 3]

• to change a list in place, use indexing/slices:

In [6]: a = [1, 2, 3]

• the key concept here is mutable vs. immutable

– mutable objects can be changed in place
– immutable objects cannot be modified once created
See also:
A very good and detailed explanation of the above issues can be found in David M. Beazley’s article
Types and Objects in Python.

2.2. Basic types 20

 Scientific Python Lectures, Edition 2024.2rc0.dev0

2.3 Control Flow

Controls the order in which the code is executed.

2.3.1 if/elif/else

>>> if 2**2 == 4:
... print("Obvious!")
...
Obvious!

Blocks are delimited by indentation

Tip: Type the following lines in your Python interpreter, and be careful to respect the indentation
depth. The Ipython shell automatically increases the indentation depth after a colon : sign; to decrease
the indentation depth, go four spaces to the left with the Backspace key. Press the Enter key twice to
leave the logical block.

>>> a = 10

>>> if a == 1:
... print(1)
... elif a == 2:
... print(2)
... else:
... print("A lot")
...
A lot

Indentation is compulsory in scripts as well. As an exercise, re-type the previous lines with the same
indentation in a script condition.py, and execute the script with run condition.py in Ipython.

2.3.2 for/range

Iterating with an index:

>>> for i in range(4):

... print(i)
0
1
2
3

But most often, it is more readable to iterate over values:

>>> for word in ('cool', 'powerful', 'readable'):

... print('Python is %s ' % word)
Python is cool
Python is powerful
Python is readable

2.3. Control Flow 21

 Scientific Python Lectures, Edition 2024.2rc0.dev0

2.3.3 while/break/continue

Typical C-style while loop (Mandelbrot problem):

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

More advanced features

break out of enclosing for/while loop:

>>> z = 1 + 1j

>>> while abs(z) < 100:

... if z.imag == 0:
... break
... z = z**2 + 1

continue the next iteration of a loop.:

>>> a = [1, 0, 2, 4]
>>> for element in a:
... if element == 0:
... continue
... print(1. / element)
1.0
0.5
0.25

2.3.4 Conditional Expressions

if <OBJECT>
Evaluates to False:
• any number equal to zero (0, 0.0, 0+0j)
• an empty container (list, tuple, set, dictionary, . . . )
• False, None
Evaluates to True:
• everything else
a == b
Tests equality, with logics:

>>> 1 == 1.
True

a is b
Tests identity: both sides are the same object:

>>> a = 1
>>> b = 1.
>>> a == b
True
(continues on next page)

2.3. Control Flow 22

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

>>> a is b
False

>>> a = 1
>>> b = 1
>>> a is b
True

a in b
For any collection b: b contains a

>>> b = [1, 2, 3]
>>> 2 in b
True
>>> 5 in b
False

If b is a dictionary, this tests that a is a key of b.

2.3.5 Advanced iteration

Iterate over any sequence

You can iterate over any sequence (string, list, keys in a dictionary, lines in a file, . . . ):

>>> vowels = 'aeiouy'

>>> for i in 'powerful':

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

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

>>> message.split() # returns a list
['Hello', 'how', 'are', 'you?']
>>> for word in message.split():
... print(word)
...
Hello
how
are
you?

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.

Warning: Not safe to modify the sequence you are iterating over.

2.3. Control Flow 23

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Keeping track of enumeration number

Common task is to iterate over a sequence while keeping track of the item number.
• Could use while loop with a counter as above. Or a for loop:

>>> words = ('cool', 'powerful', 'readable')

>>> for i in range(0, len(words)):
... print((i, words[i]))
(0, 'cool')
(1, 'powerful')
(2, 'readable')

• But, Python provides a built-in function - enumerate - for this:

>>> for index, item in enumerate(words):

... print((index, item))
(0, 'cool')
(1, 'powerful')
(2, 'readable')

Looping over a dictionary

Use items:

>>> d = {'a': 1, 'b':1.2, 'c':1j}

>>> for key, val in sorted(d.items()):

... print('Key: %s has value: %s ' % (key, val))
Key: a has value: 1
Key: b has value: 1.2
Key: c has value: 1j

Note: The ordering of a dictionary is random, thus we use sorted() which will sort on the keys.

2.3.6 List Comprehensions

Instead of creating a list by means of a loop, one can make use of a list comprehension with a rather
self-explaining syntax.

>>> [i**2 for i in range(4)]

[0, 1, 4, 9]

Exercise

Compute the decimals of Pi using the Wallis formula:

∞
∏︁ 4𝑖2
𝜋=2
𝑖=1
4𝑖2 − 1

2.3. Control Flow 24

 Scientific Python Lectures, Edition 2024.2rc0.dev0

2.4 Defining functions

2.4.1 Function definition

In [1]: def test():

...: print('in test function')
...:
...:

In [2]: test()
in test function

Warning: Function blocks must be indented as other control-flow blocks.

2.4.2 Return statement

Functions can optionally return values.

In [3]: def disk_area(radius):

...: return 3.14 * radius * radius
...:

In [4]: disk_area(1.5)
Out[4]: 7.0649999999999995

Note: By default, functions return None.

Note: Note the syntax to define a function:

• the def keyword;
• is followed by the function’s name, then
• the arguments of the function are given between parentheses followed by a colon.
• the function body;
• and return object for optionally returning values.

2.4.3 Parameters

Mandatory parameters (positional arguments)

In [5]: def double_it(x):

...: return x * 2
...:

In [6]: double_it(3)
Out[6]: 6

In [7]: double_it()
---------------------------------------------------------------------------
(continues on next page)

2.4. Defining functions 25

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

TypeError Traceback (most recent call last)
Cell In[7], line 1
----> 1 double_it()

TypeError: double_it() missing 1 required positional argument: 'x'

Optional parameters (keyword or named arguments)

In [8]: def double_it(x=2):

...: return x * 2
...:

In [9]: double_it()
Out[9]: 4

In [10]: double_it(3)
Out[10]: 6

Keyword arguments allow you to specify default values.

Warning: Default values are evaluated when the function is defined, not when it is called. This
can be problematic when using mutable types (e.g. dictionary or list) and modifying them in the
function body, since the modifications will be persistent across invocations of the function.
Using an immutable type in a keyword argument:
In [11]: bigx = 10

In [12]: def double_it(x=bigx):

....: return x * 2
....:

In [13]: bigx = 1e9 # Now really big

In [14]: double_it()
Out[14]: 20

Using an mutable type in a keyword argument (and modifying it inside the function body):
In [15]: def add_to_dict(args={'a': 1, 'b': 2}):
....: for i in args.keys():
....: args[i] += 1
....: print(args)
....:

In [16]: add_to_dict
Out[16]: <function __main__.add_to_dict(args={'a': 1, 'b': 2})>

In [17]: add_to_dict()
{'a': 2, 'b': 3}

In [18]: add_to_dict()
{'a': 3, 'b': 4}

In [19]: add_to_dict()
{'a': 4, 'b': 5}

2.4. Defining functions 26

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Tip: More involved example implementing python’s slicing:

In [20]: def slicer(seq, start=None, stop=None, step=None):
....: """Implement basic python slicing."""
....: return seq[start:stop:step]
....:

In [21]: rhyme = 'one fish, two fish, red fish, blue fish'.split()

In [22]: rhyme
Out[22]: ['one', 'fish,', 'two', 'fish,', 'red', 'fish,', 'blue', 'fish']

In [23]: slicer(rhyme)
Out[23]: ['one', 'fish,', 'two', 'fish,', 'red', 'fish,', 'blue', 'fish']

In [24]: slicer(rhyme, step=2)

Out[24]: ['one', 'two', 'red', 'blue']

In [25]: slicer(rhyme, 1, step=2)

Out[25]: ['fish,', 'fish,', 'fish,', 'fish']

In [26]: slicer(rhyme, start=1, stop=4, step=2)

Out[26]: ['fish,', 'fish,']

The order of the keyword arguments does not matter:

In [27]: slicer(rhyme, step=2, start=1, stop=4)

Out[27]: ['fish,', 'fish,']

but it is good practice to use the same ordering as the function’s definition.

Keyword arguments are a very convenient feature for defining functions with a variable number of argu-
ments, especially when default values are to be used in most calls to the function.

2.4.4 Passing by value

Tip: Can you modify the value of a variable inside a function? Most languages (C, Java, . . . ) distinguish
“passing by value” and “passing by reference”. In Python, such a distinction is somewhat artificial, and
it is a bit subtle whether your variables are going to be modified or not. Fortunately, there exist clear
rules.
Parameters to functions are references to objects, which are passed by value. When you pass a variable
to a function, python passes the reference to the object to which the variable refers (the value). Not the
variable itself.

If the value passed in a function is immutable, the function does not modify the caller’s variable. If the
value is mutable, the function may modify the caller’s variable in-place:
>>> def try_to_modify(x, y, z):
... x = 23
... y.append(42)
... z = [99] # new reference
... print(x)
... print(y)
... print(z)
(continues on next page)

2.4. Defining functions 27

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

...
>>> a = 77 # immutable variable
>>> b = [99] # mutable variable
>>> c = [28]
>>> try_to_modify(a, b, c)
23
[99, 42]
[99]
>>> print(a)
77
>>> print(b)
[99, 42]
>>> print(c)
[28]

Functions have a local variable table called a local namespace.

The variable x only exists within the function try_to_modify.

2.4.5 Global variables

Variables declared outside the function can be referenced within the function:

In [28]: x = 5

In [29]: def addx(y):

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

In [30]: addx(10)
Out[30]: 15

But these “global” variables cannot be modified within the function, unless declared global in the
function.
This doesn’t work:

In [31]: def setx(y):

....: x = y
....: print('x is %d ' % x)
....:
....:

In [32]: setx(10)
x is 10

In [33]: x
Out[33]: 5

This works:

In [34]: def setx(y):

....: global x
....: x = y
....: print('x is %d ' % x)
....:
....:
(continues on next page)

2.4. Defining functions 28

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

In [35]: setx(10)
x is 10

In [36]: x
Out[36]: 10

2.4.6 Variable number of parameters

Special forms of parameters:

• *args: any number of positional arguments packed into a tuple
• **kwargs: any number of keyword arguments packed into a dictionary

In [37]: def variable_args(*args, **kwargs):

....: print('args is', args)
....: print('kwargs is', kwargs)
....:

In [38]: variable_args('one', 'two', x=1, y=2, z=3)

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

2.4.7 Docstrings

Documentation about what the function does and its parameters. General convention:

In [39]: def funcname(params):

....: """Concise one-line sentence describing the function.
....:
....: Extended summary which can contain multiple paragraphs.
....: """
....: # function body
....: pass
....:

In [40]: funcname?
Signature: funcname(params)
Docstring:
Concise one-line sentence describing the function.
Extended summary which can contain multiple paragraphs.
File: ~/src/scientific-python-lectures/<ipython-input-13-64e466df6d64>
Type: function

Note: Docstring guidelines

For the sake of standardization, the Docstring Conventions webpage documents the semantics and con-
ventions associated with Python docstrings.
Also, the NumPy and SciPy modules have defined a precise standard for documenting scientific functions,
that you may want to follow for your own functions, with a Parameters section, an Examples section,
etc. See https://numpydoc.readthedocs.io/en/latest/format.html#docstring-standard

2.4. Defining functions 29

 Scientific Python Lectures, Edition 2024.2rc0.dev0

2.4.8 Functions are objects

Functions are first-class objects, which means they can be:

• assigned to a variable
• an item in a list (or any collection)
• passed as an argument to another function.

In [41]: va = variable_args

In [42]: va('three', x=1, y=2)

args is ('three',)
kwargs is {'x': 1, 'y': 2}

2.4.9 Methods

Methods are functions attached to objects. You’ve seen these in our examples on lists, dictionaries,
strings, etc. . .

2.4.10 Exercises

Exercise: Fibonacci sequence

Write a function that displays the n first terms of the Fibonacci sequence, defined by:
⎧
⎨ 𝑈0 = 0
𝑈1 = 1
𝑈𝑛+2 = 𝑈𝑛+1 + 𝑈𝑛
⎩

Exercise: Quicksort

Implement the quicksort algorithm, as defined by wikipedia

function quicksort(array)
var list less, greater
if length(array) < 2
return array
select and remove a pivot value pivot from array
for each x in array
if x < pivot + 1 then append x to less
else append x to greater
return concatenate(quicksort(less), pivot, quicksort(greater))

2.4. Defining functions 30

 Scientific Python Lectures, Edition 2024.2rc0.dev0

2.5 Reusing code: scripts and modules

For now, we have typed all instructions in the interpreter. For longer sets of instructions we need to
change track and write the code in text files (using a text editor), that we will call either scripts or
modules. Use your favorite text editor (provided it offers syntax highlighting for Python), or the editor
that comes with the Scientific Python Suite you may be using.

2.5.1 Scripts

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!).

The extension for Python files is .py. Write or copy-and-paste the following lines in a file called test.py

message = "Hello how are you?"

for word in message.split():
print(word)

Tip: Let us now execute the script interactively, that is inside the Ipython interpreter. This is maybe
the most common use of scripts in scientific computing.

Note: in Ipython, the syntax to execute a script is %run script.py. For example,

In [1]: %run test.py

Hello
how
are
you?

In [2]: message
Out[2]: 'Hello how are you?'

The script has been executed. Moreover the variables defined in the script (such as message) are now
available inside the interpreter’s namespace.

Tip: Other interpreters also offer the possibility to execute scripts (e.g., execfile in the plain Python
interpreter, etc.).

It is also possible In order to execute this script as a standalone program, by executing the script inside
a shell terminal (Linux/Mac console or cmd Windows console). For example, if we are in the same
directory as the test.py file, we can execute this in a console:

$ python test.py
Hello
how
are
you?

Tip: Standalone scripts may also take command-line arguments

2.5. Reusing code: scripts and modules 31

 Scientific Python Lectures, Edition 2024.2rc0.dev0

In file.py:
import sys
print(sys.argv)

$ python file.py test arguments

['file.py', 'test', 'arguments']

Warning: Don’t implement option parsing yourself. Use a dedicated module such as argparse.

2.5.2 Importing objects from modules

In [3]: import os

In [4]: os
Out[4]: <module 'os' (frozen)>

In [5]: os.listdir('.')
Out[5]:
['profile_ksttxb2h',
'.XIM-unix',
'profile_8jyd3wt0',
'systemd-private-547c90d23384455997d6600060dd5355-chrony.service-Cs0azH',
'profile_i8ku8_ln',
'profile_6c5_7yvi',
'clr-debug-pipe-589-869-in',
'profile_n4tf0uy4',
'profile_7crnvcpm',
'profile_ecjb1ld9',
'profile_m13gl1fv',
'profile_1_590dye',
'profile_v_p76c0x',
'clr-debug-pipe-1621-15129-out',
'profile_lvwe1o1z',
'profile_orrts951',
'profile_bziy8mt_',
'profile_0u5120_y',
'profile_tav_998d',
'.font-unix',
'dotnet-diagnostic-589-869-socket',
'clr-debug-pipe-589-869-out',
'profile_ujz8liq0',
'profile_rqn3apk7',
'profile_f3qz57ee',
'profile_v5h87_cl',
'profile_g8aplujp',
'systemd-private-547c90d23384455997d6600060dd5355-systemd-resolved.service-qwjgaC',
'.Test-unix',
'profile_orda1vei',
'profile_zwgjajcz',
'clr-debug-pipe-1604-14840-in',
'profile_rd6ks_2m',
'profile_hsd6fbf4',
(continues on next page)

2.5. Reusing code: scripts and modules 32

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

'.ICE-unix',
'profile_d2w42edv',
'profile_4_zbphdf',
'profile_5ph4s0ch',
'profile_b9bs2_r8',
'profile_ymfgkezo',
'profile_ylf1yz9m',
'profile_ve9ftzec',
'snap-private-tmp',
'profile_b0ttd81w',
'profile_t7xwqh0e',
'profile_vr_ay610',
'profile_559rgua5',
'profile_rhdgkmkd',
'clr-debug-pipe-1604-14840-out',
'profile_7oh0cama',
'profile_m3uuj28r',
'profile_4ks_neaw',
'profile_7j1ejay_',
'profile_boh6o1ne',
'profile_zu2iczx7',
'dotnet-diagnostic-1604-14840-socket',
'profile_xk8a882k',
'www-data-temp-aspnet-0',
'profile_lmfhc2sg',
'profile_ub0d3066',
'profile_hqwvn7hn',
'profile_wie_xyqb',
'profile_akajxynt',
'profile_c91t6jqc',
'dotnet-diagnostic-1621-15129-socket',
'profile_wihes686',
'profile_tpdh8hx5',
'.X11-unix',
'profile_3kvs0c7h',
'profile_jk9kmxtf',
'systemd-private-547c90d23384455997d6600060dd5355-systemd-logind.service-w050a8',
'systemd-private-547c90d23384455997d6600060dd5355-haveged.service-YT5gfq',
'profile_b1nj99y9',
'profile_owlfwm6o',
'profile__5ntcpq1',
'profile_xtdrih3r',
'clr-debug-pipe-1621-15129-in',
'profile_7klgdeg6',
'profile_dckcrseq',
'profile_yrvcipb_']

And also:

In [6]: from os import listdir

Importing shorthands:

In [7]: import numpy as np

2.5. Reusing code: scripts and modules 33

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Warning:
from os import *

This is called the star import and please, Do not use it

• Makes the code harder to read and understand: where do symbols come from?
• Makes it impossible to guess the functionality by the context and the name (hint: os.name is
the name of the OS), and to profit usefully from tab completion.
• Restricts the variable names you can use: os.name might override name, or vise-versa.
• Creates possible name clashes between modules.
• Makes the code impossible to statically check for undefined symbols.

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:

>>> import numpy as np # data arrays

>>> np.linspace(0, 10, 6)
array([ 0., 2., 4., 6., 8., 10.])
>>> import scipy as sp # scientific computing

2.5.3 Creating modules

Tip: If we want to write larger and better organized programs (compared to simple scripts), where
some objects are defined, (variables, functions, classes) and that we want to reuse several times, we have
to create our own modules.

Let us create a module demo contained in the file demo.py:

"A demo module."

def print_b():
"Prints b."
print("b")

def print_a():
"Prints a."
print("a")

c = 2
d = 2

Tip: In this file, we defined two functions print_a and print_b. Suppose we want to call the print_a
function from the interpreter. We could execute the file as a script, but since we just want to have access
to the function print_a, we are rather going to import it as a module. The syntax is as follows.

2.5. Reusing code: scripts and modules 34

 Scientific Python Lectures, Edition 2024.2rc0.dev0

In [8]: import demo

In [9]: demo.print_a()
a

In [10]: demo.print_b()
b

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.
Introspection
In [11]: demo?
Type: module
Base Class: <type 'module'>
String Form: <module 'demo' from 'demo.py'>
Namespace: Interactive
File: /home/varoquau/Projects/Python_talks/scipy_2009_tutorial/source/
˓→demo.py

Docstring:
A demo module.

In [12]: who
demo

In [13]: whos
Variable Type Data/Info
------------------------------
demo module <module 'demo' from 'demo.py'>

In [14]: dir(demo)
Out[14]:
['__builtins__',
'__doc__',
'__file__',
'__name__',
'__package__',
'c',
'd',
'print_a',
'print_b']

In [15]: demo.<TAB>
demo.c demo.print_a demo.py
demo.d demo.print_b demo.pyc

Importing objects from modules into the main namespace

In [16]: from demo import print_a, print_b

In [17]: whos
Variable Type Data/Info
--------------------------------
demo module <module 'demo' from 'demo.py'>
print_a function <function print_a at 0xb7421534>
print_b function <function print_b at 0xb74214c4>

(continues on next page)

2.5. Reusing code: scripts and modules 35

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

In [18]: print_a()
a

Warning: Module caching

Modules are cached: if you modify demo.py and re-import it in the old session, you will
get the old one.
Solution:
In [10]: importlib.reload(demo)

2.5.4 ‘__main__’ and module loading

Tip: Sometimes we want code to be executed when a module is run directly, but not when it is imported
by another module. if __name__ == '__main__' allows us to check whether the module is being run
directly.

File demo2.py:

def print_b():
"Prints b."
print("b")

def print_a():
"Prints a."
print("a")

# print_b() runs on import

print_b()

if __name__ == "__main__":
# print_a() is only executed when the module is run directly.
print_a()

Importing it:

In [19]: import demo2

b

In [20]: import demo2

Running it:

In [21]: %run demo2

b
a

2.5. Reusing code: scripts and modules 36

 Scientific Python Lectures, Edition 2024.2rc0.dev0

2.5.5 Scripts or modules? How to organize your code

Note: Rule of thumb

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

How modules are found and imported

When the import mymodule statement is executed, the module mymodule is searched in a given list of
directories. This list includes a list of installation-dependent default path (e.g., /usr/lib64/python3.
11) as well as the list of directories specified by the environment variable PYTHONPATH.
The list of directories searched by Python is given by the sys.path variable

In [22]: import sys

In [23]: sys.path
Out[23]:
['/home/runner/work/scientific-python-lectures/scientific-python-lectures',
'/opt/hostedtoolcache/Python/3.11.9/x64/lib/python311.zip',
'/opt/hostedtoolcache/Python/3.11.9/x64/lib/python3.11',
'/opt/hostedtoolcache/Python/3.11.9/x64/lib/python3.11/lib-dynload',
'/opt/hostedtoolcache/Python/3.11.9/x64/lib/python3.11/site-packages']

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/.venv/
lectures/lib64/python3.11/site-packages). You may use symbolic links (on Linux) to keep
the code somewhere else.
• modify the environment variable PYTHONPATH to include the directories containing the user-defined
modules.

Tip: On Linux/Unix, add the following line to a file read by the shell at startup (e.g. /etc/profile,
.profile)

export PYTHONPATH=$PYTHONPATH:/home/emma/user_defined_modules

On Windows, https://support.microsoft.com/kb/310519 explains how to handle environment vari-

ables.

• or modify the sys.path variable itself within a Python script.

Tip:

import sys
new_path = '/home/emma/user_defined_modules'
if new_path not in sys.path:
sys.path.append(new_path)

2.5. Reusing code: scripts and modules 37

 Scientific Python Lectures, Edition 2024.2rc0.dev0

This method is not very robust, however, because it makes the code less portable (user-dependent
path) and because you have to add the directory to your sys.path each time you want to import
from a module in this directory.

See also:
See https://docs.python.org/3/tutorial/modules.html for more information about modules.

2.5.6 Packages

A directory that contains many modules is called a package. A package is a module with submodules
(which can have submodules themselves, etc.). A special file called __init__.py (which may be empty)
tells Python that the directory is a Python package, from which modules can be imported.

$ ls
_build_utils/ fft/ _lib/ odr/ spatial/
cluster/ fftpack/ linalg/ optimize/ special/
conftest.py __init__.py linalg.pxd optimize.pxd special.pxd
constants/ integrate/ meson.build setup.py stats/
datasets/ interpolate/ misc/ signal/
_distributor_init.py io/ ndimage/ sparse/
$ cd ndimage
$ ls
_filters.py __init__.py _measurements.py morphology.py src/
filters.py _interpolation.py measurements.py _ni_docstrings.py tests/
_fourier.py interpolation.py meson.build _ni_support.py utils/
fourier.py LICENSE.txt _morphology.py setup.py

From Ipython:

In [24]: import scipy as sp

In [25]: sp.__file__
Out[25]: '/opt/hostedtoolcache/Python/3.11.9/x64/lib/python3.11/site-packages/scipy/__
˓→init__.py'

In [26]: sp.version.version
Out[26]: '1.13.1'

In [27]: sp.ndimage.morphology.binary_dilation?
Signature:
sp.ndimage.morphology.binary_dilation(
input,
structure=None,
iterations=1,
mask=None,
output=None,
border_value=0,
origin=0,
brute_force=False,
)
Docstring:
Multidimensional binary dilation with the given structuring element.
...

2.5. Reusing code: scripts and modules 38

 Scientific Python Lectures, Edition 2024.2rc0.dev0

2.5.7 Good practices

• Use meaningful object names

• Indentation: no choice!

Tip: Indenting is compulsory in Python! Every command block following a colon bears an
additional indentation level with respect to the previous line with a colon. One must therefore
indent after def f(): or while:. At the end of such logical blocks, one decreases the indentation
depth (and re-increases it if a new block is entered, etc.)
Strict respect of indentation is the price to pay for getting rid of { or ; characters that delineate
logical blocks in other languages. Improper indentation leads to errors such as

------------------------------------------------------------
IndentationError: unexpected indent (test.py, line 2)

All this indentation business can be a bit confusing in the beginning. However, with the clear
indentation, and in the absence of extra characters, the resulting code is very nice to read compared
to other languages.

• Indentation depth: Inside your text editor, you may choose to indent with any positive number
of spaces (1, 2, 3, 4, . . . ). However, it is considered good practice to indent with 4 spaces. You
may configure your editor to map the Tab key to a 4-space indentation.
• Style guidelines
Long lines: you should not write very long lines that span over more than (e.g.) 80 characters.
Long lines can be broken with the \ character

>>> long_line = "Here is a very very long line \

... that we break in two parts."

Spaces
Write well-spaced code: put whitespaces after commas, around arithmetic operators, etc.:

>>> a = 1 # yes
>>> a=1 # too cramped

A certain number of rules for writing “beautiful” code (and more importantly using the same
conventions as anybody else!) are given in the Style Guide for Python Code.

Quick read

If you want to do a first quick pass through the Scientific Python Lectures to learn the ecosystem, you
can directly skip to the next chapter: NumPy: creating and manipulating numerical data.
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.

2.5. Reusing code: scripts and modules 39

 Scientific Python Lectures, Edition 2024.2rc0.dev0

2.6 Input and Output

To be exhaustive, here are some information about input and output in Python. Since we will use the
NumPy methods to read and write files, you may skip this chapter at first reading.
We write or read strings to/from files (other types must be converted to strings). To write in a file:

>>> f = open('workfile', 'w') # opens the workfile file

>>> type(f)
<class '_io.TextIOWrapper'>
>>> f.write('This is a test \nand another test')
>>> f.close()

To read from a file

In [1]: f = open('workfile', 'r')

In [2]: s = f.read()

In [3]: print(s)
This is a test
and another test

In [4]: f.close()

See also:
For more details: https://docs.python.org/3/tutorial/inputoutput.html

2.6.1 Iterating over a file

In [5]: f = open('workfile', 'r')

In [6]: for line in f:

...: print(line)
...:
This is a test
and another test

In [7]: f.close()

File modes

• Read-only: r
• Write-only: w
– Note: Create a new file or overwrite existing file.
• Append a file: a
• Read and Write: r+
• Binary mode: b
– Note: Use for binary files, especially on Windows.

2.6. Input and Output 40

 Scientific Python Lectures, Edition 2024.2rc0.dev0

2.7 Standard Library

Note: Reference document for this section:

• The Python Standard Library documentation: https://docs.python.org/3/library/index.html
• Python Essential Reference, David Beazley, Addison-Wesley Professional

2.7.1 os module: operating system functionality

“A portable way of using operating system dependent functionality.”

Directory and file manipulation

Current directory:

In [1]: import os

In [2]: os.getcwd()
Out[2]: '/tmp'

List a directory:
In [3]: os.listdir(os.curdir)
Out[3]:
['profile_ksttxb2h',
'profile_crwqcy2d',
'.XIM-unix',
'profile_8jyd3wt0',
'systemd-private-547c90d23384455997d6600060dd5355-chrony.service-Cs0azH',
'profile_i8ku8_ln',
'profile_6c5_7yvi',
'clr-debug-pipe-589-869-in',
'profile_n4tf0uy4',
'profile_rwsw_k2b',
'profile_7crnvcpm',
'profile_ecjb1ld9',
'profile_m13gl1fv',
'profile_1_590dye',
'profile_v_p76c0x',
'clr-debug-pipe-1621-15129-out',
'profile_lvwe1o1z',
'profile_orrts951',
'profile_bziy8mt_',
'profile_0u5120_y',
'profile_tav_998d',
'.font-unix',
'dotnet-diagnostic-589-869-socket',
'clr-debug-pipe-589-869-out',
'profile_ujz8liq0',
'profile_rqn3apk7',
'profile_f3qz57ee',
'profile_v5h87_cl',
'profile_g8aplujp',
'systemd-private-547c90d23384455997d6600060dd5355-systemd-resolved.service-qwjgaC',
(continues on next page)

2.7. Standard Library 41

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

'profile_1fx2p0i0',
'.Test-unix',
'profile_orda1vei',
'profile_63ekqtw4',
'profile_zwgjajcz',
'clr-debug-pipe-1604-14840-in',
'profile_rd6ks_2m',
'profile_hsd6fbf4',
'.ICE-unix',
'profile_peh59e70',
'profile_d2w42edv',
'profile_4_zbphdf',
'profile_s35fv1zi',
'profile_5ph4s0ch',
'profile_b9bs2_r8',
'profile_ymfgkezo',
'profile_ylf1yz9m',
'profile_ve9ftzec',
'snap-private-tmp',
'profile_b0ttd81w',
'profile_t7xwqh0e',
'profile_vr_ay610',
'profile_559rgua5',
'profile_rhdgkmkd',
'clr-debug-pipe-1604-14840-out',
'profile_7oh0cama',
'profile_m3uuj28r',
'profile_4ks_neaw',
'profile_2yyxi31b',
'profile_7j1ejay_',
'profile_boh6o1ne',
'profile_nz20vb7_',
'profile_zu2iczx7',
'dotnet-diagnostic-1604-14840-socket',
'profile_xk8a882k',
'profile_l6786qpx',
'www-data-temp-aspnet-0',
'profile_lmfhc2sg',
'profile_ub0d3066',
'profile_hqwvn7hn',
'profile_wie_xyqb',
'profile_akajxynt',
'profile_c91t6jqc',
'dotnet-diagnostic-1621-15129-socket',
'profile_wihes686',
'profile_j7p1hjid',
'profile_tpdh8hx5',
'.X11-unix',
'profile_3kvs0c7h',
'profile_jk9kmxtf',
'systemd-private-547c90d23384455997d6600060dd5355-systemd-logind.service-w050a8',
'systemd-private-547c90d23384455997d6600060dd5355-haveged.service-YT5gfq',
'profile_b1nj99y9',
'profile_owlfwm6o',
'profile__5ntcpq1',
'profile_xtdrih3r',
(continues on next page)

2.7. Standard Library 42

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

'clr-debug-pipe-1621-15129-in',
'profile_7klgdeg6',
'profile_dckcrseq',
'profile_yrvcipb_',
'profile_ho2189dy']

Make a directory:

In [4]: os.mkdir('junkdir')

In [5]: 'junkdir' in os.listdir(os.curdir)

Out[5]: True

Rename the directory:

In [6]: os.rename('junkdir', 'foodir')

In [7]: 'junkdir' in os.listdir(os.curdir)

Out[7]: False

In [8]: 'foodir' in os.listdir(os.curdir)

Out[8]: True

In [9]: os.rmdir('foodir')

In [10]: 'foodir' in os.listdir(os.curdir)

Out[10]: False

Delete a file:

In [11]: fp = open('junk.txt', 'w')

In [12]: fp.close()

In [13]: 'junk.txt' in os.listdir(os.curdir)

Out[13]: True

In [14]: os.remove('junk.txt')

In [15]: 'junk.txt' in os.listdir(os.curdir)

Out[15]: False

os.path: path manipulations

os.path provides common operations on pathnames.

In [16]: fp = open('junk.txt', 'w')

In [17]: fp.close()

In [18]: a = os.path.abspath('junk.txt')

In [19]: a
Out[19]: '/tmp/junk.txt'

In [20]: os.path.split(a)
(continues on next page)

2.7. Standard Library 43

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

Out[20]: ('/tmp', 'junk.txt')

In [21]: os.path.dirname(a)
Out[21]: '/tmp'

In [22]: os.path.basename(a)
Out[22]: 'junk.txt'

In [23]: os.path.splitext(os.path.basename(a))
Out[23]: ('junk', '.txt')

In [24]: os.path.exists('junk.txt')
Out[24]: True

In [25]: os.path.isfile('junk.txt')
Out[25]: True

In [26]: os.path.isdir('junk.txt')
Out[26]: False

In [27]: os.path.expanduser('~/local')
Out[27]: '/home/runner/local'

In [28]: os.path.join(os.path.expanduser('~'), 'local', 'bin')

Out[28]: '/home/runner/local/bin'

Running an external command

In [29]: os.system('ls')
Out[29]: 0

Note: Alternative to os.system

A noteworthy alternative to os.system is the sh module. Which provides much more convenient ways
to obtain the output, error stream and exit code of the external command.

In [30]: import sh
In [31]: com = sh.ls()

In [31]: print(com)
basic_types.rst exceptions.rst oop.rst standard_library.rst
control_flow.rst first_steps.rst python_language.rst
demo2.py functions.rst python-logo.png
demo.py io.rst reusing_code.rst

In [32]: type(com)
Out[32]: str

2.7. Standard Library 44

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Walking a directory

os.path.walk generates a list of filenames in a directory tree.

In [33]: for dirpath, dirnames, filenames in os.walk(os.curdir):

....: for fp in filenames:
....: print(os.path.abspath(fp))
....:
....:
/tmp/clr-debug-pipe-589-869-in
/tmp/junk.txt
/tmp/clr-debug-pipe-1621-15129-out
/tmp/dotnet-diagnostic-589-869-socket
/tmp/clr-debug-pipe-589-869-out
/tmp/clr-debug-pipe-1604-14840-in
/tmp/clr-debug-pipe-1604-14840-out
/tmp/dotnet-diagnostic-1604-14840-socket
/tmp/dotnet-diagnostic-1621-15129-socket
/tmp/clr-debug-pipe-1621-15129-in
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
(continues on next page)

2.7. Standard Library 45

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/dhist
/tmp/bookmarks
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README
/tmp/README

Environment variables:

In [34]: os.environ.keys()
Out[34]: KeysView(environ({'SHELL': '/bin/bash', 'COLORTERM': 'truecolor', ...}))

In [35]: os.environ['SHELL']
Out[35]: '/bin/bash'

2.7. Standard Library 46

 Scientific Python Lectures, Edition 2024.2rc0.dev0

2.7.2 shutil: high-level file operations

The shutil provides useful file operations:

• shutil.rmtree: Recursively delete a directory tree.
• shutil.move: Recursively move a file or directory to another location.
• shutil.copy: Copy files or directories.

2.7.3 glob: Pattern matching on files

The glob module provides convenient file pattern matching.

Find all files ending in .txt:

In [36]: import glob

In [37]: glob.glob('*.txt')
Out[37]: ['junk.txt']

2.7.4 sys module: system-specific information

System-specific information related to the Python interpreter.

• Which version of python are you running and where is it installed:

In [38]: import sys

In [39]: sys.platform
Out[39]: 'linux'

In [40]: sys.version
Out[40]: '3.11.9 (main, May 9 2024, 14:13:20) [GCC 11.4.0]'

In [41]: sys.prefix
Out[41]: '/opt/hostedtoolcache/Python/3.11.9/x64'

• List of command line arguments passed to a Python script:

In [42]: sys.argv
Out[42]:
['/opt/hostedtoolcache/Python/3.11.9/x64/lib/python3.11/site-packages/sphinx/__main__.
˓→py',

'-b',
'latex',
'-d',
'build/doctrees',
'.',
'build/latex']

sys.path is a list of strings that specifies the search path for modules. Initialized from PYTHONPATH:

In [43]: sys.path
Out[43]:
['/home/runner/work/scientific-python-lectures/scientific-python-lectures',
'/opt/hostedtoolcache/Python/3.11.9/x64/lib/python311.zip',
'/opt/hostedtoolcache/Python/3.11.9/x64/lib/python3.11',
(continues on next page)

2.7. Standard Library 47

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

'/opt/hostedtoolcache/Python/3.11.9/x64/lib/python3.11/lib-dynload',
'/opt/hostedtoolcache/Python/3.11.9/x64/lib/python3.11/site-packages']

2.7.5 pickle: easy persistence

Useful to store arbitrary objects to a file. Not safe or fast!

In [44]: import pickle

In [45]: l = [1, None, 'Stan']

In [46]: with open('test.pkl', 'wb') as file:

....: pickle.dump(l, file)
....:

In [47]: with open('test.pkl', 'rb') as file:

....: out = pickle.load(file)
....:

In [48]: out
Out[48]: [1, None, 'Stan']

Exercise

Write a program to search your PYTHONPATH for the module site.py.

path_site

2.8 Exception handling in Python

It is 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.
Exceptions are raised by different kinds of errors arising when executing Python code. In your own code,
you may also catch errors, or define custom error types. You may want to look at the descriptions of the
the built-in Exceptions when looking for the right exception type.

2.8.1 Exceptions

Exceptions are raised by errors in Python:

In [1]: 1/0
---------------------------------------------------------------------------
ZeroDivisionError Traceback (most recent call last)
Cell In[1], line 1
----> 1 1/0

ZeroDivisionError: division by zero

In [2]: 1 + 'e'
(continues on next page)

2.8. Exception handling in Python 48

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

---------------------------------------------------------------------------
TypeError Traceback (most recent call last)
Cell In[2], line 1
----> 1 1 + 'e'

TypeError: unsupported operand type(s) for +: 'int' and 'str'

In [3]: d = {1:1, 2:2}

In [4]: d[3]
---------------------------------------------------------------------------
KeyError Traceback (most recent call last)
Cell In[4], line 1
----> 1 d[3]

KeyError: 3

In [5]: l = [1, 2, 3]

In [6]: l[4]
---------------------------------------------------------------------------
IndexError Traceback (most recent call last)
Cell In[6], line 1
----> 1 l[4]

IndexError: list index out of range

In [7]: l.foobar
---------------------------------------------------------------------------
AttributeError Traceback (most recent call last)
Cell In[7], line 1
----> 1 l.foobar

AttributeError: 'list' object has no attribute 'foobar'

As you can see, there are different types of exceptions for different errors.

2.8.2 Catching exceptions

try/except

In [8]: while True:

...: try:
...: x = int(input('Please enter a number: '))
...: break
...: except ValueError:
...: print('That was no valid number. Try again...')
...:
Please enter a number: a
That was no valid number. Try again...
Please enter a number: 1

In [9]: x
Out[9]: 1

2.8. Exception handling in Python 49

 Scientific Python Lectures, Edition 2024.2rc0.dev0

try/finally

In [10]: try:
....: x = int(input('Please enter a number: '))
....: finally:
....: print('Thank you for your input')
....:
Please enter a number: a
Thank you for your input
---------------------------------------------------------------------------
ValueError Traceback (most recent call last)
Cell In[10], line 2
1 try:
----> 2 x = int(input('Please enter a number: '))
3 finally:
4 print('Thank you for your input')
ValueError: invalid literal for int() with base 10: 'a'

Important for resource management (e.g. closing a file)

Easier to ask for forgiveness than for permission

In [11]: def print_sorted(collection):

....: try:
....: collection.sort()
....: except AttributeError:
....: pass # The pass statement does nothing
....: print(collection)
....:

In [12]: print_sorted([1, 3, 2])

[1, 2, 3]

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

{1, 2, 3}

In [14]: print_sorted('132')
132

2.8.3 Raising exceptions

• Capturing and reraising an exception:

In [15]: def filter_name(name):

....: try:
....: name = name.encode('ascii')
....: except UnicodeError as e:
....: if name == 'Gaël':
....: print('OK, Gaël')
....: else:
....: raise e
....: return name
....:

In [16]: filter_name('Gaël')
(continues on next page)

2.8. Exception handling in Python 50

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

OK, Gaël
Out[16]: 'Gaël'

In [17]: filter_name('Stéfan')
---------------------------------------------------------------------------
UnicodeEncodeError Traceback (most recent call last)
Cell In[17], line 1
----> 1 filter_name('Stéfan')

Cell In[15], line 8, in filter_name(name)

6 print('OK, Gaël')
7 else:
----> 8 raise e
9 return name

Cell In[15], line 3, in filter_name(name)

1 def filter_name(name):
2 try:
----> 3 name = name.encode('ascii')
4 except UnicodeError as e:
5 if name == 'Gaël':

UnicodeEncodeError: 'ascii' codec can't encode character '\xe9' in position 2:␣

˓→ordinal not in range(128)

• Exceptions to pass messages between parts of the code:

In [18]: def achilles_arrow(x):

....: if abs(x - 1) < 1e-3:
....: raise StopIteration
....: x = 1 - (1-x)/2.
....: return x
....:

In [19]: x = 0

In [20]: while True:

....: try:
....: x = achilles_arrow(x)
....: except StopIteration:
....: break
....:
....:

In [21]: x
Out[21]: 0.9990234375

Use exceptions to notify certain conditions are met (e.g. StopIteration) or not (e.g. custom error raising)

2.8. Exception handling in Python 51

 Scientific Python Lectures, Edition 2024.2rc0.dev0

2.9 Object-oriented programming (OOP)

Python supports object-oriented programming (OOP). The goals of OOP are:

• to organize the code, and
• to reuse code in similar contexts.
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
... 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')

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):

... internship = 'mandatory, from March to June'
...
>>> james = MasterStudent('james')
>>> james.internship
'mandatory, from March to June'
>>> james.set_age(23)
>>> james.age
23

The MasterStudent class inherited from the Student attributes and methods.
Thanks to classes and object-oriented programming, we can organize code with different classes corre-
sponding to different objects we encounter (an Experiment class, an Image class, a Flow class, etc.), with
their own methods and attributes. Then we can use inheritance to consider variations around a base
class and reuse code. Ex : from a Flow base class, we can create derived StokesFlow, TurbulentFlow,
PotentialFlow, etc.

2.9. Object-oriented programming (OOP) 52

 CHAPTER 3
NumPy: creating and manipulating
numerical data

Authors: Emmanuelle Gouillart, Didrik Pinte, Gaël Varoquaux, and Pauli Virtanen
This chapter gives an overview of NumPy, the core tool for performant numerical computing with Python.

3.1 The NumPy array object

Section contents

• What are NumPy and NumPy arrays?

• Creating arrays
• Basic data types
• Basic visualization
• Indexing and slicing
• Copies and views
• Fancy indexing

53
 Scientific Python Lectures, Edition 2024.2rc0.dev0

3.1.1 What are NumPy and NumPy arrays?

NumPy arrays

Python objects
• 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
• closer to hardware (efficiency)
• designed for scientific computation (convenience)
• Also known as array oriented computing

>>> import numpy as np

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

Tip: For example, An array containing:

• values of an experiment/simulation at discrete time steps
• signal recorded by a measurement device, e.g. sound wave
• pixels of an image, grey-level or colour
• 3-D data measured at different X-Y-Z positions, e.g. MRI scan
• ...

Why it is useful: Memory-efficient container that provides fast numerical operations.

In [1]: L = range(1000)

In [2]: %timeit [i**2 for i in L]

43.6 us +- 592 ns per loop (mean +- std. dev. of 7 runs, 10,000 loops each)

In [3]: a = np.arange(1000)

In [4]: %timeit a**2

892 ns +- 1.46 ns per loop (mean +- std. dev. of 7 runs, 1,000,000 loops each)

3.1. The NumPy array object 54

 Scientific Python Lectures, Edition 2024.2rc0.dev0

NumPy Reference documentation

• On the web: https://numpy.org/doc/

• Interactive help:

In [5]: np.array?
Docstring:
array(object, dtype=None, *, copy=True, order='K', subok=False, ndmin=0,
like=None)

Create an array.

Parameters
----------
object : array_like
An array, any object exposing the array interface, an object whose
``__array__`` method returns an array, or any (nested) sequence.
If object is a scalar, a 0-dimensional array containing object is
returned.
dtype : data-type, optional
The desired data-type for the array. If not given, NumPy will try to use
a default ``dtype`` that can represent the values (by applying promotion
rules when necessary.)
copy : bool, optional
If true (default), then the object is copied. Otherwise, a copy will
only be made if ``__array__`` returns a copy, if obj is a nested
sequence, or if a copy is needed to satisfy any of the other
requirements (``dtype``, ``order``, etc.).
order : {'K', 'A', 'C', 'F'}, optional
Specify the memory layout of the array. If object is not an array, the
newly created array will be in C order (row major) unless 'F' is
specified, in which case it will be in Fortran order (column major).
If object is an array the following holds.

===== ========= ===================================================

order no copy copy=True
===== ========= ===================================================
'K' unchanged F & C order preserved, otherwise most similar order
'A' unchanged F order if input is F and not C, otherwise C order
'C' C order C order
'F' F order F order
===== ========= ===================================================

When ``copy=False`` and a copy is made for other reasons, the result is
the same as if ``copy=True``, with some exceptions for 'A', see the
Notes section. The default order is 'K'.
subok : bool, optional
If True, then sub-classes will be passed-through, otherwise
the returned array will be forced to be a base-class array (default).
ndmin : int, optional
Specifies the minimum number of dimensions that the resulting
array should have. Ones will be prepended to the shape as
needed to meet this requirement.
like : array_like, optional
Reference object to allow the creation of arrays which are not
NumPy arrays. If an array-like passed in as ``like`` supports
the ``__array_function__`` protocol, the result will be defined
(continues on next page)

3.1. The NumPy array object 55

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

by it. In this case, it ensures the creation of an array object
compatible with that passed in via this argument.

.. versionadded:: 1.20.0

Returns
-------
out : ndarray
An array object satisfying the specified requirements.

See Also
--------
empty_like : Return an empty array with shape and type of input.
ones_like : Return an array of ones with shape and type of input.
zeros_like : Return an array of zeros with shape and type of input.
full_like : Return a new array with shape of input filled with value.
empty : Return a new uninitialized array.
ones : Return a new array setting values to one.
zeros : Return a new array setting values to zero.
full : Return a new array of given shape filled with value.

Notes
-----
When order is 'A' and ``object`` is an array in neither 'C' nor 'F' order,
and a copy is forced by a change in dtype, then the order of the result is
not necessarily 'C' as expected. This is likely a bug.

Examples
--------
>>> np.array([1, 2, 3])
array([1, 2, 3])

Upcasting:

>>> np.array([1, 2, 3.0])

array([ 1., 2., 3.])

More than one dimension:

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

array([[1, 2],
[3, 4]])

Minimum dimensions 2:

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

array([[1, 2, 3]])

Type provided:

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

array([ 1.+0.j, 2.+0.j, 3.+0.j])

Data-type consisting of more than one element:

(continues on next page)

3.1. The NumPy array object 56

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

>>> x = np.array([(1,2),(3,4)],dtype=[('a','<i4'),('b','<i4')])
>>> x['a']
array([1, 3])

Creating an array from sub-classes:

>>> np.array(np.mat('1 2; 3 4'))

array([[1, 2],
[3, 4]])

>>> np.array(np.mat('1 2; 3 4'), subok=True)

matrix([[1, 2],
[3, 4]])
Type: builtin_function_or_method

• Looking for something:

>>> np.lookfor('create array')

Search results for 'create array'
---------------------------------
numpy.array
Create an array.
numpy.memmap
Create a memory-map to an array stored in a *binary* file on disk.

In [6]: np.con*?
np.concatenate
np.conj
np.conjugate
np.convolve

Import conventions

The recommended convention to import NumPy is:

>>> import numpy as np

3.1.2 Creating arrays

Manual construction of arrays

• 1-D:

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

>>> a
array([0, 1, 2, 3])
>>> a.ndim
1
>>> a.shape
(4,)
>>> len(a)
4

• 2-D, 3-D, . . . :

3.1. The NumPy array object 57

 Scientific Python Lectures, Edition 2024.2rc0.dev0

>>> b = np.array([[0, 1, 2], [3, 4, 5]]) # 2 x 3 array

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

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

>>> c
array([[[1],
[2]],

[[3],
[4]]])
>>> c.shape
(2, 2, 1)

Exercise: Simple arrays

• Create a simple two dimensional array. First, redo the examples from above. And then create
your own: how about odd numbers counting backwards on the first row, and even numbers on
the second?
• Use the functions len(), numpy.shape() on these arrays. How do they relate to each other?
And to the ndim attribute of the arrays?

Functions for creating arrays

Tip: In practice, we rarely enter items one by one. . .

• Evenly spaced:

>>> a = np.arange(10) # 0 .. n-1 (!)

>>> a
array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9])
>>> b = np.arange(1, 9, 2) # start, end (exclusive), step
>>> b
array([1, 3, 5, 7])

• or by number of points:

>>> c = np.linspace(0, 1, 6) # start, end, num-points

>>> c
array([0. , 0.2, 0.4, 0.6, 0.8, 1. ])
>>> d = np.linspace(0, 1, 5, endpoint=False)
>>> d
array([0. , 0.2, 0.4, 0.6, 0.8])

• Common arrays:

3.1. The NumPy array object 58

 Scientific Python Lectures, Edition 2024.2rc0.dev0

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

>>> a
array([[1., 1., 1.],
[1., 1., 1.],
[1., 1., 1.]])
>>> b = np.zeros((2, 2))
>>> b
array([[0., 0.],
[0., 0.]])
>>> c = np.eye(3)
>>> c
array([[1., 0., 0.],
[0., 1., 0.],
[0., 0., 1.]])
>>> d = np.diag(np.array([1, 2, 3, 4]))
>>> d
array([[1, 0, 0, 0],
[0, 2, 0, 0],
[0, 0, 3, 0],
[0, 0, 0, 4]])

• np.random: random numbers (Mersenne Twister PRNG):

>>> rng = np.random.default_rng(27446968)

>>> a = rng.random(4) # uniform in [0, 1]
>>> a
array([0.64613018, 0.48984931, 0.50851229, 0.22563948])

>>> b = rng.standard_normal(4) # Gaussian

>>> b
array([-0.38250769, -0.61536465, 0.98131732, 0.59353096])

Exercise: Creating arrays using functions

• Experiment with arange, linspace, ones, zeros, eye and diag.

• Create different kinds of arrays with random numbers.
• 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?

3.1.3 Basic data types

You may have noticed that, in some instances, array elements are displayed with a trailing dot (e.g. 2.
vs 2). This is due to a difference in the data-type used:

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

>>> a.dtype
dtype('int64')

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

>>> b.dtype
dtype('float64')

Tip: Different data-types allow us to store data more compactly in memory, but most of the time we

3.1. The NumPy array object 59

 Scientific Python Lectures, Edition 2024.2rc0.dev0

simply work with floating point numbers. Note that, in the example above, NumPy auto-detects the
data-type from the input.

You can explicitly specify which data-type you want:

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

>>> c.dtype
dtype('float64')

The default data type is floating point:

>>> a = np.ones((3, 3))

>>> a.dtype
dtype('float64')

There are also other types:

Complex

>>> d = np.array([1+2j, 3+4j, 5+6*1j])

>>> d.dtype
dtype('complex128')

Bool

>>> e = np.array([True, False, False, True])

>>> e.dtype
dtype('bool')

Strings

>>> f = np.array(['Bonjour', 'Hello', 'Hallo'])

>>> f.dtype # <--- strings containing max. 7 letters
dtype('<U7')

Much more
• int32
• int64
• uint32
• uint64

3.1.4 Basic visualization

Now that we have our first data arrays, we are going to visualize them.
Start by launching IPython:

$ ipython # or ipython3 depending on your install

Or the notebook:

$ jupyter notebook

Once IPython has started, enable interactive plots:

3.1. The NumPy array object 60

 Scientific Python Lectures, Edition 2024.2rc0.dev0

>>> %matplotlib

Or, from the notebook, enable plots in the notebook:

>>> %matplotlib inline

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:

>>> import matplotlib.pyplot as plt # the tidy way

And then use (note that you have to use show explicitly if you have not enabled interactive plots with
%matplotlib):

>>> plt.plot(x, y) # line plot

>>> plt.show() # <-- shows the plot (not needed with interactive plots)

Or, if you have enabled interactive plots with %matplotlib:

>>> plt.plot(x, y) # line plot

• 1D plotting:

>>> x = np.linspace(0, 3, 20)

>>> y = np.linspace(0, 9, 20)
>>> plt.plot(x, y) # line plot
[<matplotlib.lines.Line2D object at ...>]
>>> plt.plot(x, y, 'o') # dot plot
[<matplotlib.lines.Line2D object at ...>]

• 2D arrays (such as images):

>>> rng = np.random.default_rng(27446968)

>>> image = rng.random((30, 30))
>>> plt.imshow(image, cmap=plt.cm.hot)
<matplotlib.image.AxesImage object at ...>
>>> plt.colorbar()
<matplotlib.colorbar.Colorbar object at ...>

3.1. The NumPy array object 61

 Scientific Python Lectures, Edition 2024.2rc0.dev0

See also:
More in the: matplotlib chapter

Exercise: Simple visualizations

• Plot some simple arrays: a cosine as a function of time and a 2D matrix.

• Try using the gray colormap on the 2D matrix.

3.1.5 Indexing and slicing

The items of an array can be accessed and assigned to the same way as other Python sequences (e.g.
lists):

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

Warning: Indices begin at 0, like other Python sequences (and C/C++). In contrast, in Fortran
or Matlab, indices begin at 1.

The usual python idiom for reversing a sequence is supported:

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

For multidimensional arrays, indices are tuples of integers:

>>> a = np.diag(np.arange(3))
>>> a
array([[0, 0, 0],
[0, 1, 0],
[0, 0, 2]])
>>> a[1, 1]
1
>>> a[2, 1] = 10 # third line, second column
>>> a
(continues on next page)

3.1. The NumPy array object 62

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

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

Note that the last index is not included! :

>>> a[:4]
array([0, 1, 2, 3])

All three slice components are not required: by default, start is 0, end is the last and step is 1:
>>> a[1:3]
array([1, 2])
>>> a[::2]
array([0, 2, 4, 6, 8])
>>> a[3:]
array([3, 4, 5, 6, 7, 8, 9])

A small illustrated summary of NumPy indexing and slicing. . .

>>> a[0, 3:5]
array([3, 4]) 0 1 2 3 4 5
>>> a[4:, 4:]
10 11 12 13 14 15
array([[44, 55],
[54, 55]]) 20 21 22 23 24 25
>>> a[:, 2]
30 31 32 33 34 35
a([2, 12, 22, 32, 42, 52])

>>> a[2::2, ::2] 40 41 42 43 44 45

array([[20, 22, 24],
50 51 52 53 54 55
[40, 42, 44]])

You can also combine assignment and slicing:

>>> a = np.arange(10)
>>> a[5:] = 10
>>> a
(continues on next page)

3.1. The NumPy array object 63

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

array([ 0, 1, 2, 3, 4, 10, 10, 10, 10, 10])
>>> b = np.arange(5)
>>> a[5:] = b[::-1]
>>> a
array([0, 1, 2, 3, 4, 4, 3, 2, 1, 0])

Exercise: Indexing and slicing

• Try the different flavours of slicing, using start, end and step: starting from a linspace, try to
obtain odd numbers counting backwards, and even numbers counting forwards.
• Reproduce the slices in the diagram above. You may use the following expression to create the
array:
>>> np.arange(6) + np.arange(0, 51, 10)[:, np.newaxis]
array([[ 0, 1, 2, 3, 4, 5],
[10, 11, 12, 13, 14, 15],
[20, 21, 22, 23, 24, 25],
[30, 31, 32, 33, 34, 35],
[40, 41, 42, 43, 44, 45],
[50, 51, 52, 53, 54, 55]])

Exercise: Array creation

Create the following arrays (with correct data types):

[[1, 1, 1, 1],
[1, 1, 1, 1],
[1, 1, 1, 2],
[1, 6, 1, 1]]

[[0., 0., 0., 0., 0.],

[2., 0., 0., 0., 0.],
[0., 3., 0., 0., 0.],
[0., 0., 4., 0., 0.],
[0., 0., 0., 5., 0.],
[0., 0., 0., 0., 6.]]

Par on course: 3 statements for each

Hint: Individual array elements can be accessed similarly to a list, e.g. a[1] or a[1, 2].
Hint: Examine the docstring for diag.

Exercise: Tiling for array creation

Skim through the documentation for np.tile, and use this function to construct the array:
[[4, 3, 4, 3, 4, 3],
[2, 1, 2, 1, 2, 1],
[4, 3, 4, 3, 4, 3],
[2, 1, 2, 1, 2, 1]]

3.1. The NumPy array object 64

 Scientific Python Lectures, Edition 2024.2rc0.dev0

3.1.6 Copies and views

A slicing operation creates a view on the original array, which is just a way of accessing array data.
Thus the original array is not copied in memory. You can use np.may_share_memory() to check if two
arrays share the same memory block. Note however, that this uses heuristics and may give you false
positives.
When modifying the view, the original array is modified as well:
>>> a = np.arange(10)
>>> a
array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9])
>>> b = a[::2]
>>> b
array([0, 2, 4, 6, 8])
>>> np.may_share_memory(a, b)
True
>>> b[0] = 12
>>> b
array([12, 2, 4, 6, 8])
>>> a # (!)
array([12, 1, 2, 3, 4, 5, 6, 7, 8, 9])

>>> a = np.arange(10)
>>> c = a[::2].copy() # force a copy
>>> c[0] = 12
>>> a
array([0, 1, 2, 3, 4, 5, 6, 7, 8, 9])

>>> np.may_share_memory(a, c)
False

This behavior can be surprising at first sight. . . but it allows to save both memory and time.

Worked example: Prime number sieve

Compute prime numbers in 0–99, with a sieve

• Construct a shape (100,) boolean array is_prime, filled with True in the beginning:
>>> is_prime = np.ones((100,), dtype=bool)

3.1. The NumPy array object 65

 Scientific Python Lectures, Edition 2024.2rc0.dev0

• Cross out 0 and 1 which are not primes:

>>> is_prime[:2] = 0

• For each integer j starting from 2, cross out its higher multiples:
>>> N_max = int(np.sqrt(len(is_prime) - 1))
>>> for j in range(2, N_max + 1):
... is_prime[2*j::j] = False

• Skim through help(np.nonzero), and print the prime numbers

• Follow-up:
– Move the above code into a script file named prime_sieve.py
– Run it to check it works
– Use the optimization suggested in the sieve of Eratosthenes:
1. Skip j which are already known to not be primes
2. The first number to cross out is 𝑗 2

3.1.7 Fancy indexing

Tip: NumPy arrays can be indexed with slices, but also with boolean or integer arrays (masks). This
method is called fancy indexing. It creates copies not views.

Using boolean masks

>>> rng = np.random.default_rng(27446968)

>>> a = rng.integers(0, 21, 15)
>>> a
array([ 3, 13, 12, 10, 10, 10, 18, 4, 8, 5, 6, 11, 12, 17, 3])
>>> (a % 3 == 0)
array([ True, False, True, False, False, False, True, False, False,
False, True, False, True, False, True])
>>> 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, 12, 18, 6, 12, 3])

Indexing with a mask can be very useful to assign a new value to a sub-array:

>>> a[a % 3 == 0] = -1
>>> a
array([-1, 13, -1, 10, 10, 10, -1, 4, 8, 5, -1, 11, -1, 17, -1])

3.1. The NumPy array object 66

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Indexing with an array of integers

>>> a = np.arange(0, 100, 10)

>>> a
array([ 0, 10, 20, 30, 40, 50, 60, 70, 80, 90])

Indexing can be done with an array of integers, where the same index is repeated several time:

>>> a[[2, 3, 2, 4, 2]] # note: [2, 3, 2, 4, 2] is a Python list

array([20, 30, 20, 40, 20])

New values can be assigned with this kind of indexing:

>>> a[[9, 7]] = -100

>>> a
array([ 0, 10, 20, 30, 40, 50, 60, -100, 80, -100])

Tip: When a new array is created by indexing with an array of integers, the new array has the same
shape as the array of integers:

>>> a = np.arange(10)
>>> idx = np.array([[3, 4], [9, 7]])
>>> idx.shape
(2, 2)
>>> a[idx]
array([[3, 4],
[9, 7]])

The image below illustrates various fancy indexing applications

>>> a[(0,1,2,3,4), (1,2,3,4,5)]

0 1 2 3 4 5
array([1, 12, 23, 34, 45])

>>> a[3:, [0,2,5]] 10 11 12 13 14 15

array([[30, 32, 35],
20 21 22 23 24 25
[40, 42, 45],
[50, 52, 55]])
30 31 32 33 34 35
>>> mask = np.array([1,0,1,0,0,1], dtype=bool)
>>> a[mask, 2] 40 41 42 43 44 45
array([2, 22, 52])
50 51 52 53 54 55

Exercise: Fancy indexing

• Again, reproduce the fancy indexing shown in the diagram above.

• Use fancy indexing on the left and array creation on the right to assign values into an array, for
instance by setting parts of the array in the diagram above to zero.

3.1. The NumPy array object 67

 Scientific Python Lectures, Edition 2024.2rc0.dev0

3.2 Numerical operations on arrays

Section contents

• Elementwise operations
• Basic reductions
• Broadcasting
• Array shape manipulation
• Sorting data
• Summary

3.2.1 Elementwise operations

Basic operations

With scalars:

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

>>> a + 1
array([2, 3, 4, 5])
>>> 2**a
array([ 2, 4, 8, 16])

All arithmetic operates elementwise:

>>> b = np.ones(4) + 1
>>> a - b
array([-1., 0., 1., 2.])
>>> a * b
array([2., 4., 6., 8.])

>>> j = np.arange(5)
>>> 2**(j + 1) - j
array([ 2, 3, 6, 13, 28])

These operations are of course much faster than if you did them in pure python:

>>> a = np.arange(10000)
>>> %timeit a + 1
10000 loops, best of 3: 24.3 us per loop
>>> l = range(10000)
>>> %timeit [i+1 for i in l]
1000 loops, best of 3: 861 us per loop

Warning: Array multiplication is not matrix multiplication:

>>> c = np.ones((3, 3))
>>> c * c # NOT matrix multiplication!
array([[1., 1., 1.],
[1., 1., 1.],
[1., 1., 1.]])

3.2. Numerical operations on arrays 68

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Note: Matrix multiplication:

>>> c @ c
array([[3., 3., 3.],
[3., 3., 3.],
[3., 3., 3.]])

Exercise: Elementwise operations

• Try simple arithmetic elementwise operations: add even elements with odd elements
• Time them against their pure python counterparts using %timeit.
• Generate:
– [2**0, 2**1, 2**2, 2**3, 2**4]
– a_j = 2^(3*j) - j

Other operations

Comparisons:

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

>>> b = np.array([4, 2, 2, 4])
>>> a == b
array([False, True, False, True])
>>> a > b
array([False, False, True, False])

Tip: Array-wise comparisons:

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

>>> b = np.array([4, 2, 2, 4])
>>> c = np.array([1, 2, 3, 4])
>>> np.array_equal(a, b)
False
>>> np.array_equal(a, c)
True

Logical operations:

>>> a = np.array([1, 1, 0, 0], dtype=bool)

>>> b = np.array([1, 0, 1, 0], dtype=bool)
>>> np.logical_or(a, b)
array([ True, True, True, False])
>>> np.logical_and(a, b)
array([ True, False, False, False])

Transcendental functions:

>>> a = np.arange(5)
>>> np.sin(a)
array([ 0. , 0.84147098, 0.90929743, 0.14112001, -0.7568025 ])
(continues on next page)

3.2. Numerical operations on arrays 69

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

>>> np.exp(a)
array([ 1. , 2.71828183, 7.3890561 , 20.08553692, 54.59815003])
>>> np.log(np.exp(a))
array([0., 1., 2., 3., 4.])

Shape mismatches

>>> a = np.arange(4)
>>> a + np.array([1, 2])
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
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.],
[1., 1., 0.]])

Note: The transposition is a view

The transpose returns a view of the original array:

>>> a = np.arange(9).reshape(3, 3)
>>> a.T[0, 2] = 999
>>> a.T
array([[ 0, 3, 999],
[ 1, 4, 7],
[ 2, 5, 8]])
>>> a
array([[ 0, 1, 2],
[ 3, 4, 5],
[999, 7, 8]])

Note: Linear algebra

The sub-module numpy.linalg implements basic linear algebra, such as solving linear systems, singular
value decomposition, etc. However, it is not guaranteed to be compiled using efficient routines, and thus
we recommend the use of scipy.linalg, as detailed in section Linear algebra operations: scipy.linalg

Exercise other operations

• Look at the help for np.allclose. When might this be useful?

• Look at the help for np.triu and np.tril.

3.2. Numerical operations on arrays 70

 Scientific Python Lectures, Edition 2024.2rc0.dev0

3.2.2 Basic reductions

Computing sums

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

>>> np.sum(x)
10
>>> x.sum()
10

Sum by rows and by columns:

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

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

Tip: Same idea in higher dimensions:

>>> rng = np.random.default_rng(27446968)

>>> x = rng.random((2, 2, 2))
>>> x.sum(axis=2)[0, 1]
0.73415...
>>> x[0, 1, :].sum()
0.73415...

3.2. Numerical operations on arrays 71

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Other reductions

— works the same way (and take axis=)

Extrema:

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

>>> x.min()
1
>>> x.max()
3

>>> x.argmin() # index of minimum

0
>>> x.argmax() # index of maximum
1

Logical operations:

>>> np.all([True, True, False])

False
>>> np.any([True, True, False])
True

Note: Can be used for array comparisons:

>>> a = np.zeros((100, 100))

>>> np.any(a != 0)
False
>>> np.all(a == a)
True

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

>>> b = np.array([2, 2, 3, 2])
>>> c = np.array([6, 4, 4, 5])
>>> ((a <= b) & (b <= c)).all()
True

Statistics:

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

>>> y = np.array([[1, 2, 3], [5, 6, 1]])
>>> x.mean()
1.75
>>> np.median(x)
1.5
>>> np.median(y, axis=-1) # last axis
array([2., 5.])

>>> x.std() # full population standard dev.

0.82915619758884995

. . . and many more (best to learn as you go).

Exercise: Reductions

3.2. Numerical operations on arrays 72

 Scientific Python Lectures, Edition 2024.2rc0.dev0

• Given there is a sum, what other function might you expect to see?
• What is the difference between sum and cumsum?

Worked Example: diffusion using a random walk algorithm

Tip: Let us consider a simple 1D random walk process: at each time step a walker jumps right or
left with equal probability.
We are interested in finding the typical distance from the origin of a random walker after t left or
right jumps? We are going to simulate many “walkers” to find this law, and we are going to do so
using array computing tricks: we are going to create a 2D array with the “stories” (each walker has a
story) in one direction, and the time in the other:

>>> n_stories = 1000 # number of walkers

>>> t_max = 200 # time during which we follow the walker

We randomly choose all the steps 1 or -1 of the walk:

>>> t = np.arange(t_max)
>>> rng = np.random.default_rng()
>>> steps = 2 * rng.integers(0, 1 + 1, (n_stories, t_max)) - 1 # +1 because the␣
˓→high value is exclusive

>>> np.unique(steps) # Verification: all steps are 1 or -1

array([-1, 1])

We build the walks by summing steps along the time:

>>> positions = np.cumsum(steps, axis=1) # axis = 1: dimension of time
>>> sq_distance = positions**2

We get the mean in the axis of the stories:

>>> mean_sq_distance = np.mean(sq_distance, axis=0)

Plot the results:

>>> plt.figure(figsize=(4, 3))
<Figure size ... with 0 Axes>
>>> plt.plot(t, np.sqrt(mean_sq_distance), 'g.', t, np.sqrt(t), 'y-')
[<matplotlib.lines.Line2D object at ...>, <matplotlib.lines.Line2D object at ...>]
>>> plt.xlabel(r"$t$")
Text(...'$t$')
>>> plt.ylabel(r"$\sqrt{\langle (\delta x)^2 \rangle}$")
Text(...'$\\sqrt{\\langle (\\delta x)^2 \\rangle}$')
>>> Numerical
3.2. plt.tight_layout()
operations #onprovide
arrays sufficient space for labels 73
 Scientific Python Lectures, Edition 2024.2rc0.dev0

We find a well-known result in physics: the RMS distance grows as the square root of the time!

3.2.3 Broadcasting

• Basic operations on numpy arrays (addition, etc.) are elementwise

• This works on arrays of the same size.
Nevertheless, It’s also possible to do operations on arrays of different
sizes if NumPy can transform these arrays so that they all have
the same size: this conversion is called broadcasting.
The image below gives an example of broadcasting:

Let’s verify:

3.2. Numerical operations on arrays 74

 Scientific Python Lectures, Edition 2024.2rc0.dev0

>>> 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!:

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

A useful trick:
>>> a = np.arange(0, 40, 10)
>>> a.shape
(4,)
>>> a = a[:, np.newaxis] # adds a new axis -> 2D array
>>> a.shape
(4, 1)
>>> a
array([[ 0],
[10],
[20],
[30]])
>>> a + b
array([[ 0, 1, 2],
[10, 11, 12],
[20, 21, 22],
[30, 31, 32]])

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],

3.2. Numerical operations on arrays 75

 Scientific Python Lectures, Edition 2024.2rc0.dev0

[ 303, 105, 0, 433, 568, 872, 1172, 1241, 1610, 2145],

[ 736, 538, 433, 0, 135, 439, 739, 808, 1177, 1712],
[ 871, 673, 568, 135, 0, 304, 604, 673, 1042, 1577],
[1175, 977, 872, 439, 304, 0, 300, 369, 738, 1273],
[1475, 1277, 1172, 739, 604, 300, 0, 69, 438, 973],
[1544, 1346, 1241, 808, 673, 369, 69, 0, 369, 904],
[1913, 1715, 1610, 1177, 1042, 738, 438, 369, 0, 535],
[2448, 2250, 2145, 1712, 1577, 1273, 973, 904, 535, 0]])

A lot of grid-based or network-based problems can also use broadcasting. For instance, if we want to
compute the distance from the origin of points on a 5x5 grid, we can do

>>> x, y = np.arange(5), np.arange(5)[:, np.newaxis]

>>> distance = np.sqrt(x ** 2 + y ** 2)
>>> distance
array([[0. , 1. , 2. , 3. , 4. ],
[1. , 1.41421356, 2.23606798, 3.16227766, 4.12310563],
[2. , 2.23606798, 2.82842712, 3.60555128, 4.47213595],
[3. , 3.16227766, 3.60555128, 4.24264069, 5. ],
[4. , 4.12310563, 4.47213595, 5. , 5.65685425]])

Or in color:

>>> plt.pcolor(distance)
<matplotlib.collections.PolyQuadMesh object at ...>
>>> plt.colorbar()
<matplotlib.colorbar.Colorbar object at ...>

3.2. Numerical operations on arrays 76

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Remark : the numpy.ogrid() function allows to directly create vectors x and y of the previous example,
with two “significant dimensions”:

>>> x, y = np.ogrid[0:5, 0:5]

>>> x, y
(array([[0],
[1],
[2],
[3],
[4]]), array([[0, 1, 2, 3, 4]]))
>>> x.shape, y.shape
((5, 1), (1, 5))
>>> distance = np.sqrt(x ** 2 + y ** 2)

Tip: So, np.ogrid is very useful as soon as we have to handle computations on a grid. On the other
hand, np.mgrid directly provides matrices full of indices for cases where we 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]])

See also:
Broadcasting: discussion of broadcasting in the Advanced NumPy chapter.

3.2. Numerical operations on arrays 77

 Scientific Python Lectures, Edition 2024.2rc0.dev0

3.2.4 Array shape manipulation

Flattening

>>> a = np.array([[1, 2, 3], [4, 5, 6]])

>>> a.ravel()
array([1, 2, 3, 4, 5, 6])
>>> a.T
array([[1, 4],
[2, 5],
[3, 6]])
>>> a.T.ravel()
array([1, 4, 2, 5, 3, 6])

Higher dimensions: last dimensions ravel out “first”.

Reshaping

The inverse operation to flattening:

>>> a.shape
(2, 3)
>>> b = a.ravel()
>>> b = b.reshape((2, 3))
>>> b
array([[1, 2, 3],
[4, 5, 6]])

Or,

>>> a.reshape((2, -1)) # unspecified (-1) value is inferred

array([[1, 2, 3],
[4, 5, 6]])

Warning: ndarray.reshape may return a view (cf help(np.reshape))), or copy

Tip:

>>> b[0, 0] = 99
>>> a
array([[99, 2, 3],
[ 4, 5, 6]])

Beware: reshape may also return a copy!:

>>> a = np.zeros((3, 2))

>>> b = a.T.reshape(3*2)
>>> b[0] = 9
>>> a
array([[0., 0.],
[0., 0.],
[0., 0.]])

To understand this you need to learn more about the memory layout of a NumPy array.

3.2. Numerical operations on arrays 78

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Adding a dimension

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])

>>> z
array([1, 2, 3])

>>> z[:, np.newaxis]

array([[1],
[2],
[3]])

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

Dimension shuffling

>>> a = np.arange(4*3*2).reshape(4, 3, 2)
>>> a.shape
(4, 3, 2)
>>> a[0, 2, 1]
5
>>> b = a.transpose(1, 2, 0)
>>> b.shape
(3, 2, 4)
>>> b[2, 1, 0]
5

Also creates a view:

>>> b[2, 1, 0] = -1
>>> a[0, 2, 1]
-1

Resizing

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

>>> a = np.arange(4)
>>> a.resize((8,))
>>> a
array([0, 1, 2, 3, 0, 0, 0, 0])

However, it must not be referred to somewhere else:

>>> b = a
>>> a.resize((4,))
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
ValueError: cannot resize an array that references or is referenced
by another array in this way.
Use the np.resize function or refcheck=False

3.2. Numerical operations on arrays 79

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Exercise: Shape manipulations

• Look at the docstring for reshape, especially the notes section which has some more information
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)
• Experiment with transpose for dimension shuffling.

3.2.5 Sorting data

Sorting along an axis:

>>> a = np.array([[4, 3, 5], [1, 2, 1]])

>>> b = np.sort(a, axis=1)
>>> b
array([[3, 4, 5],
[1, 1, 2]])

Note: Sorts each row separately!

In-place sort:

>>> a.sort(axis=1)
>>> a
array([[3, 4, 5],
[1, 1, 2]])

Sorting with fancy indexing:

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

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

Finding minima and maxima:

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

>>> j_max = np.argmax(a)
>>> j_min = np.argmin(a)
>>> j_max, j_min
(0, 2)

Exercise: Sorting

• Try both in-place and out-of-place sorting.

• Try creating arrays with different dtypes and sorting them.
• Use all or array_equal to check the results.
• Look at np.random.shuffle for a way to create sortable input quicker.
• Combine ravel, sort and reshape.

3.2. Numerical operations on arrays 80

 Scientific Python Lectures, Edition 2024.2rc0.dev0

• Look at the axis keyword for sort and rewrite the previous exercise.

3.2.6 Summary

What do you need to know to get started?

• Know how to create arrays : array, arange, ones, zeros.
• Know the shape of the array with array.shape, then use slicing to obtain different views of the
array: array[::2], etc. Adjust the shape of the array using reshape or flatten it with ravel.
• Obtain a subset of the elements of an array and/or modify their values with masks

>>> a[a < 0] = 0

• Know miscellaneous operations on arrays, such as finding the mean or max (array.max(), array.
mean()). No need to retain everything, but have the reflex to search in the documentation (online
docs, help(), lookfor())!!
• For advanced use: master the indexing with arrays of integers, as well as broadcasting. Know more
NumPy functions to handle various array operations.

Quick read

If you want to do a first quick pass through the Scientific Python Lectures to learn the ecosystem, you
can directly 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
come back and finish this chapter, as well as to do some more exercices.

3.3 More elaborate arrays

Section contents

• More data types

• Structured data types
• maskedarray: dealing with (propagation of) missing data

3.3.1 More data types

Casting

“Bigger” type wins in mixed-type operations:

>>> np.array([1, 2, 3]) + 1.5

array([2.5, 3.5, 4.5])

Assignment never changes the type!

3.3. More elaborate arrays 81

 Scientific Python Lectures, Edition 2024.2rc0.dev0

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

>>> a.dtype
dtype('int64')
>>> a[0] = 1.9 # <-- float is truncated to integer
>>> a
array([1, 2, 3])

Forced casts:

>>> a = np.array([1.7, 1.2, 1.6])

>>> b = a.astype(int) # <-- truncates to integer
>>> b
array([1, 1, 1])

Rounding:

>>> a = np.array([1.2, 1.5, 1.6, 2.5, 3.5, 4.5])

>>> b = np.around(a)
>>> b # still floating-point
array([1., 2., 2., 2., 4., 4.])
>>> c = np.around(a).astype(int)
>>> c
array([1, 2, 2, 2, 4, 4])

Different data type sizes

Integers (signed):

int8 8 bits
int16 16 bits
int32 32 bits (same as int on 32-bit platform)
int64 64 bits (same as int on 64-bit platform)

>>> np.array([1], dtype=int).dtype

dtype('int64')
>>> np.iinfo(np.int32).max, 2**31 - 1
(2147483647, 2147483647)

Unsigned integers:

uint8 8 bits
uint16 16 bits
uint32 32 bits
uint64 64 bits

>>> np.iinfo(np.uint32).max, 2**32 - 1

(4294967295, 4294967295)

Floating-point numbers:

3.3. More elaborate arrays 82

 Scientific Python Lectures, Edition 2024.2rc0.dev0

float16 16 bits
float32 32 bits
float64 64 bits (same as float)
float96 96 bits, platform-dependent (same as np.longdouble)
float128 128 bits, platform-dependent (same as np.longdouble)

>>> np.finfo(np.float32).eps
1.1920929e-07
>>> np.finfo(np.float64).eps
2.2204460492503131e-16

>>> np.float32(1e-8) + np.float32(1) == 1

True
>>> np.float64(1e-8) + np.float64(1) == 1
False

Complex floating-point numbers:

complex64 two 32-bit floats

complex128 two 64-bit floats
complex192 two 96-bit floats, platform-dependent
complex256 two 128-bit floats, platform-dependent

Smaller data types

If you don’t know you need special data types, then you probably don’t.
Comparison on using float32 instead of float64:
• Half the size in memory and on disk
• Half the memory bandwidth required (may be a bit faster in some operations)
In [1]: a = np.zeros((int(1e6),), dtype=np.float64)

In [2]: b = np.zeros((int(1e6),), dtype=np.float32)

In [3]: %timeit a*a

268 us +- 2.15 us per loop (mean +- std. dev. of 7 runs, 1,000 loops each)

In [4]: %timeit b*b

133 us +- 138 ns per loop (mean +- std. dev. of 7 runs, 10,000 loops each)

• But: bigger rounding errors — sometimes in surprising places (i.e., don’t use them unless you
really need them)

3.3. More elaborate arrays 83

 Scientific Python Lectures, Edition 2024.2rc0.dev0

3.3.2 Structured data types

sensor_code (4-character string)

position (float)
value (float)

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

... ('position', float), ('value', float)])
>>> samples.ndim
1
>>> samples.shape
(6,)
>>> samples.dtype.names
('sensor_code', 'position', 'value')
>>> samples[:] = [('ALFA', 1, 0.37), ('BETA', 1, 0.11), ('TAU', 1, 0.13),
... ('ALFA', 1.5, 0.37), ('ALFA', 3, 0.11), ('TAU', 1.2, 0.13)]
>>> samples
array([(b'ALFA', 1. , 0.37), (b'BETA', 1. , 0.11), (b'TAU', 1. , 0.13),
(b'ALFA', 1.5, 0.37), (b'ALFA', 3. , 0.11), (b'TAU', 1.2, 0.13)],
dtype=[('sensor_code', 'S4'), ('position', '<f8'), ('value', '<f8')])

Field access works by indexing with field names:

>>> samples['sensor_code']
array([b'ALFA', b'BETA', b'TAU', b'ALFA', b'ALFA', b'TAU'], dtype='|S4')
>>> samples['value']
array([0.37, 0.11, 0.13, 0.37, 0.11, 0.13])
>>> samples[0]
(b'ALFA', 1., 0.37)

>>> samples[0]['sensor_code'] = 'TAU'

>>> samples[0]
(b'TAU', 1., 0.37)

Multiple fields at once:

>>> samples[['position', 'value']]

array([(1. , 0.37), (1. , 0.11), (1. , 0.13), (1.5, 0.37),
(3. , 0.11), (1.2, 0.13)],
dtype={'names': ['position', 'value'], 'formats': ['<f8', '<f8'], 'offsets': [4,
˓→ 12], 'itemsize': 20})

Fancy indexing works, as usual:

>>> samples[samples['sensor_code'] == b'ALFA']

array([(b'ALFA', 1.5, 0.37), (b'ALFA', 3. , 0.11)],
dtype=[('sensor_code', 'S4'), ('position', '<f8'), ('value', '<f8')])

Note: There are a bunch of other syntaxes for constructing structured arrays, see here and here.

3.3. More elaborate arrays 84

 Scientific Python Lectures, Edition 2024.2rc0.dev0

3.3.3 maskedarray: dealing with (propagation of) missing data

• 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, --],
mask=[False, True, False, True],
fill_value=999999)

>>> y = np.ma.array([1, 2, 3, 4], mask=[0, 1, 1, 1])

>>> x + y
masked_array(data=[2, --, --, --],
mask=[False, True, True, True],
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
really do pay off in the long run:

Good practices

• Explicit variable names (no need of a comment to explain what is in the variable)
• Style: spaces after commas, around =, etc.
A certain number of rules for writing “beautiful” code (and, more importantly, using the same
conventions as everybody else!) are given in the Style Guide for Python Code and the Docstring
Conventions page (to manage help strings).
• Except some rare cases, variable names and comments in English.

3.4 Advanced operations

Section contents

• Polynomials
• Loading data files

3.4. Advanced operations 85

 Scientific Python Lectures, Edition 2024.2rc0.dev0

3.4.1 Polynomials

NumPy also contains polynomials in different bases:

For example, 3𝑥2 + 2𝑥 − 1:

>>> p = np.poly1d([3, 2, -1])

>>> p(0)
-1
>>> p.roots
array([-1. , 0.33333333])
>>> p.order
2

>>> x = np.linspace(0, 1, 20)

>>> rng = np.random.default_rng()
>>> y = np.cos(x) + 0.3*rng.random(20)
>>> p = np.poly1d(np.polyfit(x, y, 3))

>>> t = np.linspace(0, 1, 200) # use a larger number of points for smoother plotting
>>> plt.plot(x, y, 'o', t, p(t), '-')
[<matplotlib.lines.Line2D object at ...>, <matplotlib.lines.Line2D object at ...>]

See https://numpy.org/doc/stable/reference/routines.polynomials.poly1d.html for more.

More polynomials (with more bases)

NumPy also has a more sophisticated polynomial interface, which supports e.g. the Chebyshev basis.
3𝑥2 + 2𝑥 − 1:

>>> p = np.polynomial.Polynomial([-1, 2, 3]) # coefs in different order!

>>> p(0)
-1.0
>>> p.roots()
array([-1. , 0.33333333])
>>> p.degree() # In general polynomials do not always expose 'order'
2

Example using polynomials in Chebyshev basis, for polynomials in range [-1, 1]:

>>> x = np.linspace(-1, 1, 2000)

>>> rng = np.random.default_rng()
(continues on next page)

3.4. Advanced operations 86

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

>>> y = np.cos(x) + 0.3*rng.random(2000)
>>> p = np.polynomial.Chebyshev.fit(x, y, 90)

>>> plt.plot(x, y, 'r.')

[<matplotlib.lines.Line2D object at ...>]
>>> plt.plot(x, p(x), 'k-', lw=3)
[<matplotlib.lines.Line2D object at ...>]

The Chebyshev polynomials have some advantages in interpolation.

3.4.2 Loading data files

Text files

Example: populations.txt:
# year hare lynx carrot
1900 30e3 4e3 48300
1901 47.2e3 6.1e3 48200
1902 70.2e3 9.8e3 41500
1903 77.4e3 35.2e3 38200

>>> data = 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
• Using Python’s I/O functions and e.g. regexps for parsing (Python is quite well suited for this)

3.4. Advanced operations 87

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Reminder: Navigating the filesystem with IPython

In [1]: pwd # show current directory

Out[1]: '/tmp'

Images

Using Matplotlib:

>>> 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)

This saved only one channel (of RGB):

>>> plt.imshow(plt.imread('red_elephant.png'))
<matplotlib.image.AxesImage object at ...>

Other libraries:

3.4. Advanced operations 88

 Scientific Python Lectures, Edition 2024.2rc0.dev0

>>> import imageio.v3 as iio

>>> iio.imwrite('tiny_elephant.png', (img[::6,::6] * 255).astype(np.uint8))
>>> plt.imshow(plt.imread('tiny_elephant.png'), interpolation='nearest')
<matplotlib.image.AxesImage object at ...>

NumPy’s own format

NumPy has its own binary format, not portable but with efficient I/O:

>>> data = np.ones((3, 3))

>>> np.save('pop.npy', data)
>>> data3 = np.load('pop.npy')

Well-known (& more obscure) file formats

• HDF5: h5py, PyTables

• NetCDF: scipy.io.netcdf_file, netcdf4-python, . . .
• Matlab: scipy.io.loadmat, scipy.io.savemat
• MatrixMarket: scipy.io.mmread, scipy.io.mmwrite
• IDL: scipy.io.readsav
. . . if somebody uses it, there’s probably also a Python library for it.

Exercise: Text data files

Write a Python script that loads data from populations.txt:: and drop the last column and the first
5 rows. Save the smaller dataset to pop2.txt.

NumPy internals

If you are interested in the NumPy internals, there is a good discussion in Advanced NumPy.

3.4. Advanced operations 89

 Scientific Python Lectures, Edition 2024.2rc0.dev0

3.5 Some exercises

3.5.1 Array manipulations

1. Form the 2-D array (without typing it in explicitly):

[[1, 6, 11],
[2, 7, 12],
[3, 8, 13],
[4, 9, 14],
[5, 10, 15]]

and generate a new array containing its 2nd and 4th rows.
2. Divide each column of the array:

>>> import numpy as np

>>> 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 argmin 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.)

3.5.2 Picture manipulation: Framing a Face

Let’s do some manipulations on NumPy arrays by starting with an image of a raccoon. scipy provides
a 2D array of this image with the scipy.datasets.face function:

>>> import scipy as sp

>>> face = sp.datasets.face(gray=True) # 2D grayscale image

Here are a few images we will be able to obtain with our manipulations: use different colormaps, crop
the image, change some parts of the image.

• Let’s use the imshow function of matplotlib to display the image.

>>> import matplotlib.pyplot as plt

>>> face = sp.datasets.face(gray=True)
>>> plt.imshow(face)
<matplotlib.image.AxesImage object at 0x...>

• The face is displayed in false colors. A colormap must be

specified for it to be displayed in grey.

>>> plt.imshow(face, cmap=plt.cm.gray)

<matplotlib.image.AxesImage object at 0x...>

3.5. Some exercises 90

 Scientific Python Lectures, Edition 2024.2rc0.dev0

• 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.
>>> crop_face = face[100:-100, 100:-100]

• We will now frame the face with a black locket. For this, we
need to create a mask corresponding 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
>>> sy, sx = face.shape
>>> y, x = np.ogrid[0:sy, 0:sx] # x and y indices of pixels
>>> y.shape, x.shape
((768, 1), (1, 1024))
>>> centerx, centery = (660, 300) # center of the image
>>> mask = ((y - centery)**2 + (x - centerx)**2) > 230**2 # circle

then we assign the value 0 to the pixels of the image corresponding to the mask. The syntax
is extremely simple and intuitive:
>>> face[mask] = 0
>>> plt.imshow(face)
<matplotlib.image.AxesImage object at 0x...>

• Follow-up: copy all instructions of this exercise in a script called

face_locket.py then execute this script in IPython with %run face_locket.py.
Change the circle to an ellipsoid.

3.5.3 Data statistics

The data in populations.txt describes the populations of hares and lynxes (and carrots) in northern
Canada during 20 years:
>>> data = np.loadtxt('data/populations.txt')
>>> year, hares, lynxes, carrots = data.T # trick: columns to variables

>>> import matplotlib.pyplot as plt

>>> plt.axes([0.2, 0.1, 0.5, 0.8])
<Axes: >
>>> 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 ...>

3.5. Some exercises 91

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Computes and print, based on the data in populations.txt. . .

1. The mean and std of the populations of each species for the years in the period.
2. Which year each species had the largest population.
3. Which species has the largest population for each year. (Hint: argsort & fancy indexing of
np.array(['H', 'L', 'C']))
4. Which years any of the populations is above 50000. (Hint: comparisons and np.any)
5. The top 2 years for each species when they had the lowest populations. (Hint: argsort, fancy
indexing)
6. Compare (plot) the change in hare population (see help(np.gradient)) and the number of lynxes.
Check correlation (see help(np.corrcoef)).
. . . all without for-loops.
Solution: Python source file

3.5.4 Crude integral approximations

Write a function f(a, b, c) that returns 𝑎𝑏 −𝑐. Form a 24x12x6 array containing its values in parameter
ranges [0,1] x [0,1] x [0,1].
Approximate the 3-d integral
∫︁ 1 ∫︁ 1 ∫︁ 1
(𝑎𝑏 − 𝑐)𝑑𝑎 𝑑𝑏 𝑑𝑐
0 0 0

over this volume with the mean. The exact result is: ln 2 − 21 ≈ 0.1931 . . . — what is your relative error?
(Hints: use elementwise operations and broadcasting. You can make np.ogrid give a number of points
in given range with np.ogrid[0:1:20j].)
Reminder Python functions:

def f(a, b, c):

return some_result

Solution: Python source file

3.5.5 Mandelbrot set

Write a script that computes the Mandelbrot fractal. The Mandelbrot iteration:

3.5. Some exercises 92

 Scientific Python Lectures, Edition 2024.2rc0.dev0

N_max = 50
some_threshold = 50

c = x + 1j*y

z = 0
for j in range(N_max):
z = z**2 + c

Point (x, y) belongs to the Mandelbrot set if |𝑧| < some_threshold.

Do this computation by:
1. Construct a grid of c = x + 1j*y values in range [-2, 1] x [-1.5, 1.5]
2. Do the iteration
3. Form the 2-d boolean mask indicating which points are in the set
4. Save the result to an image with:

>>> import matplotlib.pyplot as plt

>>> plt.imshow(mask.T, extent=[-2, 1, -1.5, 1.5])
<matplotlib.image.AxesImage object at ...>
>>> plt.gray()
>>> plt.savefig('mandelbrot.png')

Solution: Python source file

3.5.6 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
Write a script that works with 5 states, and:
• Constructs a random matrix, and normalizes each row so that it is a transition matrix.
• Starts from a random (normalized) probability distribution p and takes 50 steps => p_50

3.5. Some exercises 93

 Scientific Python Lectures, Edition 2024.2rc0.dev0

• Computes the stationary distribution: the eigenvector of P.T with eigenvalue 1 (numerically: closest
to 1) => p_stationary
Remember to normalize the eigenvector — I didn’t. . .
• Checks if p_50 and p_stationary are equal to tolerance 1e-5
Toolbox: np.random, @, np.linalg.eig, reductions, abs(), argmin, comparisons, all, np.linalg.norm,
etc.
Solution: Python source file

3.6 Full code examples

3.6.1 Full code examples for the numpy chapter

1D plotting

Plot a basic 1D figure

import numpy as np
import matplotlib.pyplot as plt

x = np.linspace(0, 3, 20)
y = np.linspace(0, 9, 20)
plt.plot(x, y)
(continues on next page)

3.6. Full code examples 94

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

plt.plot(x, y, "o")
plt.show()

Total running time of the script: (0 minutes 0.057 seconds)

2D plotting

Plot a basic 2D figure

import numpy as np
import matplotlib.pyplot as plt

rng = np.random.default_rng()
image = rng.random((30, 30))
plt.imshow(image, cmap=plt.cm.hot)
plt.colorbar()
plt.show()

Total running time of the script: (0 minutes 0.080 seconds)

3.6. Full code examples 95

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Distances exercise

Plot distances in a grid

import numpy as np
import matplotlib.pyplot as plt

x, y = np.arange(5), np.arange(5)[:, np.newaxis]

distance = np.sqrt(x**2 + y**2)
plt.pcolor(distance)
plt.colorbar()
plt.show()

Total running time of the script: (0 minutes 0.074 seconds)

3.6. Full code examples 96

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Fitting to polynomial

Plot noisy data and their polynomial fit

import numpy as np
import matplotlib.pyplot as plt

rng = np.random.default_rng(27446968)

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

t = np.linspace(0, 1, 200)
plt.plot(x, y, "o", t, p(t), "-")
plt.show()

Total running time of the script: (0 minutes 0.056 seconds)

3.6. Full code examples 97

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Fitting in Chebyshev basis

Plot noisy data and their polynomial fit in a Chebyshev basis

import numpy as np
import matplotlib.pyplot as plt

rng = np.random.default_rng(27446968)

x = np.linspace(-1, 1, 2000)
y = np.cos(x) + 0.3 * rng.random(2000)
p = np.polynomial.Chebyshev.fit(x, y, 90)

plt.plot(x, y, "r.")
plt.plot(x, p(x), "k-", lw=3)
plt.show()

Total running time of the script: (0 minutes 0.085 seconds)

3.6. Full code examples 98

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Population exercise

Plot populations of hares, lynxes, and carrots

import numpy as np
import matplotlib.pyplot as plt

data = np.loadtxt("../../../data/populations.txt")
year, hares, lynxes, carrots = data.T

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()

Total running time of the script: (0 minutes 0.072 seconds)

3.6. Full code examples 99

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Reading and writing an elephant

Read and write images

import numpy as np
import matplotlib.pyplot as plt

original figure

plt.figure()
img = plt.imread("../../../data/elephant.png")
plt.imshow(img)

<matplotlib.image.AxesImage object at 0x7f4954da60d0>

red channel displayed in grey

plt.figure()
img_red = img[:, :, 0]
plt.imshow(img_red, cmap=plt.cm.gray)

3.6. Full code examples 100

 Scientific Python Lectures, Edition 2024.2rc0.dev0

<matplotlib.image.AxesImage object at 0x7f49549dd9d0>

lower resolution

plt.figure()
img_tiny = img[::6, ::6]
plt.imshow(img_tiny, interpolation="nearest")

plt.show()

3.6. Full code examples 101

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Total running time of the script: (0 minutes 0.296 seconds)

Mandelbrot set

Compute the Mandelbrot fractal and plot it

3.6. Full code examples 102

 Scientific Python Lectures, Edition 2024.2rc0.dev0

import numpy as np
import matplotlib.pyplot as plt
from numpy import newaxis
import warnings

def compute_mandelbrot(N_max, some_threshold, nx, ny):

# A grid of c-values
x = np.linspace(-2, 1, nx)
y = np.linspace(-1.5, 1.5, ny)

c = x[:, newaxis] + 1j * y[newaxis, :]

# Mandelbrot iteration

z = c

# The code below overflows in many regions of the x-y grid, suppress
# warnings temporarily
with warnings.catch_warnings():
warnings.simplefilter("ignore")
for j in range(N_max):
z = z**2 + c
mandelbrot_set = abs(z) < some_threshold

return mandelbrot_set

(continues on next page)

3.6. Full code examples 103

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

mandelbrot_set = compute_mandelbrot(50, 50.0, 601, 401)

plt.imshow(mandelbrot_set.T, extent=[-2, 1, -1.5, 1.5])

plt.gray()
plt.show()

Total running time of the script: (0 minutes 0.073 seconds)

Random walk exercise

Plot distance as a function of time for a random walk together with the theoretical result

import numpy as np
import matplotlib.pyplot as plt

# We create 1000 realizations with 200 steps each

n_stories = 1000
t_max = 200

t = np.arange(t_max)
# Steps can be -1 or 1 (note that randint excludes the upper limit)
rng = np.random.default_rng()
steps = 2 * rng.integers(0, 1 + 1, (n_stories, t_max)) - 1

# The time evolution of the position is obtained by successively

# summing up individual steps. This is done for each of the
# realizations, i.e. along axis 1.
positions = np.cumsum(steps, axis=1)

# Determine the time evolution of the mean square distance.

sq_distance = positions**2
mean_sq_distance = np.mean(sq_distance, axis=0)

# Plot the distance d from the origin as a function of time and

# compare with the theoretically expected result where d(t)
(continues on next page)

3.6. Full code examples 104

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

# grows as a square root of time t.
plt.figure(figsize=(4, 3))
plt.plot(t, np.sqrt(mean_sq_distance), "g.", t, np.sqrt(t), "y-")
plt.xlabel(r"$t$")
plt.ylabel(r"$\sqrt{\langle (\delta x)^2 \rangle}$")
plt.tight_layout()
plt.show()

Total running time of the script: (0 minutes 0.066 seconds)

3.6. Full code examples 105

 CHAPTER 4
Matplotlib: plotting

Thanks

Many thanks to Bill Wing and Christoph Deil for review and corrections.

Authors: Nicolas Rougier, Mike Müller, Gaël Varoquaux

Chapter contents

• Introduction
• Simple plot
• Figures, Subplots, Axes and Ticks
• Other Types of Plots: examples and exercises
• Beyond this tutorial
• Quick references
• Full code examples

106
 Scientific Python Lectures, Edition 2024.2rc0.dev0

4.1 Introduction

Tip: Matplotlib is probably the most used Python package for 2D-graphics. It provides both a quick
way to visualize data from Python and publication-quality figures in many formats. We are going to
explore matplotlib in interactive mode covering most common cases.

4.1.1 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:

For interactive matplotlib sessions, turn on the matplotlib mode

IPython console
When using the IPython console, use:

In [1]: %matplotlib

Jupyter notebook
In the notebook, insert, at the beginning of the notebook the following magic:

%matplotlib inline

4.1.2 pyplot

Tip: pyplot provides a procedural interface to the matplotlib object-oriented plotting library. It is
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.

import matplotlib.pyplot as plt

4.2 Simple plot

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.
First step is to get the data for the sine and cosine functions:

import numpy as np

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

C, S = np.cos(X), np.sin(X)

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).
To run the example, you can type them in an IPython interactive session:

4.1. Introduction 107

 Scientific Python Lectures, Edition 2024.2rc0.dev0

$ ipython --matplotlib

This brings us to the IPython prompt:

IPython 0.13 -- An enhanced Interactive Python.

? -> Introduction to IPython's features.
%magic -> Information about IPython's 'magic' % functions.
help -> Python's own help system.
object? -> Details about 'object'. ?object also works, ?? prints more.

Tip: You can also download each of the examples and run it using regular python, but you will lose
interactive data manipulation:

$ python plot_exercise_1.py

You can get source for each step by clicking on the corresponding figure.

4.2.1 Plotting with default settings

Hint: Documentation
• plot tutorial
• plot() command

Tip: Matplotlib comes with a set of default settings that allow customizing all kinds of properties. You
can control the defaults of almost every property in matplotlib: figure size and dpi, line width, color and
style, axes, axis and grid properties, text and font properties and so on.

import numpy as np
import matplotlib.pyplot as plt

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

C, S = np.cos(X), np.sin(X)

plt.plot(X, C)
plt.plot(X, S)

plt.show()

4.2. Simple plot 108

 Scientific Python Lectures, Edition 2024.2rc0.dev0

4.2.2 Instantiating defaults

Hint: Documentation
• Customizing matplotlib

In the script below, we’ve instantiated (and commented) all the figure settings that influence the ap-
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
import matplotlib.pyplot as plt

# 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)

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

C, S = np.cos(X), np.sin(X)

# Plot cosine with a blue continuous line of width 1 (pixels)

plt.plot(X, C, color="blue", linewidth=1.0, linestyle="-")

# Plot sine with a green continuous line of width 1 (pixels)

plt.plot(X, S, color="green", linewidth=1.0, linestyle="-")

# Set x limits
plt.xlim(-4.0, 4.0)

# Set x ticks
plt.xticks(np.linspace(-4, 4, 9))

# Set y limits
plt.ylim(-1.0, 1.0)

# Set y ticks
(continues on next page)

4.2. Simple plot 109

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

plt.yticks(np.linspace(-1, 1, 5))

# Save figure using 72 dots per inch

# plt.savefig("exercise_2.png", dpi=72)

# Show result on screen

plt.show()

4.2.3 Changing colors and line widths

Hint: Documentation
• Controlling line properties
• Line2D API

Tip: First step, we want to have the cosine in blue and the sine in red and a slightly thicker line for
both of them. We’ll also slightly alter the figure size to make it more horizontal.

...
plt.figure(figsize=(10, 6), dpi=80)
plt.plot(X, C, color="blue", linewidth=2.5, linestyle="-")
plt.plot(X, S, color="red", linewidth=2.5, linestyle="-")
...

4.2.4 Setting limits

4.2. Simple plot 110

 Scientific Python Lectures, Edition 2024.2rc0.dev0

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
see all data points.

...
plt.xlim(X.min() * 1.1, X.max() * 1.1)
plt.ylim(C.min() * 1.1, C.max() * 1.1)
...

4.2.5 Setting ticks

Hint: Documentation
• xticks() command
• yticks() command
• Tick container
• Tick locating and formatting

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.

...
plt.xticks([-np.pi, -np.pi/2, 0, np.pi/2, np.pi])
plt.yticks([-1, 0, +1])
...

4.2. Simple plot 111

 Scientific Python Lectures, Edition 2024.2rc0.dev0

4.2.6 Setting tick labels

Hint: Documentation
• Working with text
• xticks() command
• yticks() command
• set_xticklabels()
• set_yticklabels()

Tip: Ticks are now properly placed but their label is not very explicit. We could guess that 3.142 is 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'$+1$'])
...

4.2.7 Moving spines

Hint: Documentation
• spines API

4.2. Simple plot 112

 Scientific Python Lectures, Edition 2024.2rc0.dev0

• 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.

...
ax = plt.gca() # gca stands for 'get current axis'
ax.spines['right'].set_color('none')
ax.spines['top'].set_color('none')
ax.xaxis.set_ticks_position('bottom')
ax.spines['bottom'].set_position(('data',0))
ax.yaxis.set_ticks_position('left')
ax.spines['left'].set_position(('data',0))
...

4.2.8 Adding a legend

Hint: Documentation
• Legend guide
• legend() command
• legend API

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")

(continues on next page)

4.2. Simple plot 113

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

plt.legend(loc='upper left')
...

4.2.9 Annotate some points

Hint: Documentation
• Annotating axis
• annotate() command

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.

...

t = 2 * np.pi / 3
plt.plot([t, t], [0, np.cos(t)], color='blue', linewidth=2.5, linestyle="--")
plt.scatter([t, ], [np.cos(t), ], 50, color='blue')

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"))

plt.plot([t, t],[0, np.sin(t)], color='red', linewidth=2.5, linestyle="--")

plt.scatter([t, ],[np.sin(t), ], 50, color='red')

plt.annotate(r'$sin(\frac{2\pi}{3} )=\frac{\sqrt{3} }{2} $',

xy=(t, np.sin(t)), xycoords='data',
xytext=(+10, +30), textcoords='offset points', fontsize=16,
arrowprops=dict(arrowstyle="->", connectionstyle="arc3,rad=.2"))
...

4.2. Simple plot 114

 Scientific Python Lectures, Edition 2024.2rc0.dev0

4.2.10 Devil is in the details

Hint: Documentation
• artist API
• set_bbox() method

Tip: The tick labels are now hardly visible because of the blue and red lines. We can make them
bigger and we can also adjust their properties such that they’ll be rendered on a semi-transparent white
background. This will allow us to see both the data and the labels.

...
for label in ax.get_xticklabels() + ax.get_yticklabels():
label.set_fontsize(16)
label.set_bbox(dict(facecolor='white', edgecolor='None', alpha=0.65))
...

4.3 Figures, Subplots, Axes and Ticks

A “figure” in matplotlib means the whole window in the user interface. Within this figure there can be
“subplots”.

Tip: So far we have used implicit figure and axes creation. This is handy for fast plots. We can have
more control over the display using figure, subplot, and axes explicitly. While subplot positions the plots
in a regular grid, axes allows free placement within the figure. Both can be useful depending on your
intention. We’ve already worked with figures and subplots without explicitly calling them. When we
call plot, matplotlib calls gca() to get the current axes and gca in turn calls gcf() to get the current
figure. If there is none it calls figure() to make one, strictly speaking, to make a subplot(111). Let’s
look at the details.

4.3. Figures, Subplots, Axes and Ticks 115

 Scientific Python Lectures, Edition 2024.2rc0.dev0

4.3.1 Figures

Tip: A figure is the windows in the GUI that has “Figure #” as title. Figures are numbered starting
from 1 as opposed to the normal Python way starting from 0. This is clearly MATLAB-style. There are
several parameters that determine what the figure looks like:

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
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).

plt.close(1) # Closes figure 1

4.3.2 Subplots

Tip: With subplot you can arrange plots in a regular grid. You need to specify the number of rows and
columns and the number of the plot. Note that the gridspec command is a more powerful alternative.

4.3. Figures, Subplots, Axes and Ticks 116

 Scientific Python Lectures, Edition 2024.2rc0.dev0

4.3.3 Axes

Axes are very similar to subplots but allow placement of plots at any location in the figure. So if we
want to put a smaller plot inside a bigger one we do so with axes.

4.3. Figures, Subplots, Axes and Ticks 117

 Scientific Python Lectures, Edition 2024.2rc0.dev0

4.3.4 Ticks

Well formatted ticks are an important part of publishing-ready figures. Matplotlib provides a totally
configurable system for ticks. There are tick locators to specify where ticks should appear and tick
formatters to give ticks the appearance you want. Major and minor ticks can be located and formatted
independently from each other. Per default minor ticks are not shown, i.e. there is only an empty list
for them because it is as NullLocator (see below).

Tick Locators

Tick locators control the positions of the ticks. They are set as follows:

ax = 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.

4.4 Other Types of Plots: examples and exercises

4.4. Other Types of Plots: examples and exercises 118

 Scientific Python Lectures, Edition 2024.2rc0.dev0

4.4. Other Types of Plots: examples and exercises 119

 Scientific Python Lectures, Edition 2024.2rc0.dev0

4.4. Other Types of Plots: examples and exercises 120

 Scientific Python Lectures, Edition 2024.2rc0.dev0

4.4.1 Regular Plots

Starting from the code below, try to reproduce the graphic taking care of filled areas:

Hint: You need to use the fill_between() command.

n = 256
X = np.linspace(-np.pi, np.pi, n)
(continues on next page)

4.4. Other Types of Plots: examples and exercises 121

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

Y = np.sin(2 * X)

plt.plot(X, Y + 1, color='blue', alpha=1.00)

plt.plot(X, Y - 1, color='blue', alpha=1.00)

Click on the figure for solution.

4.4.2 Scatter Plots

Starting from the code below, try to reproduce the graphic taking care of marker size, color and trans-
parency.

Hint: Color is given by angle of (X,Y).

n = 1024
rng = np.random.default_rng()
X = rng.normal(0,1,n)
Y = rng.normal(0,1,n)

plt.scatter(X,Y)

Click on figure for solution.

4.4.3 Bar Plots

Starting from the code below, try to reproduce the graphic by adding labels for red bars.

Hint: You need to take care of text alignment.

4.4. Other Types of Plots: examples and exercises 122

 Scientific Python Lectures, Edition 2024.2rc0.dev0

n = 12
X = np.arange(n)
rng = np.random.default_rng()
Y1 = (1 - X / float(n)) * rng.uniform(0.5, 1.0, n)
Y2 = (1 - X / float(n)) * rng.uniform(0.5, 1.0, n)

plt.bar(X, +Y1, facecolor='#9999ff', edgecolor='white')

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)

Click on figure for solution.

4.4.4 Contour Plots

Starting from the code below, try to reproduce the graphic taking care of the colormap (see Colormaps
below).

Hint: You need to use the clabel() command.

def f(x, y):

return (1 - x / 2 + x ** 5 + y ** 3) * np.exp(-x ** 2 -y ** 2)

n = 256
x = np.linspace(-3, 3, n)
y = np.linspace(-3, 3, n)
X, Y = np.meshgrid(x, y)

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)

Click on figure for solution.

4.4. Other Types of Plots: examples and exercises 123

 Scientific Python Lectures, Edition 2024.2rc0.dev0

4.4.5 Imshow

Starting from the code below, try to reproduce the graphic taking care of colormap, image interpolation
and origin.

Hint: You need to take care of the origin of the image in the imshow command and use a colorbar()

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.

4.4.6 Pie Charts

Starting from the code below, try to reproduce the graphic taking care of colors and slices size.

Hint: You need to modify Z.

rng = np.random.default_rng()
Z = rng.uniform(0, 1, 20)
plt.pie(Z)

Click on the figure for the solution.

4.4. Other Types of Plots: examples and exercises 124

 Scientific Python Lectures, Edition 2024.2rc0.dev0

4.4.7 Quiver Plots

Starting from the code below, try to reproduce the graphic taking care of colors and orientations.

Hint: You need to draw arrows twice.

n = 8
X, Y = np.mgrid[0:n, 0:n]
plt.quiver(X, Y)

Click on figure for solution.

4.4.8 Grids

Starting from the code below, try to reproduce the graphic taking care of line styles.

axes = plt.gca()
axes.set_xlim(0, 4)
axes.set_ylim(0, 3)
axes.set_xticklabels([])
axes.set_yticklabels([])

Click on figure for solution.

4.4. Other Types of Plots: examples and exercises 125

 Scientific Python Lectures, Edition 2024.2rc0.dev0

4.4.9 Multi Plots

Starting from the code below, try to reproduce the graphic.

Hint: You can use several subplots with different partition.

plt.subplot(2, 2, 1)
plt.subplot(2, 2, 3)
plt.subplot(2, 2, 4)

Click on figure for solution.

4.4.10 Polar Axis

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])

N = 20
theta = np.arange(0., 2 * np.pi, 2 * np.pi / N)
rng = np.random.default_rng()
radii = 10 * rng.random(N)
width = np.pi / 4 * rng.random(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)

Click on figure for solution.

4.4. Other Types of Plots: examples and exercises 126

 Scientific Python Lectures, Edition 2024.2rc0.dev0

4.4.11 3D Plots

Starting from the code below, try to reproduce the graphic.

Hint: You need to use contourf()

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)
R = np.sqrt(X**2 + Y**2)
Z = np.sin(R)

ax.plot_surface(X, Y, Z, rstride=1, cstride=1, cmap='hot')

Click on figure for solution.

4.4.12 Text

Try to do the same from scratch !

Hint: Have a look at the matplotlib logo.

Click on figure for solution.

4.4. Other Types of Plots: examples and exercises 127

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Quick read

If you want to do a first quick pass through the Scientific Python Lectures to learn the ecosystem, you
can directly skip to the next chapter: SciPy : high-level scientific computing.
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.

4.5 Beyond this tutorial

Matplotlib benefits from extensive documentation as well as a large community of users and developers.
Here are some links of interest:

4.5.1 Tutorials

• Pyplot tutorial • Customizing your objects

• Introduction • Object containers
• Controlling line properties • Figure container
• Working with multiple figures and axes • Axes container
• Working with text • Axis containers
• Image tutorial • Tick containers
• Startup commands • Path tutorial
• Importing image data into NumPy arrays • Introduction
• Plotting NumPy arrays as images • Bézier example
• Text tutorial • Compound paths
• Text introduction • Transforms tutorial
• Basic text commands • Introduction
• Text properties and layout • Data coordinates
• Writing mathematical expressions • Axes coordinates
• Text rendering With LaTeX • Blended transformations
• Annotating text • Using offset transforms to create a shadow
• Artist tutorial effect
• Introduction • The transformation pipeline

4.5.2 Matplotlib documentation

• User guide • How-To

• FAQ • Troubleshooting
• Installation • Environment Variables
• Usage • Screenshots

4.5. Beyond this tutorial 128

 Scientific Python Lectures, Edition 2024.2rc0.dev0

4.5.3 Code documentation

The code is well documented and you can quickly access a specific command from within a python
session:

>>> import matplotlib.pyplot as plt

>>> help(plt.plot)
Help on function plot in module matplotlib.pyplot:

plot(*args: ...) -> 'list[Line2D]'

Plot y versus x as lines and/or markers.

Call signatures::

plot([x], y, [fmt], *, data=None, **kwargs)

plot([x], y, [fmt], [x2], y2, [fmt2], ..., **kwargs)
...

4.5.4 Galleries

The matplotlib gallery is also incredibly useful when you search how to render a given graphic. Each
example comes with its source.

4.5.5 Mailing lists

Finally, there is a user mailing list where you can ask for help and a developers mailing list that is more
technical.

4.6 Quick references

Here is a set of tables that show main properties and styles.

4.6. Quick references 129

 Scientific Python Lectures, Edition 2024.2rc0.dev0

4.6.1 Line properties

Property Description Appearance

alpha (or a) alpha transparency on 0-1

scale
antialiased True or False - use antialised
rendering

color (or c) matplotlib color arg

linestyle (or ls) see Line properties
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
markeredgecolor edge color if a marker is used
(mec)
markerfacecolor face color if a marker is used
(mfc)
markersize (ms) size of the marker in points

4.6.2 Line styles

4.6. Quick references 130

 Scientific Python Lectures, Edition 2024.2rc0.dev0

4.6.3 Markers

4.6.4 Colormaps

All colormaps can be reversed by appending _r. For instance, gray_r is the reverse of gray.
If you want to know more about colormaps, check the documentation on Colormaps in matplotlib.

4.6. Quick references 131

 Scientific Python Lectures, Edition 2024.2rc0.dev0

4.7 Full code examples

4.7.1 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.

Pie chart

A simple pie chart example with matplotlib.

import numpy as np
import matplotlib.pyplot as plt

n = 20
Z = np.ones(n)
Z[-1] *= 2

plt.axes([0.025, 0.025, 0.95, 0.95])

plt.pie(Z, explode=Z * 0.05, colors=[f"{ i / float(n): f} " for i in range(n)])

plt.axis("equal")
plt.xticks([])
plt.yticks()

plt.show()

Total running time of the script: (0 minutes 0.050 seconds)

4.7. Full code examples 132

 Scientific Python Lectures, Edition 2024.2rc0.dev0

A simple, good-looking plot

Demoing some simple features of matplotlib

import numpy as np
import matplotlib

matplotlib.use("Agg")
import matplotlib.pyplot as plt

fig = plt.figure(figsize=(5, 4), dpi=72)

axes = fig.add_axes([0.01, 0.01, 0.98, 0.98])
X = np.linspace(0, 2, 200)
Y = np.sin(2 * np.pi * X)
plt.plot(X, Y, lw=2)
plt.ylim(-1.1, 1.1)
plt.grid()

plt.show()

Total running time of the script: (0 minutes 0.058 seconds)

4.7. Full code examples 133

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Plotting a scatter of points

A simple example showing how to plot a scatter of points with matplotlib.

import numpy as np
import matplotlib.pyplot as plt

n = 1024
rng = np.random.default_rng()
X = rng.normal(0, 1, n)
Y = rng.normal(0, 1, n)
T = np.arctan2(Y, X)

plt.axes([0.025, 0.025, 0.95, 0.95])

plt.scatter(X, Y, s=75, c=T, alpha=0.5)

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

plt.show()

Total running time of the script: (0 minutes 0.058 seconds)

4.7. Full code examples 134

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Subplots

Show multiple subplots in matplotlib.

import matplotlib.pyplot as plt

fig = plt.figure()
fig.subplots_adjust(bottom=0.025, left=0.025, top=0.975, right=0.975)

plt.subplot(2, 1, 1)
plt.xticks([]), plt.yticks([])

plt.subplot(2, 3, 4)
plt.xticks([])
plt.yticks([])

plt.subplot(2, 3, 5)
plt.xticks([])
plt.yticks([])

plt.subplot(2, 3, 6)
plt.xticks([])
plt.yticks([])

plt.show()

Total running time of the script: (0 minutes 0.042 seconds)

4.7. Full code examples 135

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Horizontal arrangement of subplots

An example showing horizontal arrangement of subplots with matplotlib.

import matplotlib.pyplot as plt

plt.figure(figsize=(6, 4))
plt.subplot(2, 1, 1)
plt.xticks([])
plt.yticks([])
plt.text(0.5, 0.5, "subplot(2,1,1)", ha="center", va="center", size=24, alpha=0.5)

plt.subplot(2, 1, 2)
plt.xticks([])
plt.yticks([])
plt.text(0.5, 0.5, "subplot(2,1,2)", ha="center", va="center", size=24, alpha=0.5)

plt.tight_layout()
plt.show()

Total running time of the script: (0 minutes 0.049 seconds)

4.7. Full code examples 136

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Subplot plot arrangement vertical

An example showing vertical arrangement of subplots with matplotlib.

import matplotlib.pyplot as plt

plt.figure(figsize=(6, 4))
plt.subplot(1, 2, 1)
plt.xticks([])
plt.yticks([])
plt.text(0.5, 0.5, "subplot(1,2,1)", ha="center", va="center", size=24, alpha=0.5)

plt.subplot(1, 2, 2)
plt.xticks([])
plt.yticks([])
plt.text(0.5, 0.5, "subplot(1,2,2)", ha="center", va="center", size=24, alpha=0.5)

plt.tight_layout()
plt.show()

Total running time of the script: (0 minutes 0.046 seconds)

4.7. Full code examples 137

 Scientific Python Lectures, Edition 2024.2rc0.dev0

A simple plotting example

A plotting example with a few simple tweaks

import numpy as np
import matplotlib

matplotlib.use("Agg")
import matplotlib.pyplot as plt

fig = plt.figure(figsize=(5, 4), dpi=72)

axes = fig.add_axes([0.01, 0.01, 0.98, 0.98])
x = np.linspace(0, 2, 200)
y = np.sin(2 * np.pi * x)
plt.plot(x, y, lw=0.25, c="k")
plt.xticks(np.arange(0.0, 2.0, 0.1))
plt.yticks(np.arange(-1.0, 1.0, 0.1))
plt.grid()
plt.show()

Total running time of the script: (0 minutes 0.077 seconds)

4.7. Full code examples 138

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Simple axes example

This example shows a couple of simple usage of axes.

import matplotlib.pyplot as plt

plt.axes([0.1, 0.1, 0.8, 0.8])

plt.xticks([])
plt.yticks([])
plt.text(
0.6, 0.6, "axes([0.1, 0.1, 0.8, 0.8])", ha="center", va="center", size=20,␣
˓→alpha=0.5

plt.axes([0.2, 0.2, 0.3, 0.3])

plt.xticks([])
plt.yticks([])
plt.text(
0.5, 0.5, "axes([0.2, 0.2, 0.3, 0.3])", ha="center", va="center", size=16,␣
˓→alpha=0.5

plt.show()

Total running time of the script: (0 minutes 0.040 seconds)

4.7. Full code examples 139

 Scientific Python Lectures, Edition 2024.2rc0.dev0

3D plotting

A simple example of 3D plotting.

import numpy as np
import matplotlib.pyplot as plt
from mpl_toolkits.mplot3d import Axes3D

ax = plt.figure().add_subplot(projection="3d")
X = np.arange(-4, 4, 0.25)
Y = np.arange(-4, 4, 0.25)
X, Y = np.meshgrid(X, Y)
R = np.sqrt(X**2 + Y**2)
Z = np.sin(R)

ax.plot_surface(X, Y, Z, rstride=1, cstride=1, cmap=plt.cm.hot)

ax.contourf(X, Y, Z, zdir="z", offset=-2, cmap=plt.cm.hot)
ax.set_zlim(-2, 2)

plt.show()

Total running time of the script: (0 minutes 0.125 seconds)

4.7. Full code examples 140

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Imshow elaborate

An example demoing imshow and styling the figure.

import numpy as np
import matplotlib.pyplot as plt

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, int(3.5 * n))
y = np.linspace(-3, 3, int(3.0 * n))
X, Y = np.meshgrid(x, y)
Z = f(X, Y)

plt.axes([0.025, 0.025, 0.95, 0.95])

plt.imshow(Z, interpolation="nearest", cmap="bone", origin="lower")
plt.colorbar(shrink=0.92)

plt.xticks([])
plt.yticks([])
plt.show()

Total running time of the script: (0 minutes 0.130 seconds)

4.7. Full code examples 141

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Plotting a vector field: quiver

A simple example showing how to plot a vector field (quiver) with matplotlib.

import numpy as np
import matplotlib.pyplot as plt

n = 8
X, Y = np.mgrid[0:n, 0:n]
T = np.arctan2(Y - n / 2.0, X - n / 2.0)
R = 10 + np.sqrt((Y - n / 2.0) ** 2 + (X - n / 2.0) ** 2)
U, V = R * np.cos(T), R * np.sin(T)

plt.axes([0.025, 0.025, 0.95, 0.95])

plt.quiver(X, Y, U, V, R, alpha=0.5)
plt.quiver(X, Y, U, V, edgecolor="k", facecolor="None", linewidth=0.5)

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

plt.show()

Total running time of the script: (0 minutes 0.040 seconds)

4.7. Full code examples 142

 Scientific Python Lectures, Edition 2024.2rc0.dev0

A example of plotting not quite right

An “ugly” example of plotting.

import numpy as np
import matplotlib

matplotlib.use("Agg")
import matplotlib.pyplot as plt

matplotlib.rc("grid", color="black", linestyle="-", linewidth=1)

fig = plt.figure(figsize=(5, 4), dpi=72)

axes = fig.add_axes([0.01, 0.01, 0.98, 0.98], facecolor=".75")
X = np.linspace(0, 2, 40)
Y = np.sin(2 * np.pi * X)
plt.plot(X, Y, lw=0.05, c="b", antialiased=False)

plt.xticks([])
plt.yticks(np.arange(-1.0, 1.0, 0.2))
plt.grid()
ax = plt.gca()

plt.show()

Total running time of the script: (0 minutes 0.032 seconds)

4.7. Full code examples 143

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Displaying the contours of a function

An example showing how to display the contours of a function with matplotlib.

import numpy as np
import matplotlib.pyplot as plt

def f(x, y):

return (1 - x / 2 + x**5 + y**3) * np.exp(-(x**2) - y**2)

n = 256
x = np.linspace(-3, 3, n)
y = np.linspace(-3, 3, n)
X, Y = np.meshgrid(x, y)

plt.axes([0.025, 0.025, 0.95, 0.95])

plt.contourf(X, Y, f(X, Y), 8, alpha=0.75, cmap=plt.cm.hot)

C = plt.contour(X, Y, f(X, Y), 8, colors="black", linewidths=0.5)
plt.clabel(C, inline=1, fontsize=10)

plt.xticks([])
plt.yticks([])
plt.show()

Total running time of the script: (0 minutes 0.087 seconds)

4.7. Full code examples 144

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Plotting in polar coordinates

A simple example showing how to plot in polar coordinates with matplotlib.

import numpy as np
import matplotlib.pyplot as plt

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

N = 20
theta = np.arange(0.0, 2 * np.pi, 2 * np.pi / N)
rng = np.random.default_rng()
radii = 10 * rng.random(N)
width = np.pi / 4 * rng.random(N)
bars = plt.bar(theta, radii, width=width, bottom=0.0)

for r, bar in zip(radii, bars, strict=True):

bar.set_facecolor(plt.cm.jet(r / 10.0))
bar.set_alpha(0.5)

ax.set_xticklabels([])
ax.set_yticklabels([])
plt.show()

Total running time of the script: (0 minutes 0.104 seconds)

4.7. Full code examples 145

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Plot and filled plots

Simple example of plots and filling between them with matplotlib.

import numpy as np
import matplotlib.pyplot as plt

n = 256
X = np.linspace(-np.pi, np.pi, n)
Y = np.sin(2 * X)

plt.axes([0.025, 0.025, 0.95, 0.95])

plt.plot(X, Y + 1, color="blue", alpha=1.00)

plt.fill_between(X, 1, Y + 1, color="blue", alpha=0.25)

plt.plot(X, Y - 1, color="blue", alpha=1.00)

plt.fill_between(X, -1, Y - 1, (Y - 1) > -1, color="blue", alpha=0.25)
plt.fill_between(X, -1, Y - 1, (Y - 1) < -1, color="red", alpha=0.25)

plt.xlim(-np.pi, np.pi)
plt.xticks([])
plt.ylim(-2.5, 2.5)
plt.yticks([])

plt.show()

Total running time of the script: (0 minutes 0.031 seconds)

4.7. Full code examples 146

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Subplot grid

An example showing the subplot grid in matplotlib.

import matplotlib.pyplot as plt

plt.figure(figsize=(6, 4))
plt.subplot(2, 2, 1)
plt.xticks([])
plt.yticks([])
plt.text(0.5, 0.5, "subplot(2,2,1)", ha="center", va="center", size=20, alpha=0.5)

plt.subplot(2, 2, 2)
plt.xticks([])
plt.yticks([])
plt.text(0.5, 0.5, "subplot(2,2,2)", ha="center", va="center", size=20, alpha=0.5)

plt.subplot(2, 2, 3)
plt.xticks([])
plt.yticks([])

plt.text(0.5, 0.5, "subplot(2,2,3)", ha="center", va="center", size=20, alpha=0.5)

plt.subplot(2, 2, 4)
plt.xticks([])
plt.yticks([])
plt.text(0.5, 0.5, "subplot(2,2,4)", ha="center", va="center", size=20, alpha=0.5)

plt.tight_layout()
plt.show()

Total running time of the script: (0 minutes 0.075 seconds)

4.7. Full code examples 147

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Bar plots

An example of bar plots with matplotlib.

import numpy as np
import matplotlib.pyplot as plt

n = 12
X = np.arange(n)
rng = np.random.default_rng()
Y1 = (1 - X / float(n)) * rng.uniform(0.5, 1.0, n)
Y2 = (1 - X / float(n)) * rng.uniform(0.5, 1.0, n)

plt.axes([0.025, 0.025, 0.95, 0.95])

plt.bar(X, +Y1, facecolor="#9999ff", edgecolor="white")
plt.bar(X, -Y2, facecolor="#ff9999", edgecolor="white")

for x, y in zip(X, Y1, strict=True):

plt.text(x + 0.4, y + 0.05, f"{ y: .2f} ", ha="center", va="bottom")

for x, y in zip(X, Y2, strict=True):

plt.text(x + 0.4, -y - 0.05, f"{ y: .2f} ", ha="center", va="top")

plt.xlim(-0.5, n)
plt.xticks([])
plt.ylim(-1.25, 1.25)
plt.yticks([])

plt.show()

4.7. Full code examples 148

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Total running time of the script: (0 minutes 0.050 seconds)

Axes

This example shows various axes command to position matplotlib axes.

import matplotlib.pyplot as plt

plt.axes([0.1, 0.1, 0.5, 0.5])

plt.xticks([])
plt.yticks([])
plt.text(
0.1, 0.1, "axes([0.1, 0.1, 0.5, 0.5])", ha="left", va="center", size=16, alpha=0.5
)

plt.axes([0.2, 0.2, 0.5, 0.5])

plt.xticks([])
plt.yticks([])
plt.text(
0.1, 0.1, "axes([0.2, 0.2, 0.5, 0.5])", ha="left", va="center", size=16, alpha=0.5
)

plt.axes([0.3, 0.3, 0.5, 0.5])

plt.xticks([])
plt.yticks([])
plt.text(
(continues on next page)

4.7. Full code examples 149

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

0.1, 0.1, "axes([0.3, 0.3, 0.5, 0.5])", ha="left", va="center", size=16, alpha=0.5
)

plt.axes([0.4, 0.4, 0.5, 0.5])

plt.xticks([])
plt.yticks([])
plt.text(
0.1, 0.1, "axes([0.4, 0.4, 0.5, 0.5])", ha="left", va="center", size=16, alpha=0.5
)

plt.show()

Total running time of the script: (0 minutes 0.069 seconds)

Grid

Displaying a grid on the axes in matploblib.

import matplotlib.pyplot as plt

ax = plt.axes([0.025, 0.025, 0.95, 0.95])

ax.set_xlim(0, 4)
ax.set_ylim(0, 3)
ax.xaxis.set_major_locator(plt.MultipleLocator(1.0))
(continues on next page)

4.7. Full code examples 150

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

ax.xaxis.set_minor_locator(plt.MultipleLocator(0.1))
ax.yaxis.set_major_locator(plt.MultipleLocator(1.0))
ax.yaxis.set_minor_locator(plt.MultipleLocator(0.1))
ax.grid(which="major", axis="x", linewidth=0.75, linestyle="-", color="0.75")
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.set_xticklabels([])
ax.set_yticklabels([])

plt.show()

Total running time of the script: (0 minutes 0.089 seconds)

3D plotting

Demo 3D plotting with matplotlib and style the figure.

import matplotlib.pyplot as plt

from mpl_toolkits.mplot3d import axes3d

ax = plt.figure().add_subplot(projection="3d")
X, Y, Z = axes3d.get_test_data(0.05)
cset = ax.contourf(X, Y, Z)
ax.clabel(cset, fontsize=9, inline=1)
(continues on next page)

4.7. Full code examples 151

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

plt.xticks([])
plt.yticks([])
ax.set_zticks([])

ax.text2D(
-0.05,
1.05,
" 3D plots \n",
horizontalalignment="left",
verticalalignment="top",
bbox={"facecolor": "white", "alpha": 1.0},
family="DejaVu Sans",
size="x-large",
transform=plt.gca().transAxes,
)

ax.text2D(
-0.05,
0.975,
" Plot 2D or 3D data",
horizontalalignment="left",
verticalalignment="top",
family="DejaVu Sans",
size="medium",
transform=plt.gca().transAxes,
)

plt.show()

Total running time of the script: (0 minutes 0.060 seconds)

GridSpec

An example demoing gridspec

4.7. Full code examples 152

 Scientific Python Lectures, Edition 2024.2rc0.dev0

import matplotlib.pyplot as plt

from matplotlib import gridspec

plt.figure(figsize=(6, 4))
G = gridspec.GridSpec(3, 3)

axes_1 = plt.subplot(G[0, :])

plt.xticks([])
plt.yticks([])
plt.text(0.5, 0.5, "Axes 1", ha="center", va="center", size=24, alpha=0.5)

axes_2 = plt.subplot(G[1, :-1])

plt.xticks([])
plt.yticks([])
plt.text(0.5, 0.5, "Axes 2", ha="center", va="center", size=24, alpha=0.5)

axes_3 = plt.subplot(G[1:, -1])

plt.xticks([])
plt.yticks([])
plt.text(0.5, 0.5, "Axes 3", ha="center", va="center", size=24, alpha=0.5)

axes_4 = plt.subplot(G[-1, 0])

plt.xticks([])
plt.yticks([])
plt.text(0.5, 0.5, "Axes 4", ha="center", va="center", size=24, alpha=0.5)

axes_5 = plt.subplot(G[-1, -2])

plt.xticks([])
plt.yticks([])
plt.text(0.5, 0.5, "Axes 5", ha="center", va="center", size=24, alpha=0.5)

plt.tight_layout()
plt.show()

4.7. Full code examples 153

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Total running time of the script: (0 minutes 0.073 seconds)

Demo text printing

A example showing off elaborate text printing with matplotlib.

import numpy as np
import matplotlib.pyplot as plt

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} $"

)
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"$F_G = G\frac{m_1m_2} {r^2}$")

plt.axes([0.025, 0.025, 0.95, 0.95])

(continues on next page)

4.7. Full code examples 154

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

rng = np.random.default_rng()

for i in range(24):
index = rng.integers(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, 0.75)
plt.text(
x,
y,
eq,
ha="center",
va="center",
color="#11557c",
alpha=alpha,
transform=plt.gca().transAxes,
fontsize=size,
clip_on=True,
)
plt.xticks([])
plt.yticks([])

plt.show()

Total running time of the script: (0 minutes 0.503 seconds)

4.7.2 Code for the chapter’s exercises

4.7.3 Example demoing choices for an option

4.7.4 Code generating the summary figures with a title

Code for the chapter’s exercises

Exercise 1

Solution of the exercise 1 with matplotlib.

4.7. Full code examples 155

 Scientific Python Lectures, Edition 2024.2rc0.dev0

import numpy as np
import matplotlib.pyplot as plt

n = 256
X = np.linspace(-np.pi, np.pi, 256)
C, S = np.cos(X), np.sin(X)
plt.plot(X, C)
plt.plot(X, S)

plt.show()

Total running time of the script: (0 minutes 0.052 seconds)

4.7. Full code examples 156

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Exercise 4

Exercise 4 with matplotlib.

import numpy as np
import matplotlib.pyplot as plt

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

plt.subplot(111)

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

S = 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.xlim(X.min() * 1.1, X.max() * 1.1)

plt.ylim(C.min() * 1.1, C.max() * 1.1)

plt.show()

Total running time of the script: (0 minutes 0.049 seconds)

4.7. Full code examples 157

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Exercise 3

Exercise 3 with matplotlib.

import numpy as np
import matplotlib.pyplot as plt

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

plt.subplot(111)

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

C, S = np.cos(X), np.sin(X)

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

plt.plot(X, S, color="red", linewidth=2.5, linestyle="-")

plt.xlim(-4.0, 4.0)
plt.xticks(np.linspace(-4, 4, 9))

plt.ylim(-1.0, 1.0)
plt.yticks(np.linspace(-1, 1, 5))

plt.show()

Total running time of the script: (0 minutes 0.043 seconds)

4.7. Full code examples 158

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Exercise 5

Exercise 5 with matplotlib.

import numpy as np
import matplotlib.pyplot as plt

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

plt.subplot(111)

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

S = 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.xlim(X.min() * 1.1, X.max() * 1.1)

plt.xticks([-np.pi, -np.pi / 2, 0, np.pi / 2, np.pi])

plt.ylim(C.min() * 1.1, C.max() * 1.1)

plt.yticks([-1, 0, +1])

plt.show()

Total running time of the script: (0 minutes 0.035 seconds)

4.7. Full code examples 159

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Exercise 6

Exercise 6 with matplotlib.

import numpy as np
import matplotlib.pyplot as plt

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

plt.subplot(111)

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

C = np.cos(X)
S = np.sin(X)

plt.plot(X, C, color="blue", 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.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.ylim(C.min() * 1.1, C.max() * 1.1)

plt.yticks([-1, 0, +1], [r"$-1$", r"$0$", r"$+1$"])

plt.show()

Total running time of the script: (0 minutes 0.045 seconds)

4.7. Full code examples 160

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Exercise 2

Exercise 2 with matplotlib.

import numpy as np
import matplotlib.pyplot as plt

# Create a new figure of size 8x6 points, using 100 dots per inch
plt.figure(figsize=(8, 6), dpi=80)

# Create a new subplot from a grid of 1x1

plt.subplot(111)

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

C, S = np.cos(X), np.sin(X)

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

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

# Set x limits
plt.xlim(-4.0, 4.0)

# Set x ticks
plt.xticks(np.linspace(-4, 4, 9))

(continues on next page)

4.7. Full code examples 161

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

# Set y limits
plt.ylim(-1.0, 1.0)

# Set y ticks
plt.yticks(np.linspace(-1, 1, 5))

# Show result on screen

plt.show()

Total running time of the script: (0 minutes 0.045 seconds)

Exercise 7

Exercise 7 with matplotlib

import numpy as np
import matplotlib.pyplot as plt

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)

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

plt.plot(X, S, color="red", linewidth=2.5, linestyle="-")

ax = plt.gca()
(continues on next page)

4.7. Full code examples 162

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

ax.spines["right"].set_color("none")
ax.spines["top"].set_color("none")
ax.xaxis.set_ticks_position("bottom")
ax.spines["bottom"].set_position(("data", 0))
ax.yaxis.set_ticks_position("left")
ax.spines["left"].set_position(("data", 0))

plt.xlim(X.min() * 1.1, X.max() * 1.1)

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.ylim(C.min() * 1.1, C.max() * 1.1)

plt.yticks([-1, 0, +1], [r"$-1$", r"$0$", r"$+1$"])

plt.show()

Total running time of the script: (0 minutes 0.050 seconds)

Exercise 8

Exercise 8 with matplotlib.

import numpy as np
import matplotlib.pyplot as plt

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

plt.subplot(111)
(continues on next page)

4.7. Full code examples 163

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

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

C = np.cos(X)
S = np.sin(X)

plt.plot(X, C, color="blue", linewidth=2.5, linestyle="-", label="cosine")

plt.plot(X, S, color="red", linewidth=2.5, linestyle="-", label="sine")

ax = plt.gca()
ax.spines["right"].set_color("none")
ax.spines["top"].set_color("none")
ax.xaxis.set_ticks_position("bottom")
ax.spines["bottom"].set_position(("data", 0))
ax.yaxis.set_ticks_position("left")
ax.spines["left"].set_position(("data", 0))

plt.xlim(X.min() * 1.1, X.max() * 1.1)

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.ylim(C.min() * 1.1, C.max() * 1.1)

plt.yticks([-1, +1], [r"$-1$", r"$+1$"])

plt.legend(loc="upper left")

plt.show()

Total running time of the script: (0 minutes 0.054 seconds)

Exercise 9

Exercise 9 with matplotlib.

4.7. Full code examples 164

 Scientific Python Lectures, Edition 2024.2rc0.dev0

import numpy as np
import matplotlib.pyplot as plt

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)

plt.plot(X, C, color="blue", linewidth=2.5, linestyle="-", label="cosine")

plt.plot(X, S, color="red", linewidth=2.5, linestyle="-", label="sine")

ax = plt.gca()
ax.spines["right"].set_color("none")
ax.spines["top"].set_color("none")
ax.xaxis.set_ticks_position("bottom")
ax.spines["bottom"].set_position(("data", 0))
ax.yaxis.set_ticks_position("left")
ax.spines["left"].set_position(("data", 0))

plt.xlim(X.min() * 1.1, X.max() * 1.1)

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.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(
(continues on next page)

4.7. Full code examples 165

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

[
t,
],
[
np.cos(t),
],
50,
color="blue",
)
plt.annotate(
r"$sin(\frac{2\pi}{3} )=\frac{\sqrt{3} }{2} $",
xy=(t, np.sin(t)),
xycoords="data",
xytext=(+10, +30),
textcoords="offset points",
fontsize=16,
arrowprops={"arrowstyle": "->", "connectionstyle": "arc3,rad=.2"},
)

plt.plot([t, t], [0, np.sin(t)], color="red", linewidth=1.5, linestyle="--")

plt.scatter(
[
t,
],
[
np.sin(t),
],
50,
color="red",
)
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={"arrowstyle": "->", "connectionstyle": "arc3,rad=.2"},
)

plt.legend(loc="upper left")

plt.show()

Total running time of the script: (0 minutes 0.110 seconds)

4.7. Full code examples 166

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Exercise

Exercises with matplotlib.

import numpy as np
import matplotlib.pyplot as plt

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

plt.subplot(111)

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

C, S = np.cos(X), np.sin(X)

plt.plot(X, C, color="blue", linewidth=2.5, linestyle="-", label="cosine")

plt.plot(X, S, color="red", linewidth=2.5, linestyle="-", label="sine")

ax = plt.gca()
ax.spines["right"].set_color("none")
ax.spines["top"].set_color("none")
ax.xaxis.set_ticks_position("bottom")
ax.spines["bottom"].set_position(("data", 0))
ax.yaxis.set_ticks_position("left")
ax.spines["left"].set_position(("data", 0))

plt.xlim(X.min() * 1.1, X.max() * 1.1)

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.ylim(C.min() * 1.1, C.max() * 1.1)

plt.yticks([-1, 1], [r"$-1$", r"$+1$"])

(continues on next page)

4.7. Full code examples 167

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

plt.legend(loc="upper left")

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",
)
plt.annotate(
r"$sin(\frac{2\pi}{3} )=\frac{\sqrt{3} }{2} $",
xy=(t, np.sin(t)),
xycoords="data",
xytext=(10, 30),
textcoords="offset points",
fontsize=16,
arrowprops={"arrowstyle": "->", "connectionstyle": "arc3,rad=.2"},
)

plt.plot([t, t], [0, np.sin(t)], color="red", linewidth=1.5, linestyle="--")

plt.scatter(
[
t,
],
[
np.sin(t),
],
50,
color="red",
)
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={"arrowstyle": "->", "connectionstyle": "arc3,rad=.2"},
)

for label in ax.get_xticklabels() + ax.get_yticklabels():

label.set_fontsize(16)
label.set_bbox({"facecolor": "white", "edgecolor": "None", "alpha": 0.65})

plt.show()

Total running time of the script: (0 minutes 0.116 seconds)

4.7. Full code examples 168

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Example demoing choices for an option

The colors matplotlib line plots

An example demoing the various colors taken by matplotlib’s plot.

import matplotlib.pyplot as plt

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)
plt.axes([0, 0.1, 1, 0.8], frameon=False)

for i in range(1, 11):

plt.plot([i, i], [0, 1], lw=1.5)

plt.xlim(0, 11)
plt.xticks([])
plt.yticks([])
plt.show()

Total running time of the script: (0 minutes 0.013 seconds)

Linewidth

Plot various linewidth with matplotlib.

import matplotlib.pyplot as plt

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)
plt.axes([0, 0.1, 1, 0.8], frameon=False)

for i in range(1, 11):

plt.plot([i, i], [0, 1], color="b", lw=i / 2.0)

plt.xlim(0, 11)
plt.ylim(0, 1)
plt.xticks([])
plt.yticks([])

plt.show()

Total running time of the script: (0 minutes 0.014 seconds)

4.7. Full code examples 169

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Alpha: transparency

This example demonstrates using alpha for transparency.

import matplotlib.pyplot as plt

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)
plt.axes([0, 0.1, 1, 0.8], frameon=False)

for i in range(1, 11):

plt.axvline(i, linewidth=1, color="blue", alpha=0.25 + 0.75 * i / 10.0)

plt.xlim(0, 11)
plt.xticks([])
plt.yticks([])
plt.show()

Total running time of the script: (0 minutes 0.016 seconds)

Aliased versus anti-aliased

This example demonstrates aliased versus anti-aliased text.

import matplotlib.pyplot as plt

size = 128, 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)

plt.axes([0, 0, 1, 1], frameon=False)

plt.rcParams["text.antialiased"] = False
plt.text(0.5, 0.5, "Aliased", ha="center", va="center")

plt.xlim(0, 1)
plt.ylim(0, 1)
plt.xticks([])
plt.yticks([])

plt.show()

Total running time of the script: (0 minutes 0.010 seconds)

4.7. Full code examples 170

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Aliased versus anti-aliased

The example shows aliased versus anti-aliased text.

import matplotlib.pyplot as plt

size = 128, 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)
plt.axes([0, 0, 1, 1], frameon=False)

plt.rcParams["text.antialiased"] = True
plt.text(0.5, 0.5, "Anti-aliased", ha="center", va="center")

plt.xlim(0, 1)
plt.ylim(0, 1)
plt.xticks([])
plt.yticks([])

plt.show()

Total running time of the script: (0 minutes 0.011 seconds)

Marker size

Demo the marker size control in matplotlib.

import matplotlib.pyplot as plt

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)
plt.axes([0, 0, 1, 1], frameon=False)

for i in range(1, 11):

plt.plot(
[
i,
],
[
1,
],
"s",
markersize=i,
markerfacecolor="w",
markeredgewidth=0.5,
markeredgecolor="k",
(continues on next page)

4.7. Full code examples 171

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

)

plt.xlim(0, 11)
plt.xticks([])
plt.yticks([])

plt.show()

Total running time of the script: (0 minutes 0.014 seconds)

Marker edge width

Demo the marker edge widths of matplotlib’s markers.

import matplotlib.pyplot as plt

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)
plt.axes([0, 0, 1, 1], frameon=False)

for i in range(1, 11):

plt.plot(
[
i,
],
[
1,
],
"s",
markersize=5,
markeredgewidth=1 + i / 10.0,
markeredgecolor="k",
markerfacecolor="w",
)
plt.xlim(0, 11)
plt.xticks([])
plt.yticks([])

plt.show()

Total running time of the script: (0 minutes 0.014 seconds)

4.7. Full code examples 172

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Colormaps

An example plotting the matplotlib colormaps.

import numpy as np
import matplotlib.pyplot as plt

plt.rc("text", usetex=False)
a = np.outer(np.arange(0, 1, 0.01), np.ones(10))

plt.figure(figsize=(10, 5))
plt.subplots_adjust(top=0.8, bottom=0.05, left=0.01, right=0.99)
maps = [m for m in plt.cm.datad if not m.endswith("_r")]
maps.sort()
l = len(maps) + 1

for i, m in enumerate(maps):
plt.subplot(1, l, i + 1)
plt.axis("off")
plt.imshow(a, aspect="auto", cmap=plt.get_cmap(m), origin="lower")
plt.title(m, rotation=90, fontsize=10, va="bottom")

plt.show()

Total running time of the script: (0 minutes 1.352 seconds)

4.7. Full code examples 173

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Solid joint style

An example showing the different solid joint styles in matplotlib.

import numpy as np
import matplotlib.pyplot as plt

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)
plt.axes([0, 0, 1, 1], frameon=False)

plt.plot(np.arange(3), [0, 1, 0], color="blue", linewidth=8, solid_joinstyle="miter")

plt.plot(
4 + np.arange(3), [0, 1, 0], color="blue", linewidth=8, solid_joinstyle="bevel"
)
plt.plot(
8 + np.arange(3), [0, 1, 0], color="blue", linewidth=8, solid_joinstyle="round"
)

plt.xlim(0, 12)
plt.ylim(-1, 2)
plt.xticks([])
plt.yticks([])

plt.show()

Total running time of the script: (0 minutes 0.012 seconds)

Solid cap style

An example demoing the solide cap style in matplotlib.

import numpy as np
import matplotlib.pyplot as plt

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)
plt.axes([0, 0, 1, 1], frameon=False)

plt.plot(np.arange(4), np.ones(4), color="blue", linewidth=8, solid_capstyle="butt")

plt.plot(
5 + np.arange(4), np.ones(4), color="blue", linewidth=8, solid_capstyle="round"
)

(continues on next page)

4.7. Full code examples 174

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

plt.plot(
10 + np.arange(4),
np.ones(4),
color="blue",
linewidth=8,
solid_capstyle="projecting",
)

plt.xlim(0, 14)
plt.xticks([])
plt.yticks([])

plt.show()

Total running time of the script: (0 minutes 0.011 seconds)

Marker edge color

Demo the marker edge color of matplotlib’s markers.

import numpy as np
import matplotlib.pyplot as plt

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)
plt.axes([0, 0, 1, 1], frameon=False)

rng = np.random.default_rng()

for i in range(1, 11):

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([])
(continues on next page)

4.7. Full code examples 175

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

plt.show()

Total running time of the script: (0 minutes 0.014 seconds)

Marker face color

Demo the marker face color of matplotlib’s markers.

import numpy as np
import matplotlib.pyplot as plt

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)
plt.axes([0, 0, 1, 1], frameon=False)

rng = np.random.default_rng()

for i in range(1, 11):

r, g, b = np.random.uniform(0, 1, 3)
plt.plot(
[
i,
],
[
1,
],
"s",
markersize=8,
markerfacecolor=(r, g, b, 1),
markeredgewidth=0.1,
markeredgecolor=(0, 0, 0, 0.5),
)
plt.xlim(0, 11)
plt.xticks([])
plt.yticks([])
plt.show()

Total running time of the script: (0 minutes 0.014 seconds)

4.7. Full code examples 176

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Dash capstyle

An example demoing the dash capstyle.

import numpy as np
import matplotlib.pyplot as plt

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)
plt.axes([0, 0, 1, 1], frameon=False)

plt.plot(
np.arange(4),
np.ones(4),
color="blue",
dashes=[15, 15],
linewidth=8,
dash_capstyle="butt",
)

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([])
plt.yticks([])

plt.show()

Total running time of the script: (0 minutes 0.011 seconds)

4.7. Full code examples 177

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Dash join style

Example demoing the dash join style.

import numpy as np
import matplotlib.pyplot as plt

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)
plt.axes([0, 0, 1, 1], frameon=False)

plt.plot(
np.arange(3),
[0, 1, 0],
color="blue",
dashes=[12, 5],
linewidth=8,
dash_joinstyle="miter",
)
plt.plot(
4 + np.arange(3),
[0, 1, 0],
color="blue",
dashes=[12, 5],
linewidth=8,
dash_joinstyle="bevel",
)
plt.plot(
8 + np.arange(3),
[0, 1, 0],
color="blue",
dashes=[12, 5],
linewidth=8,
dash_joinstyle="round",
)

plt.xlim(0, 12)
plt.ylim(-1, 2)
plt.xticks([])
plt.yticks([])

plt.show()

Total running time of the script: (0 minutes 0.012 seconds)

4.7. Full code examples 178

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Markers

Show the different markers of matplotlib.

import numpy as np
import matplotlib.pyplot as plt

def marker(m, i):

X = i * 0.5 * np.ones(11)
Y = np.arange(11)

plt.plot(X, Y, lw=1, marker=m, ms=10, mfc=(0.75, 0.75, 1, 1), mec=(0, 0, 1, 1))

plt.text(0.5 * i, 10.25, repr(m), rotation=90, fontsize=15, va="bottom")

markers = [
0,
1,
2,
3,
4,
5,
6,
7,
"o",
"h",
"_",
"1",
"2",
"3",
"4",
"8",
"p",
"^",
"v",
"<",
">",
"|",
(continues on next page)

4.7. Full code examples 179

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

"d",
",",
"+",
"s",
"*",
"|",
"x",
"D",
"H",
".",
]

n_markers = len(markers)

size = 20 * n_markers, 300

dpi = 72.0
figsize = size[0] / float(dpi), size[1] / float(dpi)
fig = plt.figure(figsize=figsize, dpi=dpi)
plt.axes([0, 0.01, 1, 0.9], frameon=False)

for i, m in enumerate(markers):
marker(m, i)

plt.xlim(-0.2, 0.2 + 0.5 * n_markers)

plt.xticks([])
plt.yticks([])

plt.show()

Total running time of the script: (0 minutes 0.058 seconds)

Linestyles

Plot the different line styles.

4.7. Full code examples 180

 Scientific Python Lectures, Edition 2024.2rc0.dev0

import numpy as np
import matplotlib.pyplot as plt

def linestyle(ls, i):

X = i * 0.5 * np.ones(11)
Y = np.arange(11)
plt.plot(
X,
Y,
ls,
color=(0.0, 0.0, 1, 1),
lw=3,
ms=8,
mfc=(0.75, 0.75, 1, 1),
mec=(0, 0, 1, 1),
)
plt.text(0.5 * i, 10.25, ls, rotation=90, fontsize=15, va="bottom")

linestyles = [
"-",
"--",
":",
"-.",
".",
",",
"o",
"^",
"v",
"<",
">",
"s",
"+",
(continues on next page)

4.7. Full code examples 181

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

"x",
"d",
"1",
"2",
"3",
"4",
"h",
"p",
"|",
"_",
"D",
"H",
]
n_lines = len(linestyles)

size = 20 * n_lines, 300

dpi = 72.0
figsize = size[0] / float(dpi), size[1] / float(dpi)
fig = plt.figure(figsize=figsize, dpi=dpi)
plt.axes([0, 0.01, 1, 0.9], frameon=False)

for i, ls in enumerate(linestyles):
linestyle(ls, i)

plt.xlim(-0.2, 0.2 + 0.5 * n_lines)

plt.xticks([])
plt.yticks([])

plt.show()

Total running time of the script: (0 minutes 0.039 seconds)

Locators for tick on axis

An example demoing different locators to position ticks on axis for matplotlib.

4.7. Full code examples 182

 Scientific Python Lectures, Edition 2024.2rc0.dev0

import numpy as np
import matplotlib.pyplot as plt

def tickline():
plt.xlim(0, 10), plt.ylim(-1, 1), plt.yticks([])
ax = plt.gca()
ax.spines["right"].set_color("none")
ax.spines["left"].set_color("none")
ax.spines["top"].set_color("none")
ax.xaxis.set_ticks_position("bottom")
ax.spines["bottom"].set_position(("data", 0))
ax.yaxis.set_ticks_position("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)

size = 512, 40 * n_locators

dpi = 72.0
figsize = size[0] / float(dpi), size[1] / float(dpi)
fig = plt.figure(figsize=figsize, dpi=dpi)
fig.patch.set_alpha(0)

(continues on next page)

4.7. Full code examples 183

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

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=0.01, top=0.99, left=0.01, right=0.99)

plt.show()

Total running time of the script: (0 minutes 0.791 seconds)

Code generating the summary figures with a title

3D plotting vignette

Demo 3D plotting with matplotlib and decorate the figure.

import numpy as np
import matplotlib.pyplot as plt
from mpl_toolkits.mplot3d import Axes3D

fig = plt.figure()
(continues on next page)

4.7. Full code examples 184

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

ax = Axes3D(fig)
X = np.arange(-4, 4, 0.25)
Y = np.arange(-4, 4, 0.25)
X, Y = np.meshgrid(X, Y)
R = np.sqrt(X**2 + Y**2)
Z = np.sin(R)

ax.plot_surface(X, Y, Z, rstride=1, cstride=1, cmap=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,
0.93,
" 3D plots \n",
horizontalalignment="left",
verticalalignment="top",
size="xx-large",
bbox={"facecolor": "white", "alpha": 1.0},
transform=plt.gca().transAxes,
)

ax.text2D(
0.05,
0.87,
" Plot 2D or 3D data",
horizontalalignment="left",
verticalalignment="top",
size="large",
transform=plt.gca().transAxes,
)

plt.show()

Total running time of the script: (0 minutes 0.042 seconds)

Plotting in polar, decorated

An example showing how to plot in polar coordinate, and some decorations.

4.7. Full code examples 185

 Scientific Python Lectures, Edition 2024.2rc0.dev0

import numpy as np
import matplotlib.pyplot as plt

plt.subplot(1, 1, 1, polar=True)

N = 20
theta = np.arange(0.0, 2 * np.pi, 2 * np.pi / N)
rng = np.random.default_rng()
radii = 10 * rng.random(N)
width = np.pi / 4 * rng.random(N)
bars = plt.bar(theta, radii, width=width, bottom=0.0)
for r, bar in zip(radii, bars, strict=True):
bar.set_facecolor(plt.cm.jet(r / 10.0))
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",
bbox={"facecolor": "white", "alpha": 1.0},
transform=plt.gca().transAxes,
)
(continues on next page)

4.7. Full code examples 186

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

plt.text(
-0.2,
1.01,
"\n\n Plot anything using polar axis ",
horizontalalignment="left",
verticalalignment="top",
size="large",
transform=plt.gca().transAxes,
)

plt.show()

Total running time of the script: (0 minutes 0.120 seconds)

Plot example vignette

An example of plots with matplotlib, and added annotations.

import numpy as np
import matplotlib.pyplot as plt

n = 256
X = np.linspace(0, 2, n)
Y = np.sin(2 * np.pi * X)
(continues on next page)

4.7. Full code examples 187

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

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.gca()
ax.add_patch(
FancyBboxPatch(
(-0.05, 0.87),
width=0.66,
height=0.165,
clip_on=False,
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",
horizontalalignment="left",
verticalalignment="top",
size="xx-large",
transform=plt.gca().transAxes,
)

plt.text(
-0.05,
1.01,
"\n\n Plot lines and/or markers ",
horizontalalignment="left",
verticalalignment="top",
size="large",
transform=plt.gca().transAxes,
)

plt.show()

Total running time of the script: (0 minutes 0.028 seconds)

4.7. Full code examples 188

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Multiple plots vignette

Demo multiple plots and style the figure.

import matplotlib.pyplot as plt

ax = plt.subplot(2, 1, 1)
ax.set_xticklabels([])
ax.set_yticklabels([])

# Add a title and a box around it

from matplotlib.patches import FancyBboxPatch

ax = plt.gca()
ax.add_patch(
FancyBboxPatch(
(-0.05, 0.72),
width=0.66,
height=0.34,
clip_on=False,
boxstyle="square,pad=0",
zorder=3,
facecolor="white",
alpha=1.0,
transform=plt.gca().transAxes,
)
)
(continues on next page)

4.7. Full code examples 189

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

plt.text(
-0.05,
1.02,
" Multiplot: plt.subplot(...)\n",
horizontalalignment="left",
verticalalignment="top",
size="xx-large",
transform=ax.transAxes,
)
plt.text(
-0.05,
1.01,
"\n\n Plot several plots at once ",
horizontalalignment="left",
verticalalignment="top",
size="large",
transform=ax.transAxes,
)

ax = plt.subplot(2, 2, 3)
ax.set_xticklabels([])
ax.set_yticklabels([])

ax = plt.subplot(2, 2, 4)
ax.set_xticklabels([])
ax.set_yticklabels([])

plt.show()

Total running time of the script: (0 minutes 0.090 seconds)

Boxplot with matplotlib

An example of doing box plots with matplotlib

4.7. Full code examples 190

 Scientific Python Lectures, Edition 2024.2rc0.dev0

import numpy as np
import matplotlib.pyplot as plt

fig = plt.figure(figsize=(8, 5))

axes = plt.subplot(111)

n = 5
Z = np.zeros((n, 4))
X = np.linspace(0, 2, n)
rng = np.random.default_rng()
Y = rng.random((n, 4))
plt.boxplot(Y)

plt.xticks([])
plt.yticks([])

# Add a title and a box around it

from matplotlib.patches import FancyBboxPatch

ax = plt.gca()
ax.add_patch(
FancyBboxPatch(
(-0.05, 0.87),
width=0.66,
height=0.165,
clip_on=False,
boxstyle="square,pad=0",
zorder=3,
facecolor="white",
alpha=1.0,
transform=plt.gca().transAxes,
(continues on next page)

4.7. Full code examples 191

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

)
)

plt.text(
-0.05,
1.02,
" Box Plot: plt.boxplot(...)\n ",
horizontalalignment="left",
verticalalignment="top",
size="xx-large",
transform=axes.transAxes,
)

plt.text(
-0.04,
0.98,
"\n Make a box and whisker plot ",
horizontalalignment="left",
verticalalignment="top",
size="large",
transform=axes.transAxes,
)

plt.show()

Total running time of the script: (0 minutes 0.039 seconds)

Plot scatter decorated

An example showing the scatter function, with decorations.

4.7. Full code examples 192

 Scientific Python Lectures, Edition 2024.2rc0.dev0

import numpy as np
import matplotlib.pyplot as plt

n = 1024
rng = np.random.default_rng()
X = rng.normal(0, 1, n)
Y = rng.normal(0, 1, n)

T = np.arctan2(Y, X)

plt.scatter(X, Y, s=75, c=T, alpha=0.5)

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

# Add a title and a box around it

from matplotlib.patches import FancyBboxPatch

ax = plt.gca()
ax.add_patch(
FancyBboxPatch(
(-0.05, 0.87),
width=0.66,
height=0.165,
clip_on=False,
boxstyle="square,pad=0",
(continues on next page)

4.7. Full code examples 193

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

zorder=3,
facecolor="white",
alpha=1.0,
transform=plt.gca().transAxes,
)
)

plt.text(
-0.05,
1.02,
" Scatter Plot: plt.scatter(...)\n",
horizontalalignment="left",
verticalalignment="top",
size="xx-large",
transform=plt.gca().transAxes,
)

plt.text(
-0.05,
1.01,
"\n\n Make a scatter plot of x versus y ",
horizontalalignment="left",
verticalalignment="top",
size="large",
transform=plt.gca().transAxes,
)

plt.show()

Total running time of the script: (0 minutes 0.059 seconds)

Pie chart vignette

Demo pie chart with matplotlib and style the figure.

4.7. Full code examples 194

 Scientific Python Lectures, Edition 2024.2rc0.dev0

import numpy as np
import matplotlib.pyplot as plt

n = 20
X = np.ones(n)
X[-1] *= 2
plt.pie(X, explode=X * 0.05, colors=[f"{ i / float(n): f} " 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([])

# Add a title and a box around it

from matplotlib.patches import FancyBboxPatch

ax = plt.gca()
ax.add_patch(
FancyBboxPatch(
(-0.05, 0.87),
width=0.66,
height=0.165,
clip_on=False,
(continues on next page)

4.7. Full code examples 195

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

boxstyle="square,pad=0",
zorder=3,
facecolor="white",
alpha=1.0,
transform=plt.gca().transAxes,
)
)

plt.text(
-0.05,
1.02,
" Pie Chart: plt.pie(...)\n",
horizontalalignment="left",
verticalalignment="top",
size="xx-large",
transform=plt.gca().transAxes,
)

plt.text(
-0.05,
1.01,
"\n\n Make a pie chart of an array ",
horizontalalignment="left",
verticalalignment="top",
size="large",
transform=plt.gca().transAxes,
)

plt.show()

Total running time of the script: (0 minutes 0.052 seconds)

Imshow demo

Demoing imshow

4.7. Full code examples 196

 Scientific Python Lectures, Edition 2024.2rc0.dev0

import numpy as np
import matplotlib.pyplot as plt

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, 8 * n)
y = np.linspace(-3, 3, 6 * n)
X, Y = np.meshgrid(x, y)
Z = f(X, Y)
plt.imshow(Z, interpolation="nearest", cmap="bone", origin="lower")
plt.xticks([])
plt.yticks([])

# Add a title and a box around it

from matplotlib.patches import FancyBboxPatch

ax = plt.gca()
ax.add_patch(
FancyBboxPatch(
(-0.05, 0.87),
width=0.66,
height=0.165,
clip_on=False,
(continues on next page)

4.7. Full code examples 197

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

boxstyle="square,pad=0",
zorder=3,
facecolor="white",
alpha=1.0,
transform=plt.gca().transAxes,
)
)

plt.text(
-0.05,
1.02,
" Imshow: plt.imshow(...)\n",
horizontalalignment="left",
verticalalignment="top",
size="xx-large",
transform=plt.gca().transAxes,
)

plt.text(
-0.05,
1.01,
"\n\n Display an image to current axes ",
horizontalalignment="left",
verticalalignment="top",
family="DejaVu Sans",
size="large",
transform=plt.gca().transAxes,
)

plt.show()

Total running time of the script: (0 minutes 0.035 seconds)

Bar plot advanced

An more elaborate bar plot example

4.7. Full code examples 198

 Scientific Python Lectures, Edition 2024.2rc0.dev0

import numpy as np
import matplotlib.pyplot as plt

n = 16
X = np.arange(n)
Y1 = (1 - X / float(n)) * np.random.uniform(0.5, 1.0, n)
Y2 = (1 - X / float(n)) * np.random.uniform(0.5, 1.0, n)
plt.bar(X, Y1, facecolor="#9999ff", edgecolor="white")
plt.bar(X, -Y2, facecolor="#ff9999", edgecolor="white")
plt.xlim(-0.5, n)
plt.xticks([])
plt.ylim(-1, 1)
plt.yticks([])

# Add a title and a box around it

from matplotlib.patches import FancyBboxPatch

ax = plt.gca()
ax.add_patch(
FancyBboxPatch(
(-0.05, 0.87),
width=0.66,
height=0.165,
clip_on=False,
boxstyle="square,pad=0",
zorder=3,
facecolor="white",
(continues on next page)

4.7. Full code examples 199

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

alpha=1.0,
transform=plt.gca().transAxes,
)
)

plt.text(
-0.05,
1.02,
" Bar Plot: plt.bar(...)\n",
horizontalalignment="left",
verticalalignment="top",
size="xx-large",
transform=plt.gca().transAxes,
)

plt.text(
-0.05,
1.01,
"\n\n Make a bar plot with rectangles ",
horizontalalignment="left",
verticalalignment="top",
size="large",
transform=plt.gca().transAxes,
)

plt.show()

Total running time of the script: (0 minutes 0.045 seconds)

Plotting quiver decorated

An example showing quiver with decorations.

4.7. Full code examples 200

 Scientific Python Lectures, Edition 2024.2rc0.dev0

import numpy as np
import matplotlib.pyplot as plt

n = 8
X, Y = np.mgrid[0:n, 0:n]
T = np.arctan2(Y - n / 2.0, X - n / 2.0)
R = 10 + np.sqrt((Y - n / 2.0) ** 2 + (X - n / 2.0) ** 2)
U, V = R * np.cos(T), R * np.sin(T)

plt.quiver(X, Y, U, V, R, alpha=0.5)
plt.quiver(X, Y, U, V, edgecolor="k", facecolor="None", linewidth=0.5)

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

# Add a title and a box around it

from matplotlib.patches import FancyBboxPatch

ax = plt.gca()
ax.add_patch(
FancyBboxPatch(
(-0.05, 0.87),
width=0.66,
height=0.165,
clip_on=False,
(continues on next page)

4.7. Full code examples 201

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

boxstyle="square,pad=0",
zorder=3,
facecolor="white",
alpha=1.0,
transform=plt.gca().transAxes,
)
)

plt.text(
-0.05,
1.02,
" Quiver Plot: plt.quiver(...)\n",
horizontalalignment="left",
verticalalignment="top",
size="xx-large",
transform=plt.gca().transAxes,
)

plt.text(
-0.05,
1.01,
"\n\n Plot a 2-D field of arrows ",
horizontalalignment="left",
verticalalignment="top",
size="large",
transform=plt.gca().transAxes,
)

plt.show()

Total running time of the script: (0 minutes 0.043 seconds)

Display the contours of a function

An example demoing how to plot the contours of a function, with additional layout tweaks.

4.7. Full code examples 202

 Scientific Python Lectures, Edition 2024.2rc0.dev0

/home/runner/work/scientific-python-lectures/scientific-python-lectures/intro/
˓→matplotlib/examples/pretty_plots/plot_contour_ext.py:24: UserWarning: The following␣

˓→kwargs were not used by contour: 'linewidth'

C = plt.contour(X, Y, f(X, Y), 8, colors="black", linewidth=0.5)

import numpy as np
import matplotlib.pyplot as plt

def f(x, y):

return (1 - x / 2 + x**5 + y**3) * np.exp(-(x**2) - y**2)

n = 256
x = np.linspace(-3, 3, n)
y = np.linspace(-3, 3, n)
X, Y = np.meshgrid(x, y)

plt.contourf(X, Y, f(X, Y), 8, alpha=0.75, cmap=plt.cm.hot)

C = plt.contour(X, Y, f(X, Y), 8, colors="black", linewidth=0.5)
plt.clabel(C, inline=1, fontsize=10)
plt.xticks([])
plt.yticks([])
(continues on next page)

4.7. Full code examples 203

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

# Add a title and a box around it

from matplotlib.patches import FancyBboxPatch

ax = plt.gca()
ax.add_patch(
FancyBboxPatch(
(-0.05, 0.87),
width=0.66,
height=0.165,
clip_on=False,
boxstyle="square,pad=0",
zorder=3,
facecolor="white",
alpha=1.0,
transform=plt.gca().transAxes,
)
)

plt.text(
-0.05,
1.02,
" Contour Plot: plt.contour(..)\n",
horizontalalignment="left",
verticalalignment="top",
size="xx-large",
transform=plt.gca().transAxes,
)

plt.text(
-0.05,
1.01,
"\n\n Draw contour lines and filled contours ",
horizontalalignment="left",
verticalalignment="top",
size="large",
transform=plt.gca().transAxes,
)

plt.show()

Total running time of the script: (0 minutes 0.089 seconds)

4.7. Full code examples 204

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Grid elaborate

An example displaying a grid on the axes and tweaking the layout.

Text(-0.05, 1.01, '\n\n Draw ticks and grid ')

import matplotlib.pyplot as plt

from matplotlib.ticker import MultipleLocator

fig = plt.figure(figsize=(8, 6), dpi=72, facecolor="white")

axes = plt.subplot(111)
axes.set_xlim(0, 4)
axes.set_ylim(0, 3)

axes.xaxis.set_major_locator(MultipleLocator(1.0))
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([])
(continues on next page)

4.7. Full code examples 205

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

axes.set_yticklabels([])

# Add a title and a box around it

from matplotlib.patches import FancyBboxPatch

ax = plt.gca()
ax.add_patch(
FancyBboxPatch(
(-0.05, 0.87),
width=0.66,
height=0.165,
clip_on=False,
boxstyle="square,pad=0",
zorder=3,
facecolor="white",
alpha=1.0,
transform=plt.gca().transAxes,
)
)

plt.text(
-0.05,
1.02,
" Grid: plt.grid(...)\n",
horizontalalignment="left",
verticalalignment="top",
size="xx-large",
transform=axes.transAxes,
)

plt.text(
-0.05,
1.01,
"\n\n Draw ticks and grid ",
horizontalalignment="left",
verticalalignment="top",
size="large",
transform=axes.transAxes,
)

Total running time of the script: (0 minutes 0.089 seconds)

4.7. Full code examples 206

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Text printing decorated

An example showing text printing and decorating the resulting figure.

import numpy as np
import matplotlib.pyplot as plt

fig = plt.figure()
plt.xticks([])
plt.yticks([])

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} $"

)
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"$F_G = G\frac{m_1m_2} {r^2}$")

rng = np.random.default_rng()

(continues on next page)

4.7. Full code examples 207

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

for i in range(24):
index = rng.integers(0, len(eqs))
eq = eqs[index]
size = rng.uniform(12, 32)
x, y = rng.uniform(0, 1, 2)
alpha = rng.uniform(0.25, 0.75)
plt.text(
x,
y,
eq,
ha="center",
va="center",
color="#11557c",
alpha=alpha,
transform=plt.gca().transAxes,
fontsize=size,
clip_on=True,
)

# Add a title and a box around it

from matplotlib.patches import FancyBboxPatch

ax = plt.gca()
ax.add_patch(
FancyBboxPatch(
(-0.05, 0.87),
width=0.66,
height=0.165,
clip_on=False,
boxstyle="square,pad=0",
zorder=3,
facecolor="white",
alpha=1.0,
transform=plt.gca().transAxes,
)
)

plt.text(
-0.05,
1.02,
" Text: plt.text(...)\n",
horizontalalignment="left",
verticalalignment="top",
size="xx-large",
transform=plt.gca().transAxes,
)

plt.text(
-0.05,
1.01,
"\n\n Draw any kind of text ",
horizontalalignment="left",
verticalalignment="top",
size="large",
transform=plt.gca().transAxes,
(continues on next page)

4.7. Full code examples 208

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

)

plt.show()

Total running time of the script: (0 minutes 0.481 seconds)

4.7. Full code examples 209

 CHAPTER 5
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

210
 Scientific Python Lectures, Edition 2024.2rc0.dev0

• Optimization and fit: scipy.optimize

• Statistics and random numbers: scipy.stats
• Numerical integration: scipy.integrate
• Fast Fourier transforms: scipy.fft
• Signal processing: scipy.signal
• Image manipulation: scipy.ndimage
• Summary exercises on scientific computing
• Full code examples for the SciPy chapter

Warning: This tutorial is far from an introduction to numerical computing. As enumerating the
different submodules and functions in SciPy would be very boring, we concentrate instead on a few
examples to give a general idea of how to use scipy for scientific computing.

scipy is composed of task-specific sub-modules:

scipy.cluster Vector quantization / Kmeans

scipy.constants Physical and mathematical constants
scipy.fft Fourier transform
scipy.integrate Integration routines
scipy.interpolate Interpolation
scipy.io Data input and output
scipy.linalg Linear algebra routines
scipy.ndimage n-dimensional image package
scipy.odr Orthogonal distance regression
scipy.optimize Optimization
scipy.signal Signal processing
scipy.sparse Sparse matrices
scipy.spatial Spatial data structures and algorithms
scipy.special Any special mathematical functions
scipy.stats Statistics

Tip: They all depend on numpy, but are mostly independent of each other. The standard way of
importing NumPy and these SciPy modules is:

>>> import numpy as np

>>> import scipy as sp

5.1 File input/output: scipy.io

scipy.io contains functions for loading and saving data in several common formats including Matlab,
IDL, Matrix Market, and Harwell-Boeing.
Matlab files: Loading and saving:

>>> import scipy as sp

>>> a = np.ones((3, 3))
>>> sp.io.savemat('file.mat', {'a': a}) # savemat expects a dictionary
(continues on next page)

5.1. File input/output: scipy.io 211

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

>>> data = sp.io.loadmat('file.mat')
>>> data['a']
array([[1., 1., 1.],
[1., 1., 1.],
[1., 1., 1.]])

Warning: Python / Matlab mismatch: The Matlab file format does not support 1D arrays.
>>> a = np.ones(3)
>>> a
array([1., 1., 1.])
>>> a.shape
(3,)
>>> sp.io.savemat('file.mat', {'a': a})
>>> a2 = sp.io.loadmat('file.mat')['a']
>>> a2
array([[1., 1., 1.]])
>>> a2.shape
(1, 3)

Notice that the original array was a one-dimensional array, whereas the saved and reloaded array is
a two-dimensional array with a single row.
For other formats, see the scipy.io documentation.

See also:
• Load text files: numpy.loadtxt()/numpy.savetxt()
• Clever loading of text/csv files: numpy.genfromtxt()
• Fast and efficient, but NumPy-specific, binary format: numpy.save()/numpy.load()
• Basic input/output of images in Matplotlib: matplotlib.pyplot.imread()/matplotlib.pyplot.
imsave()
• More advanced input/output of images: imageio

5.2 Special functions: scipy.special

“Special” functions are functions commonly used in science and mathematics that are not considered to
be “elementary” functions. Examples include
• the gamma function, scipy.special.gamma(),
• the error function, scipy.special.erf(),
• Bessel functions, such as scipy.special.jv() (Bessel function of the first kind), and
• elliptic functions, such as scipy.special.ellipj() (Jacobi elliptic functions).
Other special functions are combinations of familiar elementary functions, but they offer better accuracy
or robustness than their naive implementations would.
Most of these function are computed elementwise and follow standard NumPy broadcasting rules when
the input arrays have different shapes. For example, scipy.special.xlog1py() is mathematically
equivalent to 𝑥 log(1 + 𝑦).

5.2. Special functions: scipy.special 212

 Scientific Python Lectures, Edition 2024.2rc0.dev0

>>> import scipy as sp

>>> x = np.asarray([1, 2])
>>> y = np.asarray([[3], [4], [5]])
>>> res = sp.special.xlog1py(x, y)
>>> res.shape
(3, 2)
>>> ref = x * np.log(1 + y)
>>> np.allclose(res, ref)
True

However, scipy.special.xlog1py() is numerically favorable for small 𝑦, when explicit addition of 1

would lead to loss of precision due to floating point truncation error.

>>> x = 2.5
>>> y = 1e-18
>>> x * np.log(1 + y)
0.0
>>> sp.special.xlog1py(x, y)
2.5e-18

Many special functions also have “logarithmized” variants. For instance, the gamma function Γ(·) is
related to the factorial function by 𝑛! = Γ(𝑛 + 1), but it extends the domain from the positive integers
to the complex plane.

>>> x = np.arange(10)
>>> np.allclose(sp.special.gamma(x + 1), sp.special.factorial(x))
True
>>> sp.special.gamma(5) < sp.special.gamma(5.5) < sp.special.gamma(6)
True

The factorial function grows quickly, and so the gamma function overflows for moderate values of the
argument. However, sometimes only the logarithm of the gamma function is needed. In such cases, we
can compute the logarithm of the gamma function directly using scipy.special.gammaln().

>>> x = [5, 50, 500]

>>> np.log(sp.special.gamma(x))
array([ 3.17805383, 144.56574395, inf])
>>> sp.special.gammaln(x)
array([ 3.17805383, 144.56574395, 2605.11585036])

Such functions can often be used when the intermediate components of a calculation would overflow
or underflow, but the final result would not. For example, suppose we wish to compute the ratio
Γ(500)/Γ(499).

>>> a = sp.special.gamma(500)
>>> b = sp.special.gamma(499)
>>> a, b
(inf, inf)

Both the numerator and denominator overflow, so performing 𝑎/𝑏 will not return the result we seek.
However, the magnitude of the result should be moderate, so the use of logarithms comes to mind.
Combining the identities log(𝑎/𝑏) = log(𝑎) − log(𝑏) and exp(log(𝑥)) = 𝑥, we get:

>>> log_a = sp.special.gammaln(500)

>>> log_b = sp.special.gammaln(499)
>>> log_res = log_a - log_b
>>> res = np.exp(log_res)
(continues on next page)

5.2. Special functions: scipy.special 213

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

>>> res
499.0000000...

Similarly, suppose we wish to compute the difference log(Γ(500) − Γ(499)). For this, we use scipy.
special.logsumexp(), which computes log(exp(𝑥)+exp(𝑦)) using a numerical trick that avoids overflow.

>>> res = sp.special.logsumexp([log_a, log_b],

... b=[1, -1]) # weights the terms of the sum
>>> res
2605.113844343...

For more information about these and many other special functions, see the documentation of scipy.
special.

5.3 Linear algebra operations: scipy.linalg

scipy.linalg provides a Python interface to efficient, compiled implementations of standard linear alge-
bra operations: the BLAS (Basic Linear Algebra Subroutines) and LAPACK (Linear Algebra PACKage)
libraries.
For example, the scipy.linalg.det() function computes the determinant of a square matrix:

>>> import scipy as sp

>>> arr = np.array([[1, 2],
... [3, 4]])
>>> sp.linalg.det(arr)
-2.0

Mathematically, the solution of a linear system 𝐴𝑥 = 𝑏 is 𝑥 = 𝐴−1 𝑏, but explicit inversion of a matrix is
numerically unstable and should be avoided. Instead, use scipy.linalg.solve():

>>> A = np.array([[1, 2],

... [2, 3]])
>>> b = np.array([14, 23])
>>> x = sp.linalg.solve(A, b)
>>> x
array([4., 5.])
>>> np.allclose(A @ x, b)
True

Linear systems with special structure can often be solved more efficiently than more general systems. For
example, systems with triangular matrices can be solved using scipy.linalg.solve_triangular():

>>> A_upper = np.triu(A)

>>> A_upper
array([[1, 2],
[0, 3]])
>>> np.allclose(sp.linalg.solve_triangular(A_upper, b, lower=False),
... sp.linalg.solve(A_upper, b))
True

scipy.linalg also features matrix factorizations/decompositions such as the singular value decomposi-
tion.

>>> A = np.array([[1, 2],

... [2, 3]])
(continues on next page)

5.3. Linear algebra operations: scipy.linalg 214

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

>>> U, s, Vh = sp.linalg.svd(A)
>>> s # singular values
array([4.23606798, 0.23606798])

The original matrix can be recovered by matrix multiplication of the factors:

>>> S = np.diag(s) # convert to diagonal matrix before matrix multiplication

>>> A2 = U @ S @ Vh
>>> np.allclose(A2, A)
True
>>> A3 = (U * s) @ Vh # more efficient: use array math broadcasting rules!
>>> np.allclose(A3, A)
True

Many other decompositions (e.g. LU, Cholesky, QR), solvers for structured linear systems (e.g. triangu-
lar, circulant), eigenvalue problem algorithms, matrix functions (e.g. matrix exponential), and routines
for special matrix creation (e.g. block diagonal, toeplitz) are available in scipy.linalg.

5.4 Interpolation: scipy.interpolate

scipy.interpolate is used for fitting a function – an “interpolant” – to experimental or computed data.

Once fit, the interpolant can be used to approximate the underlying function at intermediate points; it
can also be used to compute the integral, derivative, or inverse of the function.
Some kinds of interpolants, known as “smoothing splines”, are designed to generate smooth curves from
noisy data. For example, suppose we have the following data:

>>> rng = np.random.default_rng(27446968)

>>> measured_time = np.linspace(0, 2*np.pi, 20)
>>> function = np.sin(measured_time)
>>> noise = rng.normal(loc=0, scale=0.1, size=20)
>>> measurements = function + noise

scipy.interpolate.make_smoothing_spline() can be used to form a curve similar to the underlying

sine function.

>>> smoothing_spline = sp.interpolate.make_smoothing_spline(measured_time,␣

˓→measurements)

>>> interpolation_time = np.linspace(0, 2*np.pi, 200)

>>> smooth_results = smoothing_spline(interpolation_time)

5.4. Interpolation: scipy.interpolate 215

 Scientific Python Lectures, Edition 2024.2rc0.dev0

On the other hand, if the data are not noisy, it may be desirable to pass exactly through each point.

>>> interp_spline = sp.interpolate.make_interp_spline(measured_time, function)

>>> interp_results = interp_spline(interpolation_time)

The derivative and antiderivative methods of the result object can be used for differentiation and
integration. For the latter, the constant of integration is assumed to be zero, but we can “wrap” the
antiderivative to include a nonzero constant of integration.

>>> d_interp_spline = interp_spline.derivative()

>>> d_interp_results = d_interp_spline(interpolation_time)
>>> i_interp_spline = lambda t: interp_spline.antiderivative()(t) - 1
>>> i_interp_results = i_interp_spline(interpolation_time)

For functions that are monotonic on an interval (e.g. sin from 𝜋/2 to 3𝜋/2), we can reverse the arguments
of make_interp_spline to interpolate the inverse function. Because the first argument is expected to
be monotonically increasing, we also reverse the order of elements in the arrays with numpy.flip().

>>> i = (measured_time > np.pi/2) & (measured_time < 3*np.pi/2)

>>> inverse_spline = sp.interpolate.make_interp_spline(np.flip(function[i]),
... np.flip(measured_time[i]))
>>> inverse_spline(0)
array(3.14159265)

See the summary exercise on Maximum wind speed prediction at the Sprogø station for a more advanced
spline interpolation example, and read the SciPy interpolation tutorial and the scipy.interpolate
documentation for much more information.

5.4. Interpolation: scipy.interpolate 216

 Scientific Python Lectures, Edition 2024.2rc0.dev0

5.5 Optimization and fit: scipy.optimize

scipy.optimize provides algorithms for root finding, curve fitting, and more general optimization.

5.5.1 Root Finding

scipy.optimize.root_scalar() attempts to find a root of a specified scalar-valued function (i.e., an

argument at which the function value is zero). Like many scipy.optimize functions, the function needs
an initial guess of the solution, which the algorithm will refine until it converges or recognizes failure.
We also provide the derivative to improve the rate of convergence.

>>> def f(x):

... return (x-1)*(x-2)
>>> def df(x):
... return 2*x - 3
>>> x0 = 0 # guess
>>> res = sp.optimize.root_scalar(f, x0=x0, fprime=df)
>>> res
converged: True
flag: converged
function_calls: 12
iterations: 6
root: 1.0
method: newton

Warning: None of the functions in scipy.optimize that accept a guess are guaranteed to converge
for all possible guesses! (For example, try x0=1.5 in the example above, where the derivative of
the function is exactly zero.) If this occurs, try a different guess, adjust the options (like providing
a bracket as shown below), or consider whether SciPy offers a more appropriate method for the
problem.

Note that only one the root at 1.0 is found. By inspection, we can tell that there is a second root at
2.0. We can direct the function toward a particular root by changing the guess or by passing a bracket
that contains only the root we seek.

>>> res = sp.optimize.root_scalar(f, bracket=(1.5, 10))

>>> res.root
2.0

For multivariate problems, use scipy.optimize.root().

>>> def f(x):

... # intersection of unit circle and line from origin
... return [x[0]**2 + x[1]**2 - 1,
... x[1] - x[0]]
>>> res = sp.optimize.root(f, x0=[0, 0])
>>> np.allclose(f(res.x), 0, atol=1e-10)
True
>>> np.allclose(res.x, np.sqrt(2)/2)
True

Over-constrained problems can be solved in the least-squares sense using scipy.optimize.root() with
method='lm' (Levenberg-Marquardt).

5.5. Optimization and fit: scipy.optimize 217

 Scientific Python Lectures, Edition 2024.2rc0.dev0

>>> def f(x):

... # intersection of unit circle, line from origin, and parabola
... return [x[0]**2 + x[1]**2 - 1,
... x[1] - x[0],
... x[1] - x[0]**2]
>>> res = sp.optimize.root(f, x0=[1, 1], method='lm')
>>> res.success
True
>>> res.x
array([0.76096066, 0.66017736])

See the documentation of scipy.optimize.root_scalar() and scipy.optimize.root() for a variety

of other solution algorithms and options.

5.5.2 Curve fitting

Suppose we have data that is sinusoidal but noisy:

>>> x = np.linspace(-5, 5, num=50) # 50 values between -5 and 5

>>> noise = 0.01 * np.cos(100 * x)
>>> a, b = 2.9, 1.5
>>> y = a * np.cos(b * x) + noise

We can approximate the underlying amplitude, frequency, and phase from the data by least squares
curve fitting. To begin, we write a function that accepts the independent variable as the first argument
and all parameters to fit as separate arguments:

>>> def f(x, a, b, c):

... return a * np.sin(b * x + c)

We then use scipy.optimize.curve_fit() to find 𝑎 and 𝑏:

5.5. Optimization and fit: scipy.optimize 218

 Scientific Python Lectures, Edition 2024.2rc0.dev0

>>> params, _ = sp.optimize.curve_fit(f, x, y, p0=[2, 1, 3])

>>> params
array([2.900026 , 1.50012043, 1.57079633])
>>> ref = [a, b, np.pi/2] # what we'd expect
>>> np.allclose(params, ref, rtol=1e-3)
True

Exercise: Curve fitting of temperature data

The temperature extremes in Alaska for each month, starting in January, are given by (in
degrees Celsius):
max: 17, 19, 21, 28, 33, 38, 37, 37, 31, 23, 19, 18
min: -62, -59, -56, -46, -32, -18, -9, -13, -25, -46, -52, -58

1. Plot these temperature extremes.

2. Define a function that can describe min and max temperatures. Hint: this function
has to have a period of 1 year. Hint: include a time offset.
3. Fit this function to the data with scipy.optimize.curve_fit().
4. Plot the result. Is the fit reasonable? If not, why?
5. Is the time offset for min and max temperatures the same within the fit accuracy?
solution

5.5.3 Optimization

Suppose we wish to minimize the scalar-valued function of a single variable 𝑓 (𝑥) = 𝑥2 + 10 sin(𝑥):

>>> def f(x):

... return x**2 + 10*np.sin(x)
>>> x = np.arange(-5, 5, 0.1)
>>> plt.plot(x, f(x))
[<matplotlib.lines.Line2D object at ...>]
>>> plt.show()

We can see that the function has a local minimizer near 𝑥 = 3.8 and a global minimizer near 𝑥 = −1.3,
but the precise values cannot be determined from the plot.
The most appropriate function for this purpose is scipy.optimize.minimize_scalar(). Since we know

5.5. Optimization and fit: scipy.optimize 219

 Scientific Python Lectures, Edition 2024.2rc0.dev0

the approximate locations of the minima, we will provide bounds that restrict the search to the vicinity
of the global minimum.

>>> res = sp.optimize.minimize_scalar(f, bounds=(-2, -1))

>>> res
message: Solution found.
success: True
status: 0
fun: -7.9458233756...
x: -1.306440997...
nit: 8
nfev: 8
>>> res.fun == f(res.x)
True

If we did not already know the approximate location of the global minimum, we could use one of
SciPy’s global minimizers, such as scipy.optimize.differential_evolution(). We are required to
pass bounds, but they do not need to be tight.

>>> bounds=[(-5, 5)] # list of lower, upper bound for each variable
>>> res = sp.optimize.differential_evolution(f, bounds=bounds)
>>> res
message: Optimization terminated successfully.
success: True
fun: -7.9458233756...
x: [-1.306e+00]
nit: 6
nfev: 111
jac: [ 9.948e-06]

For multivariate optimization, a good choice for many problems is scipy.optimize.minimize(). Sup-
pose we wish to find the minimum of a quadratic function of two variables, 𝑓 (𝑥0 , 𝑥1 ) = (𝑥0 −1)2 +(𝑥1 −2)2 .

>>> def f(x):

... return (x[0] - 1)**2 + (x[1] - 2)**2

Like scipy.optimize.root(), scipy.optimize.minimize() requires a guess x0. (Note that this is the
initial value of both variables rather than the value of the variable we happened to label 𝑥0 .)

>>> res = sp.optimize.minimize(f, x0=[0, 0])

>>> res
message: Optimization terminated successfully.
success: True
status: 0
fun: 1.70578...e-16
x: [ 1.000e+00 2.000e+00]
nit: 2
jac: [ 3.219e-09 -8.462e-09]
hess_inv: [[ 9.000e-01 -2.000e-01]
[-2.000e-01 6.000e-01]]
nfev: 9
njev: 3

Maximization?

Is scipy.optimize.minimize() restricted to the solution of minimization problems? Nope! To solve

a maximization problem, simply minimize the negative of the original objective function.

5.5. Optimization and fit: scipy.optimize 220

 Scientific Python Lectures, Edition 2024.2rc0.dev0

This barely scratches the surface of SciPy’s optimization features, which include mixed integer linear
programming, constrained nonlinear programming, and the solution of assignment problems. For much
more information, see the documentation of scipy.optimize and the advanced chapter Mathematical
optimization: finding minima of functions.

Exercise: 2-D minimization

The six-hump camelback function

𝑥4 2
𝑓 (𝑥, 𝑦) = (4 − 2.1𝑥2 + )𝑥 + 𝑥𝑦 + (4𝑦 2 − 4)𝑦 2
3
has multiple local minima. Find a global minimum (there is more than one, each with the
same value of the objective function) and at least one other local minimum.
Hints:
• Variables can be restricted to −2 < 𝑥 < 2 and −1 < 𝑦 < 1.
• numpy.meshgrid() and matplotlib.pyplot.imshow() can help with visualization.
• Try minimizing with scipy.optimize.minimize() with an initial guess of (𝑥, 𝑦) =
(0, 0). Does it find the global minimum, or converge to a local minimum? What about
other initial guesses?
• Try minimizing with scipy.optimize.differential_evolution().
solution

See the summary exercise on Non linear least squares curve fitting: application to point extraction in
topographical lidar data for another, more advanced example.

5.5. Optimization and fit: scipy.optimize 221

 Scientific Python Lectures, Edition 2024.2rc0.dev0

5.6 Statistics and random numbers: scipy.stats

scipy.stats contains fundamental tools for statistics in Python.

5.6.1 Statistical Distributions

Consider a random variable distributed according to the standard normal. We draw a sample consisting
of 100000 observations from the random variable. The normalized histogram of the sample is an estimator
of the random variable’s probability density function (PDF):

>>> dist = sp.stats.norm(loc=0, scale=1) # standard normal distribution

>>> sample = dist.rvs(size=100000) # "random variate sample"
>>> plt.hist(sample, bins=50, density=True, label='normalized histogram')
>>> x = np.linspace(-5, 5)
>>> plt.plot(x, dist.pdf(x), label='PDF')
[<matplotlib.lines.Line2D object at ...>]
>>> plt.legend()
<matplotlib.legend.Legend object at ...>

Distribution objects and frozen distributions

Each of the 100+ scipy.stats distribution families is represented by an object with a __call__
method. Here, we call the scipy.stats.norm object to specify its location and scale, and it returns a
frozen distribution: a particular element of a distribution family with all parameters fixed. The frozen
distribution object has methods to compute essential functions of the particular distribution.

Suppose we knew that the sample had been drawn from a distribution belonging to the family of normal
distributions, but we did not know the particular distribution’s location (mean) and scale (standard
deviation). We perform maximum likelihood estimation of the unknown parameters using the distribution
family’s fit method:

>>> loc, scale = sp.stats.norm.fit(sample)

>>> loc
0.0015767005...
(continues on next page)

5.6. Statistics and random numbers: scipy.stats 222

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

>>> scale
0.9973396878...

Since we know the true parameters of the distribution from which the sample was drawn, we are not
surprised that these estimates are similar.

Exercise: Probability distributions

Generate 1000 random variates from a gamma distribution with a shape parameter of 1. Hint: the
shape parameter is passed as the first argument when freezing the distribution. Plot the histogram of
the sample, and overlay the distribution’s PDF. Estimate the shape parameter from the sample using
the fit method.
Extra: the distributions have many useful methods. Explore them using tab completion. Plot the
cumulative density function of the distribution, and compute the variance.

5.6.2 Sample Statistics and Hypothesis Tests

The sample mean is an estimator of the mean of the distribution from which the sample was drawn:

>>> np.mean(sample)
0.001576700508...

NumPy includes some of the most fundamental sample statistics (e.g. numpy.mean(), numpy.var(),
numpy.percentile()); scipy.stats includes many more. For instance, the geometric mean is a common
measure of central tendency for data that tends to be distributed over many orders of magnitude.

>>> sp.stats.gmean(2**sample)
1.0010934829...

SciPy also includes a variety of hypothesis tests that produce a sample statistic and a p-value. For
instance, suppose we wish to test the null hypothesis that sample was drawn from a normal distribution:

>>> res = sp.stats.normaltest(sample)

>>> res.statistic
5.20841759...
>>> res.pvalue
0.07396163283...

Here, statistic is a sample statistic that tends to be high for samples that are drawn from non-normal
distributions. pvalue is the probability of observing such a high value of the statistic for a sample
that has been drawn from a normal distribution. If the p-value is unusually small, this may be taken
as evidence that sample was not drawn from the normal distribution. Our statistic and p-value are
moderate, so the test is inconclusive.
There are many other features of scipy.stats, including circular statistics, quasi-Monte Carlo methods,
and resampling methods. For much more information, see the documentation of scipy.stats and the
advanced chapter statistics.

5.6. Statistics and random numbers: scipy.stats 223

 Scientific Python Lectures, Edition 2024.2rc0.dev0

5.7 Numerical integration: scipy.integrate

5.7.1 Quadrature
∫︀ 𝜋/2
Suppose we wish to compute the definite integral 0 sin(𝑡)𝑑𝑡 numerically. scipy.integrate.quad()
chooses one of several adaptive techniques depending on the parameters, and is therefore the recom-
mended first choice for integration of function of a single variable:

>>> integral, error_estimate = sp.integrate.quad(np.sin, 0, np.pi/2)

>>> np.allclose(integral, 1) # numerical result ~ analytical result
True
>>> abs(integral - 1) < error_estimate # actual error < estimated error
True

Other functions for numerical quadrature, including integration of multivariate functions and approxi-
mating integrals from samples, are available in scipy.integrate.

5.7.2 Initial Value Problems

scipy.integrate also features routines for integrating Ordinary Differential Equations (ODE). For
example, scipy.integrate.solve_ivp() integrates ODEs of the form:

𝑑𝑦
= 𝑓 (𝑡, 𝑦(𝑡))
𝑑𝑡
from an initial time 𝑡0 and initial state 𝑦(𝑡 = 𝑡0 ) = 𝑡0 to a final time 𝑡𝑓 or until an event occurs (e.g. a
specified state is reached).
As an introduction, consider the initial value problem given by 𝑑𝑦 𝑑𝑡 = −2𝑦 and the initial condition
𝑦(𝑡 = 0) = 1 on the interval 𝑡 = 0 . . . 4. We begin by defining a callable that computes 𝑓 (𝑡, 𝑦(𝑡)) given
the current time and state.

>>> def f(t, y):

... return -2 * y

Then, to compute y as a function of time:

>>> t_span = (0, 4) # time interval

>>> t_eval = np.linspace(*t_span) # times at which to evaluate `y`
>>> y0 = [1,] # initial state
>>> res = sp.integrate.solve_ivp(f, t_span=t_span, y0=y0, t_eval=t_eval)

and plot the result:

>>> plt.plot(res.t, res.y[0])

[<matplotlib.lines.Line2D object at ...>]
>>> plt.xlabel('t')
Text(0.5, ..., 't')
>>> plt.ylabel('y')
Text(..., 0.5, 'y')
>>> plt.title('Solution of Initial Value Problem')
Text(0.5, 1.0, 'Solution of Initial Value Problem')

5.7. Numerical integration: scipy.integrate 224

 Scientific Python Lectures, Edition 2024.2rc0.dev0

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¨ + 2𝜁𝜔0 𝑦˙ + 𝜔02 𝑦 = 0 with natural frequency 𝜔0 = 𝑘/𝑚, damping
ratio 𝜁 = 𝑐/(2𝑚𝜔0 ), spring constant 𝑘, mass 𝑚, and damping coefficient 𝑐.
Before using scipy.integrate.solve_ivp(), the 2nd order ODE needs to be transformed into a system
of first-order ODEs. Note that
𝑑𝑦 𝑑𝑦˙
= 𝑦˙ = 𝑦¨ = −(2𝜁𝜔0 𝑦˙ + 𝜔02 𝑦)
𝑑𝑡 𝑑𝑡
If we define 𝑧 = [𝑧0 , 𝑧1 ] where 𝑧0 = 𝑦 and 𝑧1 = 𝑦,˙ then the first order equation:
[︂ 𝑑𝑧0 ]︂ [︂ ]︂
𝑑𝑧 𝑑𝑡 𝑧1
= 𝑑𝑧1 =
𝑑𝑡 𝑑𝑡
−(2𝜁𝜔0 𝑧1 + 𝜔02 𝑧0 )

is equivalent to the original second order equation.

We set:

>>> m = 0.5 # kg
>>> k = 4 # N/m
>>> c = 0.4 # N s/m
>>> zeta = c / (2 * m * np.sqrt(k/m))
>>> omega = np.sqrt(k / m)

and define the function that computes 𝑧˙ = 𝑓 (𝑡, 𝑧(𝑡)):

>>> def f(t, z, zeta, omega):

... return (z[1], -2.0 * zeta * omega * z[1] - omega**2 * z[0])

Integration of the system follows:

>>> t_span = (0, 10)

>>> t_eval = np.linspace(*t_span, 100)
(continues on next page)

5.7. Numerical integration: scipy.integrate 225

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

>>> z0 = [1, 0]
>>> res = sp.integrate.solve_ivp(f, t_span, z0, t_eval=t_eval,
... args=(zeta, omega), method='LSODA')

Tip: With the option method=’LSODA’, scipy.integrate.solve_ivp() uses the LSODA (Liver-
more Solver for Ordinary Differential equations 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.

5.8 Fast Fourier transforms: scipy.fft

The scipy.fft module computes fast Fourier transforms (FFTs) and offers utilities to handle them.
Some important functions are:
• scipy.fft.fft() to compute the FFT
• scipy.fft.fftfreq() to generate the sampling frequencies
• scipy.fft.ifft() to compute the inverse FFT, from frequency space to signal space

As an illustration, a (noisy) input signal (sig), and its FFT:

>>> sig_fft = sp.fft.fft(sig)

>>> freqs = sp.fft.fftfreq(sig.size, d=time_step)

Signal FFT

As the signal comes from a real-valued function, the Fourier transform is symmetric.
The peak signal frequency can be found with freqs[power.argmax()]

5.8. Fast Fourier transforms: scipy.fft 226

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Setting the Fourier component above this frequency to zero and inverting the FFT with scipy.fft.
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.

Fully worked examples:

Crude periodicity finding (link) Gaussian image blur (link)

5.8. Fast Fourier transforms: scipy.fft 227

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Exercise: Denoise moon landing image

1. Examine the provided image moonlanding.png, which is heavily contaminated with periodic
noise. In this exercise, we aim to clean up the noise using the Fast Fourier Transform.
2. Load the image using matplotlib.pyplot.imread().
3. Find and use the 2-D FFT function in scipy.fft, and plot the spectrum (Fourier transform of)
the image. Do you have any trouble visualising the spectrum? If so, why?
4. The spectrum consists of high and low frequency components. The noise is contained in the
high-frequency part of the spectrum, so set some of those components to zero (use array slicing).
5. Apply the inverse Fourier transform to see the resulting image.
Solution

5.8. Fast Fourier transforms: scipy.fft 228

 Scientific Python Lectures, Edition 2024.2rc0.dev0

5.9 Signal processing: scipy.signal

Tip: scipy.signal is for typical signal processing: 1D, regularly-sampled signals.

Resampling scipy.signal.resample(): resample a signal to n points using FFT.

>>> t = np.linspace(0, 5, 100)

>>> x = np.sin(t)

>>> x_resampled = sp.signal.resample(x, 25)

>>> plt.plot(t, x)
[<matplotlib.lines.Line2D object at ...>]
>>> plt.plot(t[::4], x_resampled, 'ko')
[<matplotlib.lines.Line2D object at ...>]

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.

Detrending scipy.signal.detrend(): remove linear trend from signal:

5.9. Signal processing: scipy.signal 229

 Scientific Python Lectures, Edition 2024.2rc0.dev0

>>> t = np.linspace(0, 5, 100)

>>> rng = np.random.default_rng()
>>> x = t + rng.normal(size=100)

>>> x_detrended = sp.signal.detrend(x)

>>> plt.plot(t, x)
[<matplotlib.lines.Line2D object at ...>]
>>> plt.plot(t, x_detrended)
[<matplotlib.lines.Line2D object at ...>]

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).

5.9. Signal processing: scipy.signal 230

 Scientific Python Lectures, Edition 2024.2rc0.dev0

5.10 Image manipulation: scipy.ndimage

scipy.ndimage provides manipulation of n-dimensional arrays as images.

5.10.1 Geometrical transformations on images

Changing orientation, resolution, ..

>>> import scipy as sp

>>> # Load an image

>>> face = sp.datasets.face(gray=True)

>>> # Shift, rotate and zoom it

>>> shifted_face = sp.ndimage.shift(face, (50, 50))
>>> shifted_face2 = sp.ndimage.shift(face, (50, 50), mode='nearest')
>>> rotated_face = sp.ndimage.rotate(face, 30)
>>> cropped_face = face[50:-50, 50:-50]
>>> zoomed_face = sp.ndimage.zoom(face, 2)
>>> zoomed_face.shape
(1536, 2048)

>>> plt.subplot(151)
<Axes: >

>>> plt.imshow(shifted_face, cmap=plt.cm.gray)

<matplotlib.image.AxesImage object at 0x...>

>>> plt.axis('off')
(-0.5, 1023.5, 767.5, -0.5)

>>> # etc.

5.10.2 Image filtering

Generate a noisy face:

>>> import scipy as sp

>>> face = sp.datasets.face(gray=True)
>>> face = face[:512, -512:] # crop out square on right
>>> import numpy as np
>>> noisy_face = np.copy(face).astype(float)
>>> rng = np.random.default_rng()
>>> noisy_face += face.std() * 0.5 * rng.standard_normal(face.shape)

Apply a variety of filters on it:

5.10. Image manipulation: scipy.ndimage 231

 Scientific Python Lectures, Edition 2024.2rc0.dev0

>>> blurred_face = sp.ndimage.gaussian_filter(noisy_face, sigma=3)

>>> median_face = sp.ndimage.median_filter(noisy_face, size=5)
>>> wiener_face = sp.signal.wiener(noisy_face, (5, 5))

Other filters in scipy.ndimage.filters and scipy.signal can be applied to images.

Exercise

Compare histograms for the different filtered images.

5.10.3 Mathematical morphology

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.

Mathematical-morphology operations use a structuring element in order to modify geometrical structures.

Let us first generate a structuring element:

>>> el = sp.ndimage.generate_binary_structure(2, 1)
>>> el
array([[False, True, False],
[...True, True, True],
[False, True, False]])
>>> el.astype(int)
array([[0, 1, 0],
[1, 1, 1],
[0, 1, 0]])

5.10. Image manipulation: scipy.ndimage 232

 Scientific Python Lectures, Edition 2024.2rc0.dev0

• Erosion scipy.ndimage.binary_erosion()

>>> a = np.zeros((7, 7), dtype=int)

>>> a[1:6, 2:5] = 1
>>> a
array([[0, 0, 0, 0, 0, 0, 0],
[0, 0, 1, 1, 1, 0, 0],
[0, 0, 1, 1, 1, 0, 0],
[0, 0, 1, 1, 1, 0, 0],
[0, 0, 1, 1, 1, 0, 0],
[0, 0, 1, 1, 1, 0, 0],
[0, 0, 0, 0, 0, 0, 0]])
>>> sp.ndimage.binary_erosion(a).astype(a.dtype)
array([[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 1, 0, 0, 0],
[0, 0, 0, 1, 0, 0, 0],
[0, 0, 0, 1, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0]])
>>> # Erosion removes objects smaller than the structure
>>> sp.ndimage.binary_erosion(a, structure=np.ones((5,5))).astype(a.dtype)
array([[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0]])

• Dilation 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.]])
>>> sp.ndimage.binary_dilation(a).astype(a.dtype)
array([[0., 0., 0., 0., 0.],
[0., 0., 1., 0., 0.],
[0., 1., 1., 1., 0.],
[0., 0., 1., 0., 0.],
[0., 0., 0., 0., 0.]])

• Opening scipy.ndimage.binary_opening()

>>> a = np.zeros((5, 5), dtype=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]])
(continues on next page)

5.10. Image manipulation: scipy.ndimage 233

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

>>> # Opening removes small objects
>>> sp.ndimage.binary_opening(a, structure=np.ones((3, 3))).astype(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
>>> sp.ndimage.binary_opening(a).astype(int)
array([[0, 0, 0, 0, 0],
[0, 0, 1, 0, 0],
[0, 1, 1, 1, 0],
[0, 0, 1, 0, 0],
[0, 0, 0, 0, 0]])

• Closing: scipy.ndimage.binary_closing()

Exercise

Check that opening amounts to eroding, then dilating.

An opening operation removes small structures, while a closing operation fills small holes. Such opera-
tions can therefore be used to “clean” an image.

>>> a = np.zeros((50, 50))

>>> a[10:-10, 10:-10] = 1
>>> rng = np.random.default_rng()
>>> a += 0.25 * rng.standard_normal(a.shape)
>>> mask = a>=0.5
>>> opened_mask = sp.ndimage.binary_opening(mask)
>>> closed_mask = sp.ndimage.binary_closing(opened_mask)

Exercise

Check that the area of the reconstructed square is smaller than the area of the initial
square. (The opposite would occur if the closing step was performed before the opening).

For gray-valued images, eroding (resp. dilating) amounts to replacing a pixel by the minimal (resp.
maximal) value among pixels covered by the structuring element centered on the pixel of interest.

>>> a = np.zeros((7, 7), dtype=int)

>>> a[1:6, 1:6] = 3
(continues on next page)

5.10. Image manipulation: scipy.ndimage 234

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

>>> a[4, 4] = 2; a[2, 3] = 1
>>> a
array([[0, 0, 0, 0, 0, 0, 0],
[0, 3, 3, 3, 3, 3, 0],
[0, 3, 3, 1, 3, 3, 0],
[0, 3, 3, 3, 3, 3, 0],
[0, 3, 3, 3, 2, 3, 0],
[0, 3, 3, 3, 3, 3, 0],
[0, 0, 0, 0, 0, 0, 0]])
>>> sp.ndimage.grey_erosion(a, size=(3, 3))
array([[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0],
[0, 0, 1, 1, 1, 0, 0],
[0, 0, 1, 1, 1, 0, 0],
[0, 0, 3, 2, 2, 0, 0],
[0, 0, 0, 0, 0, 0, 0],
[0, 0, 0, 0, 0, 0, 0]])

5.10.4 Connected components and measurements on images

Let us first generate a nice synthetic binary image.

>>> x, y = np.indices((100, 100))

>>> sig = np.sin(2*np.pi*x/50.) * np.sin(2*np.pi*y/50.) * (1+x*y/50.**2)**2
>>> mask = sig > 1

scipy.ndimage.label() assigns a different label to each connected component:

5.10. Image manipulation: scipy.ndimage 235

 Scientific Python Lectures, Edition 2024.2rc0.dev0

>>> labels, nb = sp.ndimage.label(mask)

>>> nb
8

Now compute measurements on each connected component:

>>> areas = sp.ndimage.sum(mask, labels, range(1, labels.max()+1))

>>> areas # The number of pixels in each connected component
array([190., 45., 424., 278., 459., 190., 549., 424.])
>>> maxima = sp.ndimage.maximum(sig, labels, range(1, labels.max()+1))
>>> maxima # The maximum signal in each connected component
array([ 1.80238238, 1.13527605, 5.51954079, 2.49611818, 6.71673619,
1.80238238, 16.76547217, 5.51954079])

Extract the 4th connected component, and crop the array around it:

>>> sp.ndimage.find_objects(labels)[3]
(slice(30, 48, None), slice(30, 48, None))
>>> sl = sp.ndimage.find_objects(labels)[3]
>>> import matplotlib.pyplot as plt
>>> plt.imshow(sig[sl])
<matplotlib.image.AxesImage object at ...>

See the summary exercise on Image processing application: counting bubbles and unmolten grains for a
more advanced example.

5.11 Summary exercises on scientific computing

The summary exercises use mainly NumPy, SciPy and Matplotlib. They provide some real-life examples
of scientific computing with Python. Now that the basics of working with NumPy and SciPy have been
introduced, the interested user is invited to try these exercises.

5.11. Summary exercises on scientific computing 236

 Scientific Python Lectures, Edition 2024.2rc0.dev0

5.11.1 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
it. In the current model, it is supposed that the maximum wind speed occurring every 50 years is defined
as the upper 2% quantile.
By definition, the quantile function is the inverse of the cumulative distribution function. The latter
describes the probability distribution of an annual maxima. In the exercise, the cumulative probability
p_i for a given year i is defined as p_i = i/(N+1) with N = 21, the number of measured years. Thus
it will be possible to calculate the cumulative probability of every measured wind speed maxima. From
those experimental points, the scipy.interpolate module will be very useful for fitting the quantile function.
Finally the 50 years maxima is going to be evaluated from the cumulative probability of the 2% quantile.

Computing the cumulative probabilities

The annual wind speeds maxima have already been computed and saved in the NumPy format in the
file examples/max-speeds.npy, thus they will be loaded by using NumPy:

>>> import numpy as np

>>> max_speeds = np.load('intro/scipy/summary-exercises/examples/max-speeds.npy')
>>> years_nb = max_speeds.shape[0]

Following the cumulative probability definition p_i from the previous section, the corresponding values
will be:

>>> cprob = (np.arange(years_nb, dtype=np.float32) + 1)/(years_nb + 1)

and they are assumed to fit the given wind speeds:

>>> sorted_max_speeds = np.sort(max_speeds)

Prediction with UnivariateSpline

In this section the quantile function will be estimated by using the UnivariateSpline class which can
represent a spline from points. The default behavior is to build a spline of degree 3 and points can
have different weights according to their reliability. Variants are InterpolatedUnivariateSpline and
LSQUnivariateSpline on which errors checking is going to change. In case a 2D spline is wanted, the
BivariateSpline class family is provided. All those classes for 1D and 2D splines use the FITPACK
Fortran subroutines, that’s why a lower library access is available through the splrep and splev functions
for respectively representing and evaluating a spline. Moreover interpolation functions without the use
of FITPACK parameters are also provided for simpler use.
For the Sprogø maxima wind speeds, the UnivariateSpline will be used because a spline of degree 3
seems to correctly fit the data:

5.11. Summary exercises on scientific computing 237

 Scientific Python Lectures, Edition 2024.2rc0.dev0

>>> import scipy as sp

>>> quantile_func = sp.interpolate.UnivariateSpline(cprob, sorted_max_speeds)

The quantile function is now going to be evaluated from the full range of probabilities:

>>> nprob = np.linspace(0, 1, 100)

>>> 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. 1: Solution: Python source file

5.11. Summary exercises on scientific computing 238

 Scientific Python Lectures, Edition 2024.2rc0.dev0

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.

Fig. 2: Solution: Python source file

• The second step will be to use the Gumbell distribution on cumulative probabilities p_i defined as
-log( -log(p_i) ) for fitting a linear quantile function (remember that you can define the degree
of the UnivariateSpline). Plotting the annual maxima versus the Gumbell distribution should
give you the following figure.
• The last step will be to find 34.23 m/s for the maximum wind speed occurring every 50 years.

5.11. Summary exercises on scientific computing 239

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Fig. 3: Solution: Python source file

5.11. Summary exercises on scientific computing 240

 Scientific Python Lectures, Edition 2024.2rc0.dev0

5.11.2 Non linear least squares curve fitting: application to point extraction in
topographical lidar data

The goal of this exercise is to fit a model to some data. The data used in this tutorial are lidar data
and are described in details in the following introductory paragraph. If 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
contains peaks whose center and amplitude permit to compute the position and some characteristics of
the hit target. When the footprint of the laser beam is around 1m on the Earth surface, the beam can
hit multiple targets during the two-way propagation (for example the ground and the top of a tree or
building). The sum of the contributions of each target hit by the laser beam then produces a complex
signal with multiple peaks, each one containing information about one target.
One state of the art method to extract information from these data is to decompose them in a sum of
Gaussian functions where each function represents the contribution of a target hit by the laser beam.
Therefore, we use the scipy.optimize module to fit a waveform to one or a sum of Gaussian functions.

Loading and visualization

Load the first waveform using:

>>> import numpy as np

>>> waveform_1 = np.load('intro/scipy/summary-exercises/examples/waveform_1.npy')

and visualize it:

>>> import matplotlib.pyplot as plt

>>> t = np.arange(len(waveform_1))
>>> plt.plot(t, waveform_1)
[<matplotlib.lines.Line2D object at ...>]
>>> 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.
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.

5.11. Summary exercises on scientific computing 241

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Fitting a waveform with a simple Gaussian model

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

Model

A Gaussian function defined by

{︃ (︂ )︂2 }︃
𝑡−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)

5.11. Summary exercises on scientific computing 242

 Scientific Python Lectures, Edition 2024.2rc0.dev0

• 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,
the function to minimize is the residuals (the difference between the data and the model):

>>> def residuals(coeffs, y, t):

... return y - model(t, coeffs)

So let’s get our solution by calling scipy.optimize.leastsq() with the following arguments:
• the function to minimize
• an initial solution
• the additional arguments to pass to the function

>>> import scipy as sp

>>> t = np.arange(len(waveform_1))
>>> x, flag = sp.optimize.leastsq(residuals, x0, args=(waveform_1, t))
>>> x
array([ 2.70363, 27.82020, 15.47924, 3.05636])

And visualize the solution:

fig, ax = plt.subplots(figsize=(8, 6))

plt.plot(t, waveform_1, t, model(t, x))
plt.xlabel("Time [ns]")
plt.ylabel("Amplitude [bins]")
plt.legend(["Waveform", "Model"])
plt.show()

Remark: from scipy v0.8 and above, you should rather use scipy.optimize.curve_fit() which takes
the model and the data as arguments, so you don’t need to define the residuals any more.

Going further

• Try with a more complex waveform (for instance waveform_2.npy) that contains three significant
peaks. You must adapt the model which is now a sum of Gaussian functions instead of only one
Gaussian peak.
• In some cases, writing an explicit function to compute the Jacobian is faster than letting leastsq
estimate it numerically. Create a function to compute the Jacobian of the residuals and use it as
an input for leastsq.
• When we want to detect very small peaks in the signal, or when the initial guess is too far from a
good solution, the result given by the algorithm is often not satisfying. Adding constraints to the
parameters of the model enables to overcome such limitations. An example of a priori knowledge
we can add is the sign of our variables (which are all positive).
• See the solution.

5.11. Summary exercises on scientific computing 243

 Scientific Python Lectures, Edition 2024.2rc0.dev0

5.11. Summary exercises on scientific computing 244

 Scientific Python Lectures, Edition 2024.2rc0.dev0

• Further exercise: compare the result of scipy.optimize.leastsq() and what you can get with
scipy.optimize.fmin_slsqp() when adding boundary constraints.

5.11.3 Image processing application: counting bubbles and unmolten grains

Statement of the problem

1. Open the image file MV_HFV_012.jpg and display it. Browse through the keyword arguments
in the docstring of imshow to display the image with the “right” orientation (origin in the bottom
left corner, and not the upper left corner as for standard arrays).
This Scanning Element Microscopy image shows a glass sample (light gray matrix) with some
bubbles (on black) and unmolten sand grains (dark gray). We wish to determine the fraction
of the sample covered by these three phases, and to estimate the typical size of sand grains and
bubbles, their sizes, etc.
2. Crop the image to remove the lower panel with measure information.
3. Slightly filter the image with a median filter in order to refine its histogram. Check how the
histogram changes.

5.11. Summary exercises on scientific computing 245

 Scientific Python Lectures, Edition 2024.2rc0.dev0

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.

5.11.4 Example of solution for the image processing exercise: unmolten grains in
glass

1. Open the image file MV_HFV_012.jpg and display it. Browse through the keyword arguments
in the docstring of imshow to display the image with the “right” orientation (origin in the bottom
left corner, and not the upper left corner as for standard arrays).

>>> dat = plt.imread('data/MV_HFV_012.jpg')

5.11. Summary exercises on scientific computing 246

 Scientific Python Lectures, Edition 2024.2rc0.dev0

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 = sp.ndimage.median_filter(dat, size=(7,7))

>>> hi_dat = np.histogram(dat, bins=np.arange(256))
>>> hi_filtdat = np.histogram(filtdat, bins=np.arange(256))

4. Using the histogram of the filtered image, determine thresholds that allow to define masks for sand
pixels, glass pixels and bubble pixels. Other option (homework): write a function that determines
automatically the thresholds from the minima of the histogram.

>>> void = filtdat <= 50

>>> sand = np.logical_and(filtdat > 50, filtdat <= 114)
>>> glass = filtdat > 114

5. Display an image in which the three phases are colored with three different colors.

>>> phases = void.astype(int) + 2*glass.astype(int) + 3*sand.astype(int)

5.11. Summary exercises on scientific computing 247

 Scientific Python Lectures, Edition 2024.2rc0.dev0

6. Use mathematical morphology to clean the different phases.

>>> sand_op = sp.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 sp.ndimage.sum or np.bincount to compute the grain sizes.

>>> sand_labels, sand_nb = sp.ndimage.label(sand_op)

>>> sand_areas = np.array(sp.ndimage.sum(sand_op, sand_labels, np.arange(sand_
˓→labels.max()+1)))

>>> mask = sand_areas > 100

>>> remove_small_sand = mask[sand_labels.ravel()].reshape(sand_labels.shape)

8. Compute the mean size of bubbles.

>>> bubbles_labels, bubbles_nb = sp.ndimage.label(void)

>>> bubbles_areas = np.bincount(bubbles_labels.ravel())[1:]
>>> mean_bubble_size = bubbles_areas.mean()
>>> median_bubble_size = np.median(bubbles_areas)
(continues on next page)

5.11. Summary exercises on scientific computing 248

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

>>> mean_bubble_size, median_bubble_size
(1699.875, 65.0)

5.12 Full code examples for the SciPy chapter

5.12.1 Finding the minimum of a smooth function

Demos various methods to find the minimum of a function.

import numpy as np
import matplotlib.pyplot as plt

def f(x):
return x**2 + 10 * np.sin(x)

x = np.arange(-5, 5, 0.1)
plt.plot(x, f(x))

[<matplotlib.lines.Line2D object at 0x7f49548eef90>]

Now find the minimum with a few methods

5.12. Full code examples for the SciPy chapter 249

 Scientific Python Lectures, Edition 2024.2rc0.dev0

import scipy as sp

# The default (Nelder Mead)

print(sp.optimize.minimize(f, x0=0))

message: Optimization terminated successfully.

success: True
status: 0
fun: -7.945823375615215
x: [-1.306e+00]
nit: 5
jac: [-1.192e-06]
hess_inv: [[ 8.589e-02]]
nfev: 12
njev: 6

plt.show()

Total running time of the script: (0 minutes 0.043 seconds)

5.12.2 Resample a signal with scipy.signal.resample

scipy.signal.resample() uses FFT to resample a 1D signal.

Generate a signal with 100 data point

import numpy as np

t = np.linspace(0, 5, 100)
x = np.sin(t)

Downsample it by a factor of 4

import scipy as sp

x_resampled = sp.signal.resample(x, 25)

Plot

import matplotlib.pyplot as plt

plt.figure(figsize=(5, 4))
plt.plot(t, x, label="Original signal")
plt.plot(t[::4], x_resampled, "ko", label="Resampled signal")

plt.legend(loc="best")
plt.show()

5.12. Full code examples for the SciPy chapter 250

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Total running time of the script: (0 minutes 0.219 seconds)

5.12.3 Detrending a signal

scipy.signal.detrend() removes a linear trend.

Generate a random signal with a trend

import numpy as np

t = np.linspace(0, 5, 100)
rng = np.random.default_rng()
x = t + rng.normal(size=100)

Detrend

import scipy as sp

x_detrended = sp.signal.detrend(x)

Plot

import matplotlib.pyplot as plt

plt.figure(figsize=(5, 4))
plt.plot(t, x, label="x")
plt.plot(t, x_detrended, label="x_detrended")
plt.legend(loc="best")
plt.show()

5.12. Full code examples for the SciPy chapter 251

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Total running time of the script: (0 minutes 0.049 seconds)

5.12.4 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
import scipy as sp
import matplotlib.pyplot as plt

(continues on next page)

5.12. Full code examples for the SciPy chapter 252

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

def f(t, y):
return -2 * y

t_span = (0, 4) # time interval

t_eval = np.linspace(*t_span) # times at which to evaluate `y`
y0 = [
1,
] # initial state
res = sp.integrate.solve_ivp(f, t_span=t_span, y0=y0, t_eval=t_eval)

plt.figure(figsize=(4, 3))
plt.plot(res.t, res.y[0])
plt.xlabel("t")
plt.ylabel("y")
plt.title("Solution of Initial Value Problem")
plt.tight_layout()
plt.show()

Total running time of the script: (0 minutes 0.088 seconds)

5.12.5 Normal distribution: histogram and PDF

Explore the normal distribution: a histogram built from samples and the PDF (probability density
function).

5.12. Full code examples for the SciPy chapter 253

 Scientific Python Lectures, Edition 2024.2rc0.dev0

import numpy as np
import scipy as sp
import matplotlib.pyplot as plt

dist = sp.stats.norm(loc=0, scale=1) # standard normal distribution

sample = dist.rvs(size=100000) # "random variate sample"
plt.hist(
sample,
bins=51, # group the observations into 50 bins
density=True, # normalize the frequencies
label="normalized histogram",
)

x = np.linspace(-5, 5) # possible values of the random variable

plt.plot(x, dist.pdf(x), label="PDF")
plt.legend()
plt.show()

Total running time of the script: (0 minutes 0.107 seconds)

5.12.6 Integrate the Damped spring-mass oscillator

import numpy as np
import scipy as sp
import matplotlib.pyplot as plt

m = 0.5 # kg
k = 4 # N/m
c = 0.4 # N s/m

zeta = c / (2 * m * np.sqrt(k / m))

omega = np.sqrt(k / m)

def f(t, z, zeta, omega):

(continues on next page)

5.12. Full code examples for the SciPy chapter 254

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

return (z[1], -zeta * omega * z[1] - omega**2 * z[0])

t_span = (0, 10)

t_eval = np.linspace(*t_span, 100)
z0 = [1, 0]
res = sp.integrate.solve_ivp(
f, t_span, z0, t_eval=t_eval, args=(zeta, omega), method="LSODA"
)

plt.figure(figsize=(4, 3))
plt.plot(res.t, res.y[0], label="y")
plt.plot(res.t, res.y[1], label="dy/dt")
plt.legend(loc="best")
plt.show()

Total running time of the script: (0 minutes 0.056 seconds)

5.12.7 Comparing 2 sets of samples from Gaussians

import numpy as np
import matplotlib.pyplot as plt

# Generates 2 sets of observations

rng = np.random.default_rng(27446968)
samples1 = rng.normal(0, size=1000)
samples2 = rng.normal(1, size=1000)

(continues on next page)

5.12. Full code examples for the SciPy chapter 255

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

# Compute a histogram of the sample
bins = np.linspace(-4, 4, 30)
histogram1, bins = np.histogram(samples1, bins=bins, density=True)
histogram2, bins = np.histogram(samples2, bins=bins, density=True)

plt.figure(figsize=(6, 4))
plt.hist(samples1, bins=bins, density=True, label="Samples 1")
plt.hist(samples2, bins=bins, density=True, label="Samples 2")
plt.legend(loc="best")
plt.show()

Total running time of the script: (0 minutes 0.106 seconds)

5.12.8 Curve fitting

Demos a simple curve fitting

First generate some data

import numpy as np

# Seed the random number generator for reproducibility

rng = np.random.default_rng(27446968)

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

noise = 0.01 * np.cos(100 * x_data)
a, b = 2.9, 1.5
y_data = a * np.cos(b * x_data) + noise

# And plot it
import matplotlib.pyplot as plt

plt.figure(figsize=(6, 4))
plt.scatter(x_data, y_data)

5.12. Full code examples for the SciPy chapter 256

 Scientific Python Lectures, Edition 2024.2rc0.dev0

<matplotlib.collections.PathCollection object at 0x7f4954cc1850>

Now fit a simple sine function to the data

import scipy as sp

def test_func(x, a, b, c):

return a * np.sin(b * x + c)

params, params_covariance = sp.optimize.curve_fit(

test_func, x_data, y_data, p0=[2, 1, 3]
)

print(params)

[2.900026 1.50012043 1.57079633]

And plot the resulting curve on the data

plt.figure(figsize=(6, 4))
plt.scatter(x_data, y_data, label="Data")
plt.plot(x_data, test_func(x_data, *params), label="Fitted function")

plt.legend(loc="best")

plt.show()

5.12. Full code examples for the SciPy chapter 257

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Total running time of the script: (0 minutes 0.132 seconds)

5.12.9 Spectrogram, power spectral density

Demo spectrogram and power spectral density on a frequency chirp.

import numpy as np
import matplotlib.pyplot as plt

Generate a chirp signal

# Seed the random number generator

np.random.seed(0)

time_step = 0.01
time_vec = np.arange(0, 70, time_step)

# A signal with a small frequency chirp

sig = np.sin(0.5 * np.pi * time_vec * (1 + 0.1 * time_vec))

plt.figure(figsize=(8, 5))
plt.plot(time_vec, sig)

5.12. Full code examples for the SciPy chapter 258

 Scientific Python Lectures, Edition 2024.2rc0.dev0

[<matplotlib.lines.Line2D object at 0x7f4987c9d910>]

Compute and plot the spectrogram

The spectrum of the signal on consecutive time windows

import scipy as sp

freqs, times, spectrogram = sp.signal.spectrogram(sig)

plt.figure(figsize=(5, 4))
plt.imshow(spectrogram, aspect="auto", cmap="hot_r", origin="lower")
plt.title("Spectrogram")
plt.ylabel("Frequency band")
plt.xlabel("Time window")
plt.tight_layout()

5.12. Full code examples for the SciPy chapter 259

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Compute and plot the power spectral density (PSD)

The power of the signal per frequency band

freqs, psd = sp.signal.welch(sig)

plt.figure(figsize=(5, 4))
plt.semilogx(freqs, psd)
plt.title("PSD: power spectral density")
plt.xlabel("Frequency")
plt.ylabel("Power")
plt.tight_layout()

5.12. Full code examples for the SciPy chapter 260

 Scientific Python Lectures, Edition 2024.2rc0.dev0

plt.show()

Total running time of the script: (0 minutes 0.386 seconds)

5.12.10 Demo mathematical morphology

A basic demo of binary opening and closing.

# Generate some binary data

import numpy as np

np.random.seed(0)
a = np.zeros((50, 50))
a[10:-10, 10:-10] = 1
a += 0.25 * np.random.standard_normal(a.shape)
mask = a >= 0.5

# Apply mathematical morphology

(continues on next page)

5.12. Full code examples for the SciPy chapter 261

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

import scipy as sp

opened_mask = sp.ndimage.binary_opening(mask)
closed_mask = sp.ndimage.binary_closing(opened_mask)

# Plot
import matplotlib.pyplot as plt

plt.figure(figsize=(12, 3.5))
plt.subplot(141)
plt.imshow(a, cmap=plt.cm.gray)
plt.axis("off")
plt.title("a")

plt.subplot(142)
plt.imshow(mask, cmap=plt.cm.gray)
plt.axis("off")
plt.title("mask")

plt.subplot(143)
plt.imshow(opened_mask, cmap=plt.cm.gray)
plt.axis("off")
plt.title("opened_mask")

plt.subplot(144)
plt.imshow(closed_mask, cmap=plt.cm.gray)
plt.title("closed_mask")
plt.axis("off")

plt.subplots_adjust(wspace=0.05, left=0.01, bottom=0.01, right=0.99, top=0.99)

plt.show()

Total running time of the script: (0 minutes 0.104 seconds)

5.12.11 Plot geometrical transformations on images

Demo geometrical transformations of images.

Downloading file 'face.dat' from 'https://raw.githubusercontent.com/scipy/dataset-

˓→face/main/face.dat' to '/home/runner/.cache/scipy-data'.

5.12. Full code examples for the SciPy chapter 262

 Scientific Python Lectures, Edition 2024.2rc0.dev0

# Load some data

import scipy as sp

face = sp.datasets.face(gray=True)

# Apply a variety of transformations

import matplotlib.pyplot as plt

shifted_face = sp.ndimage.shift(face, (50, 50))

shifted_face2 = sp.ndimage.shift(face, (50, 50), mode="nearest")
rotated_face = sp.ndimage.rotate(face, 30)
cropped_face = face[50:-50, 50:-50]
zoomed_face = sp.ndimage.zoom(face, 2)
zoomed_face.shape

plt.figure(figsize=(15, 3))
plt.subplot(151)
plt.imshow(shifted_face, cmap=plt.cm.gray)
plt.axis("off")

plt.subplot(152)
plt.imshow(shifted_face2, cmap=plt.cm.gray)
plt.axis("off")

plt.subplot(153)
plt.imshow(rotated_face, cmap=plt.cm.gray)
plt.axis("off")

plt.subplot(154)
plt.imshow(cropped_face, cmap=plt.cm.gray)
plt.axis("off")

plt.subplot(155)
plt.imshow(zoomed_face, cmap=plt.cm.gray)
plt.axis("off")

plt.subplots_adjust(wspace=0.05, left=0.01, bottom=0.01, right=0.99, top=0.99)

plt.show()

Total running time of the script: (0 minutes 0.915 seconds)

5.12.12 Demo connected components

Extracting and labeling connected components in a 2D array

import numpy as np
import matplotlib.pyplot as plt

Generate some binary data

x, y = np.indices((100, 100))
sig = (
np.sin(2 * np.pi * x / 50.0)
* np.sin(2 * np.pi * y / 50.0)
(continues on next page)

5.12. Full code examples for the SciPy chapter 263

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

* (1 + x * y / 50.0**2) ** 2
)
mask = sig > 1

plt.figure(figsize=(7, 3.5))
plt.subplot(1, 2, 1)
plt.imshow(sig)
plt.axis("off")
plt.title("sig")

plt.subplot(1, 2, 2)
plt.imshow(mask, cmap=plt.cm.gray)
plt.axis("off")
plt.title("mask")
plt.subplots_adjust(wspace=0.05, left=0.01, bottom=0.01, right=0.99, top=0.9)

Label connected components

import scipy as sp

labels, nb = sp.ndimage.label(mask)

plt.figure(figsize=(3.5, 3.5))
plt.imshow(labels)
plt.title("label")
plt.axis("off")

plt.subplots_adjust(wspace=0.05, left=0.01, bottom=0.01, right=0.99, top=0.9)

5.12. Full code examples for the SciPy chapter 264

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Extract the 4th connected component, and crop the array around it

sl = sp.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.subplots_adjust(wspace=0.05, left=0.01, bottom=0.01, right=0.99, top=0.9)

plt.show()

Total running time of the script: (0 minutes 0.129 seconds)

5.12. Full code examples for the SciPy chapter 265

 Scientific Python Lectures, Edition 2024.2rc0.dev0

5.12.13 Minima and roots of a function

Demos finding minima and roots of a function.

Define the function

import numpy as np

x = np.arange(-10, 10, 0.1)

def f(x):
return x**2 + 10 * np.sin(x)

Find minima

import scipy as sp

# Global optimization
grid = (-10, 10, 0.1)
xmin_global = sp.optimize.brute(f, (grid,))
print(f"Global minima found { xmin_global} ")

# Constrain optimization
xmin_local = sp.optimize.fminbound(f, 0, 10)
print(f"Local minimum found { xmin_local} ")

Global minima found [-1.30641113]

Local minimum found 3.8374671194983834

Root finding

root = sp.optimize.root(f, 1) # our initial guess is 1

print(f"First root found { root.x} ")
root2 = sp.optimize.root(f, -2.5)
print(f"Second root found { root2.x} ")

First root found [0.]

Second root found [-2.47948183]

Plot function, minima, and roots

import matplotlib.pyplot as plt

fig = plt.figure(figsize=(6, 4))

ax = fig.add_subplot(111)

# Plot the function

ax.plot(x, f(x), "b-", label="f(x)")

# Plot the minima

(continues on next page)

5.12. Full code examples for the SciPy chapter 266

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

xmins = np.array([xmin_global[0], xmin_local])
ax.plot(xmins, f(xmins), "go", label="Minima")

# Plot the roots

roots = np.array([root.x, root2.x])
ax.plot(roots, f(roots), "kv", label="Roots")

# Decorate the figure

ax.legend(loc="best")
ax.set_xlabel("x")
ax.set_ylabel("f(x)")
ax.axhline(0, color="gray")
plt.show()

Total running time of the script: (0 minutes 0.079 seconds)

5.12. Full code examples for the SciPy chapter 267

 Scientific Python Lectures, Edition 2024.2rc0.dev0

5.12.14 Plot filtering on images

Demo filtering for denoising of images.

# Load some data

import scipy as sp

face = sp.datasets.face(gray=True)
face = face[:512, -512:] # crop out square on right

# Apply a variety of filters

import matplotlib.pyplot as plt

import numpy as np

noisy_face = np.copy(face).astype(float)
rng = np.random.default_rng()
noisy_face += face.std() * 0.5 * rng.standard_normal(face.shape)
blurred_face = sp.ndimage.gaussian_filter(noisy_face, sigma=3)
median_face = sp.ndimage.median_filter(noisy_face, size=5)
wiener_face = sp.signal.wiener(noisy_face, (5, 5))

plt.figure(figsize=(12, 3.5))
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.subplots_adjust(wspace=0.05, left=0.01, bottom=0.01, right=0.99, top=0.99)

plt.show()

5.12. Full code examples for the SciPy chapter 268

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Total running time of the script: (0 minutes 0.478 seconds)

5.12.15 Optimization of a two-parameter function

import numpy as np

# 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)

xlim = [-2, 2]
ylim = [-1, 1]
x = np.linspace(*xlim)
y = np.linspace(*ylim)
xg, yg = np.meshgrid(x, y)

A 2D image plot of the function

Simple visualization in 2D

import matplotlib.pyplot as plt

plt.figure()
plt.imshow(sixhump([xg, yg]), extent=xlim + ylim, origin="lower")
plt.colorbar()

5.12. Full code examples for the SciPy chapter 269

 Scientific Python Lectures, Edition 2024.2rc0.dev0

<matplotlib.colorbar.Colorbar object at 0x7f4955d7e150>

A 3D surface plot of the function

from mpl_toolkits.mplot3d import Axes3D

fig = plt.figure()
ax = fig.add_subplot(111, projection="3d")
surf = ax.plot_surface(
xg,
yg,
sixhump([xg, yg]),
rstride=1,
cstride=1,
cmap=plt.cm.viridis,
linewidth=0,
antialiased=False,
)

ax.set_xlabel("x")
ax.set_ylabel("y")
ax.set_zlabel("f(x, y)")
ax.set_title("Six-hump Camelback function")

5.12. Full code examples for the SciPy chapter 270

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Text(0.5, 1.0, 'Six-hump Camelback function')

Find minima

import scipy as sp

# local minimization
res_local = sp.optimize.minimize(sixhump, x0=[0, 0])

# global minimization
res_global = sp.optimize.differential_evolution(sixhump, bounds=[xlim, ylim])

plt.figure()
# Show the function in 2D
plt.imshow(sixhump([xg, yg]), extent=xlim + ylim, origin="lower")
plt.colorbar()
# Mark the minima
plt.scatter(res_local.x[0], res_local.x[1], label="local minimizer")
plt.scatter(res_global.x[0], res_global.x[1], label="global minimizer")
plt.legend()
plt.show()

5.12. Full code examples for the SciPy chapter 271

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Total running time of the script: (0 minutes 0.386 seconds)

5.12.16 Plotting and manipulating FFTs for filtering

Plot the power of the FFT of a signal and inverse FFT back to reconstruct a signal.
This example demonstrate scipy.fft.fft(), scipy.fft.fftfreq() and scipy.fft.ifft(). It imple-
ments a basic filter that is very suboptimal, and should not be used.

import numpy as np
import scipy as sp
import matplotlib.pyplot as plt

Generate the signal

# Seed the random number generator

rng = np.random.default_rng(27446968)

time_step = 0.02
period = 5.0

time_vec = np.arange(0, 20, time_step)

sig = np.sin(2 * np.pi / period * time_vec) + 0.5 * rng.normal(size=time_vec.size)

plt.figure(figsize=(6, 5))
plt.plot(time_vec, sig, label="Original signal")

5.12. Full code examples for the SciPy chapter 272

 Scientific Python Lectures, Edition 2024.2rc0.dev0

[<matplotlib.lines.Line2D object at 0x7f495545e850>]

Compute and plot the power

# The FFT of the signal

sig_fft = sp.fft.fft(sig)

# And the power (sig_fft is of complex dtype)

power = np.abs(sig_fft) ** 2

# The corresponding frequencies

sample_freq = sp.fft.fftfreq(sig.size, d=time_step)

# Plot the FFT power

plt.figure(figsize=(6, 5))
plt.plot(sample_freq, power)
plt.xlabel("Frequency [Hz]")
plt.ylabel("plower")

# Find the peak frequency: we can focus on only the positive frequencies
pos_mask = np.where(sample_freq > 0)
freqs = sample_freq[pos_mask]
peak_freq = freqs[power[pos_mask].argmax()]

# Check that it does indeed correspond to the frequency that we generate

(continues on next page)

5.12. Full code examples for the SciPy chapter 273

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

# the signal with
np.allclose(peak_freq, 1.0 / period)

# An inner plot to show the peak frequency

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

[]

Remove all the high frequencies

We now remove all the high frequencies and transform back from frequencies to signal.

high_freq_fft = sig_fft.copy()
high_freq_fft[np.abs(sample_freq) > peak_freq] = 0
filtered_sig = sp.fft.ifft(high_freq_fft)

plt.figure(figsize=(6, 5))
plt.plot(time_vec, sig, label="Original signal")
(continues on next page)

5.12. Full code examples for the SciPy chapter 274

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

plt.plot(time_vec, filtered_sig, linewidth=3, label="Filtered signal")
plt.xlabel("Time [s]")
plt.ylabel("Amplitude")

plt.legend(loc="best")

/opt/hostedtoolcache/Python/3.11.9/x64/lib/python3.11/site-packages/matplotlib/cbook.
˓→py:1762: ComplexWarning: Casting complex values to real discards the imaginary part

return math.isfinite(val)
/opt/hostedtoolcache/Python/3.11.9/x64/lib/python3.11/site-packages/matplotlib/cbook.
˓→py:1398: ComplexWarning: Casting complex values to real discards the imaginary part

return np.asarray(x, float)

<matplotlib.legend.Legend object at 0x7f4955411bd0>

Note This is actually a bad way of creating a filter: such brutal cut-off in frequency space does not
control distortion on the signal.
Filters should be created using the SciPy filter design code

plt.show()

Total running time of the script: (0 minutes 0.243 seconds)

5.12. Full code examples for the SciPy chapter 275

 Scientific Python Lectures, Edition 2024.2rc0.dev0

5.12.17 A demo of 1D interpolation

5.12. Full code examples for the SciPy chapter 276

 Scientific Python Lectures, Edition 2024.2rc0.dev0

•
# Generate data
import numpy as np

rng = np.random.default_rng(27446968)
measured_time = np.linspace(0, 2 * np.pi, 20)
function = np.sin(measured_time)
noise = rng.normal(loc=0, scale=0.1, size=20)
measurements = function + noise

# Smooth the curve and interpolate at new times

import scipy as sp

smoothing_spline = sp.interpolate.make_smoothing_spline(measured_time, measurements)

interpolation_time = np.linspace(0, 2 * np.pi, 200)
smooth_results = smoothing_spline(interpolation_time)

# Plot the data, the interpolant, and the original function

import matplotlib.pyplot as plt

plt.figure(figsize=(6, 4))
plt.plot(measured_time, measurements, ".", ms=6, label="measurements")
plt.plot(interpolation_time, smooth_results, label="smoothing spline")
plt.plot(interpolation_time, np.sin(interpolation_time), "--", label="underlying curve
˓→")

plt.legend()
plt.show()

# Fit the data exactly

interp_spline = sp.interpolate.make_interp_spline(measured_time, function)
interp_results = interp_spline(interpolation_time)

# Plot the data, the interpolant, and the original function

plt.figure(figsize=(6, 4))
(continues on next page)

5.12. Full code examples for the SciPy chapter 277

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

plt.plot(measured_time, function, ".", ms=6, label="measurements")
plt.plot(interpolation_time, interp_results, label="interpolating spline")
plt.plot(interpolation_time, np.sin(interpolation_time), "--", label="underlying curve
˓→")

plt.legend()
plt.show()

# Plot interpolant, its derivative, and its antiderivative

plt.figure(figsize=(6, 4))
t = interpolation_time
plt.plot(t, interp_spline(t), label="spline")
plt.plot(t, interp_spline.derivative()(t), label="derivative")
plt.plot(t, interp_spline.antiderivative()(t) - 1, label="antiderivative")

plt.legend()
plt.show()

Total running time of the script: (0 minutes 0.204 seconds)

5.12.18 Solutions of the exercises for SciPy

Solutions of the exercises for SciPy

Crude periodicity finding

Discover the periods in evolution of animal populations (../../../../data/populations.txt)

Load the data

import numpy as np

data = np.loadtxt("../../../../data/populations.txt")
years = data[:, 0]
populations = data[:, 1:]

Plot the data

import matplotlib.pyplot as plt

plt.figure()
plt.plot(years, populations * 1e-3)
plt.xlabel("Year")
plt.ylabel(r"Population number ($\cdot10^3$)")
plt.legend(["hare", "lynx", "carrot"], loc=1)

5.12. Full code examples for the SciPy chapter 278

 Scientific Python Lectures, Edition 2024.2rc0.dev0

<matplotlib.legend.Legend object at 0x7f495497bf10>

Plot its periods

import scipy as sp

ft_populations = sp.fft.fft(populations, axis=0)

frequencies = sp.fft.fftfreq(populations.shape[0], years[1] - years[0])
periods = 1 / frequencies

plt.figure()
plt.plot(periods, abs(ft_populations) * 1e-3, "o")
plt.xlim(0, 22)
plt.xlabel("Period")
plt.ylabel(r"Power ($\cdot10^3$)")

plt.show()

5.12. Full code examples for the SciPy chapter 279

 Scientific Python Lectures, Edition 2024.2rc0.dev0

/home/runner/work/scientific-python-lectures/scientific-python-lectures/intro/scipy/
˓→examples/solutions/plot_periodicity_finder.py:39: RuntimeWarning: divide by zero␣

˓→encountered in divide

periods = 1 / frequencies

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.
Total running time of the script: (0 minutes 0.121 seconds)

Curve fitting: temperature as a function of month of the year

We have the min and max temperatures in Alaska for each months of the year. We would like to find a
function to describe this yearly evolution.
For this, we will fit a periodic function.

5.12. Full code examples for the SciPy chapter 280

 Scientific Python Lectures, Edition 2024.2rc0.dev0

The data

import numpy as np

temp_max = np.array([17, 19, 21, 28, 33, 38, 37, 37, 31, 23, 19, 18])
temp_min = np.array([-62, -59, -56, -46, -32, -18, -9, -13, -25, -46, -52, -58])

import matplotlib.pyplot as plt

months = np.arange(12)
plt.plot(months, temp_max, "ro")
plt.plot(months, temp_min, "bo")
plt.xlabel("Month")
plt.ylabel("Min and max temperature")

Text(35.472222222222214, 0.5, 'Min and max temperature')

5.12. Full code examples for the SciPy chapter 281

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Fitting it to a periodic function

import scipy as sp

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 = sp.optimize.curve_fit(yearly_temps, months, temp_max, [20, 10, 0])

res_min, cov_min = sp.optimize.curve_fit(yearly_temps, months, temp_min, [-40, 20, 0])

Plotting the fit

days = np.linspace(0, 12, num=365)

plt.figure()
plt.plot(months, temp_max, "ro")
plt.plot(days, yearly_temps(days, *res_max), "r-")
plt.plot(months, temp_min, "bo")
plt.plot(days, yearly_temps(days, *res_min), "b-")
plt.xlabel("Month")
plt.ylabel(r"Temperature ($^\circ$C)")

plt.show()

Total running time of the script: (0 minutes 0.132 seconds)

5.12. Full code examples for the SciPy chapter 282

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Simple image blur by convolution with a Gaussian kernel

Blur an an image (../../../../data/elephant.png) using a Gaussian kernel.

Convolution is easy to perform with FFT: convolving two signals boils down to multiplying their FFTs
(and performing an inverse FFT).

import numpy as np
import scipy as sp
import matplotlib.pyplot as plt

The original image

# read image
img = plt.imread("../../../../data/elephant.png")
plt.figure()
plt.imshow(img)

<matplotlib.image.AxesImage object at 0x7f4954a3a0d0>

5.12. Full code examples for the SciPy chapter 283

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Prepare an Gaussian convolution kernel

# First a 1-D Gaussian

t = np.linspace(-10, 10, 30)
bump = np.exp(-0.1 * t**2)
bump /= np.trapz(bump) # normalize the integral to 1

# make a 2-D kernel out of it

kernel = bump[:, np.newaxis] * bump[np.newaxis, :]

Implement convolution via FFT

# Padded fourier transform, with the same shape as the image

# We use :func:`scipy.fft.fft2` to have a 2D FFT
kernel_ft = sp.fft.fft2(kernel, s=img.shape[:2], axes=(0, 1))

# convolve
img_ft = sp.fft.fft2(img, axes=(0, 1))
# the 'newaxis' is to match to color direction
img2_ft = kernel_ft[:, :, np.newaxis] * img_ft
img2 = sp.fft.ifft2(img2_ft, axes=(0, 1)).real

# clip values to range

img2 = np.clip(img2, 0, 1)

# plot output
plt.figure()
plt.imshow(img2)

5.12. Full code examples for the SciPy chapter 284

 Scientific Python Lectures, Edition 2024.2rc0.dev0

<matplotlib.image.AxesImage object at 0x7f4955a67810>

Further exercise (only if you are familiar with this stuff):

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
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
this for us, and probably do a better job: scipy.signal.fftconvolve()

# mode='same' is there to enforce the same output shape as input arrays

# (ie avoid border effects)
img3 = sp.signal.fftconvolve(img, kernel[:, :, np.newaxis], mode="same")
plt.figure()
plt.imshow(img3)

5.12. Full code examples for the SciPy chapter 285

 Scientific Python Lectures, Edition 2024.2rc0.dev0

<matplotlib.image.AxesImage object at 0x7f49601653d0>

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

plt.show()

Total running time of the script: (0 minutes 0.377 seconds)

Image denoising by FFT

Denoise an image (../../../../data/moonlanding.png) by implementing a blur with an FFT.

Implements, via FFT, the following convolution:
∫︁
𝑓1 (𝑡) = 𝑑𝑡′ 𝐾(𝑡 − 𝑡′ )𝑓0 (𝑡′ )

𝑓˜1 (𝜔) = 𝐾(𝜔)

˜ 𝑓˜0 (𝜔)

5.12. Full code examples for the SciPy chapter 286

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Read and plot the image

import numpy as np
import matplotlib.pyplot as plt

im = plt.imread("../../../../data/moonlanding.png").astype(float)

plt.figure()
plt.imshow(im, plt.cm.gray)
plt.title("Original image")

Text(0.5, 1.0, 'Original image')

Compute the 2d FFT of the input image

import scipy as sp

im_fft = sp.fft.fft2(im)

# Show the results

def plot_spectrum(im_fft):
from matplotlib.colors import LogNorm

# A logarithmic colormap
(continues on next page)

5.12. Full code examples for the SciPy chapter 287

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

plt.imshow(np.abs(im_fft), norm=LogNorm(vmin=5))
plt.colorbar()

plt.figure()
plot_spectrum(im_fft)
plt.title("Fourier transform")

Text(0.5, 1.0, 'Fourier transform')

Filter in FFT

# In the lines following, we'll make a copy of the original spectrum and
# truncate coefficients.

# Define the fraction of coefficients (in each direction) we keep

keep_fraction = 0.1

# Call ff a copy of the original transform. NumPy arrays have a copy

# method for this purpose.
im_fft2 = im_fft.copy()

# Set r and c to be the number of rows and columns of the array.

r, c = im_fft2.shape

(continues on next page)

5.12. Full code examples for the SciPy chapter 288

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

# 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")

Text(0.5, 1.0, 'Filtered Spectrum')

Reconstruct the final image

# Reconstruct the denoised image from the filtered spectrum, keep only the
# real part for display.
im_new = sp.fft.ifft2(im_fft2).real

plt.figure()
plt.imshow(im_new, plt.cm.gray)
plt.title("Reconstructed Image")

5.12. Full code examples for the SciPy chapter 289

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Text(0.5, 1.0, 'Reconstructed Image')

Easier and better: scipy.ndimage.gaussian_filter()

Implementing filtering directly with FFTs is tricky and time consuming. We can use the
Gaussian filter from scipy.ndimage

im_blur = sp.ndimage.gaussian_filter(im, 4)

plt.figure()
plt.imshow(im_blur, plt.cm.gray)
plt.title("Blurred image")

plt.show()

5.12. Full code examples for the SciPy chapter 290

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Total running time of the script: (0 minutes 0.865 seconds)

See also:
References to go further
• Some chapters of the advanced and the packages and applications parts of the SciPy lectures
• The SciPy cookbook

5.12. Full code examples for the SciPy chapter 291

 CHAPTER 6
Getting help and finding documentation

Author: Emmanuelle Gouillart

Rather than knowing all functions in NumPy and SciPy, it is important to find rapidly information
throughout the documentation and the available help. Here are some ways to get information:
• In Ipython, help function opens the docstring of the function. Only type the beginning of the
function’s name and use tab completion to display the matching functions.

In [1]: help(np.van<TAB>

In [2]: help(np.vander)
Help on _ArrayFunctionDispatcher in module numpy:

vander(x, N=None, increasing=False)

Generate a Vandermonde matrix.

The columns of the output matrix are powers of the input vector. The
order of the powers is determined by the `increasing` boolean argument.
Specifically, when `increasing` is False, the `i`-th output column is
the input vector raised element-wise to the power of ``N - i - 1``. Such
a matrix with a geometric progression in each row is named for Alexandre-
Theophile Vandermonde.

Parameters
----------
x : array_like
1-D input array.
N : int, optional
Number of columns in the output. If `N` is not specified, a square
array is returned (``N = len(x)``).
(continues on next page)

292
 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

increasing : bool, optional
Order of the powers of the columns. If True, the powers increase
from left to right, if False (the default) they are reversed.

.. versionadded:: 1.9.0

Returns
-------
out : ndarray
Vandermonde matrix. If `increasing` is False, the first column is
``x^(N-1)``, the second ``x^(N-2)`` and so forth. If `increasing` is
True, the columns are ``x^0, x^1, ..., x^(N-1)``.

See Also
--------
polynomial.polynomial.polyvander

Examples
--------
>>> x = np.array([1, 2, 3, 5])
>>> N = 3
>>> np.vander(x, N)
array([[ 1, 1, 1],
[ 4, 2, 1],
[ 9, 3, 1],
[25, 5, 1]])

>>> np.column_stack([x**(N-1-i) for i in range(N)])

array([[ 1, 1, 1],
[ 4, 2, 1],
[ 9, 3, 1],
[25, 5, 1]])

>>> x = np.array([1, 2, 3, 5])

>>> np.vander(x)
array([[ 1, 1, 1, 1],
[ 8, 4, 2, 1],
[ 27, 9, 3, 1],
[125, 25, 5, 1]])
>>> np.vander(x, increasing=True)
array([[ 1, 1, 1, 1],
[ 1, 2, 4, 8],
[ 1, 3, 9, 27],
[ 1, 5, 25, 125]])

The determinant of a square Vandermonde matrix is the product

of the differences between the values of the input vector:

>>> np.linalg.det(np.vander(x))
48.000000000000043 # may vary
>>> (5-3)*(5-2)*(5-1)*(3-2)*(3-1)*(2-1)
48

In Ipython it is not possible to open a separated window for help and documentation; however one can
always open a second Ipython shell just to display help and docstrings. . .
• Numpy’s and Scipy’s documentations can be browsed online on https://scipy.org and https://

293
 Scientific Python Lectures, Edition 2024.2rc0.dev0

numpy.org. The search button is quite useful inside the reference documentation of the two
packages.
Tutorials on various topics as well as the complete API with all docstrings are found on this website.
• Numpy’s and Scipy’s documentation is enriched and updated on a regular basis by users on a wiki
https://numpy.org/doc/stable/. As a result, some docstrings are clearer or more detailed on the
wiki, and you may want to read directly the documentation on the wiki instead of the official
documentation website. Note that anyone can create an account on the wiki and write better
documentation; this is an easy way to contribute to an open-source project and improve the tools
you are using!
• The SciPy Cookbook https://scipy-cookbook.readthedocs.io gives recipes on many common prob-
lems frequently encountered, such as fitting data points, solving ODE, etc.
• Matplotlib’s website https://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

• numpy.lookfor looks for keywords inside the docstrings of specified modules.

In [4]: np.lookfor('convolution')
Search results for 'convolution'
--------------------------------
numpy.convolve
Returns the discrete, linear convolution of two one-dimensional sequences.
numpy.ma.convolve
Returns the discrete, linear convolution of two one-dimensional sequences.
numpy.polymul
Find the product of two polynomials.
numpy.bartlett
Return the Bartlett window.
numpy.correlate
Cross-correlation of two 1-dimensional sequences.
numpy.vectorize
vectorize(pyfunc=np._NoValue, otypes=None, doc=None, excluded=None,

• If everything listed above fails (and Google doesn’t have the answer). . . don’t despair! There is
a vibrant Scientific Python community. Scientific Python is present on various platform. https:
//scientific-python.org/community/
Packages like SciPy and NumPy also have their own channels. Have a look at their respective
websites to find out how to engage with users and maintainers.

294
 Part II

Advanced topics

295
 Scientific Python Lectures, Edition 2024.2rc0.dev0

This part of the Scientific Python Lectures is dedicated to advanced usage. It strives to educate the
proficient Python coder to be an expert and tackles various specific topics.

296
 CHAPTER 7
Advanced Python Constructs

Author Zbigniew Jędrzejewski-Szmek

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
complicated programs or libraries, but not in the sense of being particularly specialized, or particularly
complicated.
It is important to underline that this chapter is purely about the language itself — about features
supported through special syntax complemented by functionality of the Python stdlib, which could not
be implemented through clever external modules.
The process of developing the Python programming language, its syntax, is very transparent; proposed
changes are evaluated from various angles and discussed via Python Enhancement Proposals — PEPs.
As a result, features described in this chapter were added after it was shown that they indeed solve real
problems and that their use is as simple as possible.

Chapter contents

• Iterators, generator expressions and generators

– Iterators
– Generator expressions
– Generators
– Bidirectional communication
– Chaining generators
• Decorators
– Replacing or tweaking the original object

297
 Scientific Python Lectures, Edition 2024.2rc0.dev0

– 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
– A while-loop removing decorator
– A plugin registration system
• Context managers
– Catching exceptions
– Using generators to define context managers

7.1 Iterators, generator expressions and generators

7.1.1 Iterators

Simplicity

Duplication of effort is wasteful, and replacing the various home-grown approaches with a standard
feature usually ends up making things more readable, and interoperable as well.
Guido van Rossum — Adding Optional Static Typing to Python

An iterator is an object adhering to the iterator protocol — basically this means that it has a next
method, which, when called, returns the next item in the sequence, and when there’s nothing to return,
raises the StopIteration exception.
An iterator object allows to loop just once. It holds the state (position) of a single iteration, or from the
other side, each loop over a sequence requires a single iterator object. This means that we can iterate
over the same sequence more than once concurrently. Separating the iteration logic from the sequence
allows us to have more than one way of iteration.
Calling the __iter__ method on a container to create an iterator object is the most straightforward way
to get hold of an iterator. The iter function does that for us, saving a few keystrokes.

>>> nums = [1, 2, 3] # note that ... varies: these are different objects
>>> iter(nums)
<...iterator object at ...>
>>> nums.__iter__()
<...iterator object at ...>
>>> nums.__reversed__()
<...reverseiterator object at ...>

>>> it = iter(nums)
>>> next(it)
1
>>> next(it)
2
>>> next(it)
3
>>> next(it)
Traceback (most recent call last):
(continues on next page)

7.1. Iterators, generator expressions and generators 298

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

File "<stdin>", line 1, in <module>
StopIteration

When used in a loop, StopIteration is swallowed and causes the loop to finish. But with explicit
invocation, we can see that once the iterator is exhausted, accessing it raises an exception.
Using the for..in loop also uses the __iter__ method. This allows us to transparently start the iteration
over a sequence. But if we already have the iterator, we want to be able to use it in an for loop in
the same way. In order to achieve this, iterators in addition to next are also required to have a method
called __iter__ which returns the iterator (self).
Support for iteration is pervasive in Python: all sequences and unordered containers in the standard
library allow this. The concept is also stretched to other things: e.g. file objects support iteration over
lines.

>>> with open("/etc/fstab") as f:

... f is f.__iter__()
...
True

The file is an iterator itself and it’s __iter__ method doesn’t create a separate object: only a single
thread of sequential access is allowed.

7.1.2 Generator expressions

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
or an expression. If round parentheses are used, then a generator iterator is created. If rectangular
parentheses are used, the process is short-circuited and we get a list.

>>> (i for i in nums)

<generator object <genexpr> at 0x...>
>>> [i for i in nums]
[1, 2, 3]
>>> list(i for i in nums)
[1, 2, 3]

The list comprehension syntax also extends to dictionary and set comprehensions. A set is cre-
ated when the generator expression is enclosed in curly braces. A dict is created when the generator
expression contains “pairs” of the form key:value:

>>> {i for i in range(3)}

{0, 1, 2}
>>> {i:i**2 for i in range(3)}
{0: 0, 1: 1, 2: 4}

One gotcha should be mentioned: in old Pythons the index variable (i) would leak, and in versions >=
3 this is fixed.

7.1. Iterators, generator expressions and generators 299

 Scientific Python Lectures, Edition 2024.2rc0.dev0

7.1.3 Generators

Generators

A generator is a function that produces a sequence of results instead of a single value.

David Beazley — A Curious Course on Coroutines and Concurrency

A third way to create iterator objects is to call a generator function. A generator is a function containing
the keyword yield. It must be noted that the mere presence of this keyword completely changes the
nature of the function: this yield statement doesn’t have to be invoked, or even reachable, but causes
the function to be marked as a generator. When a normal function is called, the instructions contained in
the body start to be executed. When a generator is called, the execution stops before the first instruction
in the body. An invocation of a generator function creates a generator object, adhering to the iterator
protocol. As with normal function invocations, concurrent and recursive invocations are allowed.
When next is called, the function is executed until the first yield. Each encountered yield statement
gives a value becomes the return value of next. After executing the yield statement, the execution of
this function is suspended.

>>> def f():

... yield 1
... yield 2
>>> f()
<generator object f at 0x...>
>>> gen = f()
>>> next(gen)
1
>>> next(gen)
2
>>> next(gen)
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
StopIteration

Let’s go over the life of the single invocation of the generator function.

>>> def f():

... print("-- start --")
... yield 3
... print("-- finish --")
... yield 4
>>> gen = f()
>>> next(gen)
-- start --
3
>>> next(gen)
-- finish --
4
>>> next(gen)
Traceback (most recent call last):
...
StopIteration

Contrary to a normal function, where executing f() would immediately cause the first print to be
executed, gen is assigned without executing any statements in the function body. Only when gen.
__next__() is invoked by next, the statements up to the first yield are executed. The second next
prints -- finish -- and execution halts on the second yield. The third next falls of the end of the

7.1. Iterators, generator expressions and generators 300

 Scientific Python Lectures, Edition 2024.2rc0.dev0

function. Since no yield was reached, an exception is raised.

What happens with the function after a yield, when the control passes to the caller? The state of
each generator is stored in the generator object. From the point of view of the generator function, is
looks almost as if it was running in a separate thread, but this is just an illusion: execution is strictly
single-threaded, but the interpreter keeps and restores the state in between the requests for the next
value.
Why are generators useful? As noted in the parts about iterators, a generator function is just a different
way to create an iterator object. Everything that can be done with yield statements, could also be
done with next methods. Nevertheless, using a function and having the interpreter perform its magic
to create an iterator has advantages. A function can be much shorter than the definition of a class with
the required next and __iter__ methods. What is more important, it is easier for the author of the
generator to understand the state which is kept in local variables, as opposed to instance attributes,
which have to be used to pass data between consecutive invocations of next on an iterator object.
A broader question is why are iterators useful? When an iterator is used to power a loop, the loop
becomes very simple. The code to initialise the state, to decide if the loop is finished, and to find the
next value is extracted into a separate place. This highlights the body of the loop — the interesting
part. In addition, it is possible to reuse the iterator code in other places.

7.1.4 Bidirectional communication

Each yield statement causes a value to be passed to the caller. This is the reason for the introduction
of generators by PEP 255. But communication in the reverse direction is also useful. One obvious way
would be some external state, either a global variable or a shared mutable object. Direct communication
is possible thanks to PEP 342. It is achieved by turning the previously boring yield statement into an
expression. When the generator resumes execution after a yield statement, the caller can call a method
on the generator object to either pass a value into the generator, which then is returned by the yield
statement, or a different method to inject an exception into the generator.
The first of the new methods is send(value), which is similar to next(), but passes value into the
generator to be used for the value of the yield expression. In fact, g.next() and g.send(None) are
equivalent.
The second of the new methods is throw(type, value=None, traceback=None) which is equivalent to:

raise type, value, traceback

at the point of the yield statement.

Unlike raise (which immediately raises an exception from the current execution point), throw() first
resumes the generator, and only then raises the exception. The word throw was picked because it
is suggestive of putting the exception in another location, and is associated with exceptions in other
languages.
What happens when an exception is raised inside the generator? It can be either raised explicitly or
when executing some statements or it can be injected at the point of a yield statement by means of
the throw() method. In either case, such an exception propagates in the standard manner: it can
be intercepted by an except or finally clause, or otherwise it causes the execution of the generator
function to be aborted and propagates in the caller.
For completeness’ sake, it’s worth mentioning that generator iterators also have a close() method,
which can be used to force a generator that would otherwise be able to provide more values to finish
immediately. It allows the generator __del__ method to destroy objects holding the state of generator.
Let’s define a generator which just prints what is passed in through send and throw.

>>> import itertools

>>> def g():
... print('--start--')
... for i in itertools.count():
(continues on next page)

7.1. Iterators, generator expressions and generators 301

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

... print('--yielding %i --' % i)
... try:
... ans = yield i
... except GeneratorExit:
... print('--closing--')
... raise
... except Exception as e:
... print('--yield raised %r --' % e)
... else:
... print('--yield returned %s --' % ans)

>>> it = g()
>>> next(it)
--start--
--yielding 0--
0
>>> it.send(11)
--yield returned 11--
--yielding 1--
1
>>> it.throw(IndexError)
--yield raised IndexError()--
--yielding 2--
2
>>> it.close()
--closing--

7.1.5 Chaining generators

Note: This is a preview of PEP 380 (not yet implemented, but accepted for Python 3.3).

Let’s say we are writing a generator and we want to yield a number of values generated by a second
generator, a subgenerator. If yielding of values is the only concern, this can be performed without
much difficulty using a loop such as

subgen = some_other_generator()
for v in subgen:
yield v

However, if the subgenerator is to interact properly with the caller in the case of calls to send(), throw()
and close(), things become considerably more difficult. The yield statement has to be guarded by a
try..except..finally structure similar to the one defined in the previous section to “debug” the generator
function. Such code is provided in PEP 380#id13, here it suffices to say that new syntax to properly
yield from a subgenerator is being introduced in Python 3.3:

yield from some_other_generator()

This behaves like the explicit loop above, repeatedly yielding values from some_other_generator until
it is exhausted, but also forwards send, throw and close to the subgenerator.

7.1. Iterators, generator expressions and generators 302

 Scientific Python Lectures, Edition 2024.2rc0.dev0

7.2 Decorators

Summary

This amazing feature appeared in the language almost apologetically and with concern that it might
not be that useful.
Bruce Eckel — An Introduction to Python Decorators

Since functions and classes are objects, they can be passed around. Since they are mutable objects, they
can be modified. The act of altering a function or class object after it has been constructed but before
is is bound to its name is called decorating.
There are two things hiding behind the name “decorator” — one is the function which does the work of
decorating, i.e. performs the real work, and the other one is the expression adhering to the decorator
syntax, i.e. an at-symbol and the name of the decorating function.
Function can be decorated by using the decorator syntax for functions:

@decorator # ❷
def function(): # ❶
pass

• A function is defined in the standard way. ❶

• An expression starting with @ placed before the function definition is the decorator ❷. The part
after @ must be a simple expression, usually this is just the name of a function or class. This part is
evaluated first, and after the function defined below is ready, the decorator is called with the newly
defined function object as the single argument. The value returned by the decorator is attached to
the original name of the function.
Decorators can be applied to functions and to classes. For classes the semantics are identical — the
original class definition is used as an argument to call the decorator and whatever is returned is assigned
under the original name.
Before the decorator syntax was implemented (PEP 318), it was possible to achieve the same effect by
assigning the function or class object to a temporary variable and then invoking the decorator explicitly
and then assigning the return value to the name of the function. This sounds like more typing, and it
is, and also the name of the decorated function doubling as a temporary variable must be used at least
three times, which is prone to errors. Nevertheless, the example above is equivalent to:

def function(): # ❶
pass
function = decorator(function) # ❷

Decorators can be stacked — the order of application is bottom-to-top, or inside-out. The semantics
are such that the originally defined function is used as an argument for the first decorator, whatever is
returned by the first decorator is used as an argument for the second decorator, . . . , and whatever is
returned by the last decorator is attached under the name of the original function.
The decorator syntax was chosen for its readability. Since the decorator is specified before the header
of the function, it is obvious that its is not a part of the function body and its clear that it can only
operate on the whole function. Because the expression is prefixed with @ is stands out and is hard to
miss (“in your face”, according to the PEP :) ). When more than one decorator is applied, each one is
placed on a separate line in an easy to read way.

7.2. Decorators 303

 Scientific Python Lectures, Edition 2024.2rc0.dev0

7.2.1 Replacing or tweaking the original object

Decorators can either return the same function or class object or they can return a completely different
object. In the first case, the decorator can exploit the fact that function and class objects are mutable
and add attributes, e.g. add a docstring to a class. A decorator might do something useful even without
modifying the object, for example register the decorated class in a global registry. In the second case,
virtually anything is possible: when something different is substituted for the original function or class,
the new object can be completely different. Nevertheless, such behaviour is not the purpose of decorators:
they are intended to tweak the decorated object, not do something unpredictable. Therefore, when a
function is “decorated” by replacing it with a different function, the new function usually calls the original
function, after doing some preparatory work. Likewise, when a class is “decorated” by replacing if with
a new class, the new class is usually derived from the original class. When the purpose of the decorator
is to do something “every time”, like to log every call to a decorated function, only the second type of
decorators can be used. On the other hand, if the first type is sufficient, it is better to use it, because it
is simpler.

7.2.2 Decorators implemented as classes and as functions

The only requirement on decorators is that they can be called with a single argument. This means that
decorators can be implemented as normal functions, or as classes with a __call__ method, or in theory,
even as lambda functions.
Let’s compare the function and class approaches. The decorator expression (the part after @) can be
either just a name, or a call. The bare-name approach is nice (less to type, looks cleaner, etc.), but is
only possible when no arguments are needed to customise the decorator. Decorators written as functions
can be used in those two cases:

>>> def simple_decorator(function):

... print("doing decoration")
... return function
>>> @simple_decorator
... def function():
... print("inside function")
doing decoration
>>> function()
inside function

>>> def decorator_with_arguments(arg):

... print("defining the decorator")
... def _decorator(function):
... # in this inner function, arg is available too
... print("doing decoration, %r " % arg)
... return function
... return _decorator
>>> @decorator_with_arguments("abc")
... def function():
... print("inside function")
defining the decorator
doing decoration, 'abc'
>>> function()
inside function

The two trivial decorators above fall into the category of decorators which return the original function.
If they were to return a new function, an extra level of nestedness would be required. In the worst case,
three levels of nested functions.

7.2. Decorators 304

 Scientific Python Lectures, Edition 2024.2rc0.dev0

>>> def replacing_decorator_with_args(arg):

... print("defining the decorator")
... def _decorator(function):
... # in this inner function, arg is available too
... print("doing decoration, %r " % arg)
... def _wrapper(*args, **kwargs):
... print("inside wrapper, %r %r " % (args, kwargs))
... return function(*args, **kwargs)
... return _wrapper
... return _decorator
>>> @replacing_decorator_with_args("abc")
... def function(*args, **kwargs):
... print("inside function, %r %r " % (args, kwargs))
... return 14
defining the decorator
doing decoration, 'abc'
>>> function(11, 12)
inside wrapper, (11, 12) {}
inside function, (11, 12) {}
14

The _wrapper function is defined to accept all positional and keyword arguments. In general we cannot
know what arguments the decorated function is supposed to accept, so the wrapper function just passes
everything to the wrapped function. One unfortunate consequence is that the apparent argument list is
misleading.
Compared to decorators defined as functions, complex decorators defined as classes are simpler. When
an object is created, the __init__ method is only allowed to return None, and the type of the created
object cannot be changed. This means that when a decorator is defined as a class, it doesn’t make
much sense to use the argument-less form: the final decorated object would just be an instance of the
decorating class, returned by the constructor call, which is not very useful. Therefore it’s enough to
discuss class-based decorators where arguments are given in the decorator expression and the decorator
__init__ method is used for decorator construction.

>>> class decorator_class(object):

... def __init__(self, arg):
... # this method is called in the decorator expression
... print("in decorator init, %s " % arg)
... self.arg = arg
... def __call__(self, function):
... # this method is called to do the job
... print("in decorator call, %s " % self.arg)
... return function
>>> deco_instance = decorator_class('foo')
in decorator init, foo
>>> @deco_instance
... def function(*args, **kwargs):
... print("in function, %s %s " % (args, kwargs))
in decorator call, foo
>>> function()
in function, () {}

Contrary to normal rules (PEP 8) decorators written as classes behave more like functions and therefore
their name often starts with a lowercase letter.
In reality, it doesn’t make much sense to create a new class just to have a decorator which returns the
original function. Objects are supposed to hold state, and such decorators are more useful when the
decorator returns a new object.

7.2. Decorators 305

 Scientific Python Lectures, Edition 2024.2rc0.dev0

>>> class replacing_decorator_class(object):

... def __init__(self, arg):
... # this method is called in the decorator expression
... print("in decorator init, %s " % arg)
... self.arg = arg
... def __call__(self, function):
... # this method is called to do the job
... print("in decorator call, %s " % self.arg)
... self.function = function
... return self._wrapper
... def _wrapper(self, *args, **kwargs):
... print("in the wrapper, %s %s " % (args, kwargs))
... return self.function(*args, **kwargs)
>>> deco_instance = replacing_decorator_class('foo')
in decorator init, foo
>>> @deco_instance
... def function(*args, **kwargs):
... print("in function, %s %s " % (args, kwargs))
in decorator call, foo
>>> function(11, 12)
in the wrapper, (11, 12) {}
in function, (11, 12) {}

A decorator like this can do pretty much anything, since it can modify the original function object and
mangle the arguments, call the original function or not, and afterwards mangle the return value.

7.2.3 Copying the docstring and other attributes of the original function

When a new function is returned by the decorator to replace the original function, an unfortunate
consequence is that the original function name, the original docstring, the original argument list are
lost. Those attributes of the original function can partially be “transplanted” to the new function
by setting __doc__ (the docstring), __module__ and __name__ (the full name of the function), and
__annotations__ (extra information about arguments and the return value of the function available in
Python 3). This can be done automatically by using functools.update_wrapper.

functools.update_wrapper(wrapper, wrapped)

“Update a wrapper function to look like the wrapped function.”

>>> import functools
>>> def replacing_decorator_with_args(arg):
... print("defining the decorator")
... def _decorator(function):
... print("doing decoration, %r " % arg)
... def _wrapper(*args, **kwargs):
... print("inside wrapper, %r %r " % (args, kwargs))
... return function(*args, **kwargs)
... return functools.update_wrapper(_wrapper, function)
... return _decorator
>>> @replacing_decorator_with_args("abc")
... def function():
... "extensive documentation"
... print("inside function")
... return 14
defining the decorator
doing decoration, 'abc'

7.2. Decorators 306

 Scientific Python Lectures, Edition 2024.2rc0.dev0

>>> function
<function function at 0x...>
>>> print(function.__doc__)
extensive documentation

One important thing is missing from the list of attributes which can be copied to the replacement
function: the argument list. The default values for arguments can be modified through the __defaults__,
__kwdefaults__ attributes, but unfortunately the argument list itself cannot be set as an attribute. This
means that help(function) will display a useless argument list which will be confusing for the user of
the function. An effective but ugly way around this problem is to create the wrapper dynamically, using
eval. This can be automated by using the external decorator module. It provides support for the
decorator decorator, which takes a wrapper and turns it into a decorator which preserves the function
signature.
To sum things up, decorators should always use functools.update_wrapper or some other means of
copying function attributes.

7.2.4 Examples in the standard library

First, it should be mentioned that there’s a number of useful decorators available in the standard library.
There are three decorators which really form a part of the language:
• classmethod causes a method to become a “class method”, which means that it can be invoked
without creating an instance of the class. When a normal method is invoked, the interpreter inserts
the instance object as the first positional parameter, self. When a class method is invoked, the
class itself is given as the first parameter, often called cls.
Class methods are still accessible through the class’ namespace, so they don’t pollute the module’s
namespace. Class methods can be used to provide alternative constructors:

class Array(object):
def __init__(self, data):
self.data = data

@classmethod
def fromfile(cls, file):
data = numpy.load(file)
return cls(data)

This is cleaner than using a multitude of flags to __init__.

• staticmethod is applied to methods to make them “static”, i.e. basically a normal function, but
accessible through the class namespace. This can be useful when the function is only needed inside
this class (its name would then be prefixed with _), or when we want the user to think of the
method as connected to the class, despite an implementation which doesn’t require this.
• property is the pythonic answer to the problem of getters and setters. A method decorated with
property becomes a getter which is automatically called on attribute access.

>>> class A(object):

... @property
... def a(self):
... "an important attribute"
... return "a value"
>>> A.a
<property object at 0x...>
>>> A().a
'a value'

7.2. Decorators 307

 Scientific Python Lectures, Edition 2024.2rc0.dev0

In this example, A.a is an read-only attribute. It is also documented: help(A) includes the
docstring for attribute a taken from the getter method. Defining a as a property allows it to be a
calculated on the fly, and has the side effect of making it read-only, because no setter is defined.
To have a setter and a getter, two methods are required, obviously:

class Rectangle(object):
def __init__(self, edge):
self.edge = edge

@property
def area(self):
"""Computed area.

Setting this updates the edge length to the proper value.

"""
return self.edge**2

@area.setter
def area(self, area):
self.edge = area ** 0.5

The way that this works, is that the property decorator replaces the getter method with a property
object. This object in turn has three methods, getter, setter, and deleter, which can be used
as decorators. Their job is to set the getter, setter and deleter of the property object (stored as
attributes fget, fset, and fdel). The getter can be set like in the example above, when creating
the object. When defining the setter, we already have the property object under area, and we add
the setter to it by using the setter method. All this happens when we are creating the class.
Afterwards, when an instance of the class has been created, the property object is special. When the
interpreter executes attribute access, assignment, or deletion, the job is delegated to the methods
of the property object.
To make everything crystal clear, let’s define a “debug” example:

>>> class D(object):

... @property
... def a(self):
... print("getting 1")
... return 1
... @a.setter
... def a(self, value):
... print("setting %r " % value)
... @a.deleter
... def a(self):
... print("deleting")
>>> D.a
<property object at 0x...>
>>> D.a.fget
<function ...>
>>> D.a.fset
<function ...>
>>> D.a.fdel
<function ...>
>>> d = D() # ... varies, this is not the same `a` function
>>> d.a
getting 1
1
>>> d.a = 2
(continues on next page)

7.2. Decorators 308

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

setting 2
>>> del d.a
deleting
>>> d.a
getting 1
1

Properties are a bit of a stretch for the decorator syntax. One of the premises of the decorator
syntax — that the name is not duplicated — is violated, but nothing better has been invented so
far. It is just good style to use the same name for the getter, setter, and deleter methods.
Some newer examples include:
• functools.lru_cache memoizes an arbitrary function maintaining a limited cache of argu-
ments:answer pairs (Python 3.2)
• functools.total_ordering is a class decorator which fills in missing ordering methods (__lt__,
__gt__, __le__, . . . ) based on a single available one.

7.2.5 Deprecation of functions

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:

class deprecated(object):
"""Print a deprecation warning once on first use of the function.

>>> @deprecated() # doctest: +SKIP

... def f():
... pass
>>> f() # doctest: +SKIP
f is deprecated
"""
def __call__(self, func):
self.func = func
self.count = 0
return self._wrapper
def _wrapper(self, *args, **kwargs):
self.count += 1
if self.count == 1:
print(self.func.__name__, 'is deprecated')
return self.func(*args, **kwargs)

It can also be implemented as a function:

def deprecated(func):
"""Print a deprecation warning once on first use of the function.

>>> @deprecated # doctest: +SKIP

... def f():
... pass
>>> f() # doctest: +SKIP
f is deprecated
"""
count = [0]
def wrapper(*args, **kwargs):
count[0] += 1
(continues on next page)

7.2. Decorators 309

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

if count[0] == 1:
print(func.__name__, 'is deprecated')
return func(*args, **kwargs)
return wrapper

7.2.6 A while-loop removing decorator

Let’s say we have function which returns a lists of things, and this list created by running a loop. If we
don’t know how many objects will be needed, the standard way to do this is something like:
def find_answers():
answers = []
while True:
ans = look_for_next_answer()
if ans is None:
break
answers.append(ans)
return answers

This is fine, as long as the body of the loop is fairly compact. Once it becomes more complicated, as
often happens in real code, this becomes pretty unreadable. We could simplify this by using yield
statements, but then the user would have to explicitly call list(find_answers()).
We can define a decorator which constructs the list for us:
def vectorized(generator_func):
def wrapper(*args, **kwargs):
return list(generator_func(*args, **kwargs))
return functools.update_wrapper(wrapper, generator_func)

Our function then becomes:

@vectorized
def find_answers():
while True:
ans = look_for_next_answer()
if ans is None:
break
yield ans

7.2.7 A plugin registration system

This is a class decorator which doesn’t modify the class, but just puts it in a global registry. It falls into
the category of decorators returning the original object:
class WordProcessor(object):
PLUGINS = []
def process(self, text):
for plugin in self.PLUGINS:
text = plugin().cleanup(text)
return text

@classmethod
def plugin(cls, plugin):
cls.PLUGINS.append(plugin)
(continues on next page)

7.2. Decorators 310

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

@WordProcessor.plugin
class CleanMdashesExtension(object):
def cleanup(self, text):
return text.replace('&mdash;', u'\N{em dash}')

Here we use a decorator to decentralise the registration of plugins. We call our decorator with a noun,
instead of a verb, because we use it to declare that our class is a plugin for WordProcessor. Method
plugin simply appends the class to the list of plugins.
A word about the plugin itself: it replaces HTML entity for em-dash with a real Unicode em-dash
character. It exploits the unicode literal notation to insert a character by using its name in the unicode
database (“EM DASH”). If the Unicode character was inserted directly, it would be impossible to
distinguish it from an en-dash in the source of a program.
See also:
More examples and reading
• PEP 318 (function and method decorator syntax)
• PEP 3129 (class decorator syntax)
• https://wiki.python.org/moin/PythonDecoratorLibrary
• https://docs.python.org/dev/library/functools.html
• https://pypi.org/project/decorator
• Bruce Eckel
– Decorators I: Introduction to Python Decorators
– Python Decorators II: Decorator Arguments
– Python Decorators III: A Decorator-Based Build System

7.3 Context managers

A context manager is an object with __enter__ and __exit__ methods which can be used in the with
statement:

with manager as var:

do_something(var)

is in the simplest case equivalent to

var = manager.__enter__()
try:
do_something(var)
finally:
manager.__exit__()

In other words, the context manager protocol defined in PEP 343 permits the extraction of the boring
part of a try..except..finally structure into a separate class leaving only the interesting do_something
block.
1. The __enter__ method is called first. It can return a value which will be assigned to var. The
as-part is optional: if it 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
successfully to the end, or it can break, continue or return, or it can throw an exception. Either

7.3. Context managers 311

 Scientific Python Lectures, Edition 2024.2rc0.dev0

way, after the block is finished, the __exit__ method is called. If an exception was thrown,
the information about the exception is passed to __exit__, which is described below in the next
subsection. In the normal case, exceptions can be ignored, just like in a finally clause, and will
be rethrown after __exit__ is finished.
Let’s say we want to make sure that a file is closed immediately after we are done writing to it:

>>> class closing(object):

... def __init__(self, obj):
... self.obj = obj
... def __enter__(self):
... return self.obj
... def __exit__(self, *args):
... self.obj.close()
>>> with closing(open('/tmp/file', 'w')) as f:
... f.write('the contents\n')

Here we have made sure that the f.close() is called when the with block is exited. Since closing files is
such a common operation, the support for this is already present in the file class. It has an __exit__
method which calls close and can be used as a context manager itself:

>>> with open('/tmp/file', 'a') as f:

... f.write('more contents\n')

The common use for try..finally is releasing resources. Various different cases are implemented
similarly: in the __enter__ phase the resource is acquired, in the __exit__ phase it is released, and the
exception, if thrown, is propagated. As with files, there’s often a natural operation to perform after the
object has been used and it is most convenient to have the support built in. With each release, Python
provides support in more places:
• all file-like objects:
– file ➥ automatically closed
– fileinput, tempfile
– bz2.BZ2File, gzip.GzipFile, tarfile.TarFile, zipfile.ZipFile
– ftplib, nntplib ➥ close connection
• locks
– multiprocessing.RLock ➥ lock and unlock
– multiprocessing.Semaphore
– memoryview ➥ automatically release
• decimal.localcontext ➥ modify precision of computations temporarily
• _winreg.PyHKEY ➥ open and close hive key
• warnings.catch_warnings ➥ kill warnings temporarily
• contextlib.closing ➥ the same as the example above, call close
• parallel programming
– concurrent.futures.ThreadPoolExecutor ➥ invoke in parallel then kill thread pool
– concurrent.futures.ProcessPoolExecutor ➥ invoke in parallel then kill process pool
– nogil ➥ solve the GIL problem temporarily (cython only :( )

7.3. Context managers 312

 Scientific Python Lectures, Edition 2024.2rc0.dev0

7.3.1 Catching exceptions

When an exception is thrown in the with-block, it is passed as arguments to __exit__. Three arguments
are used, the same as returned by sys.exc_info(): type, value, traceback. When no exception is thrown,
None is used for all three arguments. The context manager can “swallow” the exception by returning a
true value from __exit__. Exceptions can be easily ignored, because if __exit__ doesn’t use return
and just falls of the end, None is returned, a false value, and therefore the exception is rethrown after
__exit__ is finished.
The ability to catch exceptions opens interesting possibilities. A classic example comes from unit-tests
— we want to make sure that some code throws the right kind of exception:

class assert_raises(object):
# based on pytest and unittest.TestCase
def __init__(self, type):
self.type = type
def __enter__(self):
pass
def __exit__(self, type, value, traceback):
if type is None:
raise AssertionError('exception expected')
if issubclass(type, self.type):
return True # swallow the expected exception
raise AssertionError('wrong exception type')

with assert_raises(KeyError):
{}['foo']

7.3.2 Using generators to define context managers

When discussing generators, it was said that we prefer generators to iterators implemented as classes
because they are shorter, sweeter, and the state is stored as local, not instance, variables. On the other
hand, as described in Bidirectional communication, the flow of data between the generator and its caller
can be bidirectional. This includes exceptions, which can be thrown into the generator. We would like to
implement context managers as special generator functions. In fact, the generator protocol was designed
to support this use case.

@contextlib.contextmanager
def some_generator(<arguments>):
<setup>
try:
yield <value>
finally:
<cleanup>

The contextlib.contextmanager helper takes a generator and turns it into a context manager. The
generator has to obey some rules which are enforced by the wrapper function — most importantly it
must yield exactly once. The part before the yield is executed from __enter__, the block of code
protected by the context manager is executed when the generator is suspended in yield, and the rest
is executed in __exit__. If an exception is thrown, the interpreter hands it to the wrapper through
__exit__ arguments, and the wrapper function then throws it at the point of the yield statement.
Through the use of generators, the context manager is shorter and simpler.
Let’s rewrite the closing example as a generator:

@contextlib.contextmanager
def closing(obj):
(continues on next page)

7.3. Context managers 313

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

try:
yield obj
finally:
obj.close()

Let’s rewrite the assert_raises example as a generator:

@contextlib.contextmanager
def assert_raises(type):
try:
yield
except type:
return
except Exception as value:
raise AssertionError('wrong exception type')
else:
raise AssertionError('exception expected')

Here we use a decorator to turn generator functions into context managers!

7.3. Context managers 314

 CHAPTER 8
Advanced NumPy

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. . .

315
 Scientific Python Lectures, Edition 2024.2rc0.dev0

– 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
– Sharing multidimensional, typed data
– The old buffer protocol
– The old buffer protocol
– Array interface protocol
• Array siblings: chararray, maskedarray
– chararray: vectorized string operations
– masked_array missing data
– recarray: purely convenience
• Summary
• Contributing to NumPy/SciPy
– Why
– Reporting bugs
– Contributing to documentation
– Contributing features
– How to help, in general

Tip: In this section, NumPy will be imported as follows:

>>> import numpy as np

8.1 Life of ndarray

8.1.1 It’s. . .

ndarray =
block of memory + indexing scheme + data type descriptor
• raw data
• how to locate an element
• how to interpret an element

8.1. Life of ndarray 316

 Scientific Python Lectures, Edition 2024.2rc0.dev0

typedef struct PyArrayObject {

PyObject_HEAD

/* Block of memory */
char *data;

/* Data type descriptor */

PyArray_Descr *descr;

/* Indexing scheme */
int nd;
npy_intp *dimensions;
npy_intp *strides;

/* Other stuff */
PyObject *base;
int flags;
PyObject *weakreflist;
} PyArrayObject;

8.1.2 Block of memory

>>> x = np.array([1, 2, 3], dtype=np.int32)

>>> x.data
<... at ...>
>>> bytes(x.data)
b'\x01\x00\x00\x00\x02\x00\x00\x00\x03\x00\x00\x00'

Memory address of the data:

>>> x.__array_interface__['data'][0]
64803824

The whole __array_interface__:

>>> x.__array_interface__
{'data': (..., False), 'strides': None, 'descr': [('', '<i4')], 'typestr': '<i4',
˓→'shape': (3,), 'version': 3}

Reminder: two ndarrays may share the same memory:

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

>>> y = x[:-1]
>>> x[0] = 9
>>> y
array([9, 2, 3])

8.1. Life of ndarray 317

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Memory does not need to be owned by an ndarray:

>>> x = b'1234'

x is a string (in Python 3 a bytes), we can represent its data as an array of ints:

>>> y = np.frombuffer(x, dtype=np.int8)

>>> y.data
<... at ...>
>>> y.base is x
True

>>> y.flags
C_CONTIGUOUS : True
F_CONTIGUOUS : True
OWNDATA : False
WRITEABLE : False
ALIGNED : True
WRITEBACKIFCOPY : False

The owndata and writeable flags indicate status of the memory block.
See also:
array interface

8.1.3 Data types

The descriptor

dtype describes a single item in the array:

type scalar type of the data, one of:

int8, int16, float64, et al. (fixed size)
str, unicode, void (flexible size)
itemsize size of the data block
byte- byte order: big-endian > / little-endian < / not applicable |
order
fields sub-dtypes, if it’s a structured data type
shape shape of the array, if it’s a sub-array

>>> np.dtype(int).type
<class 'numpy.int64'>
>>> np.dtype(int).itemsize
8
>>> np.dtype(int).byteorder
'='

8.1. Life of ndarray 318

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Example: reading .wav files

The .wav file header:

chunk_id "RIFF"
chunk_size 4-byte unsigned little-endian integer
format "WAVE"
fmt_id "fmt "
fmt_size 4-byte unsigned little-endian integer
audio_fmt 2-byte unsigned little-endian integer
num_channels 2-byte unsigned little-endian integer
sample_rate 4-byte unsigned little-endian integer
byte_rate 4-byte unsigned little-endian integer
block_align 2-byte unsigned little-endian integer
bits_per_sample 2-byte unsigned little-endian integer
data_id "data"
data_size 4-byte unsigned little-endian integer

• 44-byte block of raw data (in the beginning of the file)

• . . . followed by data_size bytes of actual sound data.
The .wav file header as a NumPy structured data type:

>>> wav_header_dtype = np.dtype([

... ("chunk_id", (bytes, 4)), # flexible-sized scalar type, item size 4
... ("chunk_size", "<u4"), # little-endian unsigned 32-bit integer
... ("format", "S4"), # 4-byte string
... ("fmt_id", "S4"),
... ("fmt_size", "<u4"),
... ("audio_fmt", "<u2"), #
... ("num_channels", "<u2"), # .. more of the same ...
... ("sample_rate", "<u4"), #
... ("byte_rate", "<u4"),
... ("block_align", "<u2"),
... ("bits_per_sample", "<u2"),
... ("data_id", ("S1", (2, 2))), # sub-array, just for fun!
... ("data_size", "u4"),
... #
... # the sound data itself cannot be represented here:
... # it does not have a fixed size
... ])

See also:
wavreader.py

>>> wav_header_dtype['format']
dtype('S4')
>>> wav_header_dtype.fields
mappingproxy({'chunk_id': (dtype('S4'), 0), 'chunk_size': (dtype('uint32'), 4),
˓→'format': (dtype('S4'), 8), 'fmt_id': (dtype('S4'), 12), 'fmt_size': (dtype('uint32

˓→'), 16), 'audio_fmt': (dtype('uint16'), 20), 'num_channels': (dtype('uint16'), 22),

˓→'sample_rate': (dtype('uint32'), 24), 'byte_rate': (dtype('uint32'), 28), 'block_

˓→align': (dtype('uint16'), 32), 'bits_per_sample': (dtype('uint16'), 34), 'data_id':␣

˓→(dtype(('S1', (2, 2))), 36), 'data_size': (dtype('uint32'), 40)})

>>> wav_header_dtype.fields['format']
(dtype('S4'), 8)

• The first element is the sub-dtype in the structured data, corresponding to the name format

8.1. Life of ndarray 319

 Scientific Python Lectures, Edition 2024.2rc0.dev0

• The second one is its offset (in bytes) from the beginning of the item

Exercise

Mini-exercise, make a “sparse” dtype by using offsets, and only some of the fields:
>>> wav_header_dtype = np.dtype(dict(
... names=['format', 'sample_rate', 'data_id'],
... offsets=[offset_1, offset_2, offset_3], # counted from start of structure in␣
˓→bytes

... formats=list of dtypes for each of the fields,

... ))

and use that to read the sample rate, and data_id (as sub-array).

>>> f = open('data/test.wav', 'r')

>>> wav_header = np.fromfile(f, dtype=wav_header_dtype, count=1)
>>> f.close()
>>> print(wav_header)
[ ('RIFF', 17402L, 'WAVE', 'fmt ', 16L, 1, 1, 16000L, 32000L, 2, 16, [['d', 'a'], ['t
˓→', 'a']], 17366L)]

>>> wav_header['sample_rate']
array([16000], dtype=uint32)

Let’s try accessing the sub-array:

>>> wav_header['data_id']
array([[['d', 'a'],
['t', 'a']]],
dtype='|S1')
>>> wav_header.shape
(1,)
>>> wav_header['data_id'].shape
(1, 2, 2)

When accessing sub-arrays, the dimensions get added to the end!

Note: There are existing modules such as wavfile, audiolab, etc. for loading sound data. . .

Casting and re-interpretation/views

casting
• on assignment
• on array construction
• on arithmetic
• etc.
• and manually: .astype(dtype)
data re-interpretation
• manually: .view(dtype)

8.1. Life of ndarray 320

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Casting

• Casting in arithmetic, in nutshell:

– only type (not value!) of operands matters
– largest “safe” type able to represent both is picked
– scalars can “lose” to arrays in some situations
• Casting in general copies data:

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

>>> x
array([1., 2., 3., 4.])
>>> y = x.astype(np.int8)
>>> y
array([1, 2, 3, 4], dtype=int8)
>>> y + 1
array([2, 3, 4, 5], dtype=int8)
>>> y + 256
array([257, 258, 259, 260], dtype=int16)
>>> y + 256.0
array([257., 258., 259., 260.])
>>> y + np.array([256], dtype=np.int32)
array([257, 258, 259, 260], dtype=int32)

• Casting on setitem: dtype of the array is not changed on item assignment:

>>> y[:] = y + 1.5

>>> y
array([2, 3, 4, 5], dtype=int8)

Note: Exact rules: see NumPy documentation

Re-interpretation / viewing

• Data block in memory (4 bytes)

0x01 || 0x02 || 0x03 || 0x04

– 4 of uint8, OR,
– 4 of int8, OR,
– 2 of int16, OR,
– 1 of int32, OR,
– 1 of float32, OR,
– ...
How to switch from one to another?
1. Switch the dtype:

>>> x = np.array([1, 2, 3, 4], dtype=np.uint8)

>>> x.dtype = "<i2"
>>> x
(continues on next page)

8.1. Life of ndarray 321

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

array([ 513, 1027], dtype=int16)
>>> 0x0201, 0x0403
(513, 1027)

0x01 0x02 || 0x03 0x04

Note: little-endian: least significant byte is on the left in memory

2. Create a new view of type uint32, shorthand i4:

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

0x01 0x02 0x03 0x04

Note:
• .view() makes views, does not copy (or alter) the memory block
• only changes the dtype (and adjusts array shape):

>>> x[1] = 5
>>> y
array([328193], dtype=int32)
>>> y.base is x
True

Mini-exercise: data re-interpretation

See also:
view-colors.py
You have RGBA data in an array:

>>> x = np.zeros((10, 10, 4), dtype=np.int8)

>>> x[:, :, 0] = 1
>>> x[:, :, 1] = 2
>>> x[:, :, 2] = 3
>>> x[:, :, 3] = 4

where the last three dimensions are the R, B, and G, and alpha channels.
How to make a (10, 10) structured array with field names ‘r’, ‘g’, ‘b’, ‘a’ without copying data?

>>> y = ...

>>> assert (y['r'] == 1).all()

>>> assert (y['g'] == 2).all()
(continues on next page)

8.1. Life of ndarray 322

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

>>> assert (y['b'] == 3).all()
>>> assert (y['a'] == 4).all()

Solution

>>> y = x.view([('r', 'i1'),

... ('g', 'i1'),
... ('b', 'i1'),
... ('a', 'i1')]
... )[:, :, 0]

Warning: Another two arrays, each occupying exactly 4 bytes of memory:

>>> x = np.array([[1, 3], [2, 4]], dtype=np.uint8)
>>> x
array([[1, 3],
[2, 4]], dtype=uint8)
>>> y = x.transpose()
>>> y
array([[1, 2],
[3, 4]], dtype=uint8)

We view the elements of x (1 byte each) as int16 (2 bytes each):

>>> x.view(np.int16)
array([[ 769],
[1026]], dtype=int16)

What is happening here? Take a look at the bytes stored in memory by x:

>>> x.tobytes()
b'\x01\x03\x02\x04'

The \x stands for heXadecimal, so what we are seeing is:

0x01 0x03 0x02 0x04

We ask NumPy to interpret these bytes as elements of dtype int16—each of which occupies two
bytes in memory. Therefore, 0x01 0x03 becomes the first uint16 and 0x02 0x04 the second.
You may then expect to see 0x0103 (259, when converting from hexadecimal to decimal) as the first
result. But your computer likely stores most significant bytes first, and as such reads the number as
0x0301 or 769 (go on and type 0x0301 into your Python terminal to verify).
We can do the same on a copy of y (why doesn’t it work on y directly?):
>>> y.copy().view(np.int16)
array([[ 513],
[1027]], dtype=int16)

Can you explain these numbers, 513 and 1027, as well as the output shape of the resulting array?

8.1. Life of ndarray 323

 Scientific Python Lectures, Edition 2024.2rc0.dev0

8.1.4 Indexing scheme: strides

Main point

The question:

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

... [4, 5, 6],
... [7, 8, 9]], dtype=np.int8)
>>> x.tobytes('A')
b'\x01\x02\x03\x04\x05\x06\x07\x08\t'

At which byte in ``x.data`` does the item ``x[1, 2]`` begin?

The answer (in NumPy)

• strides: the number of bytes to jump to find the next element
• 1 stride per dimension

>>> x.strides
(3, 1)
>>> byte_offset = 3 * 1 + 1 * 2 # to find x[1, 2]
>>> x.flat[byte_offset]
6
>>> x[1, 2]
6

simple, flexible

C and Fortran order

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.

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

... [4, 5, 6]], dtype=np.int16, order='C')
>>> x.strides
(6, 2)
>>> x.tobytes('A')
b'\x01\x00\x02\x00\x03\x00\x04\x00\x05\x00\x06\x00'

• Need to jump 6 bytes to find the next row

• Need to jump 2 bytes to find the next column

>>> y = np.array(x, order='F')

>>> y.strides
(2, 4)
>>> y.tobytes('A')
b'\x01\x00\x04\x00\x02\x00\x05\x00\x03\x00\x06\x00'

• Need to jump 2 bytes to find the next row

• Need to jump 4 bytes to find the next column
• Similarly to higher dimensions:

8.1. Life of ndarray 324

 Scientific Python Lectures, Edition 2024.2rc0.dev0

– C: last dimensions vary fastest (= smaller strides)

– F: first dimensions vary fastest
shape = (𝑑1 , 𝑑2 , ..., 𝑑𝑛 )
strides = (𝑠1 , 𝑠2 , ..., 𝑠𝑛 )
𝑠𝐶
𝑗 = 𝑑𝑗+1 𝑑𝑗+2 ...𝑑𝑛 × itemsize

𝑠𝐹
𝑗 = 𝑑1 𝑑2 ...𝑑𝑗−1 × itemsize

Note: Now we can understand the behavior of .view():

>>> y = np.array([[1, 3], [2, 4]], dtype=np.uint8).transpose()

>>> x = y.copy()

Transposition does not affect the memory layout of the data, only strides

>>> x.strides
(2, 1)
>>> y.strides
(1, 2)

>>> x.tobytes('A')
b'\x01\x02\x03\x04'
>>> y.tobytes('A')
b'\x01\x03\x02\x04'

• the results are different when interpreted as 2 of int16

• .copy() creates new arrays in the C order (by default)

Note: In-place operations with views

Prior to NumPy version 1.13, in-place operations with views could result in incorrect results for large
arrays. Since version 1.13, NumPy includes checks for memory overlap to guarantee that results are
consistent with the non in-place version (e.g. a = a + a.T produces the same result as a += a.T).
Note however that this may result in the data being copied (as if using a += a.T.copy()), ultimately
resulting in more memory being used than might otherwise be expected for in-place operations!

Slicing with integers

• Everything can be represented by changing only shape, strides, and possibly adjusting the data
pointer!
• Never makes copies of the data

>>> x = np.array([1, 2, 3, 4, 5, 6], dtype=np.int32)

>>> y = x[::-1]
>>> y
array([6, 5, 4, 3, 2, 1], dtype=int32)
>>> y.strides
(-4,)

>>> y = x[2:]
>>> y.__array_interface__['data'][0] - x.__array_interface__['data'][0]
8
(continues on next page)

8.1. Life of ndarray 325

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

>>> x = np.zeros((10, 10, 10), dtype=float)

>>> x.strides
(800, 80, 8)
>>> x[::2,::3,::4].strides
(1600, 240, 32)

• Similarly, transposes never make copies (it just swaps strides):

>>> x = np.zeros((10, 10, 10), dtype=float)

>>> x.strides
(800, 80, 8)
>>> x.T.strides
(8, 80, 800)

But: not all reshaping operations can be represented by playing with strides:

>>> a = np.arange(6, dtype=np.int8).reshape(3, 2)

>>> b = a.T
>>> b.strides
(1, 2)

So far, so good. However:

>>> bytes(a.data)
b'\x00\x01\x02\x03\x04\x05'
>>> b
array([[0, 2, 4],
[1, 3, 5]], dtype=int8)
>>> c = b.reshape(3*2)
>>> c
array([0, 2, 4, 1, 3, 5], dtype=int8)

Here, there is no way to represent the array c given one stride and the block of memory for a. Therefore,
the reshape operation needs to make a copy here.

Example: fake dimensions with strides

Stride manipulation

>>> from numpy.lib.stride_tricks import as_strided

>>> help(as_strided)
Help on function as_strided in module numpy.lib.stride_tricks:
...

Warning: as_strided does not check that you stay inside the memory block bounds. . .

>>> x = np.array([1, 2, 3, 4], dtype=np.int16)

>>> as_strided(x, strides=(2*2, ), shape=(2, ))
array([1, 3], dtype=int16)
>>> x[::2]
array([1, 3], dtype=int16)

8.1. Life of ndarray 326

 Scientific Python Lectures, Edition 2024.2rc0.dev0

See also:
stride-fakedims.py
Exercise

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

-> array([[1, 2, 3, 4],

[1, 2, 3, 4],
[1, 2, 3, 4]], dtype=np.int8)

using only as_strided.:

Hint: byte_offset = stride[0]*index[0] + stride[1]*index[1] + ...

Spoiler
Stride can also be 0 :

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

>>> y = as_strided(x, strides=(0, 1), shape=(3, 4))
>>> y
array([[1, 2, 3, 4],
[1, 2, 3, 4],
[1, 2, 3, 4]], dtype=int8)
>>> y.base.base is x
True

Broadcasting

• Doing something useful with it: outer product of [1, 2, 3, 4] and [5, 6, 7]

>>> x = np.array([1, 2, 3, 4], dtype=np.int16)

>>> x2 = as_strided(x, strides=(0, 1*2), shape=(3, 4))
>>> x2
array([[1, 2, 3, 4],
[1, 2, 3, 4],
[1, 2, 3, 4]], dtype=int16)

>>> y = np.array([5, 6, 7], dtype=np.int16)

>>> y2 = as_strided(y, strides=(1*2, 0), shape=(3, 4))
>>> y2
array([[5, 5, 5, 5],
[6, 6, 6, 6],
[7, 7, 7, 7]], dtype=int16)

>>> x2 * y2
array([[ 5, 10, 15, 20],
[ 6, 12, 18, 24],
[ 7, 14, 21, 28]], dtype=int16)

8.1. Life of ndarray 327

 Scientific Python Lectures, Edition 2024.2rc0.dev0

. . . seems somehow familiar . . .

>>> x = np.array([1, 2, 3, 4], dtype=np.int16)

>>> y = np.array([5, 6, 7], dtype=np.int16)
>>> x[np.newaxis,:] * y[:,np.newaxis]
array([[ 5, 10, 15, 20],
[ 6, 12, 18, 24],
[ 7, 14, 21, 28]], dtype=int16)

• Internally, array broadcasting is indeed implemented using 0-strides.

More tricks: diagonals

See also:
stride-diagonals.py
Challenge
• Pick diagonal entries of the matrix: (assume C memory order):

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

... [4, 5, 6],
... [7, 8, 9]], dtype=np.int32)

>>> x_diag = as_strided(x, shape=(3,), strides=(???,))

• Pick the first super-diagonal entries [2, 6].

• And the sub-diagonals?
(Hint to the last two: slicing first moves the point where striding
starts from.)
Solution
Pick diagonals:

>>> x_diag = as_strided(x, shape=(3, ), strides=((3+1)*x.itemsize, ))

>>> x_diag
array([1, 5, 9], dtype=int32)

Slice first, to adjust the data pointer:

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

array([2, 6], dtype=int32)

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

array([4, 8], dtype=int32)

Note: Using np.diag

>>> y = np.diag(x, k=1)

>>> y
array([2, 6], dtype=int32)

However,

>>> y.flags.owndata
False

8.1. Life of ndarray 328

 Scientific Python Lectures, Edition 2024.2rc0.dev0

See also:
stride-diagonals.py
Challenge
Compute the tensor trace:

>>> x = np.arange(5*5*5*5).reshape(5, 5, 5, 5)
>>> s = 0
>>> for i in 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()

CPU cache effects

Memory layout can affect performance:

In [1]: x = np.zeros((20000,))

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

In [3]: x.shape, y.shape

Out[3]: ((20000,), (20000,))

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

5.13 us +- 14.3 ns per loop (mean +- std. dev. of 7 runs, 100,000 loops each)

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

22.3 us +- 1.43 us per loop (mean +- std. dev. of 7 runs, 10,000 loops each)

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

Out[6]: ((8,), (536,))

Smaller strides are faster?

8.1. Life of ndarray 329

 Scientific Python Lectures, Edition 2024.2rc0.dev0

• CPU pulls data from main memory to its cache in blocks

• If many array items consecutively operated on fit in a single block (small stride):
– ⇒ fewer transfers needed
– ⇒ faster
See also:
• numexpr is designed to mitigate cache effects when evaluating array expressions.
• numba is a compiler for Python code, that is aware of numpy arrays.

8.1.5 Findings in dissection

• memory block: may be shared, .base, .data

• data type descriptor: structured data, sub-arrays, byte order, casting, viewing, .astype(), .view()
• strided indexing: strides, C/F-order, slicing w/ integers, as_strided, broadcasting, stride tricks,
diag, CPU cache coherence

8.2 Universal functions

8.2.1 What they are?

• Ufunc performs and elementwise operation on all elements of an array.

Examples:

np.add, np.subtract, scipy.special.*, ...

• Automatically support: broadcasting, casting, . . .

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

8.2. Universal functions 330

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Parts of an Ufunc

1. Provided by user

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

{
/*
* int8 output = elementwise_function(int8 input_1, int8 input_2)
*
* This function must compute the ufunc for many values at once,
* in the way shown below.
*/
char *input_1 = (char*)args[0];
char *input_2 = (char*)args[1];
char *output = (char*)args[2];
int i;

for (i = 0; i < dimensions[0]; ++i) {

*output = elementwise_function(*input_1, *input_2);
input_1 += steps[0];
input_2 += steps[1];
output += steps[2];
}
}

2. The NumPy part, built by

char types[3]

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

types[1] = NPY_BYTE /* type of second input arg */
types[2] = NPY_BYTE /* type of third input arg */

PyObject *python_ufunc = PyUFunc_FromFuncAndData(

ufunc_loop,
NULL,
types,
1, /* ntypes */
2, /* num_inputs */
1, /* num_outputs */
identity_element,
name,
docstring,
unused)

• A ufunc can also support multiple different input-output type combinations.

8.2. Universal functions 331

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Making it easier

3. ufunc_loop is of very generic form, and NumPy provides pre-made ones

PyUfunc_f_f float elementwise_func(float input_1)

PyUfunc_ff_f float elementwise_func(float input_1, float input_2)
PyUfunc_d_d double elementwise_func(double input_1)
PyUfunc_dd_d double elementwise_func(double input_1, double input_2)
PyUfunc_D_D elementwise_func(npy_cdouble *input, npy_cdouble* output)
PyUfunc_DD_D elementwise_func(npy_cdouble *in1, npy_cdouble *in2, npy_cdouble*
out)

• Only elementwise_func needs to be supplied

• . . . except when your elementwise function is not in one of the above forms

8.2.2 Exercise: building an ufunc from scratch

The Mandelbrot fractal is defined by the iteration

𝑧 ← 𝑧2 + 𝑐

where 𝑐 = 𝑥 + 𝑖𝑦 is a complex number. This iteration is repeated – if 𝑧 stays finite no matter how long
the iteration runs, 𝑐 belongs to the Mandelbrot set.
• Make ufunc called mandel(z0, c) that computes:

z = z0
for k in range(iterations):
z = z*z + c

say, 100 iterations or until z.real**2 + z.imag**2 > 1000. Use it to determine which c are in
the Mandelbrot set.
• Our function is a simple one, so make use of the PyUFunc_* helpers.
• Write it in Cython
See also:
mandel.pyx, mandelplot.py

#
# Fix the parts marked by TODO
#

#
# Compile this file by (Cython >= 0.12 required because of the complex vars)
#
# cython mandel.pyx
# python setup.py build_ext -i
#
# and try it out with, in this directory,
#
# >>> import mandel
# >>> mandel.mandel(0, 1 + 2j)
#
#

(continues on next page)

8.2. Universal functions 332

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

# The elementwise function
# ------------------------

cdef void mandel_single_point(double complex *z_in,

double complex *c_in,
double complex *z_out) nogil:
#
# The Mandelbrot iteration
#

#
# Some points of note:
#
# - It's *NOT* allowed to call any Python functions here.
#
# The Ufunc loop runs with the Python Global Interpreter Lock released.
# Hence, the ``nogil``.
#
# - And so all local variables must be declared with ``cdef``
#
# - Note also that this function receives *pointers* to the data
#

cdef double complex z = z_in[0]

cdef double complex c = c_in[0]
cdef int k # the integer we use in the for loop

#
# TODO: write the Mandelbrot iteration for one point here,
# as you would write it in Python.
#
# Say, use 100 as the maximum number of iterations, and 1000
# as the cutoff for z.real**2 + z.imag**2.
#

TODO: mandelbrot iteration should go here

# Return the answer for this point

z_out[0] = z

# Boilerplate Cython definitions

#
# Pulls definitions from the NumPy C headers.
# -------------------------------------------

from numpy cimport import_array, import_ufunc

from numpy cimport (PyUFunc_FromFuncAndData,
PyUFuncGenericFunction)
from numpy cimport NPY_CDOUBLE, NP_DOUBLE, NPY_LONG

# Import all pre-defined loop functions

# you won't need all of them - keep the relevant ones

from numpy cimport (

PyUFunc_f_f_As_d_d,
(continues on next page)

8.2. Universal functions 333

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

PyUFunc_d_d,
PyUFunc_f_f,
PyUFunc_g_g,
PyUFunc_F_F_As_D_D,
PyUFunc_F_F,
PyUFunc_D_D,
PyUFunc_G_G,
PyUFunc_ff_f_As_dd_d,
PyUFunc_ff_f,
PyUFunc_dd_d,
PyUFunc_gg_g,
PyUFunc_FF_F_As_DD_D,
PyUFunc_DD_D,
PyUFunc_FF_F,
PyUFunc_GG_G)

# Required module initialization

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

import_array()
import_ufunc()

# The actual ufunc declaration

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

cdef PyUFuncGenericFunction loop_func[1]

cdef char input_output_types[3]
cdef void *elementwise_funcs[1]

#
# Reminder: some pre-made Ufunc loops:
#
# ================ =======================================================
# ``PyUfunc_f_f`` ``float elementwise_func(float input_1)``
# ``PyUfunc_ff_f`` ``float elementwise_func(float input_1, float input_2)``
# ``PyUfunc_d_d`` ``double elementwise_func(double input_1)``
# ``PyUfunc_dd_d`` ``double elementwise_func(double input_1, double input_2)``
# ``PyUfunc_D_D`` ``elementwise_func(complex_double *input, complex_double* complex_
˓→double)``

# ``PyUfunc_DD_D`` ``elementwise_func(complex_double *in1, complex_double *in2,␣

˓→complex_double* out)``

# ================ =======================================================
#
# The full list is above.
#
#
# Type codes:
#
# NPY_BOOL, NPY_BYTE, NPY_UBYTE, NPY_SHORT, NPY_USHORT, NPY_INT, NPY_UINT,
# NPY_LONG, NPY_ULONG, NPY_LONGLONG, NPY_ULONGLONG, NPY_FLOAT, NPY_DOUBLE,
# NPY_LONGDOUBLE, NPY_CFLOAT, NPY_CDOUBLE, NPY_CLONGDOUBLE, NPY_DATETIME,
# NPY_TIMEDELTA, NPY_OBJECT, NPY_STRING, NPY_UNICODE, NPY_VOID
#

(continues on next page)

8.2. Universal functions 334

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

loop_func[0] = ... TODO: suitable PyUFunc_* ...
input_output_types[0] = ... TODO ...
... TODO: fill in rest of input_output_types ...

# This thing is passed as the ``data`` parameter for the generic

# PyUFunc_* loop, to let it know which function it should call.
elementwise_funcs[0] = <void*>mandel_single_point

# Construct the ufunc:

mandel = PyUFunc_FromFuncAndData(
loop_func,
elementwise_funcs,
input_output_types,
1, # number of supported input types
TODO, # number of input args
TODO, # number of output args
0, # `identity` element, never mind this
"mandel", # function name
"mandel(z, c) -> computes z*z + c", # docstring
0 # unused
)

Reminder: some pre-made Ufunc loops:

PyUfunc_f_f float elementwise_func(float input_1)

PyUfunc_ff_f float elementwise_func(float input_1, float input_2)
PyUfunc_d_d double elementwise_func(double input_1)
PyUfunc_dd_d double elementwise_func(double input_1, double input_2)
PyUfunc_D_D elementwise_func(complex_double *input, complex_double* output)
PyUfunc_DD_D elementwise_func(complex_double *in1, complex_double *in2,
complex_double* out)

Type codes:
NPY_BOOL, NPY_BYTE, NPY_UBYTE, NPY_SHORT, NPY_USHORT, NPY_INT, NPY_UINT,
NPY_LONG, NPY_ULONG, NPY_LONGLONG, NPY_ULONGLONG, NPY_FLOAT, NPY_DOUBLE,
NPY_LONGDOUBLE, NPY_CFLOAT, NPY_CDOUBLE, NPY_CLONGDOUBLE, NPY_DATETIME,
NPY_TIMEDELTA, NPY_OBJECT, NPY_STRING, NPY_UNICODE, NPY_VOID

8.2.3 Solution: building an ufunc from scratch

# The elementwise function

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

cdef void mandel_single_point(double complex *z_in,

double complex *c_in,
double complex *z_out) nogil:
#
# The Mandelbrot iteration
#

#
# Some points of note:
(continues on next page)

8.2. Universal functions 335

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

#
# - It's *NOT* allowed to call any Python functions here.
#
# The Ufunc loop runs with the Python Global Interpreter Lock released.
# Hence, the ``nogil``.
#
# - And so all local variables must be declared with ``cdef``
#
# - Note also that this function receives *pointers* to the data;
# the "traditional" solution to passing complex variables around
#

cdef double complex z = z_in[0]

cdef double complex c = c_in[0]
cdef int k # the integer we use in the for loop

# Straightforward iteration

for k in range(100):
z = z*z + c
if z.real**2 + z.imag**2 > 1000:
break

# Return the answer for this point

z_out[0] = z

# Boilerplate Cython definitions

#
# Pulls definitions from the NumPy C headers.
# -------------------------------------------

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

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

cdef PyUFuncGenericFunction loop_func[1]

cdef char input_output_types[3]
cdef void *elementwise_funcs[1]

loop_func[0] = PyUFunc_DD_D

input_output_types[0] = NPY_CDOUBLE
input_output_types[1] = NPY_CDOUBLE
(continues on next page)

8.2. Universal functions 336

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

input_output_types[2] = NPY_CDOUBLE

elementwise_funcs[0] = <void*>mandel_single_point

mandel = PyUFunc_FromFuncAndData(
loop_func,
elementwise_funcs,
input_output_types,
1, # number of supported input types
2, # number of input args
1, # number of output args
0, # `identity` element, never mind this
"mandel", # function name
"mandel(z, c) -> computes iterated z*z + c", # docstring
0 # unused
)

"""
Plot Mandelbrot
================

Plot the Mandelbrot ensemble.

"""

import numpy as np
import mandel

x = np.linspace(-1.7, 0.6, 1000)

y = np.linspace(-1.4, 1.4, 1000)
c = x[None, :] + 1j * y[:, None]
z = mandel.mandel(c, c)

import matplotlib.pyplot as plt

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

plt.gray()
plt.show()

Note: Most of the boilerplate could be automated by these Cython modules:

8.2. Universal functions 337

 Scientific Python Lectures, Edition 2024.2rc0.dev0

https://github.com/cython/cython/wiki/MarkLodato-CreatingUfuncs

Several accepted input types

E.g. supporting both single- and double-precision versions

cdef void mandel_single_point(double complex *z_in,

double complex *c_in,
double complex *z_out) nogil:
...

cdef void mandel_single_point_singleprec(float complex *z_in,

float complex *c_in,
float complex *z_out) nogil:
...

cdef PyUFuncGenericFunction loop_funcs[2]

cdef char input_output_types[3*2]
cdef void *elementwise_funcs[1*2]

loop_funcs[0] = PyUFunc_DD_D
input_output_types[0] = NPY_CDOUBLE
input_output_types[1] = NPY_CDOUBLE
input_output_types[2] = NPY_CDOUBLE
elementwise_funcs[0] = <void*>mandel_single_point

loop_funcs[1] = PyUFunc_FF_F
input_output_types[3] = NPY_CFLOAT
input_output_types[4] = NPY_CFLOAT
input_output_types[5] = NPY_CFLOAT
elementwise_funcs[1] = <void*>mandel_single_point_singleprec

mandel = PyUFunc_FromFuncAndData(
loop_func,
elementwise_funcs,
input_output_types,
2, # number of supported input types <----------------
2, # number of input args
1, # number of output args
0, # `identity` element, never mind this
"mandel", # function name
"mandel(z, c) -> computes iterated z*z + c", # docstring
0 # unused
)

8.2. Universal functions 338

 Scientific Python Lectures, Edition 2024.2rc0.dev0

8.2.4 Generalized ufuncs

ufunc
output = elementwise_function(input)
Both output and input can be a single array element only.
generalized ufunc
output and input can be arrays with a fixed number of dimensions
For example, matrix trace (sum of diag elements):

input shape = (n, n)

output shape = () i.e. scalar

(n, n) -> ()

Matrix product:

input_1 shape = (m, n)

input_2 shape = (n, p)
output shape = (m, p)

(m, n), (n, p) -> (m, p)

• This is called the “signature” of the generalized ufunc

• The dimensions on which the g-ufunc acts, are “core dimensions”

Status in NumPy

• g-ufuncs are in NumPy already . . .

• new ones can be created with PyUFunc_FromFuncAndDataAndSignature
• most linear-algebra functions are implemented as g-ufuncs to enable working with stacked arrays:

>>> import numpy as np

>>> rng = np.random.default_rng(27446968)
>>> np.linalg.det(rng.random((3, 5, 5)))
array([ 0.01829761, -0.0077266 , -0.05336566])
>>> np.linalg._umath_linalg.det.signature
'(m,m)->()'

• matrix multiplication this way could be useful for operating on many small matrices at once
• Also see tensordot and einsum

Generalized ufunc loop

Matrix multiplication (m,n),(n,p) -> (m,p)

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

{
char *input_1 = (char*)args[0]; /* these are as previously */
char *input_2 = (char*)args[1];
char *output = (char*)args[2];

int input_1_stride_m = steps[3]; /* strides for the core dimensions */

(continues on next page)

8.2. Universal functions 339

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

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

int m = dimension[1]; /* core dimensions are added after */

int n = dimension[2]; /* the main dimension; order as in */
int p = dimension[3]; /* signature */

int i;

for (i = 0; i < dimensions[0]; ++i) {

matmul_for_strided_matrices(input_1, input_2, output,
strides for each array...);

input_1 += steps[0];
input_2 += steps[1];
output += steps[2];
}
}

8.3 Interoperability features

8.3.1 Sharing multidimensional, typed data

Suppose you
1. Write a library than handles (multidimensional) binary data,
2. Want to make it easy to manipulate the data with NumPy, or whatever other library,
3. . . . but would not like to have NumPy as a dependency.
Currently, 3 solutions:
1. the “old” buffer interface
2. the array interface
3. the “new” buffer interface (PEP 3118)

8.3.2 The old buffer protocol

• Only 1-D buffers

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

8.3. Interoperability features 340

 Scientific Python Lectures, Edition 2024.2rc0.dev0

>>> from PIL import Image

>>> data = np.zeros((200, 200, 4), dtype=np.uint8)
>>> data[:, :] = [255, 0, 0, 255] # Red
>>> # In PIL, RGBA images consist of 32-bit integers whose bytes are [RR,GG,BB,AA]
>>> data = data.view(np.int32).squeeze()
>>> img = Image.frombuffer("RGBA", (200, 200), data, "raw", "RGBA", 0, 1)
>>> img.save('test.png')

Q:
Check what happens if data is now modified, and img saved again.

8.3.3 The old buffer protocol

"""
From buffer
============

Show how to exchange data between numpy and a library that only knows
the buffer interface.
"""

import numpy as np
import Image

# Let's make a sample image, RGBA format

x = np.zeros((200, 200, 4), dtype=np.int8)

x[:, :, 0] = 254 # red

x[:, :, 3] = 255 # opaque

data = x.view(np.int32) # Check that you understand why this is OK!

img = Image.frombuffer("RGBA", (200, 200), data)

img.save("test.png")

#
# Modify the original data, and save again.
#
# It turns out that PIL, which knows next to nothing about NumPy,
# happily shares the same data.
#

x[:, :, 1] = 254
img.save("test2.png")

8.3. Interoperability features 341

 Scientific Python Lectures, Edition 2024.2rc0.dev0

8.3.4 Array interface protocol

• Multidimensional buffers
• Data type information present
• NumPy-specific approach; slowly deprecated (but not going away)
• Not integrated in Python otherwise
See also:
Documentation: https://numpy.org/doc/stable/reference/arrays.interface.html

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

>>> x.__array_interface__
{'data': (171694552, False), # memory address of data, is readonly?
'descr': [('', '<i4')], # data type descriptor
'typestr': '<i4', # same, in another form
'strides': None, # strides; or None if in C-order
'shape': (2, 2),
'version': 3,
}

::

>>> from PIL import Image

>>> img = Image.open('data/test.png')
>>> img.__array_interface__
{'version': 3,
'data': ...,
'shape': (200, 200, 4),
(continues on next page)

8.3. Interoperability features 342

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

'typestr': '|u1'}
>>> x = np.asarray(img)
>>> x.shape
(200, 200, 4)

Note: A more C-friendly variant of the array interface is also defined.

8.4 Array siblings: chararray, maskedarray

8.4.1 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

8.4.2 masked_array missing data

Masked arrays are arrays that may have missing or invalid entries.
For example, suppose we have an array where the fourth entry is invalid:

>>> x = np.array([1, 2, 3, -99, 5])

One way to describe this is to create a masked array:

>>> mx = np.ma.masked_array(x, mask=[0, 0, 0, 1, 0])

>>> mx
masked_array(data=[1, 2, 3, --, 5],
mask=[False, False, False, True, False],
fill_value=999999)

Masked mean ignores masked data:

>>> mx.mean()
2.75
>>> np.mean(mx)
2.75

Warning: Not all NumPy functions respect masks, for instance np.dot, so check the return types.

The masked_array returns a view to the original array:

8.4. Array siblings: chararray, maskedarray 343

 Scientific Python Lectures, Edition 2024.2rc0.dev0

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

The mask

You can modify the mask by assigning:

>>> mx[1] = np.ma.masked

>>> mx
masked_array(data=[1, --, 3, --, 5],
mask=[False, True, False, True, False],
fill_value=999999)

The mask is cleared on assignment:

>>> mx[1] = 9
>>> mx
masked_array(data=[1, 9, 3, --, 5],
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],
fill_value=999999)

Domain-aware functions

The masked array package also contains domain-aware functions:

>>> np.ma.log(np.array([1, 2, -1, -2, 3, -5]))

masked_array(data=[0.0, 0.693147180559..., --, --, 1.098612288668..., --],
mask=[False, False, True, True, False, True],
fill_value=1e+20)

Note: Streamlined and more seamless support for dealing with missing data in arrays is making its
way into NumPy 1.7. Stay tuned!

8.4. Array siblings: chararray, maskedarray 344

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Example: Masked statistics

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

>>> bad_years = (((year >= 1903) & (year <= 1910))

... | ((year >= 1917) & (year <= 1918)))
>>> # '&' means 'and' and '|' means 'or'
>>> populations[bad_years, 0] = np.ma.masked
>>> populations[bad_years, 1] = np.ma.masked

>>> populations.mean(axis=0)
masked_array(data=[40472.72727272727, 18627.272727272728, 42400.0],
mask=[False, False, False],
fill_value=1e+20)

>>> populations.std(axis=0)
masked_array(data=[21087.656489006717, 15625.799814240254, 3322.5062255844787],
mask=[False, False, False],
fill_value=1e+20)

Note that Matplotlib knows about masked arrays:

>>> plt.plot(year, populations, 'o-')
[<matplotlib.lines.Line2D object at ...>, ...]

8.4.3 recarray: purely convenience

>>> arr = np.array([('a', 1), ('b', 2)], dtype=[('x', 'S1'), ('y', int)])
>>> arr2 = arr.view(np.recarray)
>>> arr2.x
array([b'a', b'b'], dtype='|S1')
>>> arr2.y
array([1, 2])

8.4. Array siblings: chararray, maskedarray 345

 Scientific Python Lectures, Edition 2024.2rc0.dev0

8.5 Summary

• Anatomy of the ndarray: data, dtype, strides.

• Universal functions: elementwise operations, how to make new ones
• Ndarray subclasses
• Various buffer interfaces for integration with other tools
• Recent additions: PEP 3118, generalized ufuncs

8.6 Contributing to NumPy/SciPy

Get this tutorial: https://www.euroscipy.org/talk/882

8.6.1 Why

• “There’s a bug?”
• “I don’t understand what this is supposed to do?”
• “I have this fancy code. Would you like to have it?”
• “I’d like to help! What can I do?”

8.6.2 Reporting bugs

• Bug tracker (prefer this)

– https://github.com/numpy/numpy/issues
– https://github.com/scipy/scipy/issues
– Click the “Sign up” link to get an account
• Mailing lists (https://numpy.org/community/)
– If you’re unsure
– No replies in a week or so? Just file a bug ticket.

Good bug report

Title: numpy.random.permutations fails for non-integer arguments

I'm trying to generate random permutations, using numpy.random.permutations

When calling numpy.random.permutation with non-integer arguments

it fails with a cryptic error message::

>>> rng.permutation(12)
array([ 2, 6, 4, 1, 8, 11, 10, 5, 9, 3, 7, 0])
>>> rng.permutation(12.) #doctest: +SKIP
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "_generator.pyx", line 4844, in numpy.random._generator.Generator.permutation
numpy.exceptions.AxisError: axis 0 is out of bounds for array of dimension 0

(continues on next page)

8.5. Summary 346

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

This also happens with long arguments, and so
np.random.permutation(X.shape[0]) where X is an array fails on 64
bit windows (where shape is a tuple of longs).

It would be great if it could cast to integer or at least raise a

proper error for non-integer types.

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.

0. What are you trying to do?

1. Small code snippet reproducing the bug (if possible)
• What actually happens
• What you’d expect
2. Platform (Windows / Linux / OSX, 32/64 bits, x86/PPC, . . . )
3. Version of NumPy/SciPy

>>> print(np.__version__)
1...

Check that the following is what you expect

>>> print(np.__file__)
/...

In case you have old/broken NumPy installations lying around.

If unsure, try to remove existing NumPy installations, and reinstall. . .

8.6.3 Contributing to documentation

1. Documentation editor
• https://numpy.org/doc/stable/
• Registration
– Register an account
– Subscribe to scipy-dev mailing list (subscribers-only)
– Problem with mailing lists: you get mail
∗ But: you can turn mail delivery off
∗ “change your subscription options”, at the bottom of
https://mail.python.org/mailman3/lists/scipy-dev.python.org/
– Send a mail @ scipy-dev mailing list; ask for activation:

To: [email protected]

Hi,

I'd like to edit NumPy/SciPy docstrings. My account is XXXXX

Cheers,
N. N.

8.6. Contributing to NumPy/SciPy 347

 Scientific Python Lectures, Edition 2024.2rc0.dev0

• Check the style guide:

– https://numpy.org/doc/stable/
– Don’t be intimidated; to fix a small thing, just fix it
• Edit
2. Edit sources and send patches (as for bugs)
3. Complain on the mailing list

8.6.4 Contributing features

The contribution of features is documented on https://numpy.org/doc/stable/dev/

8.6.5 How to help, in general

• Bug fixes always welcome!

– What irks you most
– Browse the tracker
• Documentation work
– API docs: improvements to docstrings
∗ Know some SciPy module well?
– User guide
∗ https://numpy.org/doc/stable/user/
• Ask on communication channels:
– numpy-discussion list
– scipy-dev list

8.6. Contributing to NumPy/SciPy 348

 CHAPTER 9
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

349
 Scientific Python Lectures, Edition 2024.2rc0.dev0

– Debugger commands and interaction

• Debugging segmentation faults using gdb

9.1 Avoiding bugs

9.1.1 Coding best practices to avoid getting in trouble

Brian Kernighan

“Everyone knows that debugging is twice as hard as writing a program in the first place. So if you’re
as clever as you can be when you write it, how will you ever debug it?”

• We all write buggy code. Accept it. Deal with it.

• Write your code with testing and debugging in mind.
• Keep It Simple, Stupid (KISS).
– What is the simplest thing that could possibly work?
• Don’t Repeat Yourself (DRY).
– Every piece of knowledge must have a single, unambiguous, authoritative representation within
a system.
– Constants, algorithms, etc. . .
• Try to limit interdependencies of your code. (Loose Coupling)
• Give your variables, functions and modules meaningful names (not mathematics names)

9.1.2 pyflakes: fast static analysis

They are several static analysis tools in Python; to name a few:

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

9.1. Avoiding bugs 350

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Running pyflakes on the current edited file

You can bind a key to run pyflakes in the current buffer.

• In kate Menu: ‘settings -> configure kate
– In plugins enable ‘external tools’
– In external Tools’, add pyflakes:

kdialog --title "pyflakes %f ilename" --msgbox "$(pyflakes %f ilename)"

• In TextMate
Menu: TextMate -> Preferences -> Advanced -> Shell variables, add a shell variable:

TM_PYCHECKER = /Library/Frameworks/Python.framework/Versions/Current/bin/pyflakes

Then Ctrl-Shift-V is binded to a pyflakes report

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

autocmd FileType python let &mp = 'echo "*** running % ***" ; pyflakes %'
autocmd FileType tex,mp,rst,python imap <Esc>[15~ <C-O>:make!^M
autocmd FileType tex,mp,rst,python map <Esc>[15~ :make!^M
autocmd FileType tex,mp,rst,python set autowrite

• In emacs In your .emacs (binds F5 to pyflakes):

(defun pyflakes-thisfile () (interactive)

(compile (format "pyflakes %s " (buffer-file-name)))
)

(define-minor-mode pyflakes-mode
"Toggle pyflakes mode.
With no argument, this command toggles the mode.
Non-null prefix argument turns on the mode.
Null prefix argument turns off the mode."
;; The initial value.
nil
;; The indicator for the mode line.
" Pyflakes"
;; The minor mode bindings.
'( ([f5] . pyflakes-thisfile) )
)

(add-hook 'python-mode-hook (lambda () (pyflakes-mode t)))

A type-as-go spell-checker like integration

• In vim
– Use the pyflakes.vim plugin:
1. download the zip file from https://www.vim.org/scripts/script.php?script_id=2441
2. extract the files in ~/.vim/ftplugin/python
3. make sure your vimrc has filetype plugin indent on

9.1. Avoiding bugs 351

 Scientific Python Lectures, Edition 2024.2rc0.dev0

– Alternatively: use the syntastic plugin. This can be configured to use flake8 too and also
handles on-the-fly checking for many other languages.

• In emacs
Use the flymake mode with pyflakes, documented on https://www.emacswiki.org/emacs/FlyMake
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
following line to your .emacs file:

(add-hook 'python-mode-hook '(lambda () (flymake-mode)))

9.2 Debugging workflow

If you do have a non trivial bug, this is when debugging strategies kick in. There is no silver bullet. Yet,
strategies help:
For debugging a given problem, the favorable situation is when the problem is
isolated in a small number of lines of code, outside framework or application
code, with short modify-run-fail cycles
1. Make it fail reliably. Find a test case that makes the code fail every time.
2. Divide and Conquer. Once you have a failing test case, isolate the failing code.
• Which module.
• Which function.
• Which line of code.
=> isolate a small reproducible failure: a test case
3. Change one thing at a time and re-run the failing test case.
4. Use the debugger to understand what is going wrong.
5. Take notes and be patient. It may take a while.

Note: Once you have gone through this process: isolated a tight piece of code reproducing the bug
and fix the bug using this piece of code, add the corresponding code to your test suite.

9.2. Debugging workflow 352

 Scientific Python Lectures, Edition 2024.2rc0.dev0

9.3 Using the Python debugger

The python debugger, pdb: https://docs.python.org/3/library/pdb.html, allows you to inspect your code
interactively.
Specifically it allows you to:
• View the source code.
• Walk up and down the call stack.
• Inspect values of variables.
• Modify values of variables.
• Set breakpoints.

print

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

9.3.1 Invoking the debugger

Ways to launch the debugger:

1. Postmortem, launch debugger after module errors.
2. Launch the module with the debugger.
3. Call the debugger inside the module

Postmortem

Situation: You’re working in IPython and you get a traceback.

Here we debug the file index_error.py. When running it, an IndexError is raised. Type %debug and
drop into the debugger.

In [1]: %run index_error.py

---------------------------------------------------------------------------
IndexError Traceback (most recent call last)
File ~/src/scientific-python-lectures/advanced/debugging/index_error.py:10
6 print(lst[len(lst)])
9 if __name__ == "__main__":
---> 10 index_error()

File ~/src/scientific-python-lectures/advanced/debugging/index_error.py:6, in index_

˓→error()

4 def index_error():
5 lst = list("foobar")
----> 6 print(lst[len(lst)])

IndexError: list index out of range

In [2]: %debug
> /home/jarrod/src/scientific-python-lectures/advanced/debugging/index_error.
˓→py(6)index_error()

4 def index_error():
(continues on next page)

9.3. Using the Python debugger 353

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

5 lst = list("foobar")
----> 6 print(lst[len(lst)])
7
8

ipdb> list
1 """Small snippet to raise an IndexError."""
2
3
4 def index_error():
5 lst = list("foobar")
----> 6 print(lst[len(lst)])
7
8
9 if __name__ == "__main__":
10 index_error()

ipdb> len(lst)
6
ipdb> print(lst[len(lst) - 1])
r
ipdb> quit

Post-mortem debugging without IPython

In some situations you cannot use IPython, for instance to debug a script that wants to be called from
the command line. In this case, you can call the script with python -m pdb script.py:
$ python -m pdb index_error.py
> /home/jarrod/src/scientific-python-lectures/advanced/debugging/index_error.py(1)
˓→<module>()

-> """Small snippet to raise an IndexError."""

(Pdb) continue
Traceback (most recent call last):
File "/usr/lib64/python3.11/pdb.py", line 1793, in main
pdb._run(target)
File "/usr/lib64/python3.11/pdb.py", line 1659, in _run
self.run(target.code)
File "/usr/lib64/python3.11/bdb.py", line 600, in run
exec(cmd, globals, locals)
File "<string>", line 1, in <module>
File "/home/jarrod/src/scientific-python-lectures/advanced/debugging/index_error.
˓→py", line 10, in <module>

index_error()
File "/home/jarrod/src/scientific-python-lectures/advanced/debugging/index_error.
˓→py", line 6, in index_error

print(lst[len(lst)])
~~~^^^^^^^^^^
IndexError: list index out of range
Uncaught exception. Entering post mortem debugging
Running 'cont' or 'step' will restart the program
> /home/jarrod/src/scientific-python-lectures/advanced/debugging/index_error.
˓→py(6)index_error()

-> print(lst[len(lst)])
(Pdb)

9.3. Using the Python debugger 354

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Step-by-step execution

Situation: You believe a bug exists in a module but are not sure where.
For instance we are trying to debug wiener_filtering.py. Indeed the code runs, but the filtering does
not work well.
• Run the script in IPython with the debugger using %run -d wiener_filtering.py :

In [1]: %run -d wiener_filtering.py

*** Blank or comment
*** Blank or comment
*** Blank or comment
NOTE: Enter 'c' at the ipdb> prompt to continue execution.
> /home/jarrod/src/scientific-python-lectures/advanced/debugging/wiener_
˓→filtering.py(1)<module>()

----> 1 """Wiener filtering a noisy raccoon face: this module is buggy"""

2
3 import numpy as np
4 import scipy as sp
5 import matplotlib.pyplot as plt

• Set a break point at line 29 using b 29:

ipdb> n
> /home/jarrod/src/scientific-python-lectures/advanced/debugging/wiener_
˓→filtering.py(3)<module>()

1 """Wiener filtering a noisy raccoon face: this module is buggy"""

2
----> 3 import numpy as np
4 import scipy as sp
5 import matplotlib.pyplot as plt

ipdb> b 29
Breakpoint 1 at /home/jarrod/src/scientific-python-lectures/advanced/debugging/
˓→wiener_filtering.py:29

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

ipdb> c
> /home/jarrod/src/scientific-python-lectures/advanced/debugging/wiener_
˓→filtering.py(29)iterated_wiener()

27 Do not use this: this is crappy code to demo bugs!

28 """
1--> 29 noisy_img = noisy_img
30 denoised_img = local_mean(noisy_img, size=size)
31 l_var = local_var(noisy_img, size=size)

• Step into code with n(ext) and s(tep): next jumps to the next statement in the current execution
context, while step will go across execution contexts, i.e. enable exploring inside function calls:

ipdb> s
> /home/jarrod/src/scientific-python-lectures/advanced/debugging/wiener_
˓→filtering.py(30)iterated_wiener()

28 """
1 29 noisy_img = noisy_img
---> 30 denoised_img = local_mean(noisy_img, size=size)
31 l_var = local_var(noisy_img, size=size)
32 for i in range(3):
(continues on next page)

9.3. Using the Python debugger 355

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

ipdb> n
> /home/jarrod/src/scientific-python-lectures/advanced/debugging/wiener_
˓→filtering.py(31)iterated_wiener()

1 29 noisy_img = noisy_img
30 denoised_img = local_mean(noisy_img, size=size)
---> 31 l_var = local_var(noisy_img, size=size)
32 for i in range(3):
33 res = noisy_img - denoised_img

• Step a few lines and explore the local variables:

ipdb> n
> /home/jarrod/src/scientific-python-lectures/advanced/debugging/wiener_
˓→filtering.py(32)iterated_wiener()

30 denoised_img = local_mean(noisy_img, size=size)

31 l_var = local_var(noisy_img, size=size)
---> 32 for i in range(3):
33 res = noisy_img - denoised_img
34 noise = (res**2).sum() / res.size

ipdb> print(l_var)
[[2571 2782 3474 ... 3008 2922 3141]
[2105 708 475 ... 469 354 2884]
[1697 420 645 ... 273 236 2517]
...
[2437 345 432 ... 413 387 4188]
[2598 179 247 ... 367 441 3909]
[2808 2525 3117 ... 4413 4454 4385]]
ipdb> print(l_var.min())
0

Oh dear, nothing but integers, and 0 variation. Here is our bug, we are doing integer arithmetic.

Raising exception on numerical errors

When we run the wiener_filtering.py file, the following warnings are raised:
In [2]: %run wiener_filtering.py
/home/jarrod/src/scientific-python-lectures/advanced/debugging/wiener_filtering.
˓→py:35: RuntimeWarning: divide by zero encountered in divide

noise_level = 1 - noise / l_var

We can turn these warnings in exception, which enables us to do post-mortem debugging on them,
and find our problem more quickly:
In [3]: np.seterr(all='raise')
Out[3]: {'divide': 'warn', 'over': 'warn', 'under': 'ignore', 'invalid': 'warn'}

In [4]: %run wiener_filtering.py

---------------------------------------------------------------------------
FloatingPointError Traceback (most recent call last)
File ~/src/scientific-python-lectures/advanced/debugging/wiener_filtering.py:52
49 plt.matshow(face[cut], cmap=plt.cm.gray)
50 plt.matshow(noisy_face[cut], cmap=plt.cm.gray)
---> 52 denoised_face = iterated_wiener(noisy_face)
53 plt.matshow(denoised_face[cut], cmap=plt.cm.gray)
55 plt.show()

9.3. Using the Python debugger 356

 Scientific Python Lectures, Edition 2024.2rc0.dev0

File ~/src/scientific-python-lectures/advanced/debugging/wiener_filtering.py:35, in␣

˓→iterated_wiener(noisy_img, size)

33 res = noisy_img - denoised_img

34 noise = (res**2).sum() / res.size
---> 35 noise_level = 1 - noise / l_var
36 noise_level[noise_level < 0] = 0
37 denoised_img = np.int64(noise_level * res)

FloatingPointError: divide by zero encountered in divide

Other ways of starting a debugger

• Raising an exception as a poor man break point

If you find it tedious to note the line number to set a break point, you can simply raise an exception
at the point that you want to inspect and use IPython’s %debug. Note that in this case you cannot
step or continue the execution.
• Debugging test failures using nosetests
You can run nosetests --pdb to drop in post-mortem debugging on exceptions, and nosetests
--pdb-failure to inspect test failures using the debugger.
In addition, you can use the IPython interface for the debugger in nose by installing the nose plugin
ipdbplugin. You can than pass --ipdb and --ipdb-failure options to nosetests.
• Calling the debugger explicitly
Insert the following line where you want to drop in the debugger:

import pdb; pdb.set_trace()

Warning: When running nosetests, the output is captured, and thus it seems that the debugger
does not work. Simply run the nosetests with the -s flag.

Graphical debuggers and alternatives

• pudb is a good semi-graphical debugger with a text user interface in the console.
• The Visual Studio Code integrated development environment includes a debugging mode.
• The Mu editor is a simple Python editor that includes a debugging mode.

9.3.2 Debugger commands and interaction

l(list) Lists the code at the current position

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

9.3. Using the Python debugger 357

 Scientific Python Lectures, Edition 2024.2rc0.dev0

Warning: Debugger commands are not Python code

You cannot name the variables the way you want. For instance, if in you cannot override the variables
in the current frame with the same name: use different names than your local variable when
typing code in the debugger.

Getting help when in the debugger

Type h or help to access the interactive help:

ipdb> help

Documented commands (type help <topic>):

========================================
EOF commands enable ll pp s until
a condition exceptions longlist psource skip_hidden up
alias cont exit n q skip_predicates w
args context h next quit source whatis
b continue help p r step where
break d ignore pdef restart tbreak
bt debug j pdoc return u
c disable jump pfile retval unalias
cl display l pinfo run undisplay
clear down list pinfo2 rv unt

Miscellaneous help topics:

==========================
exec pdb

Undocumented commands:
======================
interact

9.4 Debugging segmentation faults using gdb

If you have a segmentation fault, you cannot debug it with pdb, as it crashes the Python interpreter
before it can drop in the debugger. Similarly, if you have a bug in C code embedded in Python, pdb is
useless. For this we turn to the gnu debugger, gdb, available on Linux.
Before we start with gdb, let us add a few Python-specific tools to it. For this we add a few macros to
our ~/.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.
To debug with gdb the Python script segfault.py, we can run the script in gdb as follows

$ gdb python
...
(gdb) run segfault.py
Starting program: /usr/bin/python segfault.py
[Thread debugging using libthread_db enabled]

Program received signal SIGSEGV, Segmentation fault.

_strided_byte_copy (dst=0x8537478 "\360\343G", outstrides=4, src=
0x86c0690 <Address 0x86c0690 out of bounds>, instrides=32, N=3,
elsize=4)
(continues on next page)

9.4. Debugging segmentation faults using gdb 358

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

at numpy/core/src/multiarray/ctors.c:365
365 _FAST_MOVE(Int32);
(gdb)

We get a segfault, and gdb captures it for post-mortem debugging in the C level stack (not the Python
call stack). We can debug the C call stack using gdb’s commands:

(gdb) up
#1 0x004af4f5 in _copy_from_same_shape (dest=<value optimized out>,
src=<value optimized out>, myfunc=0x496780 <_strided_byte_copy>,
swap=0)
at numpy/core/src/multiarray/ctors.c:748
748 myfunc(dit->dataptr, dest->strides[maxaxis],

As you can see, right now, we are in the C code of numpy. We would like to know what is the Python
code that triggers this segfault, so we go up the stack until we hit the Python execution loop:

(gdb) up
#8 0x080ddd23 in call_function (f=
Frame 0x85371ec, for file /home/varoquau/usr/lib/python2.6/site-packages/numpy/
˓→core/arrayprint.py, line 156, in _leading_trailing (a=<numpy.ndarray at remote␣

˓→0x85371b0>, _nc=<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

1630 ../Python/ceval.c: No such file or directory.

in ../Python/ceval.c
(gdb) pyframe
segfault.py (12): print_big_array

The corresponding code is:

9.4. Debugging segmentation faults using gdb 359

 Scientific Python Lectures, Edition 2024.2rc0.dev0

def make_big_array(small_array):
big_array = stride_tricks.as_strided(
small_array, shape=(int(2e6), int(2e6)), strides=(32, 32)
)
return big_array

Thus the segfault happens when printing big_array[-10:]. The reason is simply that big_array has
been allocated with its end outside the program memory.

Note: For a list of Python-specific commands defined in the gdbinit, read the source of this file.

Wrap up exercise

The following script is well documented and hopefully legible. It seeks to answer a problem of actual
interest for numerical computing, but it does not work. . . Can you debug it?
Python source code: to_debug.py

9.4. Debugging segmentation faults using gdb 360

 CHAPTER 10
Optimizing code

Donald Knuth

“Premature optimization is the root of all evil”

Author: Gaël Varoquaux

This chapter deals with strategies to make Python code go faster.

Prerequisites

• line_profiler

Chapters contents

• Optimization workflow
• Profiling Python code
– Timeit
– Profiler
– Line-profiler
• Making code go faster
– Algorithmic optimization
∗ Example of the SVD

361
 Scientific Python Lectures, Edition 2024.2rc0.dev0

• Writing faster numerical code

– Additional Links

10.1 Optimization workflow

1. Make it work: write the code in a simple legible ways.

2. Make it work reliably: write automated test cases, make really sure that your algorithm is right
and that if you break it, the tests will capture the breakage.
3. Optimize the code by profiling simple use-cases to find the bottlenecks and speeding up these
bottleneck, finding a better algorithm or implementation. Keep in mind that a trade off should
be found between profiling on a realistic example and the simplicity and speed of execution of the
code. For efficient work, it is best to work with profiling runs lasting around 10s.

10.2 Profiling Python code

No optimization without measuring!

• Measure: profiling, timing

• You’ll have surprises: the fastest code is not always what you think

10.2.1 Timeit

In IPython, use timeit (https://docs.python.org/3/library/timeit.html) to time elementary operations:

In [1]: import numpy as np

In [2]: a = np.arange(1000)

In [3]: %timeit a ** 2
867 ns +- 2.86 ns per loop (mean +- std. dev. of 7 runs, 1,000,000 loops each)

In [4]: %timeit a ** 2.1

15.3 us +- 24.9 ns per loop (mean +- std. dev. of 7 runs, 100,000 loops each)

In [5]: %timeit a * a
978 ns +- 13.9 ns per loop (mean +- std. dev. of 7 runs, 1,000,000 loops each)

Use this to guide your choice between strategies.

Note: For long running calls, using %time instead of %timeit; it is less precise but faster

10.1. Optimization workflow 362

 Scientific Python Lectures, Edition 2024.2rc0.dev0

10.2.2 Profiler

Useful when you have a large program to profile, for example the following file:

# For this example to run, you also need the 'ica.py' file

import numpy as np
import scipy as sp

from ica import fastica

# @profile # uncomment this line to run with line_profiler

def test():
rng = np.random.default_rng()
data = rng.random((5000, 100))
u, s, v = sp.linalg.svd(data)
pca = u[:, :10].T @ data
results = fastica(pca.T, whiten=False)

if __name__ == "__main__":
test()

Note: This is a combination of two unsupervised learning techniques, principal component analysis
(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
sensors. Doing a PCA first and then an ICA can be useful if you have more sensors than signals. For
more information see: the FastICA example from scikits-learn.

To run it, you also need to download the ica module. In IPython we can time the script:

In [6]: %run -t demo.py

IPython CPU timings (estimated):
User : 14.3929 s.
System: 0.256016 s.

and profile it:

In [7]: %run -p demo.py

916 function calls in 14.551 CPU seconds
Ordered by: internal time
ncalls tottime percall cumtime percall filename:lineno (function)
1 14.457 14.457 14.479 14.479 decomp.py:849 (svd)
1 0.054 0.054 0.054 0.054 {method 'random_sample' of 'mtrand.
˓→RandomState' objects}

1 0.017 0.017 0.021 0.021 function_base.py:645 (asarray_chkfinite)

54 0.011 0.000 0.011 0.000 {numpy.core._dotblas.dot}
2 0.005 0.002 0.005 0.002 {method 'any' of 'numpy.ndarray' objects}
6 0.001 0.000 0.001 0.000 ica.py:195 (gprime)
6 0.001 0.000 0.001 0.000 ica.py:192 (g)
14 0.001 0.000 0.001 0.000 {numpy.linalg.lapack_lite.dsyevd}
19 0.001 0.000 0.001 0.000 twodim_base.py:204 (diag)
1 0.001 0.001 0.008 0.008 ica.py:69 (_ica_par)
1 0.001 0.001 14.551 14.551 {execfile}
(continues on next page)

10.2. Profiling Python code 363

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

107 0.000 0.000 0.001 0.000 defmatrix.py:239 (__array_finalize__)
7 0.000 0.000 0.004 0.001 ica.py:58 (_sym_decorrelation)
7 0.000 0.000 0.002 0.000 linalg.py:841 (eigh)
172 0.000 0.000 0.000 0.000 {isinstance}
1 0.000 0.000 14.551 14.551 demo.py:1 (<module>)
29 0.000 0.000 0.000 0.000 numeric.py:180 (asarray)
35 0.000 0.000 0.000 0.000 defmatrix.py:193 (__new__)
35 0.000 0.000 0.001 0.000 defmatrix.py:43 (asmatrix)
21 0.000 0.000 0.001 0.000 defmatrix.py:287 (__mul__)
41 0.000 0.000 0.000 0.000 {numpy.core.multiarray.zeros}
28 0.000 0.000 0.000 0.000 {method 'transpose' of 'numpy.ndarray'␣
˓→objects}

1 0.000 0.000 0.008 0.008 ica.py:97 (fastica)

...

Clearly the svd (in decomp.py) is what takes most of our time, a.k.a. the bottleneck. We have to find a
way to make this step go faster, or to avoid this step (algorithmic optimization). Spending time on the
rest of the code is useless.

Profiling outside of IPython, running ``cProfile``

Similar profiling can be done outside of IPython, simply calling the built-in Python profilers cProfile
and profile.
$ 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.
This can be useful if you wish to process the profiler output with a visualization tool.

10.2.3 Line-profiler

The profiler tells us which function takes most of the time, but not where it is called.
For this, we use the line_profiler: in the source file, we decorate a few functions that we want to inspect
with @profile (no need to import it)

@profile
def test():
rng = np.random.default_rng()
data = rng.random((5000, 100))
u, s, v = linalg.svd(data)
pca = u[:, :10] @ data
results = fastica(pca.T, whiten=False)

Then we run the script using the kernprof command, with switches -l, --line-by-line and -v,
--view to use the line-by-line profiler and view the results in addition to saving them:

$ kernprof -l -v demo.py

Wrote profile results to demo.py.lprof

Timer unit: 1e-06 s

Total time: 1.27874 s

File: demo.py
Function: test at line 9
(continues on next page)

10.2. Profiling Python code 364

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

Line # Hits Time Per Hit % Time Line Contents

==============================================================
9 @profile
10 def test():
11 1 69.0 69.0 0.0 rng = np.random.default_rng()
12 1 2453.0 2453.0 0.2 data = rng.random((5000, 100))
13 1 1274715.0 1274715.0 99.7 u, s, v = sp.linalg.svd(data)
14 1 413.0 413.0 0.0 pca = u[:, :10].T @ data
15 1 1094.0 1094.0 0.1 results = fastica(pca.T,␣
˓→whiten=False)

The SVD is taking all the time. We need to optimise this line.

10.3 Making code go faster

Once we have identified the bottlenecks, we need to make the corresponding code go faster.

10.3.1 Algorithmic optimization

The first thing to look for is algorithmic optimization: are there ways to compute less, or better?
For a high-level view of the problem, a good understanding of the maths behind the algorithm helps.
However, it is not uncommon to find simple changes, like moving computation or memory allocation
outside a for loop, that bring in big gains.

Example of the SVD

In both examples above, the SVD - Singular Value Decomposition - is what takes most of the time.
Indeed, the computational cost of this algorithm is roughly 𝑛3 in the size of the input matrix.
However, in both of these example, we are not using all the output of the SVD, but only the first few rows
of its first return argument. If we use the svd implementation of SciPy, we can ask for an incomplete
version of the SVD. Note that implementations of linear algebra in SciPy are richer then those in NumPy
and should be preferred.

In [8]: %timeit np.linalg.svd(data)

1 loops, best of 3: 14.5 s per loop

In [9]: import scipy as sp

In [10]: %timeit sp.linalg.svd(data)

1 loops, best of 3: 14.2 s per loop

In [11]: %timeit sp.linalg.svd(data, full_matrices=False)

1 loops, best of 3: 295 ms per loop

In [12]: %timeit np.linalg.svd(data, full_matrices=False)

1 loops, best of 3: 293 ms per loop

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

def test():
rng = np.random.default_rng()
(continues on next page)

10.3. Making code go faster 365

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

data = rng.random((5000, 100))
u, s, v = sp.linalg.svd(data, full_matrices=False)
pca = u[:, :10].T @ data
results = fastica(pca.T, whiten=False)

In [13]: import demo

In [14]: %timeit demo.

demo.fastica demo.np demo.prof.pdf demo.py demo.pyc
demo.linalg demo.prof demo.prof.png demo.py.lprof demo.test

In [15]: %timeit demo.test()

ica.py:65: RuntimeWarning: invalid value encountered in sqrt
W = (u * np.diag(1.0/np.sqrt(s)) * u.T) * W # W = (W * W.T) ^{-1/2} * W
1 loops, best of 3: 17.5 s per loop

In [16]: import demo_opt

In [17]: %timeit demo_opt.test()

1 loops, best of 3: 208 ms per loop

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

Computational linear algebra

For certain algorithms, many of the bottlenecks will be linear algebra computations. In this case,
using the right function to solve the right problem is key. For instance, an eigenvalue problem with
a symmetric matrix is easier to solve than with a general matrix. Also, most often, you can avoid
inverting a matrix and use a less costly (and more numerically stable) operation.
Know your computational linear algebra. When in doubt, explore scipy.linalg, and use %timeit to
try out different alternatives on your data.

10.4 Writing faster numerical code

A complete discussion on advanced use of NumPy is found in chapter Advanced NumPy, or in the article
The NumPy array: a structure for efficient numerical computation by van der Walt et al. Here we discuss
only some commonly encountered tricks to make code faster.
• Vectorizing for loops
Find tricks to avoid for loops using NumPy arrays. For this, masks and indices arrays can be
useful.
• Broadcasting
Use broadcasting to do operations on arrays as small as possible before combining them.
• In place operations

In [18]: a = np.zeros(1e7)

In [19]: %timeit global a ; a = 0*a

10 loops, best of 3: 111 ms per loop

(continues on next page)

10.4. Writing faster numerical code 366

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

In [20]: %timeit global a ; a *= 0
10 loops, best of 3: 48.4 ms per loop

note: we need global a in the timeit so that it work, as it is assigning to a, and thus considers it
as a local variable.
• Be easy on the memory: use views, and not copies
Copying big arrays is as costly as making simple numerical operations on them:

In [21]: a = np.zeros(1e7)

In [22]: %timeit a.copy()

10 loops, best of 3: 124 ms per loop

In [23]: %timeit a + 1
10 loops, best of 3: 112 ms per loop

• Beware of cache effects

Memory access is cheaper when it is grouped: accessing a big array in a continuous way is much
faster than random access. This implies amongst other things that smaller strides are faster
(see CPU cache effects):

In [24]: c = np.zeros((1e4, 1e4), order='C')

In [25]: %timeit c.sum(axis=0)

1 loops, best of 3: 3.89 s per loop

In [26]: %timeit c.sum(axis=1)

1 loops, best of 3: 188 ms per loop

In [27]: c.strides
Out[27]: (80000, 8)

This is the reason why Fortran ordering or C ordering may make a big difference on operations:

In [28]: rng = np.random.default_rng()

In [29]: a = rng.random((20, 2**18))

In [30]: b = rng.random((20, 2**18))

In [31]: %timeit b @ a.T

8.1 ms +- 14.6 us per loop (mean +- std. dev. of 7 runs, 100 loops each)

In [32]: c = np.ascontiguousarray(a.T)

In [33]: %timeit b @ c
8.04 ms +- 23.8 us per loop (mean +- std. dev. of 7 runs, 100 loops each)

Note that copying the data to work around this effect may not be worth it:

In [34]: %timeit c = np.ascontiguousarray(a.T)

16.5 ms +- 192 us per loop (mean +- std. dev. of 7 runs, 100 loops each)

Using numexpr can be useful to automatically optimize code for such effects.
• Use compiled code

10.4. Writing faster numerical code 367

 Scientific Python Lectures, Edition 2024.2rc0.dev0

The last resort, once you are sure that all the high-level optimizations have been explored, is to
transfer the hot spots, i.e. the few lines or functions in which most of the time is spent, to compiled
code. For compiled code, the preferred option is to use Cython: it is easy to transform exiting
Python code in compiled code, and with a good use of the NumPy support yields efficient code on
NumPy arrays, for instance by unrolling loops.

Warning: For all the above: profile and time your choices. Don’t base your optimization on
theoretical considerations.

10.4.1 Additional Links

• If you need to profile memory usage, you could try the memory_profiler
• If you need to profile down into C extensions, you could try using gperftools from Python with
yep.
• If you would like to track performance of your code across time, i.e. as you make new commits to
your repository, you could try: asv
• If you need some interactive visualization why not try RunSnakeRun

10.4. Writing faster numerical code 368

 CHAPTER 11
Sparse Arrays in SciPy

Author: Robert Cimrman

11.1 Introduction

(dense) matrix is:

• mathematical object
• data structure for storing a 2D array of values
important features:
• memory allocated once for all items
– usually a contiguous chunk, think NumPy ndarray
• fast access to individual items (*)

369
 Scientific Python Lectures, Edition 2024.2rc0.dev0

11.1.1 Why Sparse Matrices?

• the memory grows like n**2 for dense matrix

• small example (double precision matrix):

>>> import numpy as np

>>> import matplotlib.pyplot as plt
>>> x = np.linspace(0, 1e6, 10)
>>> plt.plot(x, 8.0 * (x**2) / 1e6, lw=5)
[<matplotlib.lines.Line2D object at ...>]
>>> plt.xlabel('size n')
Text(...'size n')
>>> plt.ylabel('memory [MB]')
Text(...'memory [MB]')

11.1.2 Sparse Matrices vs. Sparse Matrix Storage Schemes

• sparse matrix is a matrix, which is almost empty

• storing all the zeros is wasteful -> store only nonzero items
• think compression
• pros: huge memory savings
• cons: slow access to individual items, but it depends on actual storage scheme.

11.1.3 Typical Applications

• solution of partial differential equations (PDEs)

– the finite element method
– mechanical engineering, electrotechnics, physics, . . .
• graph theory
– nonzero at (i, j) means that node i is connected to node j
• natural language processing
– nonzero at (i, j) means that the document i contains the word j
• ...

11.1.4 Prerequisites

• numpy
• scipy
• matplotlib (optional)
• ipython (the enhancements come handy)

11.1. Introduction 370

 Scientific Python Lectures, Edition 2024.2rc0.dev0

11.1.5 Sparsity Structure Visualization

• spy() from matplotlib

• example plots:

11.1. Introduction 371

 Scientific Python Lectures, Edition 2024.2rc0.dev0

11.2 Storage Schemes

• seven sparse array types in scipy.sparse:

1. csr_array: Compressed Sparse Row format
2. csc_array: Compressed Sparse Column format
3. bsr_array: Block Sparse Row format
4. lil_array: List of Lists format
5. dok_array: Dictionary of Keys format
6. coo_array: COOrdinate format (aka IJV, triplet format)
7. dia_array: DIAgonal format
• each suitable for some tasks
• many employ sparsetools C++ module by Nathan Bell
• assume the following is imported:

>>> import numpy as np

>>> import scipy as sp
>>> import matplotlib.pyplot as plt

• warning for Numpy users:

– passing a sparse array object to NumPy functions that expect ndarray/matrix does not
work. Use sparse functions.
– the older csr_matrix classes use ‘*’ for matrix multiplication (dot product) and
‘A.multiply(B)’ for elementwise multiplication.
– the newer csr_array uses ‘@’ for dot product and ‘*’ for elementwise multiplication
– sparse arrays can be 1D or 2D, but not nD for n > 2 (unlike Numpy arrays).

11.2.1 Common Methods

• all scipy.sparse array classes are subclasses of sparray

– default implementation of arithmetic operations
∗ always converts to CSR
∗ subclasses override for efficiency
– shape, data type, set/get
– indices of nonzero values in the array
– format conversion, interaction with NumPy (toarray())
– ...
• attributes:
– mtx.T - transpose (same as mtx.transpose())
– mtx.real - real part of complex matrix
– mtx.imag - imaginary part of complex matrix
– mtx.size - the number of nonzeros (same as self.getnnz())
– mtx.shape - the number of rows and columns (tuple)

11.2. Storage Schemes 372

 Scientific Python Lectures, Edition 2024.2rc0.dev0

• data and indices usually stored in 1D NumPy arrays

11.2.2 Sparse Array Classes

Diagonal Format (DIA)

• very simple scheme

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

Examples

• create some DIA arrays:

>>> data = np.array([[1, 2, 3, 4]]).repeat(3, axis=0)

>>> data
array([[1, 2, 3, 4],
[1, 2, 3, 4],
[1, 2, 3, 4]])
>>> offsets = np.array([0, -1, 2])
>>> mtx = sp.sparse.dia_array((data, offsets), shape=(4, 4))
>>> mtx
<4x4 sparse array of type '<... 'numpy.int64'>'
with 9 stored elements (3 diagonals) in DIAgonal format>
>>> mtx.toarray()
array([[1, 0, 3, 0],
[1, 2, 0, 4],
(continues on next page)

11.2. Storage Schemes 373

 Scientific Python Lectures, Edition 2024.2rc0.dev0

(continued from previous page)

[0, 2, 3, 0],
[0, 0, 3, 4]])

>>> data = np.arange(12).reshape((3, 4)) + 1

>>> data
array([[ 1, 2, 3, 4],
[ 5, 6, 7, 8],
[ 9, 10, 11, 12]])
>>> mtx = sp.sparse.dia_array((data, offsets), shape=(4, 4))
>>> mtx.data
array([[ 1, 2, 3, 4],
[ 5, 6, 7, 8],
[ 9, 10, 11, 12]])
>>> mtx.offsets
array([ 0, -1, 2], dtype=int32)
>>> print(mtx)
(0, 0) 1
(1, 1) 2
(2, 2) 3
(3, 3) 4
(1, 0) 5
(2, 1) 6
(3, 2) 7
(0, 2) 11
(1, 3) 12
>>> mtx.toarray()
array([[ 1, 0, 11, 0],
[ 5, 2, 0, 12],
[ 0, 6, 3, 0],
[ 0, 0, 7, 4]])

• explanation with a scheme:

offset: row

2: 9
1: --10------
0: 1 . 11 .
-1: 5 2 . 12
-2: . 6 3 .
-3: . . 7 4
---------8

• matrix-vector multiplication

>>> vec = np.ones((4, ))

>>> vec
array([1., 1., 1., 1.])
>>> mtx @ vec
array([12., 19., 9., 11.])
>>> (mtx * vec).toarray()
array([[ 1., 0., 11., 0.],
[ 5., 2., 0., 12.],
[ 0., 6., 3., 0.],
[ 0., 0., 7., 4.]])

11.2. Storage Schemes 374

 Scientific Python Lectures, Edition 2024.2rc0.dev0

List of Lists Format (LIL)

• row-based linked list

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

Examples

• create an empty LIL array:

>>> mtx = sp.sparse.lil_array((4, 5))

• prepare random data:

>>> rng = np.random.default_rng(27446968)

>>> data = np.round(rng.random((2, 3)))
>>> data
array([[1., 0., 1.],
[0., 0., 1.]])

• assign the data using fancy indexing:

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

>>> mtx
<4x5 sparse array of type '<... 'numpy.float64'>'
with 3 stored elements in List of Lists format>
>>> print(mtx)
(0, 1) 1.0
(0, 3) 1.0
(1, 3) 1.0
>>> mtx.toarray()
array([[0., 1., 0., 1., 0.],
[0., 0., 0., 1., 0.],
[0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0.]])
>>> mtx.toarray()
array([[0., 1., 0., 1., 0.],
[0., 0., 0., 1., 0.],
[0., 0., 0., 0., 0.],
[0., 0., 0., 0., 0.]])

11.2. Storage Schemes 375

 Scientific Python Lectures, Edition 2024.2rc0.dev0

• more slicing and indexing:

>>> mtx = sp.sparse.lil_array([[0, 1, 2, 0], [3, 0, 1, 0], [1, 0, 0, 1]])

>>> mtx.toarray()
array([[0, 1, 2, 0],
[3, 0, 1, 0],
[1, 0, 0, 1]]...)
>>> print(mtx)
(0, 1) 1
(0, 2) 2
(1, 0) 3
(1, 2) 1
(2, 0) 1
(2, 3) 1
>>> mtx[:2, :]
<2x4 sparse array of type '<... 'numpy.int64'>'
with 4 stored elements in List of Lists format>
>>> mtx[:2, :].toarray()
array([[0, 1, 2, 0],
[3, 0, 1, 0]]...)
>>> mtx[1:2, [0,2]].toarray()
array([[3, 1]]...)
>>> mtx.toarray()
array([[0, 1, 2, 0],
[3, 0, 1, 0],
[1, 0, 0, 1]]...)

Dictionary of Keys Format (DOK)

• subclass of Python dict

– keys are (row, column) index tuples (no duplicate entries allowed)
– values are corresponding non-zero values
• efficient for constructing sparse arrays incrementally
• constructor accepts:
– dense array/matrix
– sparse array/matrix
– shape tuple (create empty array)
• efficient O(1) access to individual elements
• flexible slicing, changing sparsity structure is efficient
• can be efficiently converte