The Python interpreter

Remi Lehe

Lawrence Berkeley National Laboratory (LBNL)

US Particle Accelerator School (USPAS) Summer Session

Self-Consistent Simulations of Beam and Plasma Systems
S. M. Lund, J.-L. Vay, R. Lehe & D. Winklehner
Colorado State U, Ft. Collins, CO, 13-17 June, 2016
Python interpreter: Outline

1 Overview of the Python language

2 Python, numpy and matplotlib

3 Reusing code: functions, modules, classes

4 Faster computation: Forthon

Overview Scientific Python Reusing code Forthon

Overview of the Python programming language

Interpreted language (i.e. not compiled)

→ Interactive, but not optimal for computational speed
Readable and non-verbose
No need to declare variables
Indentation is enforced
Free and open-source
+ Large community of open-souce packages
Well adapted for scientific and data analysis applications
Many excellent packages, esp. numerical computation (numpy),
scientific applications (scipy), plotting (matplotlib), data
analysis (pandas, scikit-learn)

3
Overview Scientific Python Reusing code Forthon

Interfaces to the Python language

Scripting Interactive shell
Code written in a file, with a Obtained by typing python or
text editor (gedit, vi, emacs) (better) ipython
Execution via command line Commands are typed in and
(python + filename) executed one by one

Convenient for exploratory work,

Convenient for long-term code debugging, rapid feedback, etc... 4
Overview Scientific Python Reusing code Forthon

Interfaces to the Python language

IPython (a.k.a Jupyter) notebook

Notebook interface, similar to
Mathematica.
Intermediate between scripting
and interactive shell, through
reusable cells
Obtained by typing jupyter
notebook, opens in your web
browser

Convenient for exploratory work,

scientific analysis and reports

5
Overview Scientific Python Reusing code Forthon

Overview of the Python language

This lecture
Reminder of the main points of the Scipy lecture notes through an
example problem.

Example problem: Euler’s method

Use Euler’s method to numerically integrate, between t = 0 and 10:
d x(t)
= x(t) cos(t) with x(0) = 1
dt
Compare it with the exact solution: x(t) = esin(t)

Reminder: In this case, Euler’s method gives:

ti = i∆t
xi = xi−1 + ∆t × xi−1 cos(ti−1 )

x1 x2 x3 ...
x0 = 1 t
∆t

6
Python interpreter: Outline

1 Overview of the Python language

2 Python, numpy and matplotlib

3 Reusing code: functions, modules, classes

4 Faster computation: Forthon

Overview Scientific Python Reusing code Forthon

Example problem: Structure of the code

Storage in memory:

t t0 t1 t2 t3 t4 t5 t6 t7 t8 t9

x x0 x1 x2 x3 x4 x5 x6 x7 x8 x9

t
∆t

For loop:
Repeatedly apply: xi = xi−1 + ∆t × xi−1 cos(ti−1 )

8
Overview Scientific Python Reusing code Forthon

Numpy arrays
Numpy arrays
Provide efficient memory storage and computation,
for large number of elements of the same type.

Standard import: import numpy as np

Creation of numpy arrays:
np.arange, np.zeros, np.random.rand, np.empty, etc...
(In ipython, use e.g. np.arange? to read the documentation)
Individual elements are accessed with square brackets:
x[i] (1D array), y[i,j,k] (3D array)
For an array with N elements, the indices start at 0
(included) and end at N-1 (included)
Subsets of the array are accessed using slicing syntax:
x[ start index : end index : step ] ; in particular:
x[ start index : end index ] : slicing with step 1 by default
x[ : end index ] : slicing with start index 0 by default
x[ start index : -1 ] : slicing up to the last-but-one element 9
Overview Scientific Python Reusing code Forthon

For loops
For loop
Repeatedly perform a given operation
(e.g. apply the same operation to every element of a numpy array)

Syntax:
for i in range( start index , end index , step ):
Perform some operation that depends on i

Indentation and the use of column (:) are key.

The range function can be used with 1, 2 or 3 arguments:
range(N): loop from index 0 to index N-1 (included)
range(i,N): loop from index i to index N-1 (included)
range(i,N,k): loop from index i to index N-1 with step k
In the above, range can also be replaced by a list or any iterable.

10
Overview Scientific Python Reusing code Forthon

Numpy and for loops: task

Task 1
In a text editor, write a python script (named euler.py) which:
Sets the number of integration steps to N = 200,
and the timestep to dt = 10./N
Initializes the array t (with N elements) using np.arange so that

ti = i∆t

Initializes the array x (with N elements) using np.empty and

setting the initial point x[0] to 1.
Loops through the array x and applies Euler’s method:
(Here, the loop should start at i = 1, not i = 0)

xi = xi−1 + ∆t × xi−1 cos(ti−1 )

Run the script (python euler.py), to check that there is no error. 11

Overview Scientific Python Reusing code Forthon

Comparison with the exact solution

t t0 t1 t2 t3 t4 t5 t6 t7 t8 t9

x x0 x1 x2 x3 x4 x5 x6 x7 x8 x9

x exact esin(t0 ) esin(t1 ) esin(t2 ) esin(t3 ) esin(t4 ) esin(t5 ) esin(t6 ) esin(t7 ) esin(t8 ) esin(t9 )

We wish to compare the two results by:

Calculating the RMS error:
v
u N −1
u1 X
RM S = t (xi − xexact,i )2
N i=0

Plotting x and xexact versus t.

12
Overview Scientific Python Reusing code Forthon

Numpy arrays: element-wise operations

Element-wise operation
Operation that is repeated for each element of an array
and does not depend on previous/next elements.

e.g. xexact,i = esin(ti ) ∀i ∈ [0, N − 1]

Could be done with a for loop:

for i in range(N):
x exact[i] = np.exp( np.sin( t[i] ) )
But is computationally faster with numpy vector syntax:
x exact = np.exp( np.sin( t ) )
Numpy vector syntax also works for the element-wise operations:
+, -, *, /, ** (power), np.sqrt (square-root), np.log, etc...

13
Overview Scientific Python Reusing code Forthon

Numpy arrays: reduction operations

Reduction operation
Operation that extracts a single scalar from a full array

e.g. N
X −1
S= yi
i=0

Again, could be done with a for loop:

S = 0
for i in range(N):
S = S + y[i]
But is computationally faster with numpy reduction methods
S = np.sum( y )
Other reduction operations:
np.product, np.max, np.mean, etc... (for real or integer arrays)
np.any, np.all, etc... (for boolean arrays)
14
Overview Scientific Python Reusing code Forthon

Plotting package: matplotlib

Other Python plotting packages: pygist, bokeh, seaborn, bqplot, ...

Pros of matplotlib Cons of matplotlib

Publication-quality figures Slow
Extremely versatile and customizable Sometimes verbose
Standard plotting package in the Limited interactivity
Python community

Standard import: import matplotlib.pyplot as plt

Basic plotting commands:
plt.plot( t, x ) (plots 1darray x as a function of 1darray t)
Show the image to the screen:
plt.show() (unneeded when using ipython --matplotlib)
Save the figure to a file:
plt.savefig( file name )
15
Overview Scientific Python Reusing code Forthon

Numpy and matplotlib: task

Task 2
In a text editor, add the following features to euler.py:
Create the array x exact so that xexact,i = esin(ti )
Calculate the RMS error, without using any for loop:
v
u N −1
u1 X
RM S = t (xi − xexact,i )2
N i=0

Use the print statement, to show the value of the RMS error
Plot x and xexact as a function of t on the same figure,
and show it to the screen. (Use plot(t, x exact, ’--’) to
show the exact solution with dashed lines.)
Run the script (python euler.py), to check that it works.

16
Python interpreter: Outline

1 Overview of the Python language

2 Python, numpy and matplotlib

3 Reusing code: functions, modules, classes

4 Faster computation: Forthon

Overview Scientific Python Reusing code Forthon

Reusing code for the example problem

Example problem
Compare the results of Euler’s method for different values of N
(and thus of dt) on the same plot.

→ Not possible with the code from task 2

(unless we copy and paste a lot of code)

We need to make the code more abstract and reusable:

Define functions that depend on N and initialize the arrays,
perform Euler integration, and plot the results.
Place these functions inside a module so that they can be
imported and used elsewhere.

18
Overview Scientific Python Reusing code Forthon

Functions
Example for function definition Key syntax: def, () and :,
the body is indented
The “docstring” is optional.
Users can see it in ipython
with geometric sum? or
help(geometric sum)
Here, b has a default value,
which is used when only 2
arguments are given
Functions can also return
Example for function call several objects
(e.g. return(x, a, b))
or nothing
(no return statement)
Similarly, functions can be
defined with no arguments 19
Overview Scientific Python Reusing code Forthon

Modules

Module Example module

Defines variables to be imported In file geometric.py:
by other Python sessions.

Any Python script can be

treated as a module.
numpy is a set of modules.
The section
if name == ’ main ’:
is executed if the script is run
(e.g. python geometric.py) Example import and use
but not when it is imported In e.g. ipython:
(import geometric as gm) import geometric as gm
S = gm.geometric sum(8, 2)

20
Overview Scientific Python Reusing code Forthon

Importing modules
Different import styles:
import geometric
→ S = geometric.geometric sum(8,2)
import geometric as gm
→ S = gm.geometric sum(8,2)
from geometric import geometric sum
or from geometric import * (imports all variables)
→ S = geometric sum(8,2)

The source file of the module needs to be:

in the same directory
or in the default Python path
(case of installed packages like numpy, matplotlib or even warp)

21
Overview Scientific Python Reusing code Forthon

Functions and modules: task

Task 3
Reorganize the script euler.py so as to make it a reusable module:
Start with the import statements (numpy and matplotlib)
Write a function with signature initialize arrays(N, T=10.)
which sets dt = T/N, initializes t and x, and returns t, x, dt
Write a function euler integration( t, x, dt, N ), which
fills the array x (this function does not return anything)
Write a function evaluate result(t, x, N), which computes
the exact result, prints the RMS error, and plots the arrays
Finally, create a section if name == ’ main ’:, in which
you set N = 200, and call the 3 functions successively
Type python euler.py to check that the final section runs.

22
Overview Scientific Python Reusing code Forthon

Functions and modules: task

Task 4
Use the module that you wrote, inside ipython
In the shell, type ipython --matplotlib
Then, inside ipython, type from euler import *
Then set N1 = 100, N2 = 200 and create the corresponding
variables t1,x1,dt1 and t2,x2,dt2 with initialize array.
Then call euler integration and evaluate result on each set
of arrays and values. Compare the results.
(NB: Do not hesitate to use tab completion in ipython)

Although the code works, note that it is tedious to:

create 4 different variables with a suffix 1 or 2
pass these variables as arguments to the different functions
This is solved by object-oriented programming and classes.
23
Overview Scientific Python Reusing code Forthon

Classes: initialization and attributes

Example of class definition Classes are “containers”:
Variables are encapsulated
together as attributes of an
instance of the class.
Creation of an instance
(e.g. EulerSolver(100))
executes the code in init .
Accessing attributes
replace self by the name of
Example of use the instance.
Predefined syntax:
Use the keywords class,
(object): and init
Note that init takes self
as first argument when
defined, but this is skipped
when creating an instance. 24
Overview Scientific Python Reusing code Forthon

Classes: methods
Example of class definition Methods are functions
which can access the
attributes of a class.
→ The attributes do not
need to be passed as
arguments.
Syntax for definition
Pass self as first
arguments, then use
self. to access attributes
Syntax for calling
Prefix with name of the
instance, then skip self
in arguments
Example of use

25
Overview Scientific Python Reusing code Forthon

Classes: task
Task 5
Rewrite euler.py so as to define a class EulerSolver
Replace the function initialize arrays by a corresponding
method init (self, N, T=10.) This method should define
N, x, t, dt as attributes.
Replace the functions euler integration and evaluate result
by methods with the same name respectively. These methods
should take no argument (besides self), but should use the
attributes through the self. syntax.
In the section if name == ’ main ’:, type the code:
solver = EulerSolver( 200 )
solver.euler integration()
solver.evaluate result()
Execute the file (python euler.py) to check that it works.
Then in ipython, compare again N=100 and N=200.
26
Python interpreter: Outline

1 Overview of the Python language

2 Python, numpy and matplotlib

3 Reusing code: functions, modules, classes

4 Faster computation: Forthon

Overview Scientific Python Reusing code Forthon

Faster computation
Problem
Large for loops are slow in Python.

Example:

Solution
If the operation is of type element-wise or reduction:
Use numpy syntax
Otherwise, rewrite the for loop in a compiled language
(e.g. Fortran, C) and link it to the rest of the Python code

→ High-level control with Python (modularity, interactivity)

→ Low-level number-crunching with e.g. Fortran or C (efficiency)
28
Overview Scientific Python Reusing code Forthon

Faster computation: Forthon

Forthon
Generates links between Fortran and Python
Open-source, created by D. P. Grote (LLNL)
https://github.com/dpgrote/Forthon
Heavily used in Warp for low-level number crunching

On the user side:

Write Fortran subroutines and modules in a .F file
Write a .v file to tell which variables to link to Python
Compile with Forthon → produces a Python module
Import the module in Python and use the linked variables

NB: Other similar solutions exist: f2py (links Fortran code), Cython
(generates and links C code), Numba (compiles Python code), etc...
29
Overview Scientific Python Reusing code Forthon

Faster computation: task

Task 6
Download and decompress the code from
http://github.com/RemiLehe/uspas exercise/raw/master/Forthon task.tgz
The files acc euler.F and acc euler.v are the files needed by
Forthon, while euler.py is the code from task 5.
The Fortran file acc euler.F contains an error in the line that
starts with x(i) = . Spot it and correct it.
Compile the code with Forthon by typing make in the shell.
A new file acc eulerpy.so should be created.
At the beginning of the file euler.py, add
from acc eulerpy import forthon integration then create a
new method acc euler integration(self), which calls
forthon integration (see acc euler.F for its signature).
In ipython, create an instance with N=10**6, and compare the
runtime of euler integration and acc euler integration
30
Overview Scientific Python Reusing code Forthon

References

Scipy lecture notes:

http://www.scipy-lectures.org/ (G. Varoquaux et al., 2015)

Python tutorial:
https://docs.python.org/3/tutorial/ (Python Software
foundation, 2016)

Forthon:
https://github.com/dpgrote/Forthon (D. Grote et al., 2016)

31