Scientific Software Engineering

i.e. What not to do and perhaps what you should do

---

Based on materials by Katy Huff, Rachel Slaybaugh, and Anthony Scopatz

Presented by Kyle Mandli

Some Preliminaries

Imports and Doc-Strings

import numpy

def unicorn_awesome(x, y):
    """ Calculate unicorn awesome based on parrots x and eagles y.
        
        Input
        -----
        - `x` (int) - Number of unicorn awesome
        - `y` (int) - Number of stupid eagles

        Output
        ------
        (float) Amount of unicorn awesome on your minivan
    """
    return numpy.exp(y - x)

Try-Except Blocks

try:
    D = {'stuff':1}
    print D['not stuff']
except KeyError as e: 
    print "Whoops!"
    print e

Onto Testing

What is testing?

Software testing is a process by which one or more expected behaviors and results from a piece of software are exercised and confirmed. Well chosen tests will confirm expected code behavior for the extreme boundaries of the input domains, output ranges, parametric combinations, and other behavioral edge cases.

Why test software?

Unless you write flawless, bug-free, perfectly accurate, fully precise, and predictable code every time, you must test your code in order to trust it enough to answer in the affirmative to at least a few of the following questions:

• Does your code work?
• Always?
• Does it do what you think it does? (Patriot Missile Failure)
• Does it continue to work after changes are made?
• Does it continue to work after system configurations or libraries are upgraded?
• Does it respond properly for a full range of input parameters?
• What about edge or corner cases?
• What's the limit on that input parameter?
• How will it affect your publications?

Verification

Verification is the process of asking, "Have we built the software correctly?" That is, is the code bug free, precise, accurate, and repeatable?

Validation

Validation is the process of asking, "Have we built the right software?" That is, is the code designed in such a way as to produce the answers we are interested in, data we want, etc.

Uncertainty Quantification

Uncertainty Quantification is the process of asking, "Given that our algorithm may not be deterministic, was our execution within acceptable error bounds?" This is particularly important for anything which uses random numbers, eg Monte Carlo methods.

Where are tests?

Say we have an averaging function:

def mean(numlist):
    total = sum(numlist)
    length = len(numlist)
    return total/length

Tests could be implemented as runtime exceptions in the function:

def mean(numlist):
    try:
        total = sum(numlist)
        length = len(numlist)
    except TypeError:
        raise TypeError("The number list was not a list of numbers.")
    except:
        print "There was a problem evaluating the number list."
    return total/length

Sometimes tests they are functions alongside the function definitions they are testing.

def mean(numlist):
    try:
        total = sum(numlist)
        length = len(numlist)
    except TypeError:
        raise TypeError("The number list was not a list of numbers.")
    except:
        print "There was a problem evaluating the number list."
    return total/length


def test_mean():
    assert mean([0, 0, 0, 0]) == 0
    assert mean([0, 200]) == 100
    assert mean([0, -200]) == -100
    assert mean([0]) == 0


def test_floating_mean():
    assert mean([1, 2]) == 1.5

Sometimes they are in an executable independent of the main executable.

def mean(numlist):
    try:
        total = sum(numlist)
        length = len(numlist)
    except TypeError:
        raise TypeError("The number list was not a list of numbers.")
    except:
        print "There was a problem evaluating the number list."
    return total/length

Where, in a different file exists a test module:

import mean

def test_mean():
    assert mean([0, 0, 0, 0]) == 0
    assert mean([0, 200]) == 100
    assert mean([0, -200]) == -100
    assert mean([0]) == 0


def test_floating_mean():
    assert mean([1, 2]) == 1.5

When should we test?

The three right answers are:

• ALWAYS!
• EARLY!
• OFTEN!

The longer answer is that testing either before or after your software is written will improve your code, but testing after your program is used for something important is too late.

If we have a robust set of tests, we can run them before adding something new and after adding something new. If the tests give the same results (as appropriate), we can have some assurance that we didn't wreak anything. The same idea applies to making changes in your system configuration, updating support codes, etc.

Another important feature of testing is that it helps you remember what all the parts of your code do. If you are working on a large project over three years and you end up with 200 classes, it may be hard to remember what the widget class does in detail. If you have a test that checks all of the widget's functionality, you can look at the test to remember what it's supposed to do.

Who should test?

In a collaborative coding environment, where many developers contribute to the same code base, developers should be responsible individually for testing the functions they create and collectively for testing the code as a whole.

Professionals often test their code, and take pride in test coverage, the percent of their functions that they feel confident are comprehensively tested.

How are tests written?

The type of tests that are written is determined by the testing framework you adopt. Don't worry, there are a lot of choices.

Types of Tests

Exceptions: Exceptions can be thought of as type of runtime test. They alert the user to exceptional behavior in the code. Often, exceptions are related to functions that depend on input that is unknown at compile time. Checks that occur within the code to handle exceptional behavior that results from this type of input are called Exceptions.

Unit Tests: Unit tests are a type of test which test the fundamental units of a program's functionality. Often, this is on the class or function level of detail. However what defines a code unit is not formally defined.

To test functions and classes, the interfaces (API) - rather than the implementation - should be tested. Treating the implementation as a black box, we can probe the expected behavior with boundary cases for the inputs.

System Tests: System level tests are intended to test the code as a whole. As opposed to unit tests, system tests ask for the behavior as a whole. This sort of testing involves comparison with other validated codes, analytical solutions, etc.

Regression Tests: A regression test ensures that new code does not change anything. If you change the default answer, for example, or add a new question, you'll need to make sure that missing entries are still found and fixed.

Integration Tests: Integration tests query the ability of the code to integrate well with the system configuration and third party libraries and modules. This type of test is essential for codes that depend on libraries which might be updated independently of your code or when your code might be used by a number of users who may have various versions of libraries.

Test Suites: Putting a series of unit tests into a collection of modules creates, a test suite. Typically the suite as a whole is executed (rather than each test individually) when verifying that the code base still functions after changes have been made.

Elements of a Test

Behavior: The behavior you want to test. For example, you might want to test the fun() function.

Expected Result: This might be a single number, a range of numbers, a new fully defined object, a system state, an exception, etc. When we run the fun() function, we expect to generate some fun. If we don't generate any fun, the fun() function should fail its test. Alternatively, if it does create some fun, the fun() function should pass this test. The the expected result should known a priori. For numerical functions, this is result is ideally analytically determined even if the function being tested isn't.

Assertions: Require that some conditional be true. If the conditional is false, the test fails.

Fixtures: Sometimes you have to do some legwork to create the objects that are necessary to run one or many tests. These objects are called fixtures as they are not really part of the test themselves but rather involve getting the computer into the appropriate state.

For example, since fun varies a lot between people, the fun() function is a method of the Person class. In order to check the fun function, then, we need to create an appropriate Person object on which to run fun().

Setup and teardown: Creating fixtures is often done in a call to a setup function. Deleting them and other cleanup is done in a teardown function.

The Big Picture: Putting all this together, the testing algorithm is often:

setup()
test()
teardown()

But, sometimes it's the case that your tests change the fixtures. If so, it's better for the setup() and teardown() functions to occur on either side of each test. In that case, the testing algorithm should be:

setup()
test1()
teardown()

setup()
test2()
teardown()

setup()
test3()
teardown()

---

Nose: A Python Testing Framework

The testing framework we'll discuss today is called nose. However, there are several other testing frameworks available in most language. Most notably there is JUnit in Java which can arguably attributed to inventing the testing framework.

Where do nose tests live?

Nose tests are files that begin with Test-, Test_, test-, or test_. Specifically, these satisfy the testMatch regular expression [Tt]est[-_]. (You can also teach nose to find tests by declaring them in the unittest.TestCase subclasses chat you create in your code. You can also create test functions which are not unittest.TestCase subclasses if they are named with the configured testMatch regular expression.)

Nose Test Syntax

To write a nose test, we make assertions.

assert should_be_true()
assert not should_not_be_true()

Additionally, nose itself defines number of assert functions which can be used to test more specific aspects of the code base.

from nose.tools import *

assert_equal(a, b)
assert_almost_equal(a, b)
assert_true(a)
assert_false(a)
assert_raises(exception, func, *args, **kwargs)
assert_is_instance(a, b)
# and many more!

Moreover, numpy offers similar testing functions for arrays:

from numpy.testing import *

assert_array_equal(a, b)
assert_array_almost_equal(a, b)
# etc.

Exercise: Writing tests for mean()

1. Write a mean() function in your group.
2. Write tests for your mean() function called test_mean.py.

Feel free to refer to the tests and function illustrated earlier.

Hint: Think about what form your input could take and what you should do to handle it. Also, think about the type of the elements in the list. What should be done if you pass a list of integers? What if you pass a list of strings?

Example:

nosetests test_mean.py

Test Driven Development

Test driven development (TDD) is a philosophy whereby the developer creates code by writing the tests first. That is to say you write the tests before writing the associated code!

This is an iterative process whereby you write a test then write the minimum amount code to make the test pass. If a new feature is needed, another test is written and the code is expanded to meet this new use case. This continues until the code does what is needed.

TDD operates on the YAGNI principle (You Ain't Gonna Need It). People who diligently follow TDD swear by its effectiveness. This development style was put forth most strongly by Kent Beck in 2002.

A TDD Example

Say you want to write a fib() function which generates values of the Fibonacci sequence of given indexes. You would - of course - start by writing the test, possibly testing a single value:

from nose.tools import assert_equal

from pisa import fib

def test_fib1():
    obs = fib(2)
    exp = 1
    assert_equal(obs, exp)

You would then go ahead and write the actual function:

def fib(n):
    # you snarky so-and-so
    return 1

And that is it right?! Well, not quite. This implementation fails for most other values. Adding tests we see that:

def test_fib1():
    obs = fib(2)
    exp = 1
    assert_equal(obs, exp)


def test_fib2():
    obs = fib(0)
    exp = 0
    assert_equal(obs, exp)

    obs = fib(1)
    exp = 1
    assert_equal(obs, exp)

This extra test now requires that we bother to implement at least the initial values:

def fib(n):
    # a little better
    if n == 0 or n == 1:
        return n
    return 1

However, this function still falls over for 2 < n. Time for more tests!

def test_fib1():
    obs = fib(2)
    exp = 1
    assert_equal(obs, exp)


def test_fib2():
    obs = fib(0)
    exp = 0
    assert_equal(obs, exp)

    obs = fib(1)
    exp = 1
    assert_equal(obs, exp)


def test_fib3():
    obs = fib(3)
    exp = 2
    assert_equal(obs, exp)

    obs = fib(6)
    exp = 8
    assert_equal(obs, exp)

At this point, we had better go ahead and try do the right thing...

def fib(n):
    # finally, some math
    if n == 0 or n == 1:
        return n
    else:
        return fib(n - 1) + fib(n - 2)

Here it becomes very tempting to take an extended coffee break or possibly a power lunch. But then you remember those pesky negative numbers and floats. Perhaps the right thing to do here is to just be undefined.

def test_fib1():
    obs = fib(2)
    exp = 1
    assert_equal(obs, exp)


def test_fib2():
    obs = fib(0)
    exp = 0
    assert_equal(obs, exp)

    obs = fib(1)
    exp = 1
    assert_equal(obs, exp)


def test_fib3():
    obs = fib(3)
    exp = 2
    assert_equal(obs, exp)

    obs = fib(6)
    exp = 8
    assert_equal(obs, exp)


def test_fib3():
    obs = fib(13.37)
    exp = NotImplemented
    assert_equal(obs, exp)

    obs = fib(-9000)
    exp = NotImplemented
    assert_equal(obs, exp)

This means that it is time to add the appropriate case to the function itself:

def fib(n):
    # sequence and you shall find
    if n < 0 or int(n) != n:
        return NotImplemented
    elif n == 0 or n == 1:
        return n
    else:
        return fib(n - 1) + fib(n - 2)

Quality Assurance Exercise

Can you think of other tests to make for the fibonacci function? I promise there are at least two.

Implement one new test in test_fib.py, run nosetests, and if it fails, implement a more robust function for that case.

And thus - finally - we have a robust function together with working tests!

Exercise

The Problem: In 2D or 3D, we have two points (p1 and p2) which define a line segment. Additionally there exists experimental data which can be anywhere in the domain. Find the data point which is closest to the line segment.

In the close_line.py file there are four different implementations which all solve this problem. You can read more about them here. However, there are no tests! Please write from scratch a test_close_line.py file which tests the closest_data_to_line() functions.

Hint: you can use one implementation function to test another. Below is some sample data to help you get started.

import numpy as np

p1 = np.array([0.0, 0.0])
p2 = np.array([1.0, 1.0])
data = np.array([[0.3, 0.6], [0.25, 0.5], [1.0, 0.75]])

Continuous Integration

**Continuous integration is the practice of checking that the tests in **a package all pass on a regular basis. Usually this implies that **the master branch of a project is regular built, installed and **tested as often as it makes sense to do so. Often this is also done **when a developer submits a patch or pull request.

Cheese Shop

Add Cheese to the Cheese shop

**Want to test the cheese as we integate into the shop **

Aron wants cheese but we have not checked it yet!

If we test before the Stilton patch is applied then we would have known and Aron would not be unhappy.

Here is an example of integration with github through the Travis-CI server. Note that the Travis-CI servers are checking this pull reqeust for use before it is applied to the master branch.