Extracted from:

Python Companion to Data Science

Collect → Organize → Explore → Predict → Value

This PDF file contains pages extracted from Python Companion to Data Science,
published by the Pragmatic Bookshelf. For more information or to purchase a
paperback or PDF copy, please visit http://www.pragprog.com.
Note: This extract contains some colored text (particularly in code listing). This
is available only in online versions of the books. The printed versions are black
and white. Pagination might vary between the online and printed versions; the
content is otherwise identical.
Copyright © 2016 The Pragmatic Programmers, LLC.

All rights reserved.

No part of this publication may be reproduced, stored in a retrieval system, or transmitted,

in any form, or by any means, electronic, mechanical, photocopying, recording, or otherwise,
without the prior consent of the publisher.

The Pragmatic Bookshelf

Raleigh, North Carolina
Python Companion to Data Science
Collect → Organize → Explore → Predict → Value

Dmitry Zinoviev

The Pragmatic Bookshelf

Raleigh, North Carolina
Many of the designations used by manufacturers and sellers to distinguish their products
are claimed as trademarks. Where those designations appear in this book, and The Pragmatic
Programmers, LLC was aware of a trademark claim, the designations have been printed in
initial capital letters or in all capitals. The Pragmatic Starter Kit, The Pragmatic Programmer,
Pragmatic Programming, Pragmatic Bookshelf, PragProg and the linking g device are trade-
marks of The Pragmatic Programmers, LLC.
Every precaution was taken in the preparation of this book. However, the publisher assumes
no responsibility for errors or omissions, or for damages that may result from the use of
information (including program listings) contained herein.
Our Pragmatic books, screencasts, and audio books can help you and your team create
better software and have more fun. Visit us at https://pragprog.com.

The team that produced this book includes:

Katharine Dvorak (editor)
Potomac Indexing, LLC (index)
Nicole Abramowitz (copyedit)
Gilson Graphics (layout)
Janet Furlow (producer)

For sales, volume licensing, and support, please contact [email protected].

For international rights, please contact [email protected].

Copyright © 2016 The Pragmatic Programmers, LLC.

All rights reserved.

No part of this publication may be reproduced, stored in a retrieval system, or transmitted,

in any form, or by any means, electronic, mechanical, photocopying, recording, or otherwise,
without the prior consent of the publisher.

Printed in the United States of America.

ISBN-13: 978-1-68050-184-1
Encoded using the finest acid-free high-entropy binary digits.
Book version: P1.0—August 2016
 To my beautiful and most intelligent wife
Anna; to our children: graceful ballerina
Eugenia and romantic gamer Roman; and to
my first data science class of summer 2015.
It is a capital mistake to theorize before one has data.
➤ Sir Arthur Conan Doyle, British writer

CHAPTER 5

Working with Tabular Numeric Data

Often raw data comes from all kinds of text documents. Quite often the text
actually represents numbers. Excel and CSV spreadsheets and especially
database tables may contain millions or billions of numerical records. Core
Python is an excellent text-processing tool, but it sometimes fails to deliver
adequate numeric performance. That’s where numpy comes to the rescue.

NumPy—Numeric Python (imported as numpy)—is an interface to a family of

efficient and parallelizable functions that implement high-performance
numerical operations. The module numpy provides a new Python data structure
—array—and a toolbox of array-specific functions, as well as support for
random numbers, data aggregation, linear algebra, Fourier transform, and
other goodies.

Bridge to Terabytia
If your program needs access to huge amounts of numerical data
—terabytes and more—you can’t avoid using the module h5py.1
The module is a portal to the HDF5 binary data format that works
with a lot of third-party software, such as IDL and MATLAB. h5py
imitates familiar numpy and Python mechanisms, such as arrays
and dictionaries. Once you know how to use numpy, you can go
straight to h5py—but not in this book.

In this chapter, you’ll learn how to create numpy arrays of different shapes and
from different sources, reshape and slice arrays, add array indexes, and apply
arithmetic, logic, and aggregation functions to some or all array elements.

1. www.h5py.org

• Click HERE to purchase this book now. discuss

 •8

Unit 21

Creating Arrays

numpy arrays are more compact and faster than native Python lists, especially
in multidimensional cases. However, unlike lists, arrays are homogeneous:
you cannot mix and match array items that belong to different data types.

There are several ways to create a numpy array. The function array() creates an
array from array-like data. The data can be a list, a tuple, or another array.
numpy infers the type of the array elements from the data, unless you explicitly
pass the dtype parameter. numpy supports close to twenty data types, such as
bool_, int64, uint64, float64, and <U32 (for Unicode strings).

When numpy creates an array, it doesn’t copy the data from the source to the
new array, but it links to it for efficiency reasons. This means that, by default,
a numpy array is a view of its underlying data, not a copy of it. If the underlying
data object changes, the array data changes, too. If this behavior is undesirable
(which it always is, unless the amount of data to copy is prohibitively large),
pass copy=True to the constructor.

Lists Are Arrays, and Arrays Are Arrays, Too

Contrary to their name, Python “lists” are actually implemented
as arrays, not as lists. No storage is reserved for forward pointers,
because no forward pointers are used. Large Python lists require
only about 13% more storage than “real” numpy arrays. However,
Python executes some simple built-in operations, such as sum(),
five to ten times faster on lists than on arrays. Ask yourself if you
really need any numpy-specific features before you start a numpy
project!

Let’s create our first array—a silly array of the first ten positive integer numbers:
import numpy as np
numbers = np.array(range(1, 11), copy=True)

➾ array([1, 2, 3, 4, 5, 6, 7, 8, 9, 10])

• Click HERE to purchase this book now. discuss

 Creating Arrays •9

The functions ones(), zeros(), and empty() construct arrays of all ones, all zeros,
and all uninitialized entries, respectively. They then take a required shape
parameter—a list or tuple of array dimensions.
ones = np.ones([2, 4], dtype=np.float64)

➾ array([[ 1., 1., 1., 1.],

➾ [ 1., 1., 1., 1.]])

zeros = np.zeros([2, 4], dtype=np.float64)

➾ array([[ 0., 0., 0., 0.],

➾ [ 0., 0., 0., 0.]])

empty = np.empty([2, 4], dtype=np.float64)

# The array content is not always zeros!

➾ array([[ 0., 0., 0., 0.],

➾ [ 0., 0., 0., 0.]])

numpy stores the number of dimensions, the shape, and the data type of an
array in the attributes ndim, shape, and dtype.
ones.shape # The original shape, unless reshaped

➾ (2, 4)

numbers.ndim # Same as len(numbers.shape)

➾ 1

zeros.dtype

➾ dtype('float64')

The function eye(N, M=None, k=0, dtype=np.float) constructs an N×M eye-dentity

matrix with ones on the kth diagonal and zeros elsewhere. Positive k denotes
the diagonals above the main diagonal. When M is None (default), M=N.
eye = np.eye(3, k=1)

➾ array([[ 0., 1., 0.],

➾ [ 0., 0., 1.],
➾ [ 0., 0., 0.]])

When you need to multiply several matrices, use an identity matrix as the
initial value of the accumulator in the multiplication chain.

• Click HERE to purchase this book now. discuss

 • 10

In addition to the good old built-in range(), numpy has its own, more efficient
way of generating arrays of regularly spaced numbers: the function arange([start,]
stop[, step,], dtype=None).

np_numbers = np.arange(2, 5, 0.25)

➾ array([ 2. , 2.25, 2.5 , 2.75, 3. , 3.25, 3.5 , 3.75, 4. ,

➾ 4.25, 4.5 , 4.75])

Just like with range(), the value of stop can be smaller than start—but then step
must be negative, and the order of numbers in the array is decreasing.

numpy records the type of items at the time of array creation, but the type is
not immutable: you can change it later by calling the astype(dtype, casting="unsafe",
copy=True) function. In the case of type narrowing (converting to a more specific
data type), some information may be lost. This is true about any narrowing,
not just in numpy.
np_inumbers = np_numbers.astype(np.int)

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

Most numpy operations (such as transposing, which is discussed in the next

unit on page 11) return a view, not a copy of the original array. To preserve
original data, the function copy() creates a copy of an existing array. Any
changes to the original array don’t affect the copy, but if your array has one
billion items, think twice before copying it.
np_inumbers_copy = np_inumbers.copy()

Let’s now proceed to more advanced operations.

• Click HERE to purchase this book now. discuss

 Transposing and Reshaping • 11

Unit 22

Transposing and Reshaping

Unlike the great monuments of the past, numpy arrays are not carved in stone.
They can easily change their shape and orientation without being called
opportunists. Let’s build a one-dimensional array of some S&P stock symbols
and twist it in every possible way:
# Some S&P stock symbols
sap = np.array(["MMM", "ABT", "ABBV", "ACN", "ACE", "ATVI", "ADBE", "ADT"])

➾ array('MMM', 'ABT', 'ABBV', 'ACN', 'ACE', 'ATVI', 'ADBE', 'ADT'],

➾ dtype='<U4')

The function reshape(d0, d1, ...) changes the shape of an existing array. The
arguments define the new dimensions. The total number of items in the old
and new shapes must be equal: the conservation law still holds in numpyland!
sap2d = sap.reshape(2, 4)

➾ array([['MMM', 'ABT', 'ABBV', 'ACN'],

➾ ['ACE', 'ATVI', 'ADBE', 'ADT']],
➾ dtype='<U4')

sap3d = sap.reshape(2, 2, 2)

➾ array([[['MMM', 'ABT'],
➾ ['ABBV', 'ACN']],
➾
➾ [['ACE', 'ATVI'],
➾ ['ADBE', 'ADT']]],
➾ dtype='<U4')

To transpose an array, you don’t even need to call a function: the value of
the attribute T is the transposed view of the array (for a one-dimensional
array, data.T==data; for a two-dimensional array, the rows and the columns
are swapped).
sap2d.T

➾ array([['MMM', 'ACE'],
➾ ['ABT', 'ATVI'],
➾ ['ABBV', 'ADBE'],
➾ ['ACN', 'ADT']],
➾ dtype='<U4')

• Click HERE to purchase this book now. discuss

 • 12

Essentially, the attribute T shows us the matrix through a relabeling cross-

hair: axis number 0 (“vertical”) becomes axis number 1 (“horizontal”), and
the other way around. The function swapaxes() is a more general version of the
T. It transposes a multidimensional array by swapping any two axes that you
pass as the parameters. Naturally, passing the axes 0 and 1 for a two-
dimensional array simply transposes the array, just as before.
sap3d.swapaxes(1, 2)

➾ array([[['MMM', 'ABBV'],
➾ ['ABT', 'ACN']],
➾
➾ [['ACE', 'ADBE'],
➾ ['ATVI', 'ADT']]],
➾ dtype='<U4')

The function transpose() is even more general than swapaxes() (despite its name
implying similarity to the T attribute). transpose() permutes some or all axes of
a multidimensional array according to its parameter, which must be a tuple.
In the following example, the first axis remains “vertical,” but the other two
axes are swapped.
sap3d.transpose((0, 2, 1))

➾ array([[['MMM', 'ABBV'],
➾ ['ABT', 'ACN']],
➾
➾ [['ACE', 'ADBE'],
➾ ['ATVI', 'ADT']]],
➾ dtype='<U4')

Incidentally, the result is the same as for swapaxes(1, 2)!

• Click HERE to purchase this book now. discuss

 Indexing and Slicing • 13

Unit 23

Indexing and Slicing

numpy arrays support the same indexing [i] and slicing [i:j] operations as Python
lists. In addition, they implement Boolean indexing: you can use an array of
Boolean values as an index, and the result of the selection is the array of
items of the original array for which the Boolean index is True. Often the
Boolean array is in turn a result of broadcasting. You can use Boolean
indexing on the right-hand side (for selection) and on the left-hand side (for
assignment).

Suppose your data sponsor told you that all data in the data set dirty is strictly
non-negative. This means that any negative value is not a true value but an
error, and you must replace it with something that makes more sense—say,
with a zero. This operation is called data cleaning. To clean the dirty data, locate
the offending values and substitute them with reasonable alternatives.
dirty = np.array([9, 4, 1, -0.01, -0.02, -0.001])
whos_dirty = dirty < 0 # Boolean array, to be used as Boolean index

➾ array([False, False, False, True, True, True], dtype=bool)

dirty[whos_dirty] = 0 # Change all negative values to 0

➾ array([9, 4, 1, 0, 0, 0])

You can combine several Boolean expressions with the operators | (logical or),
& (logical and), and - (logical not). Which of the items in the following list are
between -0.5 and 0.5? Ask numpy!
linear = np.arange(-1, 1.1, 0.2)
(linear <= 0.5) & (linear >= -0.5)

➾ array([False, False, False, True, True, True, True, True, False,

➾ False, False], dtype=bool)

• Click HERE to purchase this book now. discuss

 • 14

Boolean and “Boolean”

Relational operators (such as < and ==) have lower precedence
than the bit-wise operators &, |, and ! that represent “Boolean”
operators on numpy arrays. This is very confusing, because “normal”
Python Boolean operators—or, and, and not—have lower precedence
than the relational operators. You must enclose array comparisons
in parentheses to ensure that numpy evaluates them first.

Another cool feature of numpy arrays is “smart” indexing and “smart” slicing,
whereby an index is not a scalar but an array or list of indexes. The result of
the selection is the array of items that are referenced in the index. Let’s select
the second, the third, and the last stock symbols from our S&P list. (That’s
“smart” indexing.)
sap[[1, 2, -1]]

➾ array(['ABT', 'BBV', 'ADT'],

➾ dtype='<U4')

Why not extract all rows in the middle column from the reshaped array?
(That’s “smart” slicing.) In fact, you can do it two ways:
sap2d[:, [1]]

➾ array([['ABT'],
➾ ['ATVI']],
➾ dtype='<U4')

sap2d[:, 1]

➾ array(['ABT', 'ATVI'],
➾ dtype='<U4')

Python is famous for giving us many similar problem-solving tools and not
forgiving us for choosing a wrong one. Compare the two selections shown
earlier. The first selection is a two-dimensional matrix; the second one is a
one-dimensional array. Depending on what you wanted to extract, one of
them is wrong. Oops. Make sure you check that you’ve gotten what you
wanted.

• Click HERE to purchase this book now. discuss