Statistics Machine Learning Python Draft

Statistics and Machine Learning in
Python
Release 0.3 beta
Edouard Duchesnay, Tommy Löfstedt, Feki Younes
Oct 29, 2020

CONTENTS
1 Introduction 1
1.1 Python ecosystem for data-science . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Introduction to Machine Learning . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3 Data analysis methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2 Python language 9
2.1 Import libraries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2 Basic operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.3 Data types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.4 Execution control statements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.5 Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.6 List comprehensions, iterators, etc. . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.7 Regular expression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.8 System programming . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.9 Scripts and argument parsing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.10 Networking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.11 Modules and packages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.12 Object Oriented Programming (OOP) . . . . . . . . . . . . . . . . . . . . . . . . 32
2.13 Style guide for Python programming . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.14 Documenting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.15 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3 Scientific Python 37
3.1 Numpy: arrays and matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.2 Pandas: data manipulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.3 Matplotlib: data visualization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
4 Statistics 75
4.1 Univariate statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
4.2 Lab 1: Brain volumes study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
4.3 Multivariate statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
4.4 Time Series in python . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
5 Machine Learning 159

5.1 Dimension reduction and feature extraction . . . . . . . . . . . . . . . . . . . . . 159
5.2 Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
5.3 Linear methods for regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
5.4 Linear classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
5.5 Non linear learning algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
i
5.6 Resampling Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219
5.7 Ensemble learning: bagging, boosting and stacking . . . . . . . . . . . . . . . . . 232
5.8 Gradient descent . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247
6 Deep Learning 257

6.1 Backpropagation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
6.2 Multilayer Perceptron (MLP) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271
6.3 Convolutional neural network . . . . . . . . . . . . . . . . . . . . . . . . . . . . 291
6.4 Transfer Learning Tutorial . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 319
7 Indices and tables 329
ii
CHAPTER
ONE
INTRODUCTION
1.1 Python ecosystem for data-science
1.1.1 Python language
• Interpreted
• Garbage collector (do not prevent from memory leak)
• Dynamically-typed language (Java is statically typed)
1.1.2 Anaconda
Anaconda is a python distribution that ships most of python tools and libraries
Installation
1. Download anaconda (Python 3.x) http://continuum.io/downloads
2. Install it, on Linux
bash Anaconda3-2.4.1-Linux-x86_64.sh
3. Add anaconda path in your PATH variable in your .bashrc file:
export PATH="${HOME}/anaconda3/bin:$PATH"
Managing with ``conda``

Update conda package and environment manager to current version
conda update conda
Install additional packages. Those commands install qt back-end (Fix a temporary issue to run
spyder)
conda install pyqt

conda install PyOpenGL
conda update --all
Install seaborn for graphics
1
Statistics and Machine Learning in Python, Release 0.3 beta
conda install seaborn

# install a specific version from anaconda chanel
conda install -c anaconda pyqt=4.11.4
List installed packages
conda list
Search available packages
conda search pyqt

conda search scikit-learn
Environments
• A conda environment is a directory that contains a specific collection of conda packages
that you have installed.
• Control packages environment for a specific purpose: collaborating with someone else,
delivering an application to your client,
• Switch between environments
List of all environments
:: conda info –envs
1. Create new environment
2. Activate
3. Install new package
conda create --name test

# Or
conda env create -f environment.yml
source activate test
conda info --envs
conda list
conda search -f numpy
conda install numpy
Miniconda
Anaconda without the collection of (>700) packages. With Miniconda you download only the
packages you want with the conda command: conda install PACKAGENAME
1. Download anaconda (Python 3.x) https://conda.io/miniconda.html
2. Install it, on Linux
bash Miniconda3-latest-Linux-x86_64.sh
3. Add anaconda path in your PATH variable in your .bashrc file:
export PATH=${HOME}/miniconda3/bin:$PATH
4. Install required packages
2 Chapter 1. Introduction
conda install -y scipy

conda install -y pandas
conda install -y matplotlib
conda install -y statsmodels
conda install -y scikit-learn
conda install -y sqlite
conda install -y spyder
conda install -y jupyter
1.1.3 Commands
python: python interpreter. On the dos/unix command line execute wholes file:
python file.py
Interactive mode:
python
Quite with CTL-D

ipython: advanced interactive python interpreter:
ipython
Quite with CTL-D

pip alternative for packages management (update -U in user directory --user):
pip install -U --user seaborn
For neuroimaging:
pip install -U --user nibabel

pip install -U --user nilearn
spyder: IDE (integrated development environment):

• Syntax highlighting.
• Code introspection for code completion (use TAB).
• Support for multiple Python consoles (including IPython).
• Explore and edit variables from a GUI.
• Debugging.
• Navigate in code (go to function definition) CTL.
3 or 4 panels:
text editor help/variable explorer

ipython interpreter
Shortcuts: - F9 run line/selection
1.1. Python ecosystem for data-science 3

1.1.4 Libraries
scipy.org: https://www.scipy.org/docs.html
Numpy: Basic numerical operation. Matrix operation plus some basic solvers.:
import numpy as np
X = np.array([[1, 2], [3, 4]])
#v = np.array([1, 2]).reshape((2, 1))
v = np.array([1, 2])
np.dot(X, v) # no broadcasting
X * v # broadcasting
np.dot(v, X)
X - X.mean(axis=0)
Scipy: general scientific libraries with advanced solver:
import scipy
import scipy.linalg
scipy.linalg.svd(X, full_matrices=False)
Matplotlib: visualization:
import numpy as np
import matplotlib.pyplot as plt
#%matplotlib qt
x = np.linspace(0, 10, 50)
sinus = np.sin(x)
plt.plot(x, sinus)
plt.show()
Pandas: Manipulation of structured data (tables). input/output excel files, etc.

Statsmodel: Advanced statistics
Scikit-learn: Machine learning
li- Arrays data, Structured Solvers: Solvers: Stats: Stats: Machine

brary Num. comp, data, I/O basic advanced basic ad- learning
I/O vanced
Numpy X X
Scipy X X X
Pan- X
das
Stat- X X
mod-
els
Scikit- X
learn
1.2 Introduction to Machine Learning
1.2.1 Machine learning within data science
Machine learning covers two main types of data analysis:

1. Exploratory analysis: Unsupervised learning. Discover the structure within the data.
E.g.: Experience (in years in a company) and salary are correlated.
2. Predictive analysis: Supervised learning. This is sometimes described as “learn from
the past to predict the future”. Scenario: a company wants to detect potential future
clients among a base of prospects. Retrospective data analysis: we go through the data
constituted of previous prospected companies, with their characteristics (size, domain,
localization, etc. . . ). Some of these companies became clients, others did not. The ques-
tion is, can we possibly predict which of the new companies are more likely to become
clients, based on their characteristics based on previous observations? In this example,
the training data consists of a set of n training samples. Each sample, 𝑥𝑖 , is a vector of p
input features (company characteristics) and a target feature (𝑦𝑖 ∈ {𝑌 𝑒𝑠, 𝑁 𝑜} (whether
they became a client or not).
1.2. Introduction to Machine Learning 5

1.2.2 IT/computing science tools
• High Performance Computing (HPC)

• Data flow, data base, file I/O, etc.
• Python: the programming language.
• Numpy: python library particularly useful for handling of raw numerical data (matrices,
mathematical operations).
• Pandas: input/output, manipulation structured data (tables).
1.2.3 Statistics and applied mathematics
• Linear model.
• Non parametric statistics.
• Linear algebra: matrix operations, inversion, eigenvalues.
1.3 Data analysis methodology
1. Formalize customer’s needs into a learning problem:

• A target variable: supervised problem.
– Target is qualitative: classification.
– Target is quantitative: regression.
• No target variable: unsupervised problem
– Vizualisation of high-dimensional samples: PCA, manifolds learning, etc.
– Finding groups of samples (hidden structure): clustering.
2. Ask question about the datasets
• Number of samples
• Number of variables, types of each variable.
3. Define the sample
• For prospective study formalize the experimental design: inclusion/exlusion cri-
teria. The conditions that define the acquisition of the dataset.
• For retrospective study formalize the experimental design: inclusion/exlusion
criteria. The conditions that define the selection of the dataset.
4. In a document formalize (i) the project objectives; (ii) the required learning dataset (more
specifically the input data and the target variables); (iii) The conditions that define the ac-
quisition of the dataset. In this document, warn the customer that the learned algorithms
may not work on new data acquired under different condition.
5. Read the learning dataset.
6. (i) Sanity check (basic descriptive statistics); (ii) data cleaning (impute missing data,
recoding); Final Quality Control (QC) perform descriptive statistics and think ! (re-
move possible confounding variable, etc.).
7. Explore data (visualization, PCA) and perform basic univariate statistics for association
between the target an input variables.
8. Perform more complex multivariate-machine learning.
9. Model validation using a left-out-sample strategy (cross-validation, etc.).
10. Apply on new data.
1.3. Data analysis methodology 7

CHAPTER
TWO
PYTHON LANGUAGE
Note: Click here to download the full example code
Source Kevin Markham https://github.com/justmarkham/python-reference
2.1 Import libraries
# 'generic import' of math module

import math
math.sqrt(25)
# import a function
from math import sqrt
sqrt(25) # no longer have to reference the module
# import multiple functions at once

from math import cos, floor
# import all functions in a module (generally discouraged)

# from os import *
# define an alias
import numpy as np
# show all functions in math module

content = dir(math)
2.2 Basic operations
# Numbers
10 + 4 # add (returns 14)
10 - 4 # subtract (returns 6)
10 * 4 # multiply (returns 40)
10 ** 4 # exponent (returns 10000)
10 / 4 # divide (returns 2 because both types are 'int')
10 / float(4) # divide (returns 2.5)
5 % 4 # modulo (returns 1) - also known as the remainder
(continues on next page)
9
(continued from previous page)

10 / 4 # true division (returns 2.5)
10 // 4 # floor division (returns 2)
# Boolean operations
# comparisons (these return True)
5 > 3
5 >= 3
5 != 3
5 == 5
# boolean operations (these return True)

5 > 3 and 6 > 3
5 > 3 or 5 < 3
not False
False or not False and True # evaluation order: not, and, or
Out:
True
2.3 Data types
# determine the type of an object

type(2) # returns 'int'
type(2.0) # returns 'float'
type('two') # returns 'str'
type(True) # returns 'bool'
type(None) # returns 'NoneType'
# check if an object is of a given type

isinstance(2.0, int) # returns False
isinstance(2.0, (int, float)) # returns True
# convert an object to a given type

float(2)
int(2.9)
str(2.9)
# zero, None, and empty containers are converted to False

bool(0)
bool(None)
bool('') # empty string
bool([]) # empty list
bool({}) # empty dictionary
# non-empty containers and non-zeros are converted to True

bool(2)
bool('two')
bool([2])
Out:
10 Chapter 2. Python language

True
2.3.1 Lists
Different objects categorized along a certain ordered sequence, lists are ordered, iterable, mu-
table (adding or removing objects changes the list size), can contain multiple data types.
# create an empty list (two ways)
empty_list = []
empty_list = list()
# create a list
simpsons = ['homer', 'marge', 'bart']
# examine a list
simpsons[0] # print element 0 ('homer')
len(simpsons) # returns the length (3)
# modify a list (does not return the list)

simpsons.append('lisa') # append element to end
simpsons.extend(['itchy', 'scratchy']) # append multiple elements to end
simpsons.insert(0, 'maggie') # insert element at index 0 (shifts everything␣
˓→right)
simpsons.remove('bart') # searches for first instance and removes it
simpsons.pop(0) # removes element 0 and returns it
del simpsons[0] # removes element 0 (does not return it)
simpsons[0] = 'krusty' # replace element 0
# concatenate lists (slower than 'extend' method)

neighbors = simpsons + ['ned','rod','todd']
# find elements in a list

simpsons.count('lisa') # counts the number of instances
simpsons.index('itchy') # returns index of first instance
# list slicing [start:end:stride]

weekdays = ['mon','tues','wed','thurs','fri']
weekdays[0] # element 0
weekdays[0:3] # elements 0, 1, 2
weekdays[:3] # elements 0, 1, 2
weekdays[3:] # elements 3, 4
weekdays[-1] # last element (element 4)
weekdays[::2] # every 2nd element (0, 2, 4)
weekdays[::-1] # backwards (4, 3, 2, 1, 0)
# alternative method for returning the list backwards

list(reversed(weekdays))
# sort a list in place (modifies but does not return the list)
simpsons.sort()
simpsons.sort(reverse=True) # sort in reverse
simpsons.sort(key=len) # sort by a key
# return a sorted list (but does not modify the original list)
sorted(simpsons)
2.3. Data types 11


sorted(simpsons, reverse=True)
sorted(simpsons, key=len)
# create a second reference to the same list

num = [1, 2, 3]
same_num = num
same_num[0] = 0 # modifies both 'num' and 'same_num'
# copy a list (three ways)

new_num = num.copy()
new_num = num[:]
new_num = list(num)
# examine objects
id(num) == id(same_num) # returns True
id(num) == id(new_num) # returns False
num is same_num # returns True
num is new_num # returns False
num == same_num # returns True
num == new_num # returns True (their contents are equivalent)
# conatenate +, replicate *
[1, 2, 3] + [4, 5, 6]
["a"] * 2 + ["b"] * 3
Out:
['a', 'a', 'b', 'b', 'b']
2.3.2 Tuples
Like lists, but their size cannot change: ordered, iterable, immutable, can contain multiple data
types
# create a tuple
digits = (0, 1, 'two') # create a tuple directly
digits = tuple([0, 1, 'two']) # create a tuple from a list
zero = (0,) # trailing comma is required to indicate it's a tuple
# examine a tuple
digits[2] # returns 'two'
len(digits) # returns 3
digits.count(0) # counts the number of instances of that value (1)
digits.index(1) # returns the index of the first instance of that value (1)
# elements of a tuple cannot be modified

# digits[2] = 2 # throws an error
# concatenate tuples
digits = digits + (3, 4)
# create a single tuple with elements repeated (also works with lists)
(3, 4) * 2 # returns (3, 4, 3, 4)


# tuple unpacking
bart = ('male', 10, 'simpson') # create a tuple
2.3.3 Strings
A sequence of characters, they are iterable, immutable
# create a string
s = str(42) # convert another data type into a string
s = 'I like you'
# examine a string
s[0] # returns 'I'
len(s) # returns 10
# string slicing like lists

s[:6] # returns 'I like'
s[7:] # returns 'you'
s[-1] # returns 'u'
# basic string methods (does not modify the original string)

s.lower() # returns 'i like you'
s.upper() # returns 'I LIKE YOU'
s.startswith('I') # returns True
s.endswith('you') # returns True
s.isdigit() # returns False (returns True if every character in the string is a␣
˓→digit)
s.find('like') # returns index of first occurrence (2), but doesn't support regex
s.find('hate') # returns -1 since not found
s.replace('like','love') # replaces all instances of 'like' with 'love'
# split a string into a list of substrings separated by a delimiter

s.split(' ') # returns ['I','like','you']
s.split() # same thing
s2 = 'a, an, the'
s2.split(',') # returns ['a',' an',' the']
# join a list of strings into one string using a delimiter

stooges = ['larry','curly','moe']
' '.join(stooges) # returns 'larry curly moe'
# concatenate strings
s3 = 'The meaning of life is'
s4 = '42'
s3 + ' ' + s4 # returns 'The meaning of life is 42'
s3 + ' ' + str(42) # same thing
# remove whitespace from start and end of a string

s5 = ' ham and cheese '
s5.strip() # returns 'ham and cheese'
# string substitutions: all of these return 'raining cats and dogs'

'raining %s and %s' % ('cats','dogs') # old way
'raining {} and {}'.format('cats','dogs') # new way
2.3. Data types 13


'raining {arg1} and {arg2}'.format(arg1='cats',arg2='dogs') # named arguments
# string formatting
# more examples: http://mkaz.com/2012/10/10/python-string-format/
'pi is {:.2f}'.format(3.14159) # returns 'pi is 3.14'
Out:
'pi is 3.14'
2.3.4 Strings 2/2
Normal strings allow for escaped characters
print('first line\nsecond line')
Out:
first line
second line
raw strings treat backslashes as literal characters
print(r'first line\nfirst line')
Out:
first line\nfirst line
sequece of bytes are not strings, should be decoded before some operations
s = b'first line\nsecond line'

print(s)
print(s.decode('utf-8').split())
Out:
b'first line\nsecond line'

['first', 'line', 'second', 'line']
2.3.5 Dictionaries
Dictionaries are structures which can contain multiple data types, and is ordered with key-value
pairs: for each (unique) key, the dictionary outputs one value. Keys can be strings, numbers, or
tuples, while the corresponding values can be any Python object. Dictionaries are: unordered,
iterable, mutable
# create an empty dictionary (two ways)

empty_dict = {}
empty_dict = dict()

# create a dictionary (two ways)

family = {'dad':'homer', 'mom':'marge', 'size':6}
family = dict(dad='homer', mom='marge', size=6)
# convert a list of tuples into a dictionary

list_of_tuples = [('dad','homer'), ('mom','marge'), ('size', 6)]
family = dict(list_of_tuples)
# examine a dictionary
family['dad'] # returns 'homer'
len(family) # returns 3
family.keys() # returns list: ['dad', 'mom', 'size']
family.values() # returns list: ['homer', 'marge', 6]
family.items() # returns list of tuples:
# [('dad', 'homer'), ('mom', 'marge'), ('size', 6)]
'mom' in family # returns True
'marge' in family # returns False (only checks keys)
# modify a dictionary (does not return the dictionary)

family['cat'] = 'snowball' # add a new entry
family['cat'] = 'snowball ii' # edit an existing entry
del family['cat'] # delete an entry
family['kids'] = ['bart', 'lisa'] # value can be a list
family.pop('dad') # removes an entry and returns the value ('homer')
family.update({'baby':'maggie', 'grandpa':'abe'}) # add multiple entries
# accessing values more safely with 'get'

family['mom'] # returns 'marge'
family.get('mom') # same thing
try:
family['grandma'] # throws an error
except KeyError as e:
print("Error", e)
family.get('grandma') # returns None

family.get('grandma', 'not found') # returns 'not found' (the default)
# accessing a list element within a dictionary

family['kids'][0] # returns 'bart'
family['kids'].remove('lisa') # removes 'lisa'
# string substitution using a dictionary

'youngest child is %(baby)s' % family # returns 'youngest child is maggie'
Out:
Error 'grandma'
'youngest child is maggie'
2.3. Data types 15

2.3.6 Sets
Like dictionaries, but with unique keys only (no corresponding values). They are: unordered, it-
erable, mutable, can contain multiple data types made up of unique elements (strings, numbers,
or tuples)
# create an empty set

empty_set = set()
# create a set
languages = {'python', 'r', 'java'} # create a set directly
snakes = set(['cobra', 'viper', 'python']) # create a set from a list
# examine a set
len(languages) # returns 3
'python' in languages # returns True
# set operations
languages & snakes # returns intersection: {'python'}
languages | snakes # returns union: {'cobra', 'r', 'java', 'viper', 'python'}
languages - snakes # returns set difference: {'r', 'java'}
snakes - languages # returns set difference: {'cobra', 'viper'}
# modify a set (does not return the set)

languages.add('sql') # add a new element
languages.add('r') # try to add an existing element (ignored, no error)
languages.remove('java') # remove an element
try:
languages.remove('c') # try to remove a non-existing element (throws an error)
except KeyError as e:
print("Error", e)
languages.discard('c') # removes an element if present, but ignored otherwise

languages.pop() # removes and returns an arbitrary element
languages.clear() # removes all elements
languages.update('go', 'spark') # add multiple elements (can also pass a list or set)
# get a sorted list of unique elements from a list

sorted(set([9, 0, 2, 1, 0])) # returns [0, 1, 2, 9]
Out:
Error 'c'
[0, 1, 2, 9]

2.4 Execution control statements
2.4.1 Conditional statements
x = 3
# if statement
if x > 0:
print('positive')
# if/else statement
if x > 0:
print('positive')
else:
print('zero or negative')
# if/elif/else statement
if x > 0:
print('positive')
elif x == 0:
print('zero')
else:
print('negative')
# single-line if statement (sometimes discouraged)

if x > 0: print('positive')
# single-line if/else statement (sometimes discouraged)

# known as a 'ternary operator'
sign = 'positive' if x > 0 else 'zero or negative'
Out:
positive
positive
positive
positive
2.4.2 Loops
Loops are a set of instructions which repeat until termination conditions are met. This can
include iterating through all values in an object, go through a range of values, etc
# range returns a list of integers
range(0, 3) # returns [0, 1, 2]: includes first value but excludes second value
range(3) # same thing: starting at zero is the default
range(0, 5, 2) # returns [0, 2, 4]: third argument specifies the 'stride'
# for loop
fruits = ['apple', 'banana', 'cherry']
for i in range(len(fruits)):
print(fruits[i].upper())
# alternative for loop (recommended style)

for fruit in fruits:
2.4. Execution control statements 17


print(fruit.upper())
# use range when iterating over a large sequence to avoid actually creating the integer␣
˓→list in memory
v = 0
for i in range(10 ** 6):
v += 1
Out:
APPLE
BANANA
CHERRY
APPLE
BANANA
CHERRY
2.4.3 Exercice: count words in a sentence
quote = """
our incomes are like our shoes; if too small they gall and pinch us
but if too large they cause us to stumble and to trip
"""
count = {k:0 for k in set(quote.split())}

for word in quote.split():
count[word] += 1
# iterate through two things at once (using tuple unpacking)

family = {'dad':'homer', 'mom':'marge', 'size':6}
for key, value in family.items():
print(key, value)
# use enumerate if you need to access the index value within the loop
for index, fruit in enumerate(fruits):
print(index, fruit)
# for/else loop
for fruit in fruits:
if fruit == 'banana':
print("Found the banana!")
break # exit the loop and skip the 'else' block
else:
# this block executes ONLY if the for loop completes without hitting 'break'
print("Can't find the banana")
# while loop
count = 0
while count < 5:
print("This will print 5 times")
count += 1 # equivalent to 'count = count + 1'
Out:

dad homer
mom marge
size 6
0 apple
1 banana
2 cherry
Can't find the banana
Found the banana!
This will print 5 times
2.4.4 Exceptions handling
dct = dict(a=[1, 2], b=[4, 5])
key = 'c'
try:
dct[key]
except:
print("Key %s is missing. Add it with empty value" % key)
dct['c'] = []
print(dct)
Out:
Key c is missing. Add it with empty value

{'a': [1, 2], 'b': [4, 5], 'c': []}
2.5 Functions
Functions are sets of instructions launched when called upon, they can have multiple input
values and a return value
# define a function with no arguments and no return values

def print_text():
print('this is text')
# call the function

print_text()
# define a function with one argument and no return values

def print_this(x):
print(x)
# call the function

print_this(3) # prints 3
n = print_this(3) # prints 3, but doesn't assign 3 to n
# because the function has no return statement
2.5. Functions 19
#
def add(a, b):
return a + b
add(2, 3)
add("deux", "trois")
add(["deux", "trois"], [2, 3])
# define a function with one argument and one return value

def square_this(x):
return x ** 2
# include an optional docstring to describe the effect of a function

def square_this(x):
"""Return the square of a number."""
return x ** 2
# call the function

square_this(3) # prints 9
var = square_this(3) # assigns 9 to var, but does not print 9
# default arguments
def power_this(x, power=2):
return x ** power
power_this(2) # 4
power_this(2, 3) # 8
# use 'pass' as a placeholder if you haven't written the function body

def stub():
pass
# return two values from a single function

def min_max(nums):
return min(nums), max(nums)
# return values can be assigned to a single variable as a tuple

nums = [1, 2, 3]
min_max_num = min_max(nums) # min_max_num = (1, 3)
# return values can be assigned into multiple variables using tuple unpacking
min_num, max_num = min_max(nums) # min_num = 1, max_num = 3
Out:
this is text
3
3

2.6 List comprehensions, iterators, etc.
2.6.1 List comprehensions
Process which affects whole lists without iterating through loops. For more: http://
python-3-patterns-idioms-test.readthedocs.io/en/latest/Comprehensions.html
# for loop to create a list of cubes
nums = [1, 2, 3, 4, 5]
cubes = []
for num in nums:
cubes.append(num**3)
# equivalent list comprehension

cubes = [num**3 for num in nums] # [1, 8, 27, 64, 125]
# for loop to create a list of cubes of even numbers

cubes_of_even = []
for num in nums:
if num % 2 == 0:
cubes_of_even.append(num**3)

# syntax: [expression for variable in iterable if condition]
cubes_of_even = [num**3 for num in nums if num % 2 == 0] # [8, 64]
# for loop to cube even numbers and square odd numbers

cubes_and_squares = []
for num in nums:
if num % 2 == 0:
cubes_and_squares.append(num**3)
else:
cubes_and_squares.append(num**2)
# equivalent list comprehension (using a ternary expression)

# syntax: [true_condition if condition else false_condition for variable in iterable]
cubes_and_squares = [num**3 if num % 2 == 0 else num**2 for num in nums] # [1, 8, 9,␣
˓→64, 25]
# for loop to flatten a 2d-matrix

matrix = [[1, 2], [3, 4]]
items = []
for row in matrix:
for item in row:
items.append(item)

items = [item for row in matrix
for item in row] # [1, 2, 3, 4]
# set comprehension
fruits = ['apple', 'banana', 'cherry']
unique_lengths = {len(fruit) for fruit in fruits} # {5, 6}
# dictionary comprehension
fruit_lengths = {fruit:len(fruit) for fruit in fruits} # {'apple': 5, 'banana
˓→': 6, 'cherry': 6} (continues on next page)
2.6. List comprehensions, iterators, etc. 21

Exercise: upper-case names and add 1 year to all simpsons
simpsons = {'Homer':45, 'Marge':45, 'Bart':10, 'Lisa':10}
{k.upper(): v + 1 for k, v in simpsons.items()}
Out:
{'HOMER': 46, 'MARGE': 46, 'BART': 11, 'LISA': 11}
2.7 Regular expression
import re
# 1. Compile regular expression with a patetrn

regex = re.compile("^.+(sub-.+)_(ses-.+)_(mod-.+)")
2. Match compiled RE on string

Capture the pattern ànyprefixsub-<subj id>_ses-<session id>_<modality>`
strings = ["abcsub-033_ses-01_mod-mri", "defsub-044_ses-01_mod-mri", "ghisub-055_ses-02_

˓→mod-ctscan" ]
print([regex.findall(s)[0] for s in strings])
Out:
[('sub-033', 'ses-01', 'mod-mri'), ('sub-044', 'ses-01', 'mod-mri'), ('sub-055', 'ses-02',

˓→ 'mod-ctscan')]
Match methods on compiled regular expression
Method/Attribute Purpose
match(string) Determine if the RE matches at the beginning of the string.
search(string) Scan through a string, looking for any location where this RE matches.
findall(string) Find all substrings where the RE matches, and returns them as a list.
finditer(string) Find all substrings where the RE matches, and returns them as an itera-
tor.
2. Replace compiled RE on string
regex = re.compile("(sub-[^_]+)") # match (sub-...)_

print([regex.sub("SUB-", s) for s in strings])
regex.sub("SUB-", "toto")
Out:

['abcSUB-_ses-01_mod-mri', 'defSUB-_ses-01_mod-mri', 'ghiSUB-_ses-02_mod-ctscan']
'toto'
Replace all non-alphanumeric characters in a string

re.sub('[^0-9a-zA-Z]+', '', 'h^&ell`.,|o w]{+orld')
Out:
'helloworld'
2.8 System programming
2.8.1 Operating system interfaces (os)
import os
Current working directory

# Get the current working directory
cwd = os.getcwd()
print(cwd)
# Set the current working directory

os.chdir(cwd)
Out:
/home/ed203246/git/pystatsml/python_lang
Temporary directory
import tempfile
tmpdir = tempfile.gettempdir()
Join paths
mytmpdir = os.path.join(tmpdir, "foobar")
# list containing the names of the entries in the directory given by path.
os.listdir(tmpdir)
Out:
['tracker-extract-files.16094', 'pymp-b2pv57wx', 'pymp-ewr8p5l2', 'systemd-private-
˓→cd5e03034b9b4cc4abcad7be9bb39d90-fwupd.service-waEboi', 'systemd-private-
˓→cd5e03034b9b4cc4abcad7be9bb39d90-systemd-timesyncd.service-XpeGLh', 'dropbox-antifreeze-
˓→MFb6ch', 'snap.chromium', 'snap.libreoffice', 'config-err-OKsGNU', '.font-unix', 'plop2
˓→', '.X11-unix', 'spyder-ed203246', 'net-export', 'VMwareDnD', '.org.chromium.Chromium.
˓→mMsN15', 'foobar', '.ICE-unix', '.XIM-unix', 'pymp-70seh7of', 'vmware-root', 'systemd-

˓→private-cd5e03034b9b4cc4abcad7be9bb39d90-systemd-logind.service-npESri', 'v8-compile-
˓→cache-16094', 'README2.md', 'systemd-private-cd5e03034b9b4cc4abcad7be9bb39d90-colord.
˓→service-M4EH8f', 'hsperfdata_ed203246', 'systemd-private-
˓→cd5e03034b9b4cc4abcad7be9bb39d90-bolt.service-pAaoJh', 'tracker-extract-files.132', '.
2.8. System programming
˓→X1001-lock', 'snap.zoom-client', 'systemd-private-cd5e03034b9b4cc4abcad7be9bb39d90-
23
˓→switcheroo-control.service-Ub4qhg', 'skype-105971', 'systemd-private-
˓→cd5e03034b9b4cc4abcad7be9bb39d90-systemd-resolved.service-smewXi', '.org.chromium.
˓→Chromium.Jr6bX2', 'systemd-private-cd5e03034b9b4cc4abcad7be9bb39d90-upower.service-
Create a directory
if not os.path.exists(mytmpdir):
os.mkdir(mytmpdir)
os.makedirs(os.path.join(tmpdir, "foobar", "plop", "toto"), exist_ok=True)
2.8.2 File input/output
filename = os.path.join(mytmpdir, "myfile.txt")

print(filename)
# Write
lines = ["Dans python tout est bon", "Enfin, presque"]
## write line by line

fd = open(filename, "w")
fd.write(lines[0] + "\n")
fd.write(lines[1]+ "\n")
fd.close()
## use a context manager to automatically close your file

with open(filename, 'w') as f:
for line in lines:
f.write(line + '\n')
# Read
## read one line at a time (entire file does not have to fit into memory)
f = open(filename, "r")
f.readline() # one string per line (including newlines)
f.readline() # next line
f.close()
## read one line at a time (entire file does not have to fit into memory)
f = open(filename, 'r')
f.readline() # one string per line (including newlines)
f.readline() # next line
f.close()
## read the whole file at once, return a list of lines

f.readlines() # one list, each line is one string
f.close()
## use list comprehension to duplicate readlines without reading entire file at once
[line for line in f]
f.close()
## use a context manager to automatically close your file

with open(filename, 'r') as f:
lines = [line for line in f]

Out:
/tmp/foobar/myfile.txt
2.8.3 Explore, list directories
Walk
import os
WD = os.path.join(tmpdir, "foobar")
for dirpath, dirnames, filenames in os.walk(WD):

print(dirpath, dirnames, filenames)
Out:
/tmp/foobar ['plop'] ['myfile.txt']

/tmp/foobar/plop ['toto'] ['myfile.txt']
/tmp/foobar/plop/toto [] []
glob, basename and file extension
import tempfile
import glob
filenames = glob.glob(os.path.join(tmpdir, "*", "*.txt"))

print(filenames)
# take basename then remove extension

basenames = [os.path.splitext(os.path.basename(f))[0] for f in filenames]
print(basenames)
Out:
['/tmp/plop2/myfile.txt', '/tmp/foobar/myfile.txt']
['myfile', 'myfile']
shutil - High-level file operations
import shutil
src = os.path.join(tmpdir, "foobar", "myfile.txt")

dst = os.path.join(tmpdir, "foobar", "plop", "myfile.txt")
print("copy %s to %s" % (src, dst))
shutil.copy(src, dst)
print("File %s exists ?" % dst, os.path.exists(dst))
src = os.path.join(tmpdir, "foobar", "plop")

dst = os.path.join(tmpdir, "plop2")
print("copy tree %s under %s" % (src, dst))
2.8. System programming 25

try:
shutil.copytree(src, dst)
shutil.rmtree(dst)
shutil.move(src, dst)
except (FileExistsError, FileNotFoundError) as e:
pass
Out:
copy /tmp/foobar/myfile.txt to /tmp/foobar/plop/myfile.txt

File /tmp/foobar/plop/myfile.txt exists ? True
copy tree /tmp/foobar/plop under /tmp/plop2
2.8.4 Command execution with subprocess
• For more advanced use cases, the underlying Popen interface can be used directly.
• Run the command described by args.
• Wait for command to complete
• return a CompletedProcess instance.
• Does not capture stdout or stderr by default. To do so, pass PIPE for the stdout and/or
stderr arguments.
import subprocess
# doesn't capture output

p = subprocess.run(["ls", "-l"])
print(p.returncode)
# Run through the shell.

subprocess.run("ls -l", shell=True)
# Capture output
out = subprocess.run(["ls", "-a", "/"], stdout=subprocess.PIPE, stderr=subprocess.STDOUT)
# out.stdout is a sequence of bytes that should be decoded into a utf-8 string
print(out.stdout.decode('utf-8').split("\n")[:5])
Out:
0
['.', '..', 'bin', 'boot', 'cdrom']

2.8.5 Multiprocessing and multithreading
Process
A process is a name given to a program instance that has been loaded into memory
and managed by the operating system.
Process = address space + execution context (thread of control)
Process address space (segments):
• Code.
• Data (static/global).
• Heap (dynamic memory allocation).
• Stack.
Execution context:
• Data registers.
• Stack pointer (SP).
• Program counter (PC).
• Working Registers.
OS Scheduling of processes: context switching (ie. save/load Execution context)
Pros/cons
• Context switching expensive.
• (potentially) complex data sharing (not necessary true).
• Cooperating processes - no need for memory protection (separate address
spaces).
• Relevant for parrallel computation with memory allocation.
Threads
• Threads share the same address space (Data registers): access to code, heap
and (global) data.
• Separate execution stack, PC and Working Registers.
Pros/cons
• Faster context switching only SP, PC and Working Registers.
• Can exploit fine-grain concurrency
• Simple data sharing through the shared address space.
• Precautions have to be taken or two threads will write to the same memory at
the same time. This is what the global interpreter lock (GIL) is for.
• Relevant for GUI, I/O (Network, disk) concurrent operation
In Python
• The threading module uses threads.
2.8. System programming 27

• The multiprocessing module uses processes.

Multithreading
import time
import threading
def list_append(count, sign=1, out_list=None):

if out_list is None:
out_list = list()
for i in range(count):
out_list.append(sign * i)
sum(out_list) # do some computation
return out_list
size = 10000 # Number of numbers to add
out_list = list() # result is a simple list

thread1 = threading.Thread(target=list_append, args=(size, 1, out_list, ))
thread2 = threading.Thread(target=list_append, args=(size, -1, out_list, ))
startime = time.time()
# Will execute both in parallel
thread1.start()
thread2.start()
# Joins threads back to the parent process
thread1.join()
thread2.join()
print("Threading ellapsed time ", time.time() - startime)
print(out_list[:10])
Out:
Threading ellapsed time 0.945124626159668

[0, 1, 2, 3, 4, 5, 6, 7, 8, 9]
Multiprocessing
import multiprocessing
# Sharing requires specific mecanism

out_list1 = multiprocessing.Manager().list()
p1 = multiprocessing.Process(target=list_append, args=(size, 1, None))
out_list2 = multiprocessing.Manager().list()
p2 = multiprocessing.Process(target=list_append, args=(size, -1, None))
p1.start()
p2.start()
p1.join()
p2.join()
print("Multiprocessing ellapsed time ", time.time() - startime)
# print(out_list[:10]) is not availlable
Out:

Multiprocessing ellapsed time 0.31537318229675293
Sharing object between process with Managers

Managers provide a way to create data which can be shared between different processes, in-
cluding sharing over a network between processes running on different machines. A manager
object controls a server process which manages shared objects.
import multiprocessing
import time
size = int(size / 100) # Number of numbers to add
# Sharing requires specific mecanism

out_list = multiprocessing.Manager().list()
p1 = multiprocessing.Process(target=list_append, args=(size, 1, out_list))
p2 = multiprocessing.Process(target=list_append, args=(size, -1, out_list))
p1.start()
p2.start()
p1.join()
p2.join()
print(out_list[:10])
print("Multiprocessing with shared object ellapsed time ", time.time() - startime)
Out:
[0, 1, 2, 3, 4, 0, 5, -1, 6, -2]

Multiprocessing with shared object ellapsed time 0.4048492908477783
2.9 Scripts and argument parsing
Example, the word count script
import os
import os.path
import argparse
import re
import pandas as pd
if __name__ == "__main__":
# parse command line options
output = "word_count.csv"
parser = argparse.ArgumentParser()
parser.add_argument('-i', '--input',
help='list of input files.',
nargs='+', type=str)
parser.add_argument('-o', '--output',
help='output csv file (default %s)' % output,
2.9. Scripts and argument parsing 29


type=str, default=output)
options = parser.parse_args()
if options.input is None :
parser.print_help()
raise SystemExit("Error: input files are missing")
else:
filenames = [f for f in options.input if os.path.isfile(f)]
# Match words
regex = re.compile("[a-zA-Z]+")
count = dict()
for filename in filenames:
fd = open(filename, "r")
for line in fd:
for word in regex.findall(line.lower()):
if not word in count:
count[word] = 1
else:
count[word] += 1
fd = open(options.output, "w")
# Pandas
df = pd.DataFrame([[k, count[k]] for k in count], columns=["word", "count"])
df.to_csv(options.output, index=False)
2.10 Networking
# TODO
2.10.1 FTP
# Full FTP features with ftplib

import ftplib
ftp = ftplib.FTP("ftp.cea.fr")
ftp.login()
ftp.cwd('/pub/unati/people/educhesnay/pystatml')
ftp.retrlines('LIST')
fd = open(os.path.join(tmpdir, "README.md"), "wb")

ftp.retrbinary('RETR README.md', fd.write)
fd.close()
ftp.quit()
# File download urllib

import urllib.request
ftp_url = 'ftp://ftp.cea.fr/pub/unati/people/educhesnay/pystatml/README.md'
urllib.request.urlretrieve(ftp_url, os.path.join(tmpdir, "README2.md"))
Out:

-rw-r--r-- 1 ftp ftp 3019 Oct 16 2019 README.md

-rw-r--r-- 1 ftp ftp 9628432 Oct 14 07:11␣
˓→StatisticsMachineLearningPythonDraft.pdf
-rw-r--r-- 1 ftp ftp 9798485 Jul 08 07:48␣
˓→StatisticsMachineLearningPythonDraft_202007.pdf
('/tmp/README2.md', <email.message.Message object at 0x7f14ab507510>)
2.10.2 HTTP
# TODO
2.10.3 Sockets
# TODO
2.10.4 xmlrpc
# TODO
2.11 Modules and packages
A module is a Python file. A package is a directory which MUST contain a special file called
__init__.py
To import, extend variable PYTHONPATH:
export PYTHONPATH=path_to_parent_python_module:${PYTHONPATH}
Or
import sys
sys.path.append("path_to_parent_python_module")
The __init__.py file can be empty. But you can set which modules the package exports as the
API, while keeping other modules internal, by overriding the __all__ variable, like so:
parentmodule/__init__.py file:
from . import submodule1

from . import submodule2
from .submodule3 import function1

from .submodule3 import function2
__all__ = ["submodule1", "submodule2",

"function1", "function2"]
User can import:
2.11. Modules and packages 31

import parentmodule.submodule1
import parentmodule.function1
Python Unit Testing
2.12 Object Oriented Programming (OOP)
Sources
• http://python-textbok.readthedocs.org/en/latest/Object_Oriented_Programming.html
Principles
• Encapsulate data (attributes) and code (methods) into objects.
• Class = template or blueprint that can be used to create objects.
• An object is a specific instance of a class.
• Inheritance: OOP allows classes to inherit commonly used state and behaviour from other
classes. Reduce code duplication
• Polymorphism: (usually obtained through polymorphism) calling code is agnostic as to
whether an object belongs to a parent class or one of its descendants (abstraction, modu-
larity). The same method called on 2 objects of 2 different classes will behave differently.
import math
class Shape2D:
def area(self):
raise NotImplementedError()
# __init__ is a special method called the constructor
# Inheritance + Encapsulation
class Square(Shape2D):
def __init__(self, width):
self.width = width
def area(self):
return self.width ** 2
class Disk(Shape2D):
def __init__(self, radius):
self.radius = radius
def area(self):
return math.pi * self.radius ** 2
shapes = [Square(2), Disk(3)]
# Polymorphism
print([s.area() for s in shapes])
s = Shape2D()
try:


s.area()
except NotImplementedError as e:
print("NotImplementedError")
Out:
[4, 28.274333882308138]
NotImplementedError
2.13 Style guide for Python programming
See PEP 8
• Spaces (four) are the preferred indentation method.
• Two blank lines for top level function or classes definition.
• One blank line to indicate logical sections.
• Never use: from lib import *
• Bad: Capitalized_Words_With_Underscores
• Function and Variable Names: lower_case_with_underscores
• Class Names: CapitalizedWords (aka: CamelCase)
2.14 Documenting
See Documenting Python Documenting = comments + docstrings (Python documentation

string)
• Docstrings are use as documentation for the class, module, and packages. See it as “living
documentation”.
• Comments are used to explain non-obvious portions of the code. “Dead documentation”.
Docstrings for functions (same for classes and methods):
def my_function(a, b=2):

"""
This function ...
Parameters
----------
a : float
First operand.
b : float, optional
Second operand. The default is 2.
Returns
-------
Sum of operands.
2.13. Style guide for Python programming 33


Example
-------
>>> my_function(3)
5
"""
# Add a with b (this is a comment)
return a + b
print(help(my_function))
Out:
Help on function my_function in module __main__:
my_function(a, b=2)
This function ...
Parameters
----------
a : float
First operand.
b : float, optional
Second operand. The default is 2.
Returns
-------
Sum of operands.
Example
-------
>>> my_function(3)
5
None
Docstrings for scripts:

At the begining of a script add a pream:
"""
Created on Thu Nov 14 12:08:41 CET 2019
@author: [email protected]
Some description
"""

2.15 Exercises
2.15.1 Exercise 1: functions
Create a function that acts as a simple calulator If the operation is not specified, default to
addition If the operation is misspecified, return an prompt message Ex: calc(4,5,"multiply")
returns 20 Ex: calc(3,5) returns 8 Ex: calc(1, 2, "something") returns error message
2.15.2 Exercise 2: functions + list + loop
Given a list of numbers, return a list where all adjacent duplicate elements have been reduced
to a single element. Ex: [1, 2, 2, 3, 2] returns [1, 2, 3, 2]. You may create a new list or
modify the passed in list.
Remove all duplicate values (adjacent or not) Ex: [1, 2, 2, 3, 2] returns [1, 2, 3]
2.15.3 Exercise 3: File I/O
1. Copy/paste the BSD 4 clause license (https://en.wikipedia.org/wiki/BSD_licenses) into a

text file. Read, the file and count the occurrences of each word within the file. Store the words’
occurrence number in a dictionary.
2. Write an executable python command count_words.py that parse a list of input files provided
after --input parameter. The dictionary of occurrence is save in a csv file provides by --output.
with default value word_count.csv. Use: - open - regular expression - argparse (https://docs.
python.org/3/howto/argparse.html)
2.15.4 Exercise 4: OOP
1. Create a class Employee with 2 attributes provided in the constructor: name,

years_of_service. With one method salary with is obtained by 1500 + 100 *
years_of_service.
2. Create a subclass Manager which redefine salary method 2500 + 120 * years_of_service.
3. Create a small dictionary-nosed database where the key is the employee’s name. Populate
the database with: samples = Employee(‘lucy’, 3), Employee(‘john’, 1), Manager(‘julie’,
10), Manager(‘paul’, 3)
4. Return a table of made name, salary rows, i.e. a list of list [[name, salary]]
5. Compute the average salary
Total running time of the script: ( 0 minutes 3.072 seconds)
2.15. Exercises 35

CHAPTER
THREE
SCIENTIFIC PYTHON
3.1 Numpy: arrays and matrices
NumPy is an extension to the Python programming language, adding support for large, multi-
dimensional (numerical) arrays and matrices, along with a large library of high-level mathe-
matical functions to operate on these arrays.
Sources:
• Kevin Markham: https://github.com/justmarkham
import numpy as np
3.1.1 Create arrays
Create ndarrays from lists. note: every element must be the same type (will be converted if
possible)
data1 = [1, 2, 3, 4, 5] # list

arr1 = np.array(data1) # 1d array
data2 = [range(1, 5), range(5, 9)] # list of lists
arr2 = np.array(data2) # 2d array
arr2.tolist() # convert array back to list
Out:
[[1, 2, 3, 4], [5, 6, 7, 8]]
create special arrays
np.zeros(10)
np.zeros((3, 6))
np.ones(10)
np.linspace(0, 1, 5) # 0 to 1 (inclusive) with 5 points
np.logspace(0, 3, 4) # 10^0 to 10^3 (inclusive) with 4 points
Out:
37
array([ 1., 10., 100., 1000.])
arange is like range, except it returns an array (not a list)
int_array = np.arange(5)
float_array = int_array.astype(float)
3.1.2 Examining arrays
arr1.dtype # float64
arr2.dtype # int32
arr2.ndim # 2
arr2.shape # (2, 4) - axis 0 is rows, axis 1 is columns
arr2.size # 8 - total number of elements
len(arr2) # 2 - size of first dimension (aka axis)
Out:
3.1.3 Reshaping
arr = np.arange(10, dtype=float).reshape((2, 5))

print(arr.shape)
print(arr.reshape(5, 2))
Out:
(2, 5)
[[0. 1.]
[2. 3.]
[4. 5.]
[6. 7.]
[8. 9.]]
Add an axis
a = np.array([0, 1])
a_col = a[:, np.newaxis]
print(a_col)
#or
a_col = a[:, None]
Out:
[[0]
[1]]
Transpose
print(a_col.T)
Out:
38 Chapter 3. Scientific Python

[[0 1]]
Flatten: always returns a flat copy of the orriginal array
arr_flt = arr.flatten()
arr_flt[0] = 33
print(arr_flt)
print(arr)
Out:
[33. 1. 2. 3. 4. 5. 6. 7. 8. 9.]
[[0. 1. 2. 3. 4.]
[5. 6. 7. 8. 9.]]
Ravel: returns a view of the original array whenever possible.
arr_flt = arr.ravel()
arr_flt[0] = 33
print(arr_flt)
print(arr)
Out:
[33. 1. 2. 3. 4. 5. 6. 7. 8. 9.]
[[33. 1. 2. 3. 4.]
[ 5. 6. 7. 8. 9.]]
3.1.4 Summary on axis, reshaping/flattening and selection
Numpy internals: By default Numpy use C convention, ie, Row-major language: The matrix is
stored by rows. In C, the last index changes most rapidly as one moves through the array as
stored in memory.
For 2D arrays, sequential move in the memory will:
• iterate over rows (axis 0)
– iterate over columns (axis 1)
For 3D arrays, sequential move in the memory will:
• iterate over plans (axis 0)
– iterate over rows (axis 1)
* iterate over columns (axis 2)

x = np.arange(2 * 3 * 4)
print(x)
Out:
[ 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23]
Reshape into 3D (axis 0, axis 1, axis 2)
3.1. Numpy: arrays and matrices 39

x = x.reshape(2, 3, 4)
print(x)
Out:
[[[ 0 1 2 3]
[ 4 5 6 7]
[ 8 9 10 11]]
[[12 13 14 15]
[16 17 18 19]
[20 21 22 23]]]
Selection get first plan
print(x[0, :, :])
Out:
[[ 0 1 2 3]
[ 4 5 6 7]
[ 8 9 10 11]]
Selection get first rows
print(x[:, 0, :])
Out:
[[ 0 1 2 3]
[12 13 14 15]]
Selection get first columns
print(x[:, :, 0])
Out:

[[ 0 4 8]
[12 16 20]]
Ravel
print(x.ravel())
Out:
[ 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23]
3.1.5 Stack arrays
Stack flat arrays in columns
a = np.array([0, 1])
b = np.array([2, 3])
ab = np.stack((a, b)).T
print(ab)
# or
np.hstack((a[:, None], b[:, None]))
Out:
[[0 2]
[1 3]]
array([[0, 2],
[1, 3]])
3.1.6 Selection
Single item
arr = np.arange(10, dtype=float).reshape((2, 5))
arr[0] # 0th element (slices like a list)

arr[0, 3] # row 0, column 3: returns 4
arr[0][3] # alternative syntax
Out:
3.0

Slicing
Syntax: start:stop:step with start (default 0) stop (default last) step (default 1)
arr[0, :] # row 0: returns 1d array ([1, 2, 3, 4])

arr[:, 0] # column 0: returns 1d array ([1, 5])
arr[:, :2] # columns strictly before index 2 (2 first columns)
arr[:, 2:] # columns after index 2 included
arr2 = arr[:, 1:4] # columns between index 1 (included) and 4 (excluded)
print(arr2)
Out:
[[1. 2. 3.]
[6. 7. 8.]]
Slicing returns a view (not a copy)
arr2[0, 0] = 33
print(arr2)
print(arr)
Out:
[[33. 2. 3.]
[ 6. 7. 8.]]
[[ 0. 33. 2. 3. 4.]
[ 5. 6. 7. 8. 9.]]
Row 0: reverse order
print(arr[0, ::-1])
# The rule of thumb here can be: in the context of lvalue indexing (i.e. the indices are␣
˓→placed in the left hand side value of an assignment), no view or copy of the array is␣
˓→created (because there is no need to). However, with regular values, the above rules␣
˓→for creating views does apply.
Out:
[ 4. 3. 2. 33. 0.]
Fancy indexing: Integer or boolean array indexing
Fancy indexing returns a copy not a view.

Integer array indexing
arr2 = arr[:, [1,2,3]] # return a copy

print(arr2)
arr2[0, 0] = 44
print(arr2)
print(arr)
Out:

[[33. 2. 3.]
[ 6. 7. 8.]]
[[44. 2. 3.]
[ 6. 7. 8.]]
[[ 0. 33. 2. 3. 4.]
[ 5. 6. 7. 8. 9.]]
Boolean arrays indexing
arr2 = arr[arr > 5] # return a copy
print(arr2)
arr2[0] = 44
print(arr2)
print(arr)
Out:
[33. 6. 7. 8. 9.]
[44. 6. 7. 8. 9.]
[[ 0. 33. 2. 3. 4.]
[ 5. 6. 7. 8. 9.]]
However, In the context of lvalue indexing (left hand side value of an assignment) Fancy autho-
rizes the modification of the original array
arr[arr > 5] = 0
print(arr)
Out:
[[0. 0. 2. 3. 4.]
[5. 0. 0. 0. 0.]]
Boolean arrays indexing continues
names = np.array(['Bob', 'Joe', 'Will', 'Bob'])

names == 'Bob' # returns a boolean array
names[names != 'Bob'] # logical selection
(names == 'Bob') | (names == 'Will') # keywords "and/or" don't work with boolean arrays
names[names != 'Bob'] = 'Joe' # assign based on a logical selection
np.unique(names) # set function
Out:
array(['Bob', 'Joe'], dtype='<U4')

3.1.7 Vectorized operations
nums = np.arange(5)
nums * 10 # multiply each element by 10
nums = np.sqrt(nums) # square root of each element
np.ceil(nums) # also floor, rint (round to nearest int)
np.isnan(nums) # checks for NaN
nums + np.arange(5) # add element-wise
np.maximum(nums, np.array([1, -2, 3, -4, 5])) # compare element-wise
# Compute Euclidean distance between 2 vectors

vec1 = np.random.randn(10)
vec2 = np.random.randn(10)
dist = np.sqrt(np.sum((vec1 - vec2) ** 2))
# math and stats

rnd = np.random.randn(4, 2) # random normals in 4x2 array
rnd.mean()
rnd.std()
rnd.argmin() # index of minimum element
rnd.sum()
rnd.sum(axis=0) # sum of columns
rnd.sum(axis=1) # sum of rows
# methods for boolean arrays

(rnd > 0).sum() # counts number of positive values
(rnd > 0).any() # checks if any value is True
(rnd > 0).all() # checks if all values are True
# random numbers
np.random.seed(12234) # Set the seed
np.random.rand(2, 3) # 2 x 3 matrix in [0, 1]
np.random.randn(10) # random normals (mean 0, sd 1)
np.random.randint(0, 2, 10) # 10 randomly picked 0 or 1
Out:
array([0, 0, 0, 1, 1, 0, 1, 1, 1, 1])
3.1.8 Broadcasting
Sources: https://docs.scipy.org/doc/numpy-1.13.0/user/basics.broadcasting.html Implicit con-

version to allow operations on arrays of different sizes. - The smaller array is stretched or
“broadcasted” across the larger array so that they have compatible shapes. - Fast vectorized
operation in C instead of Python. - No needless copies.

Rules
Starting with the trailing axis and working backward, Numpy compares arrays dimensions.
• If two dimensions are equal then continues
• If one of the operand has dimension 1 stretches it to match the largest one
• When one of the shapes runs out of dimensions (because it has less dimensions than
the other shape), Numpy will use 1 in the comparison process until the other shape’s
dimensions run out as well.
Fig. 1: Source: http://www.scipy-lectures.org
a = np.array([[ 0, 0, 0],
[10, 10, 10],
[20, 20, 20],
[30, 30, 30]])
b = np.array([0, 1, 2])
print(a + b)
Out:
[[ 0 1 2]
[10 11 12]
[20 21 22]
[30 31 32]]

Examples
Shapes of operands A, B and result:
A (2d array): 5 x 4
B (1d array): 1
Result (2d array): 5 x 4
A (2d array): 5 x 4
B (1d array): 4
Result (2d array): 5 x 4
A (3d array): 15 x 3 x 5
B (3d array): 15 x 1 x 5
Result (3d array): 15 x 3 x 5
A (3d array): 15 x 3 x 5
B (2d array): 3 x 5
A (3d array): 15 x 3 x 5
B (2d array): 3 x 1
3.1.9 Exercises
Given the array:
X = np.random.randn(4, 2) # random normals in 4x2 array
• For each column find the row index of the minimum value.
• Write a function standardize(X) that return an array whose columns are centered and
scaled (by std-dev).
3.2 Pandas: data manipulation
It is often said that 80% of data analysis is spent on the cleaning and small, but important,
aspect of data manipulation and cleaning with Pandas.
Sources:
• Kevin Markham: https://github.com/justmarkham
• Pandas doc: http://pandas.pydata.org/pandas-docs/stable/index.html
Data structures

• Series is a one-dimensional labeled array capable of holding any data type (inte-
gers, strings, floating point numbers, Python objects, etc.). The axis labels are col-
lectively referred to as the index. The basic method to create a Series is to call
pd.Series([1,3,5,np.nan,6,8])
• DataFrame is a 2-dimensional labeled data structure with columns of potentially different
types. You can think of it like a spreadsheet or SQL table, or a dict of Series objects. It
stems from the R data.frame() object.
import pandas as pd
import numpy as np
3.2.1 Create DataFrame
columns = ['name', 'age', 'gender', 'job']
user1 = pd.DataFrame([['alice', 19, "F", "student"],

['john', 26, "M", "student"]],
columns=columns)
user2 = pd.DataFrame([['eric', 22, "M", "student"],

['paul', 58, "F", "manager"]],
columns=columns)
user3 = pd.DataFrame(dict(name=['peter', 'julie'],

age=[33, 44], gender=['M', 'F'],
job=['engineer', 'scientist']))
print(user3)
Out:
name age gender job

0 peter 33 M engineer
1 julie 44 F scientist
3.2.2 Combining DataFrames
Concatenate DataFrame
user1.append(user2)
users = pd.concat([user1, user2, user3])
print(users)
Out:
name age gender job

0 alice 19 F student
1 john 26 M student
0 eric 22 M student
1 paul 58 F manager
3.2. Pandas: data manipulation 47


0 peter 33 M engineer
1 julie 44 F scientist
Join DataFrame
user4 = pd.DataFrame(dict(name=['alice', 'john', 'eric', 'julie'],

height=[165, 180, 175, 171]))
print(user4)
Out:
name height
0 alice 165
1 john 180
2 eric 175
3 julie 171
Use intersection of keys from both frames
merge_inter = pd.merge(users, user4, on="name")
print(merge_inter)
Out:
name age gender job height

0 alice 19 F student 165
1 john 26 M student 180
2 eric 22 M student 175
3 julie 44 F scientist 171
Use union of keys from both frames
users = pd.merge(users, user4, on="name", how='outer')

print(users)
Out:

0 alice 19 F student 165.0
1 john 26 M student 180.0
2 eric 22 M student 175.0
3 paul 58 F manager NaN
4 peter 33 M engineer NaN
5 julie 44 F scientist 171.0

Reshaping by pivoting
“Unpivots” a DataFrame from wide format to long (stacked) format,
staked = pd.melt(users, id_vars="name", var_name="variable", value_name="value")

print(staked)
Out:
name variable value

0 alice age 19
1 john age 26
2 eric age 22
3 paul age 58
4 peter age 33
5 julie age 44
6 alice gender F
7 john gender M
8 eric gender M
9 paul gender F
10 peter gender M
11 julie gender F
12 alice job student
13 john job student
14 eric job student
15 paul job manager
16 peter job engineer
17 julie job scientist
18 alice height 165
19 john height 180
20 eric height 175
21 paul height NaN
22 peter height NaN
23 julie height 171
“pivots” a DataFrame from long (stacked) format to wide format,
print(staked.pivot(index='name', columns='variable', values='value'))
Out:
variable age gender height job

name
alice 19 F 165 student
eric 22 M 175 student
john 26 M 180 student
julie 44 F 171 scientist
paul 58 F NaN manager
peter 33 M NaN engineer

3.2.3 Summarizing
# examine the users data
users # print the first 30 and last 30 rows

type(users) # DataFrame
users.head() # print the first 5 rows
users.tail() # print the last 5 rows
users.index # "the index" (aka "the labels")

users.columns # column names (which is "an index")
users.dtypes # data types of each column
users.shape # number of rows and columns
users.values # underlying numpy array
users.info() # concise summary (includes memory usage as of pandas 0.15.0)
Out:
<class 'pandas.core.frame.DataFrame'>
Int64Index: 6 entries, 0 to 5
Data columns (total 5 columns):
# Column Non-Null Count Dtype
--- ------ -------------- -----
0 name 6 non-null object
1 age 6 non-null int64
2 gender 6 non-null object
3 job 6 non-null object
4 height 4 non-null float64
dtypes: float64(1), int64(1), object(3)
memory usage: 288.0+ bytes
3.2.4 Columns selection
users['gender'] # select one column

type(users['gender']) # Series
users.gender # select one column using the DataFrame
# select multiple columns

users[['age', 'gender']] # select two columns
my_cols = ['age', 'gender'] # or, create a list...
users[my_cols] # ...and use that list to select columns
type(users[my_cols]) # DataFrame

3.2.5 Rows selection (basic)
iloc is strictly integer position based
df = users.copy()
df.iloc[0] # first row
df.iloc[0, 0] # first item of first row
df.iloc[0, 0] = 55
for i in range(users.shape[0]):
row = df.iloc[i]
row.age *= 100 # setting a copy, and not the original frame data.
print(df) # df is not modified
Out:
/home/ed203246/anaconda3/lib/python3.7/site-packages/pandas/core/generic.py:5168:␣
˓→SettingWithCopyWarning:
A value is trying to be set on a copy of a slice from a DataFrame
See the caveats in the documentation: https://pandas.pydata.org/pandas-docs/stable/user_

˓→guide/indexing.html#returning-a-view-versus-a-copy
self[name] = value
0 55 19 F student 165.0
ix supports mixed integer and label based access.
df = users.copy()
df.loc[0] # first row
df.loc[0, "age"] # first item of first row
df.loc[0, "age"] = 55
for i in range(df.shape[0]):
df.loc[i, "age"] *= 10
print(df) # df is modified
Out:


3.2.6 Rows selection (filtering)
simple logical filtering
users[users.age < 20] # only show users with age < 20

young_bool = users.age < 20 # or, create a Series of booleans...
young = users[young_bool] # ...and use that Series to filter rows
users[users.age < 20].job # select one column from the filtered results
print(young)
Out:

Advanced logical filtering
users[users.age < 20][['age', 'job']] # select multiple columns

users[(users.age > 20) & (users.gender == 'M')] # use multiple conditions
users[users.job.isin(['student', 'engineer'])] # filter specific values
3.2.7 Sorting
df = users.copy()
df.age.sort_values() # only works for a Series

df.sort_values(by='age') # sort rows by a specific column
df.sort_values(by='age', ascending=False) # use descending order instead
df.sort_values(by=['job', 'age']) # sort by multiple columns
df.sort_values(by=['job', 'age'], inplace=True) # modify df
print(df)
Out:

3.2.8 Descriptive statistics
Summarize all numeric columns
print(df.describe())
Out:

age height
count 6.000000 4.000000
mean 33.666667 172.750000
std 14.895189 6.344289
min 19.000000 165.000000
25% 23.000000 169.500000
50% 29.500000 173.000000
75% 41.250000 176.250000
max 58.000000 180.000000
Summarize all columns
print(df.describe(include='all'))
print(df.describe(include=['object'])) # limit to one (or more) types
Out:

count 6 6.000000 6 6 4.000000
unique 6 NaN 2 4 NaN
top eric NaN F student NaN
freq 1 NaN 3 3 NaN
mean NaN 33.666667 NaN NaN 172.750000
std NaN 14.895189 NaN NaN 6.344289
min NaN 19.000000 NaN NaN 165.000000
25% NaN 23.000000 NaN NaN 169.500000
50% NaN 29.500000 NaN NaN 173.000000
75% NaN 41.250000 NaN NaN 176.250000
max NaN 58.000000 NaN NaN 180.000000
name gender job
count 6 6 6
unique 6 2 4
top eric F student
freq 1 3 3
Statistics per group (groupby)
print(df.groupby("job").mean())
print(df.groupby("job")["age"].mean())
print(df.groupby("job").describe(include='all'))
Out:
age height
job
engineer 33.000000 NaN
manager 58.000000 NaN
scientist 44.000000 171.000000
student 22.333333 173.333333
job
engineer 33.000000
manager 58.000000
scientist 44.000000
student 22.333333


Name: age, dtype: float64
name age ..
˓→. gender height
count unique top freq mean std min 25% 50% 75% max count unique top ..
˓→. 50% 75% max count unique top freq mean std min 25% 50% ␣
˓→75% max
job ..
˓→.
engineer 1 1 peter 1 NaN NaN NaN NaN NaN NaN NaN 1.0 NaN NaN ..
˓→. NaN NaN NaN 0.0 NaN NaN NaN NaN NaN NaN NaN NaN ␣
˓→NaN NaN
manager 1 1 paul 1 NaN NaN NaN NaN NaN NaN NaN 1.0 NaN NaN ..
˓→. NaN NaN NaN 0.0 NaN NaN NaN NaN NaN NaN NaN NaN ␣
˓→NaN NaN
scientist 1 1 julie 1 NaN NaN NaN NaN NaN NaN NaN 1.0 NaN NaN ..
˓→. NaN NaN NaN 1.0 NaN NaN NaN 171.000000 NaN 171.0 171.0 171.0 ␣
˓→171.0 171.0
student 3 3 eric 1 NaN NaN NaN NaN NaN NaN NaN 3.0 NaN NaN ..
˓→. NaN NaN NaN 3.0 NaN NaN NaN 173.333333 7.637626 165.0 170.0 175.0 ␣
˓→177.5 180.0
[4 rows x 44 columns]
Groupby in a loop
for grp, data in df.groupby("job"):

print(grp, data)
Out:
engineer name age gender job height

manager name age gender job height
scientist name age gender job height
student name age gender job height
3.2.9 Quality check
Remove duplicate data
df = users.append(df.iloc[0], ignore_index=True)
print(df.duplicated()) # Series of booleans

# (True if a row is identical to a previous row)
df.duplicated().sum() # count of duplicates
df[df.duplicated()] # only show duplicates
df.age.duplicated() # check a single column for duplicates
df.duplicated(['age', 'gender']).sum() # specify columns for finding duplicates
df = df.drop_duplicates() # drop duplicate rows

Out:
0 False
1 False
2 False
3 False
4 False
5 False
6 True
dtype: bool
Missing data
# Missing values are often just excluded

df = users.copy()
df.describe(include='all') # excludes missing values
# find missing values in a Series

df.height.isnull() # True if NaN, False otherwise
df.height.notnull() # False if NaN, True otherwise
df[df.height.notnull()] # only show rows where age is not NaN
df.height.isnull().sum() # count the missing values
# find missing values in a DataFrame

df.isnull() # DataFrame of booleans
df.isnull().sum() # calculate the sum of each column
Out:
name 0
age 0
gender 0
job 0
height 2
dtype: int64
Strategy 1: drop missing values

df.dropna() # drop a row if ANY values are missing
df.dropna(how='all') # drop a row only if ALL values are missing
Strategy 2: fill in missing values

df.height.mean()
df = users.copy()
df.loc[df.height.isnull(), "height"] = df["height"].mean()
print(df)
Out:


3 paul 58 F manager 172.75
4 peter 33 M engineer 172.75
3.2.10 Rename values
df = users.copy()
print(df.columns)
df.columns = ['age', 'genre', 'travail', 'nom', 'taille']
df.travail = df.travail.map({ 'student':'etudiant', 'manager':'manager',

'engineer':'ingenieur', 'scientist':'scientific'})
# assert df.travail.isnull().sum() == 0
df['travail'].str.contains("etu|inge")
Out:
Index(['name', 'age', 'gender', 'job', 'height'], dtype='object')
0 NaN
1 NaN
2 NaN
3 NaN
4 NaN
5 NaN
Name: travail, dtype: object
3.2.11 Dealing with outliers
size = pd.Series(np.random.normal(loc=175, size=20, scale=10))

# Corrupt the first 3 measures
size[:3] += 500
Based on parametric statistics: use the mean
Assume random variable follows the normal distribution Exclude data outside 3 standard-
deviations: - Probability that a sample lies within 1 sd: 68.27% - Probability that a sample
lies within 3 sd: 99.73% (68.27 + 2 * 15.73)
size_outlr_mean = size.copy()
size_outlr_mean[((size - size.mean()).abs() > 3 * size.std())] = size.mean()
print(size_outlr_mean.mean())
Out:

251.72690773066762
Based on non-parametric statistics: use the median
Median absolute deviation (MAD), based on the median, is a robust non-parametric statistics.
https://en.wikipedia.org/wiki/Median_absolute_deviation
mad = 1.4826 * np.median(np.abs(size - size.median()))

size_outlr_mad = size.copy()
size_outlr_mad[((size - size.median()).abs() > 3 * mad)] = size.median()

print(size_outlr_mad.mean(), size_outlr_mad.median())
Out:
179.66768954321032 184.46486745682998
3.2.12 File I/O
csv
import tempfile, os.path

csv_filename = os.path.join(tmpdir, "users.csv")
users.to_csv(csv_filename, index=False)
other = pd.read_csv(csv_filename)
Read csv from url
url = 'https://raw.github.com/neurospin/pystatsml/master/datasets/salary_table.csv'
salary = pd.read_csv(url)
Excel
xls_filename = os.path.join(tmpdir, "users.xlsx")

users.to_excel(xls_filename, sheet_name='users', index=False)
pd.read_excel(xls_filename, sheet_name='users')
# Multiple sheets
with pd.ExcelWriter(xls_filename) as writer:
users.to_excel(writer, sheet_name='users', index=False)
df.to_excel(writer, sheet_name='salary', index=False)
pd.read_excel(xls_filename, sheet_name='users')
pd.read_excel(xls_filename, sheet_name='salary')

SQL (SQLite)
import pandas as pd
import sqlite3
db_filename = os.path.join(tmpdir, "users.db")
Connect
conn = sqlite3.connect(db_filename)
Creating tables with pandas
salary.to_sql("salary", conn, if_exists="replace")
Push modifications
cur = conn.cursor()
values = (100, 14000, 5, 'Bachelor', 'N')
cur.execute("insert into salary values (?, ?, ?, ?, ?)", values)
conn.commit()
Reading results into a pandas DataFrame
salary_sql = pd.read_sql_query("select * from salary;", conn)

print(salary_sql.head())
pd.read_sql_query("select * from salary;", conn).tail()

pd.read_sql_query('select * from salary where salary>25000;', conn)
pd.read_sql_query('select * from salary where experience=16;', conn)
pd.read_sql_query('select * from salary where education="Master";', conn)
Out:
index salary experience education management

0 0 13876 1 Bachelor Y
1 1 11608 1 Ph.D N
2 2 18701 1 Ph.D Y
3 3 11283 1 Master N
4 4 11767 1 Ph.D N
3.2.13 Exercises
Data Frame
1. Read the iris dataset at ‘https://github.com/neurospin/pystatsml/tree/master/datasets/

iris.csv’
2. Print column names
3. Get numerical columns

4. For each species compute the mean of numerical columns and store it in a stats table
like:
species sepal_length sepal_width petal_length petal_width

0 setosa 5.006 3.428 1.462 0.246
1 versicolor 5.936 2.770 4.260 1.326
2 virginica 6.588 2.974 5.552 2.026
Missing data
Add some missing data to the previous table users:
df = users.copy()
df.loc[[0, 2], "age"] = None
df.loc[[1, 3], "gender"] = None
1. Write a function fillmissing_with_mean(df) that fill all missing value of numerical col-
umn with the mean of the current columns.
2. Save the original users and “imputed” frame in a single excel file “users.xlsx” with 2 sheets:
original, imputed.
3.3 Matplotlib: data visualization
Sources - Nicolas P. Rougier: http://www.labri.fr/perso/nrougier/teaching/matplotlib - https:

//www.kaggle.com/benhamner/d/uciml/iris/python-data-visualizations
3.3.1 Basic plots
import numpy as np
# inline plot (for jupyter)

%matplotlib inline
x = np.linspace(0, 10, 50)

sinus = np.sin(x)
plt.plot(x, sinus)
plt.show()
3.3. Matplotlib: data visualization 59

plt.plot(x, sinus, "o")

plt.show()
# use plt.plot to get color / marker abbreviations
# Rapid multiplot
cosinus = np.cos(x)
plt.plot(x, sinus, "-b", x, sinus, "ob", x, cosinus, "-r", x, cosinus, "or")
plt.xlabel('this is x!')
plt.ylabel('this is y!')
plt.title('My First Plot')
plt.show()

# Step by step
plt.plot(x, sinus, label='sinus', color='blue', linestyle='--', linewidth=2)
plt.plot(x, cosinus, label='cosinus', color='red', linestyle='-', linewidth=2)
plt.legend()
plt.show()

3.3.2 Scatter (2D) plots
Load dataset
import pandas as pd
try:
salary = pd.read_csv("../datasets/salary_table.csv")
except:
df = salary
Simple scatter with colors
colors = colors_edu = {'Bachelor':'r', 'Master':'g', 'Ph.D':'blue'}

plt.scatter(df['experience'], df['salary'], c=df['education'].apply(lambda x: colors[x]),␣
˓→s=100)
<matplotlib.collections.PathCollection at 0x7f2d82e0be10>
Scatter plot with colors and symbols
## Figure size
plt.figure(figsize=(6,5))
## Define colors / sumbols manually

symbols_manag = dict(Y='*', N='.')
colors_edu = {'Bachelor':'r', 'Master':'g', 'Ph.D':'b'}
## group by education x management => 6 groups



for values, d in salary.groupby(['education','management']):
edu, manager = values
plt.scatter(d['experience'], d['salary'], marker=symbols_manag[manager], color=colors_
˓→edu[edu],
s=150, label=manager+"/"+edu)
## Set labels
plt.xlabel('Experience')
plt.ylabel('Salary')
plt.legend(loc=4) # lower right
plt.show()
3.3.3 Saving Figures
### bitmap format

plt.plot(x, sinus)
plt.savefig("sinus.png")
plt.close()
# Prefer vectorial format (SVG: Scalable Vector Graphics) can be edited with
# Inkscape, Adobe Illustrator, Blender, etc.
plt.plot(x, sinus)
plt.savefig("sinus.svg")
plt.close()
# Or pdf
plt.plot(x, sinus)


plt.savefig("sinus.pdf")
plt.close()
3.3.4 Seaborn
Sources: - http://stanford.edu/~mwaskom/software/seaborn - https://elitedatascience.com/

python-seaborn-tutorial
If needed, install using: pip install -U --user seaborn
Boxplot
Box plots are non-parametric: they display variation in samples of a statistical population with-
out making any assumptions of the underlying statistical distribution.
import seaborn as sns
sns.boxplot(x="education", y="salary", hue="management", data=salary)
<AxesSubplot:xlabel='education', ylabel='salary'>
sns.boxplot(x="management", y="salary", hue="education", data=salary)

sns.stripplot(x="management", y="salary", hue="education", data=salary, jitter=True,␣
˓→dodge=True, linewidth=1)# Jitter and split options separate datapoints according to␣
˓→group"
<AxesSubplot:xlabel='management', ylabel='salary'>

Fig. 2: title

### Density plot with one figure containing multiple axis

One figure can contain several axis, whose contain the graphic elements
# Set up the matplotlib figure: 3 x 1 axis
f, axes = plt.subplots(3, 1, figsize=(9, 9), sharex=True)
i = 0
for edu, d in salary.groupby(['education']):
sns.distplot(d.salary[d.management == "Y"], color="b", bins=10, label="Manager",␣
˓→ax=axes[i])
sns.distplot(d.salary[d.management == "N"], color="r", bins=10, label="Employee",␣
˓→ax=axes[i])
axes[i].set_title(edu)
axes[i].set_ylabel('Density')
i += 1
ax = plt.legend()
/home/ed203246/anaconda3/lib/python3.7/site-packages/seaborn/distributions.
˓→py:2551: FutureWarning: distplot is a deprecated function and will be␣
˓→removed in a future version. Please adapt your code to use either displot (a␣
˓→figure-level function with similar flexibility) or histplot (an axes-level␣
˓→function for histograms).
warnings.warn(msg, FutureWarning)



Violin plot (distribution)
ax = sns.violinplot(x="salary", data=salary)

Tune bandwidth
ax = sns.violinplot(x="salary", data=salary, bw=.15)
ax = sns.violinplot(x="management", y="salary", hue="education", data=salary)

Tips dataset One waiter recorded information about each tip he received over a period of a few
months working in one restaurant. He collected several variables:

#sns.set(style="whitegrid")
tips = sns.load_dataset("tips")
print(tips.head())
ax = sns.violinplot(x=tips["total_bill"])
total_bill tip sex smoker day time size

0 16.99 1.01 Female No Sun Dinner 2
1 10.34 1.66 Male No Sun Dinner 3
4 24.59 3.61 Female No Sun Dinner 4

Group by day
ax = sns.violinplot(x="day", y="total_bill", data=tips, palette="muted")
Group by day and color by time (lunch vs dinner)
ax = sns.violinplot(x="day", y="total_bill", hue="time", data=tips, palette="muted",␣

˓→split=True)

Pairwise scatter plots
g = sns.PairGrid(salary, hue="management")
g.map_diag(plt.hist)
g.map_offdiag(plt.scatter)
ax = g.add_legend()

3.3.5 Time series

sns.set(style="darkgrid")
# Load an example dataset with long-form data

fmri = sns.load_dataset("fmri")
# Plot the responses for different events and regions
ax = sns.pointplot(x="timepoint", y="signal",
hue="region", style="event",
data=fmri)
# version 0.9
# sns.lineplot(x="timepoint", y="signal",
# hue="region", style="event",
# data=fmri)


CHAPTER
FOUR
STATISTICS
4.1 Univariate statistics
Basics univariate statistics are required to explore dataset:

• Discover associations between a variable of interest and potential predictors. It is strongly
recommended to start with simple univariate methods before moving to complex multi-
variate predictors.
• Assess the prediction performances of machine learning predictors.
• Most of the univariate statistics are based on the linear model which is one of the main
model in machine learning.
4.1.1 Estimators of the main statistical measures
Mean
Properties of the expected value operator E(·) of a random variable 𝑋
𝐸(𝑋 + 𝑐) = 𝐸(𝑋) + 𝑐 (4.1)

𝐸(𝑋 + 𝑌 ) = 𝐸(𝑋) + 𝐸(𝑌 ) (4.2)
𝐸(𝑎𝑋) = 𝑎𝐸(𝑋) (4.3)
The estimator 𝑥
¯ on a sample of size 𝑛: 𝑥 = 𝑥1 , ..., 𝑥𝑛 is given by
1 ∑︁
𝑥
¯= 𝑥𝑖
𝑛
𝑖
¯ is itself a random variable with properties:

𝑥
• 𝐸(¯ ¯,
𝑥) = 𝑥
𝑉 𝑎𝑟(𝑋)
• 𝑉 𝑎𝑟(¯
𝑥) = 𝑛 .
75
Variance
𝑉 𝑎𝑟(𝑋) = 𝐸((𝑋 − 𝐸(𝑋))2 ) = 𝐸(𝑋 2 ) − (𝐸(𝑋))2

The estimator is
1 ∑︁
𝜎𝑥2 = ¯)2
(𝑥𝑖 − 𝑥
𝑛−1
𝑖
Note here the subtracted 1 degree of freedom (df) in the divisor. In standard statistical practice,
𝑑𝑓 = 1 provides an unbiased estimator of the variance of a hypothetical infinite population.
With 𝑑𝑓 = 0 it instead provides a maximum likelihood estimate of the variance for normally
distributed variables.
Standard deviation
√︀
𝑆𝑡𝑑(𝑋) = 𝑉 𝑎𝑟(𝑋)
√︀
The estimator is simply 𝜎𝑥 = 𝜎𝑥2 .
Covariance
𝐶𝑜𝑣(𝑋, 𝑌 ) = 𝐸((𝑋 − 𝐸(𝑋))(𝑌 − 𝐸(𝑌 ))) = 𝐸(𝑋𝑌 ) − 𝐸(𝑋)𝐸(𝑌 ).

Properties:
Cov(𝑋, 𝑋) = Var(𝑋)
Cov(𝑋, 𝑌 ) = Cov(𝑌, 𝑋)
Cov(𝑐𝑋, 𝑌 ) = 𝑐 Cov(𝑋, 𝑌 )
Cov(𝑋 + 𝑐, 𝑌 ) = Cov(𝑋, 𝑌 )
The estimator with 𝑑𝑓 = 1 is

1 ∑︁
𝜎𝑥𝑦 = (𝑥𝑖 − 𝑥
¯)(𝑦𝑖 − 𝑦¯).
𝑛−1
𝑖
Correlation
𝐶𝑜𝑣(𝑋, 𝑌 )
𝐶𝑜𝑟(𝑋, 𝑌 ) =
𝑆𝑡𝑑(𝑋)𝑆𝑡𝑑(𝑌 )
The estimator is
𝜎𝑥𝑦
𝜌𝑥𝑦 = .
𝜎𝑥 𝜎𝑦
76 Chapter 4. Statistics
Standard Error (SE)
The standard error (SE) is the standard deviation (of the sampling distribution) of a statistic:
𝑆𝑡𝑑(𝑋)
𝑆𝐸(𝑋) = √ .
𝑛
It is most commonly considered for the mean with the estimator
𝑆𝐸(𝑥) = 𝑆𝑡𝑑(𝑋) = 𝜎𝑥¯ (4.4)

𝜎𝑥
=√ . (4.5)
𝑛
Exercises
• Generate 2 random samples: 𝑥 ∼ 𝑁 (1.78, 0.1) and 𝑦 ∼ 𝑁 (1.66, 0.1), both of size 10.
• Compute 𝑥¯, 𝜎𝑥 , 𝜎𝑥𝑦 (xbar, xvar, xycov) using only the np.sum() operation. Explore
the np. module to find out which numpy functions performs the same computations and
compare them (using assert) with your previous results.
4.1.2 Main distributions
Normal distribution
The normal distribution, noted 𝒩 (𝜇, 𝜎) with parameters: 𝜇 mean (location) and 𝜎 > 0 std-dev.
Estimators: 𝑥
¯ and 𝜎𝑥 .
The normal distribution, noted 𝒩 , is useful because of the central limit theorem (CLT) which
states that: given certain conditions, the arithmetic mean of a sufficiently large number of iter-
ates of independent random variables, each with a well-defined expected value and well-defined
variance, will be approximately normally distributed, regardless of the underlying distribution.
import numpy as np
from scipy.stats import norm
%matplotlib inline
mu = 0 # mean
variance = 2 #variance
sigma = np.sqrt(variance) #standard deviation",
x = np.linspace(mu-3*variance,mu+3*variance, 100)
plt.plot(x, norm.pdf(x, mu, sigma))
[<matplotlib.lines.Line2D at 0x7f3c95a58b50>]
4.1. Univariate statistics 77

The Chi-Square distribution
The chi-square or 𝜒2𝑛 distribution with 𝑛 degrees of freedom (df) is the distribution of a sum of
the squares of 𝑛 independent standard normal random variables 𝒩 (0, 1). Let 𝑋 ∼ 𝒩 (𝜇, 𝜎 2 ),
then, 𝑍 = (𝑋 − 𝜇)/𝜎 ∼ 𝒩 (0, 1), then:
• The squared standard 𝑍 2 ∼ 𝜒21 (one df).
∑︀𝑛
• The distribution of sum of squares of 𝑛 normal random variables: 𝑖 𝑍𝑖2 ∼ 𝜒2𝑛
The sum of two 𝜒2 RV with 𝑝 and 𝑞 df is a 𝜒2 RV with 𝑝 + 𝑞 df. This is useful when sum-
ming/subtracting sum of squares.
The 𝜒2 -distribution is used to model errors measured as sum of squares or the distribution of
the sample variance.
The Fisher’s F-distribution
The 𝐹 -distribution, 𝐹𝑛,𝑝 , with 𝑛 and 𝑝 degrees of freedom is the ratio of two independent 𝜒2
variables. Let 𝑋 ∼ 𝜒2𝑛 and 𝑌 ∼ 𝜒2𝑝 then:
𝑋/𝑛
𝐹𝑛,𝑝 =
𝑌 /𝑝
The 𝐹 -distribution plays a central role in hypothesis testing answering the question: Are two
variances equals?, is the ratio or two errors significantly large ?.
import numpy as np
from scipy.stats import f
%matplotlib inline
fvalues = np.linspace(.1, 5, 100)

# pdf(x, df1, df2): Probability density function at x of F.

plt.plot(fvalues, f.pdf(fvalues, 1, 30), 'b-', label="F(1, 30)")
plt.plot(fvalues, f.pdf(fvalues, 5, 30), 'r-', label="F(5, 30)")
plt.legend()
# cdf(x, df1, df2): Cumulative distribution function of F.

# ie.
proba_at_f_inf_3 = f.cdf(3, 1, 30) # P(F(1,30) < 3)
# ppf(q, df1, df2): Percent point function (inverse of cdf) at q of F.

f_at_proba_inf_95 = f.ppf(.95, 1, 30) # q such P(F(1,30) < .95)
assert f.cdf(f_at_proba_inf_95, 1, 30) == .95
# sf(x, df1, df2): Survival function (1 - cdf) at x of F.

proba_at_f_sup_3 = f.sf(3, 1, 30) # P(F(1,30) > 3)
assert proba_at_f_inf_3 + proba_at_f_sup_3 == 1
# p-value: P(F(1, 30)) < 0.05

low_proba_fvalues = fvalues[fvalues > f_at_proba_inf_95]
plt.fill_between(low_proba_fvalues, 0, f.pdf(low_proba_fvalues, 1, 30),
alpha=.8, label="P < 0.05")
plt.show()

The Student’s 𝑡-distribution
Let 𝑀 ∼ 𝒩 (0, 1) and 𝑉 ∼ 𝜒2𝑛 . The 𝑡-distribution, 𝑇𝑛 , with 𝑛 degrees of freedom is the ratio:
𝑀
𝑇𝑛 = √︀
𝑉 /𝑛
The distribution of the difference between an estimated parameter and its true (or assumed)
value divided by the standard deviation of the estimated parameter (standard error) follow a
𝑡-distribution. Is this parameters different from a given value?
4.1.3 Hypothesis Testing
Examples
• Test a proportion: Biased coin ? 200 heads have been found over 300 flips, is it coins
biased ?
• Test the association between two variables.
– Exemple height and sex: In a sample of 25 individuals (15 females, 10 males), is
female height is different from male height ?
– Exemple age and arterial hypertension: In a sample of 25 individuals is age height
correlated with arterial hypertension ?
Steps
1. Model the data
2. Fit: estimate the model parameters (frequency, mean, correlation, regression coeficient)
3. Compute a test statistic from model the parameters.
4. Formulate the null hypothesis: What would be the (distribution of the) test statistic if the
observations are the result of pure chance.
5. Compute the probability (𝑝-value) to obtain a larger value for the test statistic by chance
(under the null hypothesis).
Flip coin: Simplified example
Biased coin ? 2 heads have been found over 3 flips, is it coins biased ?
1. Model the data: number of heads follow a Binomial disctribution.
2. Compute model parameters: N=3, P = the frequency of number of heads over the number
of flip: 2/3.
3. Compute a test statistic, same as frequency.
4. Under the null hypothesis the distribution of the number of tail is:
1 2 3 count #heads
0
H 1
H 1
H 1
H H 2
H H 2
H H 2
H H H 3
8 possibles configurations, probabilities of differents values for 𝑝 are: 𝑥 measure the number of
success.
• 𝑃 (𝑥 = 0) = 1/8
• 𝑃 (𝑥 = 1) = 3/8
• 𝑃 (𝑥 = 2) = 3/8
• 𝑃 (𝑥 = 3) = 1/8
plt.bar([0, 1, 2, 3], [1/8, 3/8, 3/8, 1/8], width=0.9)

_ = plt.xticks([0, 1, 2, 3], [0, 1, 2, 3])
plt.xlabel("Distribution of the number of head over 3 flip under the null hypothesis")
Text(0.5, 0, 'Distribution of the number of head over 3 flip under the null hypothesis')
3. Compute the probability (𝑝-value) to observe a value larger or equal that 2 under the null
hypothesis ? This probability is the 𝑝-value:
𝑃 (𝑥 ≥ 2|𝐻0 ) = 𝑃 (𝑥 = 2) + 𝑃 (𝑥 = 3) = 3/8 + 1/8 = 4/8 = 1/2

Flip coin: Real Example
Biased coin ? 60 heads have been found over 100 flips, is it coins biased ?
1. Model the data: number of heads follow a Binomial disctribution.
2. Compute model parameters: N=100, P=60/100.
3. Compute a test statistic, same as frequency.
4. Compute a test statistic: 60/100.
5. Under the null hypothesis the distribution of the number of tail (𝑘) follow the binomial
distribution of parameters N=100, P=0.5:
(︂ )︂
100
𝑃 𝑟(𝑋 = 𝑘|𝐻0 ) = 𝑃 𝑟(𝑋 = 𝑘|𝑛 = 100, 𝑝 = 0.5) = 0.5𝑘 (1 − 0.5)(100−𝑘) .
𝑘
100 (︂ )︂
∑︁ 100
𝑃 (𝑋 = 𝑘 ≥ 60|𝐻0 ) = 0.5𝑘 (1 − 0.5)(100−𝑘)
𝑘
𝑘=60
60 (︂ )︂
∑︁ 100
=1− 0.5𝑘 (1 − 0.5)(100−𝑘) , the cumulative distribution function.
𝑘
𝑘=1
Use tabulated binomial distribution
import scipy.stats
#tobs = 2.39687663116 # assume the t-value

succes = np.linspace(30, 70, 41)
plt.plot(succes, scipy.stats.binom.pmf(succes, 100, 0.5), 'b-', label="Binomial(100, 0.5)
˓→")
upper_succes_tvalues = succes[succes > 60]
plt.fill_between(upper_succes_tvalues, 0, scipy.stats.binom.pmf(upper_succes_tvalues, 100,
˓→ 0.5), alpha=.8, label="p-value")
_ = plt.legend()
pval = 1 - scipy.stats.binom.cdf(60, 100, 0.5)

print(pval)
0.01760010010885238
Random sampling of the Binomial distribution under the null hypothesis
sccess_h0 = scipy.stats.binom.rvs(100, 0.5, size=10000, random_state=4)

print(sccess_h0)
pval_rnd = np.sum(sccess_h0 >= 60) / (len(sccess_h0) + 1)

print("P-value using monte-carlo sampling of the Binomial distribution under H0=", pval_
˓→rnd)
[60 52 51 ... 45 51 44]

P-value using monte-carlo sampling of the Binomial distribution under H0= 0.
˓→025897410258974102
One sample 𝑡-test
The one-sample 𝑡-test is used to determine whether a sample comes from a population with a
specific mean. For example you want to test if the average height of a population is 1.75 𝑚.
Assumptions
1. Independence of residuals (𝜀𝑖 ). This assumptions must be satisfied.

2. Normality of residuals. Approximately normally distributed can be accepted.
Remarks: Although the parent population does not need to be normally distributed, the dis-
tribution of the population of sample means, 𝑥, is assumed to be normal. By the central limit
theorem, if the sampling of the parent population is independent then the sample means will
be approximately normal.

1 Model the data
Assume that height is normally distributed: 𝑋 ∼ 𝒩 (𝜇, 𝜎), ie:
height𝑖 = average height over the population + error𝑖 (4.6)

𝑥𝑖 = 𝑥
¯ + 𝜀𝑖 (4.7)
The 𝜀𝑖 are called the residuals
2 Fit: estimate the model parameters
¯, 𝑠𝑥 are the estimators of 𝜇, 𝜎.

𝑥
3 Compute a test statistic
In testing the null hypothesis that the population mean is equal to a specified value 𝜇0 = 1.75,
one uses the statistic:
difference of means √
𝑡= 𝑛 (4.8)
std-dev √of noise
𝑡 = effect size 𝑛 (4.9)
¯ − 𝜇0 √
𝑥
𝑡= 𝑛 (4.10)
𝑠𝑥
4 Compute the probability of the test statistic under the null hypotheis. This require to have
the distribution of the t statistic under 𝐻0 .
Example
Given the following samples, we will test whether its true mean is 1.75.
Warning, when computing the std or the variance, set ddof=1. The default value, ddof=0, leads
to the biased estimator of the variance.
import numpy as np
x = [1.83, 1.83, 1.73, 1.82, 1.83, 1.73, 1.99, 1.85, 1.68, 1.87]
xbar = np.mean(x) # sample mean

mu0 = 1.75 # hypothesized value
s = np.std(x, ddof=1) # sample standard deviation
n = len(x) # sample size
print(xbar)
tobs = (xbar - mu0) / (s / np.sqrt(n))

print(tobs)
1.816
2.3968766311585883
The :math:`p`-value is the probability to observe a value 𝑡 more extreme than the observed
one 𝑡𝑜𝑏𝑠 under the null hypothesis 𝐻0 : 𝑃 (𝑡 > 𝑡𝑜𝑏𝑠 |𝐻0 )
import scipy.stats as stats

#tobs = 2.39687663116 # assume the t-value

tvalues = np.linspace(-10, 10, 100)
plt.plot(tvalues, stats.t.pdf(tvalues, n-1), 'b-', label="T(n-1)")
upper_tval_tvalues = tvalues[tvalues > tobs]
plt.fill_between(upper_tval_tvalues, 0, stats.t.pdf(upper_tval_tvalues, n-1), alpha=.8,␣
˓→label="p-value")
_ = plt.legend()
4.1.4 Testing pairwise associations
Univariate statistical analysis: explore association betweens pairs of variables.

• In statistics, a categorical variable or factor is a variable that can take on one of a limited,
and usually fixed, number of possible values, thus assigning each individual to a particular
group or “category”. The levels are the possibles values of the variable. Number of levels
= 2: binomial; Number of levels > 2: multinomial. There is no intrinsic ordering to the
categories. For example, gender is a categorical variable having two categories (male and
female) and there is no intrinsic ordering to the categories. For example, Sex (Female,
Male), Hair color (blonde, brown, etc.).
• An ordinal variable is a categorical variable with a clear ordering of the levels. For
example: drinks per day (none, small, medium and high).
• A continuous or quantitative variable 𝑥 ∈ R is one that can take any value in a range of
possible values, possibly infinite. E.g.: salary, experience in years, weight.

What statistical test should I use?

See: http://www.ats.ucla.edu/stat/mult_pkg/whatstat/
Fig. 1: Statistical tests
## Pearson correlation test: test association between two quantitative variables

Test the correlation coefficient of two quantitative variables. The test calculates a Pearson cor-
relation coefficient and the 𝑝-value for testing non-correlation.
Let 𝑥 and 𝑦 two quantitative variables, where 𝑛 samples were obeserved. The linear correlation
coeficient is defined as :
∑︀𝑛
(𝑥𝑖 − 𝑥¯)(𝑦𝑖 − 𝑦¯)
𝑟 = √︀∑︀𝑛 𝑖=1 √︀∑︀𝑛 .
2 ¯)2
𝑖=1 (𝑥𝑖 − 𝑥¯) 𝑖=1 (𝑦𝑖 − 𝑦
√ 𝑟
Under 𝐻0 , the test statistic 𝑡 = 𝑛 − 2 √1−𝑟 2
follow Student distribution with 𝑛 − 2 degrees of
freedom.
import numpy as np
n = 50
x = np.random.normal(size=n)
y = 2 * x + np.random.normal(size=n)

# Compute with scipy
cor, pval = stats.pearsonr(x, y)
print(cor, pval)
0.9421665358030606 1.9724847076189737e-24
4.1.5 Two sample (Student) 𝑡-test: compare two means
Fig. 2: Two-sample model
The two-sample 𝑡-test (Snedecor and Cochran, 1989) is used to determine if two population
means are equal. There are several variations on this test. If data are paired (e.g. 2 measures,
before and after treatment for each individual) use the one-sample 𝑡-test of the difference. The
variances of the two samples may be assumed to be equal (a.k.a. homoscedasticity) or unequal
(a.k.a. heteroscedasticity).
Assumptions
1. Independence of residuals (𝜀𝑖 ). This assumptions must be satisfied.

2. Normality of residuals. Approximately normally distributed can be accepted.
3. Homosedasticity use T-test, Heterosedasticity use Welch t-test.
1. Model the data
Assume that the two random variables are normally distributed: 𝑦1 ∼ 𝒩 (𝜇1 , 𝜎1 ), 𝑦2 ∼
𝒩 (𝜇2 , 𝜎2 ).

2. Fit: estimate the model parameters
Estimate means and variances: 𝑦¯1 , 𝑠2𝑦1 , 𝑦¯2 , 𝑠2𝑦2 .
3. 𝑡-test
The general principle is
difference of means
𝑡= (4.11)
standard dev of error
difference of means
= (4.12)
its standard error
𝑦¯1 − 𝑦¯2 √
= √︀∑︀ 𝑛−2 (4.13)
𝜀2
𝑦¯1 − 𝑦¯2
= (4.14)
𝑠𝑦¯1 −𝑦¯2
Since 𝑦1 and 𝑦2 are independant:
𝑠2𝑦1 𝑠2𝑦
𝑠2𝑦¯1 −𝑦¯2 = 𝑠2𝑦¯1 + 𝑠2𝑦¯2 = + 2 (4.15)
𝑛1 𝑛2
thus (4.16)
√︃
𝑠2𝑦1 𝑠2𝑦
𝑠𝑦¯1 −𝑦¯2 = + 2 (4.17)
𝑛1 𝑛2
Equal or unequal sample sizes, unequal variances (Welch’s 𝑡-test)
Welch’s 𝑡-test defines the 𝑡 statistic as

𝑦¯1 − 𝑦¯2
𝑡 = √︁ 2 .
𝑠𝑦1 𝑠2𝑦2
𝑛1 + 𝑛2
To compute the 𝑝-value one needs the degrees of freedom associated with this variance estimate.
It is approximated using the Welch–Satterthwaite equation:
(︂ )︂2
𝑠2𝑦1 𝑠2𝑦2
𝑛1 + 𝑛2
𝜈≈ 𝑠4𝑦1 𝑠4𝑦2
.
𝑛21 (𝑛1 −1)
+ 𝑛22 (𝑛2 −1)
Equal or unequal sample sizes, equal variances
If we assume equal variance (ie, 𝑠2𝑦1 = 𝑠2𝑦1 = 𝑠2 ), where 𝑠2 is an estimator of the common
variance of the two samples:
𝑠2𝑦1 (𝑛1 − 1) + 𝑠2𝑦2 (𝑛2 − 1)

𝑠2 = (4.18)
𝑛1 + 𝑛2 − 2
∑︀𝑛1 2
∑︀𝑛2 2
𝑖 (𝑦1𝑖 − 𝑦¯1 ) + 𝑗 (𝑦2𝑗 − 𝑦¯2 )
= (4.19)
(𝑛1 − 1) + (𝑛2 − 1)
then
√︃ √︂
𝑠2 𝑠2 1 1
𝑠𝑦¯1 −𝑦¯2 = + =𝑠 +
𝑛1 𝑛2 𝑛1 𝑛2
Therefore, the 𝑡 statistic, that is used to test whether the means are different is:
𝑦¯ − 𝑦¯2
𝑡= √︁1 ,
𝑠 · 𝑛11 + 1
𝑛2
Equal sample sizes, equal variances
If we simplify the problem assuming equal samples of size 𝑛1 = 𝑛2 = 𝑛 we get
𝑦¯1 − 𝑦¯2 √
𝑡= √ · 𝑛 (4.20)
𝑠 2
√
≈ effect size · 𝑛 (4.21)
difference of means √
≈ · 𝑛 (4.22)
standard deviation of the noise
Example
Given the following two samples, test whether their means are equal using the standard t-test,
assuming equal variance.
height = np.array([ 1.83, 1.83, 1.73, 1.82, 1.83, 1.73, 1.99, 1.85, 1.68, 1.87,
1.66, 1.71, 1.73, 1.64, 1.70, 1.60, 1.79, 1.73, 1.62, 1.77])
grp = np.array(["M"] * 10 + ["F"] * 10)

print(stats.ttest_ind(height[grp == "M"], height[grp == "F"], equal_var=True))
Ttest_indResult(statistic=3.5511519888466885, pvalue=0.00228208937112721)

4.1.6 ANOVA 𝐹 -test (quantitative ~ categorial (>=2 levels))
Analysis of variance (ANOVA) provides a statistical test of whether or not the means of several
(k) groups are equal, and therefore generalizes the 𝑡-test to more than two groups. ANOVAs
are useful for comparing (testing) three or more means (groups or variables) for statistical
significance. It is conceptually similar to multiple two-sample 𝑡-tests, but is less conservative.
Here we will consider one-way ANOVA with one independent variable, ie one-way anova.
Wikipedia:
• Test if any group is on average superior, or inferior, to the others versus the null hypothesis
that all four strategies yield the same mean response
• Detect any of several possible differences.
• The advantage of the ANOVA 𝐹 -test is that we do not need to pre-specify which strategies
are to be compared, and we do not need to adjust for making multiple comparisons.
• The disadvantage of the ANOVA 𝐹 -test is that if we reject the null hypothesis, we do not
know which strategies can be said to be significantly different from the others.
Assumptions
1. The samples are randomly selected in an independent manner from the k populations.
2. All k populations have distributions that are approximately normal. Check by plotting
groups distribution.
3. The k population variances are equal. Check by plotting groups distribution.
1. Model the data
Is there a difference in Petal Width in species from iris dataset. Let 𝑦1 , 𝑦2 and 𝑦3 be Petal Width
in three species.
Here we assume (see assumptions) that the three populations were sampled from three random
variables that are normally distributed. I.e., 𝑌1 ∼ 𝑁 (𝜇1 , 𝜎1 ), 𝑌2 ∼ 𝑁 (𝜇2 , 𝜎2 ) and 𝑌3 ∼ 𝑁 (𝜇3 , 𝜎3 ).
Estimate means and variances: 𝑦¯𝑖 , 𝜎𝑖 , ∀𝑖 ∈ {1, 2, 3}.
3. 𝐹 -test
The formula for the one-way ANOVA F-test statistic is
Explained variance
𝐹 = (4.23)
Unexplained variance
Between-group variability 𝑠2
= = 2𝐵 . (4.24)
Within-group variability 𝑠𝑊
The “explained variance”, or “between-group variability” is

∑︁
𝑠2𝐵 = 𝑦𝑖· − 𝑦¯)2 /(𝐾 − 1),
𝑛𝑖 (¯
𝑖
where 𝑦¯𝑖· denotes the sample mean in the 𝑖th group, 𝑛𝑖 is the number of observations in the 𝑖th
group, 𝑦¯ denotes the overall mean of the data, and 𝐾 denotes the number of groups.
The “unexplained variance”, or “within-group variability” is
∑︁
𝑠2𝑊 = (𝑦𝑖𝑗 − 𝑦¯𝑖· )2 /(𝑁 − 𝐾),
𝑖𝑗
where 𝑦𝑖𝑗 is the 𝑗th observation in the 𝑖th out of 𝐾 groups and 𝑁 is the overall sample size.
This 𝐹 -statistic follows the 𝐹 -distribution with 𝐾 − 1 and 𝑁 − 𝐾 degrees of freedom under the
null hypothesis. The statistic will be large if the between-group variability is large relative to
the within-group variability, which is unlikely to happen if the population means of the groups
all have the same value.
Note that when there are only two groups for the one-way ANOVA F-test, 𝐹 = 𝑡2 where 𝑡 is the
Student’s 𝑡 statistic.

import statsmodels.api as sm
from statsmodels.formula.api import ols
# Load iris datset

iris = sm.datasets.get_rdataset("iris").data
iris.columns = [s.replace('.', '_') for s in iris.columns]
# Group means
means = iris.groupby("Species").mean().reset_index()
print(means)
# Group Stds (equal variances ?)

stds = iris.groupby("Species").std().reset_index()
print(stds)
# Plot groups
ax = sns.violinplot(x="Species", y="Sepal_Length", data=iris)
ax = sns.swarmplot(x="Species", y="Sepal_Length", data=iris,
color="white")
ax = sns.swarmplot(x="Species", y="Sepal_Length", color="black", data=means, size=10)
# ANOVA
lm = ols('Sepal_Length ~ Species', data=iris).fit()
sm.stats.anova_lm(lm, typ=2) # Type 2 ANOVA DataFrame
Species Sepal_Length Sepal_Width Petal_Length Petal_Width

0 setosa 5.006 3.428 1.462 0.246
1 versicolor 5.936 2.770 4.260 1.326
2 virginica 6.588 2.974 5.552 2.026
Species Sepal_Length Sepal_Width Petal_Length Petal_Width
0 setosa 0.352490 0.379064 0.173664 0.105386
1 versicolor 0.516171 0.313798 0.469911 0.197753
2 virginica 0.635880 0.322497 0.551895 0.274650

4.1.7 Chi-square, 𝜒2 (categorial ~ categorial)
Computes the chi-square, 𝜒2 , statistic and 𝑝-value for the hypothesis test of independence of
frequencies in the observed contingency table (cross-table). The observed frequencies are tested
against an expected contingency table obtained by computing expected frequencies based on
the marginal sums under the assumption of independence.
Example: 20 participants: 10 exposed to some chemical product and 10 non exposed (exposed
= 1 or 0). Among the 20 participants 10 had cancer 10 not (cancer = 1 or 0). 𝜒2 tests the
association between those two variables.
import numpy as np
import pandas as pd
# Dataset:
# 15 samples:
# 10 first exposed
exposed = np.array([1] * 10 + [0] * 10)
# 8 first with cancer, 10 without, the last two with.
cancer = np.array([1] * 8 + [0] * 10 + [1] * 2)
crosstab = pd.crosstab(exposed, cancer, rownames=['exposed'],

colnames=['cancer'])
print("Observed table:")
print("---------------")
print(crosstab)
chi2, pval, dof, expected = stats.chi2_contingency(crosstab)

print("Statistics:")
print("-----------")
print("Chi2 = %f, pval = %f" % (chi2, pval))
print("Expected table:")

print("---------------")
print(expected)
Observed table:
---------------
cancer 0 1
exposed
0 8 2
1 2 8
Statistics:
-----------
Chi2 = 5.000000, pval = 0.025347
Expected table:
---------------
[[5. 5.]
[5. 5.]]
Computing expected cross-table
# Compute expected cross-table based on proportion

exposed_marg = crosstab.sum(axis=0)
exposed_freq = exposed_marg / exposed_marg.sum()
cancer_marg = crosstab.sum(axis=1)
cancer_freq = cancer_marg / cancer_marg.sum()
print('Exposed frequency? Yes: %.2f' % exposed_freq[0],

'No: %.2f' % exposed_freq[1])
print('Cancer frequency? Yes: %.2f' % cancer_freq[0],
'No: %.2f' % cancer_freq[1])
print('Expected frequencies:')
print(np.outer(exposed_freq, cancer_freq))
print('Expected cross-table (frequencies * N): ')

print(np.outer(exposed_freq, cancer_freq) * len(exposed))
Exposed frequency? Yes: 0.50 No: 0.50

Cancer frequency? Yes: 0.50 No: 0.50
Expected frequencies:
[[0.25 0.25]
[0.25 0.25]]
Expected cross-table (frequencies * N):
[[5. 5.]
[5. 5.]]

4.1.8 Non-parametric test of pairwise associations
Spearman rank-order correlation (quantitative ~ quantitative)
The Spearman correlation is a non-parametric measure of the monotonicity of the relationship

between two datasets.
When to use it? Observe the data distribution: - presence of outliers - the distribution of the
residuals is not Gaussian.
Like other correlation coefficients, this one varies between -1 and +1 with 0 implying no cor-
relation. Correlations of -1 or +1 imply an exact monotonic relationship. Positive correlations
imply that as 𝑥 increases, so does 𝑦. Negative correlations imply that as 𝑥 increases, 𝑦 decreases.
import numpy as np
x = np.array([44.4, 45.9, 41.9, 53.3, 44.7, 44.1, 50.7, 45.2, 46, 47, 48, 60.1])
y = np.array([2.6, 3.1, 2.5, 5.0, 3.6, 4.0, 5.2, 2.8, 4, 4.1, 4.5, 3.8])
plt.plot(x, y, "bo")
# Non-Parametric Spearman
cor, pval = stats.spearmanr(x, y)
print("Non-Parametric Spearman cor test, cor: %.4f, pval: %.4f" % (cor, pval))
# "Parametric Pearson cor test

cor, pval = stats.pearsonr(x, y)
print("Parametric Pearson cor test: cor: %.4f, pval: %.4f" % (cor, pval))
Non-Parametric Spearman cor test, cor: 0.7110, pval: 0.0095

Parametric Pearson cor test: cor: 0.5263, pval: 0.0788
Wilcoxon signed-rank test (quantitative ~ cte)
Source: https://en.wikipedia.org/wiki/Wilcoxon_signed-rank_test
The Wilcoxon signed-rank test is a non-parametric statistical hypothesis test used when com-
paring two related samples, matched samples, or repeated measurements on a single sample
to assess whether their population mean ranks differ (i.e. it is a paired difference test). It is
equivalent to one-sample test of the difference of paired samples.
It can be used as an alternative to the paired Student’s 𝑡-test, 𝑡-test for matched pairs, or the 𝑡-
test for dependent samples when the population cannot be assumed to be normally distributed.
When to use it? Observe the data distribution: - presence of outliers - the distribution of the
residuals is not Gaussian
It has a lower sensitivity compared to 𝑡-test. May be problematic to use when the sample size is
small.
Null hypothesis 𝐻0 : difference between the pairs follows a symmetric distribution around zero.

n = 20
# Buisness Volume time 0
bv0 = np.random.normal(loc=3, scale=.1, size=n)
# Buisness Volume time 1
bv1 = bv0 + 0.1 + np.random.normal(loc=0, scale=.1, size=n)
# create an outlier
bv1[0] -= 10
# Paired t-test
print(stats.ttest_rel(bv0, bv1))
# Wilcoxon
print(stats.wilcoxon(bv0, bv1))
Ttest_relResult(statistic=0.7508746146807028, pvalue=0.4619272512287318)
WilcoxonResult(statistic=28.0, pvalue=0.002712249755859375)
Mann–Whitney 𝑈 test (quantitative ~ categorial (2 levels))
In statistics, the Mann–Whitney 𝑈 test (also called the Mann–Whitney–Wilcoxon, Wilcoxon

rank-sum test or Wilcoxon–Mann–Whitney test) is a nonparametric test of the null hypothesis
that two samples come from the same population against an alternative hypothesis, especially
that a particular population tends to have larger values than the other.
It can be applied on unknown distributions contrary to e.g. a 𝑡-test that has to be applied only
on normal distributions, and it is nearly as efficient as the 𝑡-test on normal distributions.

n = 20
# Buismess Volume group 0
bv0 = np.random.normal(loc=1, scale=.1, size=n)
# Buismess Volume group 1



bv1 = np.random.normal(loc=1.2, scale=.1, size=n)
# create an outlier
bv1[0] -= 10
# Two-samples t-test
print(stats.ttest_ind(bv0, bv1))
# Wilcoxon
print(stats.mannwhitneyu(bv0, bv1))
Ttest_indResult(statistic=0.5170316045185565, pvalue=0.6081306040701799)
MannwhitneyuResult(statistic=45.0, pvalue=1.4624324663684467e-05)
4.1.9 Linear model
Fig. 3: Linear model

Given 𝑛 random samples (𝑦𝑖 , 𝑥1𝑖 , . . . , 𝑥𝑝𝑖 ), 𝑖 = 1, . . . , 𝑛, the linear regression models the relation
between the observations 𝑦𝑖 and the independent variables 𝑥𝑝𝑖 is formulated as
𝑦𝑖 = 𝛽0 + 𝛽1 𝑥1𝑖 + · · · + 𝛽𝑝 𝑥𝑝𝑖 + 𝜀𝑖 𝑖 = 1, . . . , 𝑛
• The 𝛽’s are the model parameters, ie, the regression coeficients.
• 𝛽0 is the intercept or the bias.
• 𝜀𝑖 are the residuals.
• An independent variable (IV). It is a variable that stands alone and isn’t changed by
the other variables you are trying to measure. For example, someone’s age might be an
independent variable. Other factors (such as what they eat, how much they go to school,
how much television they watch) aren’t going to change a person’s age. In fact, when
you are looking for some kind of relationship between variables you are trying to see if
the independent variable causes some kind of change in the other variables, or dependent
variables. In Machine Learning, these variables are also called the predictors.
• A dependent variable. It is something that depends on other factors. For example, a test
score could be a dependent variable because it could change depending on several factors
such as how much you studied, how much sleep you got the night before you took the
test, or even how hungry you were when you took it. Usually when you are looking for
a relationship between two things you are trying to find out what makes the dependent
variable change the way it does. In Machine Learning this variable is called a target
variable.
Assumptions
1. Independence of residuals (𝜀𝑖 ). This assumptions must be satisfied

2. Normality of residuals (𝜀𝑖 ). Approximately normally distributed can be accepted.
Regression diagnostics: testing the assumptions of linear regression
Simple regression: test association between two quantitative variables
Using the dataset “salary”, explore the association between the dependant variable (e.g. Salary)
and the independent variable (e.g.: Experience is quantitative), considering only non-managers.
import pandas as pd
%matplotlib inline
salary = salary[salary.management == 'N']
1. Model the data
Model the data on some hypothesis e.g.: salary is a linear function of the experience.
salary𝑖 = 𝛽0 + 𝛽 experience𝑖 + 𝜖𝑖 ,
more generally
𝑦𝑖 = 𝛽0 + 𝛽 𝑥𝑖 + 𝜖𝑖
This can be rewritten in the matrix form using the design matrix made of values of independant
variable and the intercept:
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
𝑦1 1 𝑥1 𝜖1
⎢𝑦2 ⎥ ⎢1 𝑥2 ⎥ [︂ ]︂ ⎢𝜖2 ⎥
⎢𝑦3 ⎥ = ⎢1 𝑥3 ⎥ 𝛽0 + ⎢𝜖3 ⎥
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎣𝑦4 ⎦ ⎣1 𝑥4 ⎦ 𝛽1
⎢ ⎥ ⎢ ⎥ ⎢ ⎥
⎣𝜖4 ⎦
𝑦5 1 𝑥5 𝜖5
• 𝛽: the slope or coefficient or parameter of the model,

• 𝛽0 : the intercept or bias is the second parameter of the model,
• 𝜖𝑖 : is the 𝑖th error, or residual with 𝜖 ∼ 𝒩 (0, 𝜎 2 ).
The simple regression is equivalent to the Pearson correlation.

The goal it so estimate 𝛽, 𝛽0 and 𝜎 2 .

Minimizes the mean squared error (MSE) or the Sum squared error ∑︀ (SSE). The so-called
Ordinary Least Squares (OLS) finds 𝛽, 𝛽0 that minimizes the 𝑆𝑆𝐸 = 𝑖 𝜖2𝑖
∑︁
𝑆𝑆𝐸 = (𝑦𝑖 − 𝛽 𝑥𝑖 − 𝛽0 )2
𝑖
Recall from calculus that an extreme point can be found by computing where the derivative is
zero, i.e. to find the intercept, we perform the steps:
𝜕𝑆𝑆𝐸 ∑︁
= (𝑦𝑖 − 𝛽 𝑥𝑖 − 𝛽0 ) = 0
𝜕𝛽0
𝑖
∑︁ ∑︁
𝑦𝑖 = 𝛽 𝑥𝑖 + 𝑛 𝛽0
𝑖 𝑖
𝑛 𝑦¯ = 𝑛 𝛽 𝑥
¯ + 𝑛 𝛽0
𝛽0 = 𝑦¯ − 𝛽 𝑥
¯
To find the regression coefficient, we perform the steps:
𝜕𝑆𝑆𝐸 ∑︁
= 𝑥𝑖 (𝑦𝑖 − 𝛽 𝑥𝑖 − 𝛽0 ) = 0
𝜕𝛽
𝑖
Plug in 𝛽0 :
∑︁
𝑥𝑖 (𝑦𝑖 − 𝛽 𝑥𝑖 − 𝑦¯ + 𝛽 𝑥
¯) = 0
𝑖
∑︁ ∑︁ ∑︁
𝑥𝑖 𝑦𝑖 − 𝑦¯ 𝑥𝑖 = 𝛽 (𝑥𝑖 − 𝑥
¯)
𝑖 𝑖 𝑖
Divide both sides by 𝑛:

1 ∑︁ 1 ∑︁
𝑥𝑖 𝑦𝑖 − 𝑦¯𝑥¯= 𝛽 (𝑥𝑖 − 𝑥
¯)
𝑛 𝑛
𝑖 𝑖
1 ∑︀
𝑥𝑖 𝑦𝑖 − 𝑦¯𝑥
¯ 𝐶𝑜𝑣(𝑥, 𝑦)
𝛽 = 1 ∑︀𝑖
𝑛
= .
𝑛 𝑖 (𝑥𝑖 − 𝑥
¯) 𝑉 𝑎𝑟(𝑥)
from scipy import stats

import numpy as np
y, x = salary.salary, salary.experience
beta, beta0, r_value, p_value, std_err = stats.linregress(x,y)
print("y = %f x + %f, r: %f, r-squared: %f,\np-value: %f, std_err: %f"
% (beta, beta0, r_value, r_value**2, p_value, std_err))
print("Regression line with the scatterplot")

yhat = beta * x + beta0 # regression line
plt.plot(x, yhat, 'r-', x, y,'o')
plt.xlabel('Experience (years)')
plt.ylabel('Salary')
plt.show()
print("Using seaborn")
sns.regplot(x="experience", y="salary", data=salary)
y = 452.658228 x + 10785.911392, r: 0.965370, r-squared: 0.931939,

p-value: 0.000000, std_err: 24.970021
Regression line with the scatterplot
Using seaborn
<AxesSubplot:xlabel='experience', ylabel='salary'>

Regression measures of goodness of fit (losses)
R-squared
The goodness of fit of a statistical model describes how well it fits a set of observations. Mea-
sures of goodness of fit typically summarize the discrepancy between observed values and the
values expected under the model in question. We will consider the explained variance also
known as the coefficient of determination, denoted 𝑅2 pronounced R-squared.
The total sum of squares, 𝑆𝑆tot is the sum of the sum of squares explained by the regression,
𝑆𝑆reg , plus the sum of squares of residuals unexplained by the regression, 𝑆𝑆res , also called the
SSE, i.e. such that
𝑆𝑆tot = 𝑆𝑆reg + 𝑆𝑆res
Fig. 4: title
The mean of 𝑦 is
1 ∑︁
𝑦¯ = 𝑦𝑖 .
𝑛
𝑖
The total sum of squares is the total squared sum of deviations from the mean of 𝑦, i.e.
∑︁
𝑆𝑆tot = (𝑦𝑖 − 𝑦¯)2
𝑖
The regression sum of squares, also called the explained sum of squares:
∑︁
𝑆𝑆reg = 𝑦𝑖 − 𝑦¯)2 ,
(ˆ
𝑖
where 𝑦ˆ𝑖 = 𝛽𝑥𝑖 + 𝛽0 is the estimated value of salary 𝑦ˆ𝑖 given a value of experience 𝑥𝑖 .
The sum of squares of the residuals, also called the residual sum of squares (RSS) is:
∑︁
𝑆𝑆res = (𝑦𝑖 − 𝑦ˆ𝑖 )2 .
𝑖
𝑅2 is the explained sum of squares of errors. It is the variance explain by the regression divided
by the total variance, i.e.
explained SS 𝑆𝑆reg 𝑆𝑆𝑟𝑒𝑠
𝑅2 = = =1− .
total SS 𝑆𝑆𝑡𝑜𝑡 𝑆𝑆𝑡𝑜𝑡

Test
ˆ 2 = 𝑆𝑆res /(𝑛 − 2) be an estimator of the variance of 𝜖. The 2 in the denominator stems
Let 𝜎
from the 2 estimated parameters: intercept and coefficient.
𝑆𝑆res
• Unexplained variance: ^2
𝜎
∼ 𝜒2𝑛−2
𝑆𝑆
• Explained variance: 𝜎^ 2reg ∼ 𝜒21 . The single degree of freedom comes from the difference
between 𝑆𝑆
^2
𝜎
tot
(∼ 𝜒2𝑛−1 ) and 𝑆𝑆
^2
𝜎
res
(∼ 𝜒2𝑛−2 ), i.e. (𝑛 − 1) − (𝑛 − 2) degree of freedom.
The Fisher statistics of the ratio of two variances:
Explained variance 𝑆𝑆reg /1
𝐹 = = ∼ 𝐹 (1, 𝑛 − 2)
Unexplained variance 𝑆𝑆res /(𝑛 − 2)
Using the 𝐹 -distribution, compute the probability of observing a value greater than 𝐹 under
𝐻0 , i.e.: 𝑃 (𝑥 > 𝐹 |𝐻0 ), i.e. the survival function (1 − Cumulative Distribution Function) at 𝑥 of
the given 𝐹 -distribution.
res = y - (x * beta + beta0)
y_mu = np.mean(y)
ss_tot = np.sum((y - y_mu) ** 2)
ss_res = np.sum(res ** 2)
print("r-squared: %.3f" % (1 - ss_res / ss_tot))
r-squared: 0.932
MSE and MAE
• MSE: Mean Squared Error

• MAE: Mean Absolute Error
print("MSE %.3f" % np.mean(res ** 2))

print("MAE %.3f" % np.mean(np.abs(res)))
MSE 459923.068
MAE 602.265
Multiple regression
Theory
Muliple Linear Regression is the most basic supervised learning algorithm.

Given: a set of training data {𝑥1 , ..., 𝑥𝑁 } with corresponding targets {𝑦1 , ..., 𝑦𝑁 }.
In linear regression, we assume that the model that generates the data involves only a linear
combination of the input variables, i.e.
𝑦𝑖 = 𝛽0 + 𝛽1 𝑥𝑖1 + ... + 𝛽𝑃 𝑥𝑖𝑃 + 𝜀𝑖 ,

or, simplified
𝑃 −1
𝛽𝑗 𝑥𝑗𝑖 + 𝜀𝑖 .
∑︁
𝑦𝑖 = 𝛽0 +
𝑗=1
Extending each sample with an intercept, 𝑥𝑖 := [1, 𝑥𝑖 ] ∈ 𝑅𝑃 +1 allows us to use a more general
notation based on linear algebra and write it as a simple dot product:
𝑦𝑖 = x𝑇𝑖 𝛽 + 𝜀𝑖 ,
where 𝛽 ∈ 𝑅𝑃 +1 is a vector of weights that define the 𝑃 + 1 parameters of the model. From
now we have 𝑃 regressors + the intercept.
Using the matrix notation:
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
𝑦1 1 𝑥11 ... 𝑥1𝑃 ⎡ 𝛽 ⎤ 𝜖1
0
⎢𝑦2 ⎥ ⎢1 𝑥21 ... 𝑥2𝑃 ⎥ ⎢ 𝛽 ⎥ ⎢𝜖2 ⎥
⎥ ⎢
⎢ ⎥ ⎢
⎢𝑦3 ⎥ = ⎢1 ⎢ 1⎥ ⎢ ⎥
𝑥31 ... 𝑥3𝑃 ⎥
⎥⎢ .. ⎥ + ⎢𝜖3 ⎥
𝑥4𝑃 ⎦ .
⎢ ⎥ ⎢ ⎦ ⎣ ⎥
⎣𝑦4 ⎦ ⎣1 𝑥41 ...
⎣
𝜖4 ⎦
𝑦5 1 𝑥5 ... 𝑥5 𝛽 𝑃 𝜖5
Let 𝑋 = [𝑥𝑇0 , ..., 𝑥𝑇𝑁 ] be the (𝑁 × 𝑃 + 1) design matrix of 𝑁 samples of 𝑃 input features with
one column of one and let be 𝑦 = [𝑦1 , ..., 𝑦𝑁 ] be a vector of the 𝑁 targets.
𝑦 = 𝑋𝛽 + 𝜀
Minimize the Mean Squared Error MSE loss:

𝑁
1 ∑︁
𝑀 𝑆𝐸(𝛽) == (𝑦𝑖 − x𝑇𝑖 𝛽)2
𝑁
𝑖=1
Using the matrix notation, the mean squared error (MSE) loss can be rewritten:
1
𝑀 𝑆𝐸(𝛽) = ||𝑦 − 𝑋𝛽||22 .
𝑁
The 𝛽 that minimises the MSE can be found by:
(︂ )︂
1
∇𝛽 ||𝑦 − 𝑋𝛽||22 =0 (4.25)
𝑁
1
∇𝛽 (𝑦 − 𝑋𝛽)𝑇 (𝑦 − 𝑋𝛽) = 0 (4.26)
𝑁
1
∇𝛽 (𝑦 𝑇 𝑦 − 2𝛽 𝑇 𝑋 𝑇 𝑦 + 𝛽 𝑇 𝑋 𝑇 𝑋𝛽) = 0 (4.27)
𝑁
−2𝑋 𝑇 𝑦 + 2𝑋 𝑇 𝑋𝛽 = 0 (4.28)
𝑇 𝑇
𝑋 𝑋𝛽 = 𝑋 𝑦 (4.29)
𝛽 = (𝑋 𝑇 𝑋)−1 𝑋 𝑇 𝑦, (4.30)
where (𝑋 𝑇 𝑋)−1 𝑋 𝑇 is a pseudo inverse of 𝑋.

Fit with numpy
import numpy as np
from scipy import linalg
np.random.seed(seed=42) # make the example reproducible
# Dataset
N, P = 50, 4
X = np.random.normal(size= N * P).reshape((N, P))
## Our model needs an intercept so we add a column of 1s:
X[:, 0] = 1
print(X[:5, :])
betastar = np.array([10, 1., .5, 0.1])

e = np.random.normal(size=N)
y = np.dot(X, betastar) + e
# Estimate the parameters

Xpinv = linalg.pinv2(X)
betahat = np.dot(Xpinv, y)
print("Estimated beta:\n", betahat)
[[ 1. -0.1382643 0.64768854 1.52302986]

[ 1. -0.23413696 1.57921282 0.76743473]
[ 1. 0.54256004 -0.46341769 -0.46572975]
[ 1. -1.91328024 -1.72491783 -0.56228753]
[ 1. 0.31424733 -0.90802408 -1.4123037 ]]
Estimated beta:
[10.14742501 0.57938106 0.51654653 0.17862194]
4.1.10 Linear model with statsmodels
Sources: http://statsmodels.sourceforge.net/devel/examples/
Multiple regression
Interface with statsmodels
## Fit and summary:

model = sm.OLS(y, X).fit()
print(model.summary())
# prediction of new values

ypred = model.predict(X)
# residuals + prediction == true values

assert np.all(ypred + model.resid == y)
OLS Regression Results

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


Dep. Variable: y R-squared: 0.363
Model: OLS Adj. R-squared: 0.322
Method: Least Squares F-statistic: 8.748
Date: Thu, 29 Oct 2020 Prob (F-statistic): 0.000106
Time: 17:54:46 Log-Likelihood: -71.271
No. Observations: 50 AIC: 150.5
Df Residuals: 46 BIC: 158.2
Df Model: 3
Covariance Type: nonrobust
==============================================================================
coef std err t P>|t| [0.025 0.975]
------------------------------------------------------------------------------
const 10.1474 0.150 67.520 0.000 9.845 10.450
x1 0.5794 0.160 3.623 0.001 0.258 0.901
x2 0.5165 0.151 3.425 0.001 0.213 0.820
x3 0.1786 0.144 1.240 0.221 -0.111 0.469
==============================================================================
Omnibus: 2.493 Durbin-Watson: 2.369
Prob(Omnibus): 0.288 Jarque-Bera (JB): 1.544
Skew: 0.330 Prob(JB): 0.462
Kurtosis: 3.554 Cond. No. 1.27
==============================================================================
Notes:
[1] Standard Errors assume that the covariance matrix of the errors is correctly␣
˓→specified.
Interface with Pandas
Use R language syntax for data.frame. For an additive model: 𝑦𝑖 = 𝛽 0 + 𝑥1𝑖 𝛽 1 + 𝑥2𝑖 𝛽 2 + 𝜖𝑖 ≡ y ~
x1 + x2.
import statsmodels.formula.api as smf
df = pd.DataFrame(np.column_stack([X, y]), columns=['inter', 'x1','x2', 'x3', 'y'])

print(df.columns, df.shape)
# Build a model excluding the intercept, it is implicit
model = smf.ols("y~x1 + x2 + x3", df).fit()
Index(['inter', 'x1', 'x2', 'x3', 'y'], dtype='object') (50, 5)

==============================================================================
Dep. Variable: y R-squared: 0.363
Date: Thu, 29 Oct 2020 Prob (F-statistic): 0.000106
No. Observations: 50 AIC: 150.5
Df Residuals: 46 BIC: 158.2
Df Model: 3
==============================================================================


coef std err t P>|t| [0.025 0.975]
------------------------------------------------------------------------------
Intercept 10.1474 0.150 67.520 0.000 9.845 10.450
x1 0.5794 0.160 3.623 0.001 0.258 0.901
x2 0.5165 0.151 3.425 0.001 0.213 0.820
x3 0.1786 0.144 1.240 0.221 -0.111 0.469
==============================================================================
Skew: 0.330 Prob(JB): 0.462
==============================================================================
Notes:
˓→specified.
Multiple regression with categorical independent variables or factors: Analysis of covariance

(ANCOVA)
Analysis of covariance (ANCOVA) is a linear model that blends ANOVA and linear regression.
ANCOVA evaluates whether population means of a dependent variable (DV) are equal across
levels of a categorical independent variable (IV) often called a treatment, while statistically
controlling for the effects of other quantitative or continuous variables that are not of primary
interest, known as covariates (CV).
import pandas as pd
%matplotlib inline
try:
df = pd.read_csv("../datasets/salary_table.csv")
except:
df = pd.read_csv(url)
import statsmodels.formula.api as smf

import statsmodels.stats.api as sms
lm = smf.ols('salary ~ experience', df).fit()

df["residuals"] = lm.resid
print("Jarque-Bera normality test p-value %.5f" % sms.jarque_bera(oneway.resid)[1])
ax = sns.displot(df, x='residuals', kind="kde", fill=True)

ax = sns.displot(df, x='residuals', kind="kde", hue='management', fill=True)
Jarque-Bera normality test p-value 0.00359


Normality assumption of the residuals can be rejected (p-value < 0.05). There is an efect of the
“management” factor, take it into account.
One-way AN(C)OVA
• ANOVA: one categorical independent variable, i.e. one factor.

• ANCOVA: ANOVA with some covariates.
oneway = smf.ols('salary ~ management + experience', df).fit()

df["residuals"] = oneway.resid
sns.displot(df, x='residuals', kind="kde", fill=True)
print(sm.stats.anova_lm(oneway, typ=2))
print("Jarque-Bera normality test p-value %.3f" % sms.jarque_bera(oneway.resid)[1])
sum_sq df F PR(>F)
management 5.755739e+08 1.0 183.593466 4.054116e-17
experience 3.334992e+08 1.0 106.377768 3.349662e-13
Residual 1.348070e+08 43.0 NaN NaN

Distribution of residuals is still not normal but closer to normality. Both management and
experience are significantly associated with salary.
Two-way AN(C)OVA
Ancova with two categorical independent variables, i.e. two factors.
twoway = smf.ols('salary ~ education + management + experience', df).fit()

df["residuals"] = twoway.resid
sns.displot(df, x='residuals', kind="kde", fill=True)
print(sm.stats.anova_lm(twoway, typ=2))
print("Jarque-Bera normality test p-value %.3f" % sms.jarque_bera(twoway.resid)[1])
sum_sq df F PR(>F)
education 9.152624e+07 2.0 43.351589 7.672450e-11
management 5.075724e+08 1.0 480.825394 2.901444e-24
experience 3.380979e+08 1.0 320.281524 5.546313e-21
Residual 4.328072e+07 41.0 NaN NaN

Normality assumtion cannot be rejected. Assume it. Education, management and experience
are significantly associated with salary.
Comparing two nested models
oneway is nested within twoway. Comparing two nested models tells us if the additional predic-
tors (i.e. education) of the full model significantly decrease the residuals. Such comparison can
be done using an 𝐹 -test on residuals:
print(twoway.compare_f_test(oneway)) # return F, pval, df
(43.35158945918107, 7.672449570495418e-11, 2.0)
twoway is significantly better than one way

Factor coding
See http://statsmodels.sourceforge.net/devel/contrasts.html
By default Pandas use “dummy coding”. Explore:
print(twoway.model.data.param_names)
print(twoway.model.data.exog[:10, :])
['Intercept', 'education[T.Master]', 'education[T.Ph.D]', 'management[T.Y]', 'experience']

[[1. 0. 0. 1. 1.]
[1. 0. 1. 0. 1.]
[1. 0. 1. 1. 1.]
[1. 1. 0. 0. 1.]
[1. 0. 1. 0. 1.]
[1. 1. 0. 1. 2.]
[1. 1. 0. 0. 2.]
[1. 0. 0. 0. 2.]
[1. 0. 1. 0. 2.]
[1. 1. 0. 0. 3.]]
Contrasts and post-hoc tests
# t-test of the specific contribution of experience:

ttest_exp = twoway.t_test([0, 0, 0, 0, 1])
ttest_exp.pvalue, ttest_exp.tvalue
print(ttest_exp)
# Alternatively, you can specify the hypothesis tests using a string

twoway.t_test('experience')
# Post-hoc is salary of Master different salary of Ph.D?

# ie. t-test salary of Master = salary of Ph.D.
print(twoway.t_test('education[T.Master] = education[T.Ph.D]'))
Test for Constraints

==============================================================================
coef std err t P>|t| [0.025 0.975]
------------------------------------------------------------------------------
c0 546.1840 30.519 17.896 0.000 484.549 607.819
==============================================================================
Test for Constraints
==============================================================================
coef std err t P>|t| [0.025 0.975]
------------------------------------------------------------------------------
c0 147.8249 387.659 0.381 0.705 -635.069 930.719
==============================================================================

4.1.11 Multiple comparisons
import numpy as np
np.random.seed(seed=42) # make example reproducible
# Dataset
n_samples, n_features = 100, 1000
n_info = int(n_features/10) # number of features with information
n1, n2 = int(n_samples/2), n_samples - int(n_samples/2)
snr = .5
Y = np.random.randn(n_samples, n_features)
grp = np.array(["g1"] * n1 + ["g2"] * n2)
# Add some group effect for Pinfo features

Y[grp=="g1", :n_info] += snr
#
tvals, pvals = np.full(n_features, np.NAN), np.full(n_features, np.NAN)
for j in range(n_features):
tvals[j], pvals[j] = stats.ttest_ind(Y[grp=="g1", j], Y[grp=="g2", j],
equal_var=True)
fig, axis = plt.subplots(3, 1)#, sharex='col')
axis[0].plot(range(n_features), tvals, 'o')

axis[0].set_ylabel("t-value")
axis[1].plot(range(n_features), pvals, 'o')

axis[1].axhline(y=0.05, color='red', linewidth=3, label="p-value=0.05")
#axis[1].axhline(y=0.05, label="toto", color='red')
axis[1].set_ylabel("p-value")
axis[1].legend()
axis[2].hist([pvals[n_info:], pvals[:n_info]],
stacked=True, bins=100, label=["Negatives", "Positives"])
axis[2].set_xlabel("p-value histogram")
axis[2].set_ylabel("density")
axis[2].legend()
plt.tight_layout()

Note that under the null hypothesis the distribution of the p-values is uniform.
Statistical measures:
• True Positive (TP) equivalent to a hit. The test correctly concludes the presence of an
effect.
• True Negative (TN). The test correctly concludes the absence of an effect.
• False Positive (FP) equivalent to a false alarm, Type I error. The test improperly con-
cludes the presence of an effect. Thresholding at 𝑝-value < 0.05 leads to 47 FP.
• False Negative (FN) equivalent to a miss, Type II error. The test improperly concludes the
absence of an effect.
P, N = n_info, n_features - n_info # Positives, Negatives

TP = np.sum(pvals[:n_info ] < 0.05) # True Positives
FP = np.sum(pvals[n_info: ] < 0.05) # False Positives
print("No correction, FP: %i (expected: %.2f), TP: %i" % (FP, N * 0.05, TP))
No correction, FP: 47 (expected: 45.00), TP: 71
Bonferroni correction for multiple comparisons
The Bonferroni correction is based on the idea that if an experimenter is testing 𝑃 hypothe-
ses, then one way of maintaining the familywise error rate (FWER) is to test each individual
hypothesis at a statistical significance level of 1/𝑃 times the desired maximum overall level.
So, if the desired significance level for the whole family of tests is 𝛼 (usually 0.05), then the
Bonferroni correction would test each individual hypothesis at a significance level of 𝛼/𝑃 . For
example, if a trial is testing 𝑃 = 8 hypotheses with a desired 𝛼 = 0.05, then the Bonferroni
correction would test each individual hypothesis at 𝛼 = 0.05/8 = 0.00625.

import statsmodels.sandbox.stats.multicomp as multicomp

_, pvals_fwer, _, _ = multicomp.multipletests(pvals, alpha=0.05,
method='bonferroni')
TP = np.sum(pvals_fwer[:n_info ] < 0.05) # True Positives
FP = np.sum(pvals_fwer[n_info: ] < 0.05) # False Positives
print("FWER correction, FP: %i, TP: %i" % (FP, TP))
FWER correction, FP: 0, TP: 6
The False discovery rate (FDR) correction for multiple comparisons
FDR-controlling procedures are designed to control the expected proportion of rejected null
hypotheses that were incorrect rejections (“false discoveries”). FDR-controlling procedures pro-
vide less stringent control of Type I errors compared to the familywise error rate (FWER) con-
trolling procedures (such as the Bonferroni correction), which control the probability of at least
one Type I error. Thus, FDR-controlling procedures have greater power, at the cost of increased
rates of Type I errors.
import statsmodels.sandbox.stats.multicomp as multicomp

_, pvals_fdr, _, _ = multicomp.multipletests(pvals, alpha=0.05,
method='fdr_bh')
TP = np.sum(pvals_fdr[:n_info ] < 0.05) # True Positives
FP = np.sum(pvals_fdr[n_info: ] < 0.05) # False Positives
print("FDR correction, FP: %i, TP: %i" % (FP, TP))
FDR correction, FP: 3, TP: 20
4.1.12 Exercises
Simple linear regression and correlation (application)
Load the dataset: birthwt Risk Factors Associated with Low Infant Birth Weight at https://raw.
github.com/neurospin/pystatsml/master/datasets/birthwt.csv
1. Test the association of mother’s (bwt) age and birth weight using the correlation test and
linear regeression.
2. Test the association of mother’s weight (lwt) and birth weight using the correlation test
and linear regeression.
3. Produce two scatter plot of: (i) age by birth weight; (ii) mother’s weight by birth weight.
Conclusion ?

Simple linear regression (maths)
Considering the salary and the experience of the salary table. https://raw.github.com/
neurospin/pystatsml/master/datasets/salary_table.csv
Compute:
• Estimate the model paramters 𝛽, 𝛽0 using scipy stats.linregress(x,y)
• Compute the predicted values 𝑦ˆ
Compute:
• 𝑦¯: y_mu
• 𝑆𝑆tot : ss_tot
• 𝑆𝑆reg : ss_reg
• 𝑆𝑆res : ss_res
• Check partition of variance formula based on sum of squares by using assert np.
allclose(val1, val2, atol=1e-05)
• Compute 𝑅2 and compare it with the r_value above
• Compute the 𝐹 score
• Compute the 𝑝-value:
• Plot the 𝐹 (1, 𝑛) distribution for 100 𝑓 values within [10, 25]. Draw 𝑃 (𝐹 (1, 𝑛) > 𝐹 ),
i.e. color the surface defined by the 𝑥 values larger than 𝐹 below the 𝐹 (1, 𝑛).
• 𝑃 (𝐹 (1, 𝑛) > 𝐹 ) is the 𝑝-value, compute it.
Multiple regression
Considering the simulated data used below:

1. What are the dimensions of pinv(𝑋)?
2. Compute the MSE between the predicted values and the true values.
import numpy as np
from scipy import linalg
np.random.seed(seed=42) # make the example reproducible
# Dataset
N, P = 50, 4
X = np.random.normal(size= N * P).reshape((N, P))
## Our model needs an intercept so we add a column of 1s:
X[:, 0] = 1
print(X[:5, :])
betastar = np.array([10, 1., .5, 0.1])

e = np.random.normal(size=N)
y = np.dot(X, betastar) + e
# Estimate the parameters

Xpinv = linalg.pinv2(X)


betahat = np.dot(Xpinv, y)
print("Estimated beta:\n", betahat)
[[ 1. -0.1382643 0.64768854 1.52302986]

[ 1. -0.23413696 1.57921282 0.76743473]
[ 1. 0.54256004 -0.46341769 -0.46572975]
[ 1. -1.91328024 -1.72491783 -0.56228753]
[ 1. 0.31424733 -0.90802408 -1.4123037 ]]
Estimated beta:
[10.14742501 0.57938106 0.51654653 0.17862194]
Two sample t-test (maths)
Given the following two sample, test whether their means are equals.
height = np.array([ 1.83, 1.83, 1.73, 1.82, 1.83,

1.73,1.99, 1.85, 1.68, 1.87,
1.66, 1.71, 1.73, 1.64, 1.70,
1.60, 1.79, 1.73, 1.62, 1.77])
grp = np.array(["M"] * 10 + ["F"] * 10)
• Compute the means/std-dev per groups.

• Compute the 𝑡-value (standard two sample t-test with equal variances).
• Compute the 𝑝-value.
• The 𝑝-value is one-sided: a two-sided test would test P(T > tval) and P(T < -tval).
What would the two sided 𝑝-value be?
• Compare the two-sided 𝑝-value with the one obtained by stats.ttest_ind using assert
np.allclose(arr1, arr2).
Two sample t-test (application)
Risk Factors Associated with Low Infant Birth Weight: https://raw.github.com/neurospin/

pystatsml/master/datasets/birthwt.csv
1. Explore the data
2. Recode smoke factor
3. Compute the means/std-dev per groups.
4. Plot birth weight by smoking (box plot, violin plot or histogram)
5. Test the effect of smoking on birth weight

Two sample t-test and random permutations
Generate 100 samples following the model:
𝑦 =𝑔+𝜀
Where the noise 𝜀 ∼ 𝑁 (1, 1) and 𝑔 ∈ {0, 1} is a group indicator variable with 50 ones and 50
zeros.
• Write a function tstat(y, g) that compute the two samples t-test of y splited in two
groups defined by g.
• Sample the t-statistic distribution under the null hypothesis using random permutations.
• Assess the p-value.
Univariate associations (developpement)
Write a function univar_stat(df, target, variables) that computes the parametric statistics
and 𝑝-values between the target variable (provided as as string) and all variables (provided
as a list of string) of the pandas DataFrame df. The target is a quantitative variable but vari-
ables may be quantitative or qualitative. The function returns a DataFrame with four columns:
variable, test, value, p_value.
Apply it to the salary dataset available at https://raw.github.com/neurospin/pystatsml/master/
datasets/salary_table.csv, with target being S: salaries for IT staff in a corporation.
Multiple comparisons
This exercise has 2 goals: apply you knowledge of statistics using vectorized numpy operations.
Given the dataset provided for multiple comparisons, compute the two-sample 𝑡-test (assuming
equal variance) for each (column) feature of the Y array given the two groups defined by grp
variable. You should return two vectors of size n_features: one for the 𝑡-values and one for the
𝑝-values.
ANOVA
Perform an ANOVA dataset described bellow

• Compute between and within variances
• Compute 𝐹 -value: fval
• Compare the 𝑝-value with the one obtained by stats.f_oneway using assert np.
allclose(arr1, arr2)
# dataset
mu_k = np.array([1, 2, 3]) # means of 3 samples
sd_k = np.array([1, 1, 1]) # sd of 3 samples
n_k = np.array([10, 20, 30]) # sizes of 3 samples
grp = [0, 1, 2] # group labels
n = np.sum(n_k)
label = np.hstack([[k] * n_k[k] for k in [0, 1, 2]])

y = np.zeros(n)
for k in grp:
y[label == k] = np.random.normal(mu_k[k], sd_k[k], n_k[k])

fval, pval = stats.f_oneway(y[label == 0], y[label == 1], y[label == 2])
4.2 Lab 1: Brain volumes study
The study provides the brain volumes of grey matter (gm), white matter (wm) and cerebrospinal
fluid) (csf) of 808 anatomical MRI scans.
4.2.1 Manipulate data
Set the working directory within a directory called “brainvol”

Create 2 subdirectories: data that will contain downloaded data and reports for results of the
analysis.
import os
import os.path
import pandas as pd
import tempfile
import urllib.request
WD = os.path.join(tempfile.gettempdir(), "brainvol")
os.makedirs(WD, exist_ok=True)
#os.chdir(WD)
# use cookiecutter file organization

# https://drivendata.github.io/cookiecutter-data-science/
os.makedirs(os.path.join(WD, "data"), exist_ok=True)
#os.makedirs("reports", exist_ok=True)
Fetch data
• Demographic data demo.csv (columns: participant_id, site, group, age, sex) and tissue
volume data: group is Control or Patient. site is the recruiting site.
• Gray matter volume gm.csv (columns: participant_id, session, gm_vol)
• White matter volume wm.csv (columns: participant_id, session, wm_vol)
• Cerebrospinal Fluid csf.csv (columns: participant_id, session, csf_vol)
base_url = 'https://raw.github.com/neurospin/pystatsml/master/datasets/brain_volumes/%s'
data = dict()
for file in ["demo.csv", "gm.csv", "wm.csv", "csf.csv"]:
4.2. Lab 1: Brain volumes study 117


urllib.request.urlretrieve(base_url % file, os.path.join(WD, "data", file))
demo = pd.read_csv(os.path.join(WD, "data", "demo.csv"))

gm = pd.read_csv(os.path.join(WD, "data", "gm.csv"))
wm = pd.read_csv(os.path.join(WD, "data", "wm.csv"))
csf = pd.read_csv(os.path.join(WD, "data", "csf.csv"))
print("tables can be merge using shared columns")

print(gm.head())
Out:
tables can be merge using shared columns

participant_id session gm_vol
0 sub-S1-0002 ses-01 0.672506
1 sub-S1-0002 ses-02 0.678772
2 sub-S1-0002 ses-03 0.665592
3 sub-S1-0004 ses-01 0.890714
4 sub-S1-0004 ses-02 0.881127
Merge tables according to participant_id
brain_vol = pd.merge(pd.merge(pd.merge(demo, gm), wm), csf)

assert brain_vol.shape == (808, 9)
Drop rows with missing values
brain_vol = brain_vol.dropna()
assert brain_vol.shape == (766, 9)
Compute Total Intra-cranial volume tiv_vol = gm_vol + csf_vol + wm_vol.
brain_vol["tiv_vol"] = brain_vol["gm_vol"] + brain_vol["wm_vol"] + brain_vol["csf_vol"]
Compute tissue fractions gm_f = gm_vol / tiv_vol, wm_f = wm_vol / tiv_vol.
brain_vol["gm_f"] = brain_vol["gm_vol"] / brain_vol["tiv_vol"]

brain_vol["wm_f"] = brain_vol["wm_vol"] / brain_vol["tiv_vol"]
Save in a excel file brain_vol.xlsx
brain_vol.to_excel(os.path.join(WD, "data", "brain_vol.xlsx"),

sheet_name='data', index=False)
4.2.2 Descriptive Statistics
Load excel file brain_vol.xlsx
import os
import pandas as pd
import statsmodels.formula.api as smfrmla

brain_vol = pd.read_excel(os.path.join(WD, "data", "brain_vol.xlsx"),

sheet_name='data')
# Round float at 2 decimals when printing
pd.options.display.float_format = '{:,.2f}'.format
Descriptive statistics Most of participants have several MRI sessions (column session) Select
on rows from session one “ses-01”
brain_vol1 = brain_vol[brain_vol.session == "ses-01"]

# Check that there are no duplicates
assert len(brain_vol1.participant_id.unique()) == len(brain_vol1.participant_id)
Global descriptives statistics of numerical variables
desc_glob_num = brain_vol1.describe()
print(desc_glob_num)
Out:
age gm_vol wm_vol csf_vol tiv_vol gm_f wm_f

count 244.00 244.00 244.00 244.00 244.00 244.00 244.00
mean 34.54 0.71 0.44 0.31 1.46 0.49 0.30
std 12.09 0.08 0.07 0.08 0.17 0.04 0.03
min 18.00 0.48 0.05 0.12 0.83 0.37 0.06
25% 25.00 0.66 0.40 0.25 1.34 0.46 0.28
50% 31.00 0.70 0.43 0.30 1.45 0.49 0.30
75% 44.00 0.77 0.48 0.37 1.57 0.52 0.31
max 61.00 1.03 0.62 0.63 2.06 0.60 0.36
Global Descriptive statistics of categorical variable
desc_glob_cat = brain_vol1[["site", "group", "sex"]].describe(include='all')

print(desc_glob_cat)
print("Get count by level")

desc_glob_cat = pd.DataFrame({col: brain_vol1[col].value_counts().to_dict()
for col in ["site", "group", "sex"]})
Out:
site group sex

count 244 244 244
unique 7 2 2
top S7 Patient M
freq 65 157 155
Get count by level
site group sex
S7 65.00 nan nan
S5 62.00 nan nan
S8 59.00 nan nan
S3 29.00 nan nan
S4 15.00 nan nan
S1 13.00 nan nan


S6 1.00 nan nan
Patient nan 157.00 nan
Control nan 87.00 nan
M nan nan 155.00
F nan nan 89.00
Remove the single participant from site 6
brain_vol = brain_vol[brain_vol.site != "S6"]

brain_vol1 = brain_vol[brain_vol.session == "ses-01"]
desc_glob_cat = pd.DataFrame({col: brain_vol1[col].value_counts().to_dict()
for col in ["site", "group", "sex"]})
Out:
site group sex

S7 65.00 nan nan
S5 62.00 nan nan
S8 59.00 nan nan
S3 29.00 nan nan
S4 15.00 nan nan
S1 13.00 nan nan
Patient nan 157.00 nan
Control nan 86.00 nan
M nan nan 155.00
F nan nan 88.00
Descriptives statistics of numerical variables per clinical status
desc_group_num = brain_vol1[["group", 'gm_vol']].groupby("group").describe()

print(desc_group_num)
Out:
gm_vol
count meanstd min 25% 50% 75% max
group
Control 86.00 0.72 0.09 0.48 0.66 0.71 0.78 1.03
Patient 157.00 0.70 0.08 0.53 0.65 0.70 0.76 0.90
4.2.3 Statistics
Objectives:
1. Site effect of gray matter atrophy
2. Test the association between the age and gray matter atrophy in the control and patient
population independently.
3. Test for differences of atrophy between the patients and the controls
4. Test for interaction between age and clinical status, ie: is the brain atrophy process in
patient population faster than in the control population.
5. The effect of the medication in the patient population.

import statsmodels.formula.api as smfrmla
import scipy.stats
1 Site effect on Grey Matter atrophy

The model is Oneway Anova gm_f ~ site The ANOVA test has important assumptions that must
be satisfied in order for the associated p-value to be valid.
• The samples are independent.
• Each sample is from a normally distributed population.
• The population standard deviations of the groups are all equal. This property is known as
homoscedasticity.
Plot
sns.violinplot("site", "gm_f", data=brain_vol1)
Out:
/home/ed203246/anaconda3/lib/python3.7/site-packages/seaborn/_decorators.py:43:␣
˓→FutureWarning: Pass the following variables as keyword args: x, y. From version 0.12,␣
˓→the only valid positional argument will be `data`, and passing other arguments without␣
˓→an explicit keyword will result in an error or misinterpretation.
FutureWarning

<AxesSubplot:xlabel='site', ylabel='gm_f'>
Stats with scipy
fstat, pval = scipy.stats.f_oneway(*[brain_vol1.gm_f[brain_vol1.site == s]

for s in brain_vol1.site.unique()])
print("Oneway Anova gm_f ~ site F=%.2f, p-value=%E" % (fstat, pval))
Out:
Oneway Anova gm_f ~ site F=14.82, p-value=1.188136E-12
Stats with statsmodels
anova = smfrmla.ols("gm_f ~ site", data=brain_vol1).fit()

# print(anova.summary())
print("Site explains %.2f%% of the grey matter fraction variance" %
(anova.rsquared * 100))
print(sm.stats.anova_lm(anova, typ=2))
Out:
Site explains 23.82% of the grey matter fraction variance

sum_sq df F PR(>F)
site 0.11 5.00 14.82 0.00
Residual 0.35 237.00 nan nan
2. Test the association between the age and gray matter atrophy in the control and patient
population independently.
Plot
sns.lmplot("age", "gm_f", hue="group", data=brain_vol1)
brain_vol1_ctl = brain_vol1[brain_vol1.group == "Control"]

brain_vol1_pat = brain_vol1[brain_vol1.group == "Patient"]

Out:

FutureWarning
Stats with scipy
print("--- In control population ---")

beta, beta0, r_value, p_value, std_err = \
scipy.stats.linregress(x=brain_vol1_ctl.age, y=brain_vol1_ctl.gm_f)
print("gm_f = %f * age + %f" % (beta, beta0))

print("Corr: %f, r-squared: %f, p-value: %f, std_err: %f"\
% (r_value, r_value**2, p_value, std_err))
print("--- In patient population ---")

beta, beta0, r_value, p_value, std_err = \
scipy.stats.linregress(x=brain_vol1_pat.age, y=brain_vol1_pat.gm_f)
print("gm_f = %f * age + %f" % (beta, beta0))

print("Corr: %f, r-squared: %f, p-value: %f, std_err: %f"\
% (r_value, r_value**2, p_value, std_err))
print("Decrease seems faster in patient than in control population")

Out:
--- In control population ---

gm_f = -0.001181 * age + 0.529829
Corr: -0.325122, r-squared: 0.105704, p-value: 0.002255, std_err: 0.000375
--- In patient population ---
gm_f = -0.001899 * age + 0.556886
Corr: -0.528765, r-squared: 0.279592, p-value: 0.000000, std_err: 0.000245
Decrease seems faster in patient than in control population
print("--- In control population ---")

lr = smfrmla.ols("gm_f ~ age", data=brain_vol1_ctl).fit()
print(lr.summary())
print("Age explains %.2f%% of the grey matter fraction variance" %
(lr.rsquared * 100))
print("--- In patient population ---")

lr = smfrmla.ols("gm_f ~ age", data=brain_vol1_pat).fit()
print(lr.summary())
print("Age explains %.2f%% of the grey matter fraction variance" %
(lr.rsquared * 100))
Out:
--- In control population ---

==============================================================================
Dep. Variable: gm_f R-squared: 0.106
Date: lun., 12 oct. 2020 Prob (F-statistic): 0.00226
Time: 00:32:37 Log-Likelihood: 159.34
No. Observations: 86 AIC: -314.7
Df Residuals: 84 BIC: -309.8
Df Model: 1
==============================================================================
coef std err t P>|t| [0.025 0.975]
------------------------------------------------------------------------------
Intercept 0.5298 0.013 40.350 0.000 0.504 0.556
age -0.0012 0.000 -3.151 0.002 -0.002 -0.000
==============================================================================
Skew: 0.233 Prob(JB): 0.676
Kurtosis: 2.962 Cond. No. 111.
==============================================================================
Notes:
˓→specified.
Age explains 10.57% of the grey matter fraction variance
--- In patient population ---
==============================================================================


Dep. Variable: gm_f R-squared: 0.280
Date: lun., 12 oct. 2020 Prob (F-statistic): 1.09e-12
Time: 00:32:37 Log-Likelihood: 289.38
No. Observations: 157 AIC: -574.8
Df Residuals: 155 BIC: -568.7
Df Model: 1
==============================================================================
coef std err t P>|t| [0.025 0.975]
------------------------------------------------------------------------------
Intercept 0.5569 0.009 60.817 0.000 0.539 0.575
age -0.0019 0.000 -7.756 0.000 -0.002 -0.001
==============================================================================
Skew: 0.230 Prob(JB): 0.396
Kurtosis: 3.268 Cond. No. 111.
==============================================================================
Notes:
˓→specified.
Age explains 27.96% of the grey matter fraction variance
Before testing for differences of atrophy between the patients ans the controls Preliminary tests
for age x group effect (patients would be older or younger than Controls)
Plot
sns.violinplot("group", "age", data=brain_vol1)

Out:

FutureWarning
<AxesSubplot:xlabel='group', ylabel='age'>
Stats with scipy

print(scipy.stats.ttest_ind(brain_vol1_ctl.age, brain_vol1_pat.age))
Out:
Ttest_indResult(statistic=-1.2155557697674162, pvalue=0.225343592508479)

print(smfrmla.ols("age ~ group", data=brain_vol1).fit().summary())
print("No significant difference in age between patients and controls")
Out:
==============================================================================
Dep. Variable: age R-squared: 0.006


Date: lun., 12 oct. 2020 Prob (F-statistic): 0.225
No. Observations: 243 AIC: 1903.
Df Residuals: 241 BIC: 1910.
Df Model: 1
====================================================================================
coef std err t P>|t| [0.025 0.975]
------------------------------------------------------------------------------------
Intercept 33.2558 1.305 25.484 0.000 30.685 35.826
group[T.Patient] 1.9735 1.624 1.216 0.225 -1.225 5.172
==============================================================================
Skew: 0.569 Prob(JB): 3.16e-05
==============================================================================
Notes:
˓→specified.
No significant difference in age between patients and controls
Preliminary tests for sex x group (more/less males in patients than in Controls)
crosstab = pd.crosstab(brain_vol1.sex, brain_vol1.group)

print("Obeserved contingency table")
print(crosstab)
chi2, pval, dof, expected = scipy.stats.chi2_contingency(crosstab)
print("Chi2 = %f, pval = %f" % (chi2, pval))

print("Expected contingency table under the null hypothesis")
print(expected)
print("No significant difference in sex between patients and controls")
Out:
Obeserved contingency table

group Control Patient
sex
F 33 55
M 53 102
Chi2 = 0.143253, pval = 0.705068
Expected contingency table under the null hypothesis
[[ 31.14403292 56.85596708]
[ 54.85596708 100.14403292]]
No significant difference in sex between patients and controls
3. Test for differences of atrophy between the patients and the controls
print(sm.stats.anova_lm(smfrmla.ols("gm_f ~ group", data=brain_vol1).fit(), typ=2))


Out:
sum_sq df F PR(>F)
group 0.00 1.00 0.01 0.92
This model is simplistic we should adjust for age and site
print(sm.stats.anova_lm(smfrmla.ols(
"gm_f ~ group + age + site", data=brain_vol1).fit(), typ=2))
Out:
sum_sq df F PR(>F)
group 0.00 1.00 1.82 0.18
site 0.11 5.00 19.79 0.00
age 0.09 1.00 86.86 0.00
4. Test for interaction between age and clinical status, ie: is the brain atrophy process in
patient population faster than in the control population.
ancova = smfrmla.ols("gm_f ~ group:age + age + site", data=brain_vol1).fit()

print(sm.stats.anova_lm(ancova, typ=2))
print("= Parameters =")

print(ancova.params)
print("%.3f%% of grey matter loss per year (almost %.1f%% per decade)" %\
(ancova.params.age * 100, ancova.params.age * 100 * 10))
print("grey matter loss in patients is accelerated by %.3f%% per decade" %

(ancova.params['group[T.Patient]:age'] * 100 * 10))
Out:
sum_sq df F PR(>F)
site 0.11 5.00 20.28 0.00
age 0.10 1.00 89.37 0.00
group:age 0.00 1.00 3.28 0.07
= Parameters =
Intercept 0.52
site[T.S3] 0.01
site[T.S4] 0.03
site[T.S5] 0.01
site[T.S7] 0.06
site[T.S8] 0.02
age -0.00
group[T.Patient]:age -0.00
dtype: float64
-0.148% of grey matter loss per year (almost -1.5% per decade)
grey matter loss in patients is accelerated by -0.232% per decade

4.3 Multivariate statistics
Multivariate statistics includes all statistical techniques for analyzing samples made of two or
more variables. The data set (a 𝑁 × 𝑃 matrix X) is a collection of 𝑁 independent samples
column vectors [x1 , . . . , x𝑖 , . . . , x𝑁 ] of length 𝑃
⎡ 𝑇 ⎤ ⎡ ⎤ ⎡ ⎤
−x1 − 𝑥11 · · · 𝑥1𝑗 · · · 𝑥1𝑃 𝑥11 . . . 𝑥1𝑃
⎢ .. ⎥ ⎢ .. .. .. ⎥ ⎢ .. .. ⎥
⎢ . ⎥ ⎢ . . . ⎥⎥ ⎢ . . ⎥
⎢
⎢ 𝑇 ⎥ ⎢ ⎥
X = ⎢−x𝑖 −⎥ = ⎢ 𝑥𝑖1 · · · 𝑥𝑖𝑗 · · · 𝑥𝑖𝑃 ⎥ = ⎢
⎢ ⎥ ⎢ ⎥ ⎢ X ⎥ .
⎢ .. ⎥ ⎢ .. .. .. ⎥ ⎢ .. .. ⎥
⎥
⎣ . ⎦ ⎣ . . . ⎦ ⎣ . . ⎦
𝑇
−x𝑃 − 𝑥𝑁 1 · · · 𝑥𝑁 𝑗 · · · 𝑥𝑁 𝑃 𝑥𝑁 1 . . . 𝑥𝑁 𝑃 𝑁 ×𝑃
4.3.1 Linear Algebra
Euclidean norm and distance
The Euclidean norm of a vector a ∈ R𝑃 is denoted

⎯
⎸ 𝑃
⎸∑︁
‖a‖2 = ⎷ 𝑎𝑖 2
𝑖
The Euclidean distance between two vectors a, b ∈ R𝑃 is

⎯
⎸ 𝑃
⎸∑︁
‖a − b‖2 = ⎷ (𝑎𝑖 − 𝑏𝑖 )2
𝑖
Dot product and projection
Source: Wikipedia
Algebraic definition
The dot product, denoted ’‘·” of two 𝑃 -dimensional vectors a = [𝑎1 , 𝑎2 , ..., 𝑎𝑃 ] and a =
[𝑏1 , 𝑏2 , ..., 𝑏𝑃 ] is defined as
⎡ ⎤
𝑏1
⎢ .. ⎥
]︀ ⎢ . ⎥
⎢ ⎥
∑︁
𝑇 𝑇
[︀
a·b=a b= 𝑎𝑖 𝑏𝑖 = 𝑎1 . . . a . . . 𝑎𝑃 ⎢ ⎢ b ⎥.
⎥
.
𝑖
⎣ .. ⎦
⎢ ⎥
𝑏𝑃
The Euclidean norm of a vector can be computed using the dot product, as
√
‖a‖2 = a · a.
Geometric definition: projection
4.3. Multivariate statistics 129

In Euclidean space, a Euclidean vector is a geometrical object that possesses both a magnitude
and a direction. A vector can be pictured as an arrow. Its magnitude is its length, and its
direction is the direction that the arrow points. The magnitude of a vector a is denoted by ‖a‖2 .
The dot product of two Euclidean vectors a and b is defined by
a · b = ‖a‖2 ‖b‖2 cos 𝜃,
where 𝜃 is the angle between a and b.

In particular, if a and b are orthogonal, then the angle between them is 90° and
a · b = 0.
At the other extreme, if they are codirectional, then the angle between them is 0° and
a · b = ‖a‖2 ‖b‖2
This implies that the dot product of a vector a by itself is
a · a = ‖a‖22 .
The scalar projection (or scalar component) of a Euclidean vector a in the direction of a Eu-
clidean vector b is given by
𝑎𝑏 = ‖a‖2 cos 𝜃,
where 𝜃 is the angle between a and b.

In terms of the geometric definition of the dot product, this can be rewritten
a·b
𝑎𝑏 = ,
‖b‖2
Fig. 5: Projection.
import numpy as np
np.random.seed(42)
a = np.random.randn(10)
b = np.random.randn(10)
np.dot(a, b)

-4.085788532659924
4.3.2 Mean vector
The mean (𝑃 × 1) column-vector 𝜇 whose estimator is

⎡ ⎤ ⎡ ⎤
𝑥𝑖1 𝑥
¯1
⎢ .. ⎥ ⎢ .. ⎥
𝑁 ⎢ . ⎥
𝑁
⎥ ⎢ . ⎥
⎢ ⎥
1 ∑︁ 1 ∑︁ ⎢⎢ 𝑥𝑖𝑗 ⎥ = ⎢ 𝑥
x̄ = xi = ⎢ ⎥ ⎢ ¯𝑗 ⎥ .
⎥
𝑁 𝑁 . .
𝑖=1 𝑖=1
⎣ .. ⎦ ⎣ .. ⎦
⎢ ⎥ ⎢ ⎥
𝑥𝑖𝑃 𝑥¯𝑃
4.3.3 Covariance matrix
• The covariance matrix ΣXX is a symmetric positive semi-definite matrix whose element
in the 𝑗, 𝑘 position is the covariance between the 𝑗 𝑡ℎ and 𝑘 𝑡ℎ elements of a random vector
i.e. the 𝑗 𝑡ℎ and 𝑘 𝑡ℎ columns of X.
• The covariance matrix generalizes the notion of covariance to multiple dimensions.
• The covariance matrix describe the shape of the sample distribution around the mean
assuming an elliptical distribution:
ΣXX = 𝐸(X − 𝐸(X))𝑇 𝐸(X − 𝐸(X)),

whose estimator SXX is a 𝑃 × 𝑃 matrix given by
1
SXX = (X − 1x̄𝑇 )𝑇 (X − 1x̄𝑇 ).
𝑁 −1
If we assume that X is centered, i.e. X is replaced by X − 1x̄𝑇 then the estimator is
⎡ ⎤ ⎡ ⎤
𝑥11 · · · 𝑥𝑁 1 ⎡ ⎤ 𝑠1 . . . 𝑠1𝑘 𝑠1𝑃
⎥ 𝑥11 · · · 𝑥1𝑘 𝑥1𝑃 .. .. ⎥
1 1 ⎢ ⎢ 𝑥1𝑗 · · · 𝑥𝑁 𝑗 ⎥ ⎢ .. .. .. ⎥ = ⎢ . 𝑠𝑗𝑘
⎢
SXX = 𝑇
X X= ⎢ .. . ⎥
⎥,
.. . . .
𝑁 −1 𝑁 −1⎣ .
⎥ ⎣ ⎦ ⎢
. ⎦ ⎣ 𝑠𝑘 𝑠𝑘𝑃 ⎦
𝑥𝑁 1 · · · 𝑥𝑁 𝑘 𝑥𝑁 𝑃
𝑥1𝑃 · · · 𝑥𝑁 𝑃 𝑠𝑃
where
𝑁
1 1 ∑︁
𝑠𝑗𝑘 = 𝑠𝑘𝑗 = xj 𝑇 x k = 𝑥𝑖𝑗 𝑥𝑖𝑘
𝑁 −1 𝑁 −1
𝑖=1
is an estimator of the covariance between the 𝑗 𝑡ℎ and 𝑘 𝑡ℎ variables.
## Avoid warnings and force inline plot

%matplotlib inline
import warnings
warnings.filterwarnings("ignore")
##
import numpy as np
import scipy


import pystatsml.plot_utils
import seaborn as sns # nice color
np.random.seed(42)
colors = sns.color_palette()
mean, Cov, X = [None] * 4, [None] * 4, [None] * 4

mean[0] = np.array([-2.5, 2.5])
Cov[0] = np.array([[1, 0],
[0, 1]])
mean[1] = np.array([2.5, 2.5])

Cov[1] = np.array([[1, .5],
[.5, 1]])
mean[2] = np.array([-2.5, -2.5])

Cov[2] = np.array([[1, .9],
[.9, 1]])
mean[3] = np.array([2.5, -2.5])

Cov[3] = np.array([[1, -.9],
[-.9, 1]])
# Generate dataset
for i in range(len(mean)):
X[i] = np.random.multivariate_normal(mean[i], Cov[i], n_samples)
# Plot
for i in range(len(mean)):
# Points
plt.scatter(X[i][:, 0], X[i][:, 1], color=colors[i], label="class %i" % i)
# Means
plt.scatter(mean[i][0], mean[i][1], marker="o", s=200, facecolors='w',
edgecolors=colors[i], linewidth=2)
# Ellipses representing the covariance matrices
pystatsml.plot_utils.plot_cov_ellipse(Cov[i], pos=mean[i], facecolor='none',
linewidth=2, edgecolor=colors[i])
plt.axis('equal')
_ = plt.legend(loc='upper left')

4.3.4 Correlation matrix
import numpy as np
import pandas as pd
url = 'https://python-graph-gallery.com/wp-content/uploads/mtcars.csv'
# Compute the correlation matrix

corr = df.corr()
# Generate a mask for the upper triangle

mask = np.zeros_like(corr, dtype=np.bool)
mask[np.triu_indices_from(mask)] = True
f, ax = plt.subplots(figsize=(5.5, 4.5))
cmap = sns.color_palette("RdBu_r", 11)
# Draw the heatmap with the mask and correct aspect ratio
_ = sns.heatmap(corr, mask=None, cmap=cmap, vmax=1, center=0,
square=True, linewidths=.5, cbar_kws={"shrink": .5})

Re-order correlation matrix using AgglomerativeClustering
# convert correlation to distances

d = 2 * (1 - np.abs(corr))
from sklearn.cluster import AgglomerativeClustering

clustering = AgglomerativeClustering(n_clusters=3, linkage='single', affinity="precomputed
˓→").fit(d)
lab=0
clusters = [list(corr.columns[clustering.labels_==lab]) for lab in set(clustering.labels_

˓→)]
print(clusters)
reordered = np.concatenate(clusters)
R = corr.loc[reordered, reordered]
f, ax = plt.subplots(figsize=(5.5, 4.5))
# Draw the heatmap with the mask and correct aspect ratio
_ = sns.heatmap(R, mask=None, cmap=cmap, vmax=1, center=0,
square=True, linewidths=.5, cbar_kws={"shrink": .5})
[['mpg', 'cyl', 'disp', 'hp', 'wt', 'qsec', 'vs', 'carb'], ['am', 'gear'], ['drat']]

4.3.5 Precision matrix
In statistics, precision is the reciprocal of the variance, and the precision matrix is the matrix
inverse of the covariance matrix.
It is related to partial correlations that measures the degree of association between two vari-
ables, while controlling the effect of other variables.
import numpy as np
Cov = np.array([[1.0, 0.9, 0.9, 0.0, 0.0, 0.0],

[0.9, 1.0, 0.9, 0.0, 0.0, 0.0],
[0.9, 0.9, 1.0, 0.0, 0.0, 0.0],
[0.0, 0.0, 0.0, 1.0, 0.9, 0.0],
[0.0, 0.0, 0.0, 0.9, 1.0, 0.0],
[0.0, 0.0, 0.0, 0.0, 0.0, 1.0]])
print("# Precision matrix:")

Prec = np.linalg.inv(Cov)
print(Prec.round(2))
print("# Partial correlations:")

Pcor = np.zeros(Prec.shape)
Pcor[::] = np.NaN
for i, j in zip(*np.triu_indices_from(Prec, 1)):

Pcor[i, j] = - Prec[i, j] / np.sqrt(Prec[i, i] * Prec[j, j])
print(Pcor.round(2))

# Precision matrix:
[[ 6.79 -3.21 -3.21 0. 0. 0. ]
[-3.21 6.79 -3.21 0. 0. 0. ]
[-3.21 -3.21 6.79 0. 0. 0. ]
[ 0. -0. -0. 5.26 -4.74 -0. ]
[ 0. 0. 0. -4.74 5.26 0. ]
[ 0. 0. 0. 0. 0. 1. ]]
# Partial correlations:
[[ nan 0.47 0.47 -0. -0. -0. ]
[ nan nan 0.47 -0. -0. -0. ]
[ nan nan nan -0. -0. -0. ]
[ nan nan nan nan 0.9 0. ]
[ nan nan nan nan nan -0. ]
[ nan nan nan nan nan nan]]
4.3.6 Mahalanobis distance
• The Mahalanobis distance is a measure of the distance between two points x and 𝜇 where
the dispersion (i.e. the covariance structure) of the samples is taken into account.
• The dispersion is considered through covariance matrix.
This is formally expressed as
√︁
𝐷𝑀 (x, 𝜇) = (x − 𝜇)𝑇 Σ−1 (x − 𝜇).
Intuitions
• Distances along the principal directions of dispersion are contracted since they correspond
to likely dispersion of points.
• Distances othogonal to the principal directions of dispersion are dilated since they corre-
spond to unlikely dispersion of points.
For example
√
𝐷𝑀 (1) = 1𝑇 Σ−1 1.
ones = np.ones(Cov.shape[0])
d_euc = np.sqrt(np.dot(ones, ones))
d_mah = np.sqrt(np.dot(np.dot(ones, Prec), ones))
print("Euclidean norm of ones=%.2f. Mahalanobis norm of ones=%.2f" % (d_euc, d_mah))
Euclidean norm of ones=2.45. Mahalanobis norm of ones=1.77
The first dot product that distances along the principal directions of dispersion are contracted:
print(np.dot(ones, Prec))
[0.35714286 0.35714286 0.35714286 0.52631579 0.52631579 1. ]

import numpy as np
import scipy
%matplotlib inline
np.random.seed(40)
mean = np.array([0, 0])

Cov = np.array([[1, .8],
[.8, 1]])
samples = np.random.multivariate_normal(mean, Cov, 100)
x1 = np.array([0, 2])
x2 = np.array([2, 2])
plt.scatter(samples[:, 0], samples[:, 1], color=colors[0])

plt.scatter(mean[0], mean[1], color=colors[0], s=200, label="mean")
plt.scatter(x1[0], x1[1], color=colors[1], s=200, label="x1")
plt.scatter(x2[0], x2[1], color=colors[2], s=200, label="x2")
# plot covariance ellipsis

pystatsml.plot_utils.plot_cov_ellipse(Cov, pos=mean, facecolor='none',
linewidth=2, edgecolor=colors[0])
# Compute distances
d2_m_x1 = scipy.spatial.distance.euclidean(mean, x1)
d2_m_x2 = scipy.spatial.distance.euclidean(mean, x2)
Covi = scipy.linalg.inv(Cov)
dm_m_x1 = scipy.spatial.distance.mahalanobis(mean, x1, Covi)
dm_m_x2 = scipy.spatial.distance.mahalanobis(mean, x2, Covi)
# Plot distances
vm_x1 = (x1 - mean) / d2_m_x1
vm_x2 = (x2 - mean) / d2_m_x2
jitter = .1
plt.plot([mean[0] - jitter, d2_m_x1 * vm_x1[0] - jitter],
[mean[1], d2_m_x1 * vm_x1[1]], color='k')
plt.plot([mean[0] - jitter, d2_m_x2 * vm_x2[0] - jitter],
[mean[1], d2_m_x2 * vm_x2[1]], color='k')
plt.plot([mean[0] + jitter, dm_m_x1 * vm_x1[0] + jitter],

[mean[1], dm_m_x1 * vm_x1[1]], color='r')
plt.plot([mean[0] + jitter, dm_m_x2 * vm_x2[0] + jitter],
[mean[1], dm_m_x2 * vm_x2[1]], color='r')
plt.legend(loc='lower right')
plt.text(-6.1, 3,
'Euclidian: d(m, x1) = %.1f<d(m, x2) = %.1f' % (d2_m_x1, d2_m_x2), color='k')
plt.text(-6.1, 3.5,
'Mahalanobis: d(m, x1) = %.1f>d(m, x2) = %.1f' % (dm_m_x1, dm_m_x2), color='r')
plt.axis('equal')
print('Euclidian d(m, x1) = %.2f < d(m, x2) = %.2f' % (d2_m_x1, d2_m_x2))
print('Mahalanobis d(m, x1) = %.2f > d(m, x2) = %.2f' % (dm_m_x1, dm_m_x2))

Euclidian d(m, x1) = 2.00 < d(m, x2) = 2.83

Mahalanobis d(m, x1) = 3.33 > d(m, x2) = 2.11
If the covariance matrix is the identity matrix, the Mahalanobis distance reduces to the Eu-
clidean distance. If the covariance matrix is diagonal, then the resulting distance measure is
called a normalized Euclidean distance.
More generally, the Mahalanobis distance is a measure of the distance between a point x and a
distribution 𝒩 (x|𝜇, Σ). It is a multi-dimensional generalization of the idea of measuring how
many standard deviations away x is from the mean. This distance is zero if x is at the mean,
and grows as x moves away from the mean: along each principal component axis, it measures
the number of standard deviations from x to the mean of the distribution.
4.3.7 Multivariate normal distribution
The distribution, or probability density function (PDF) (sometimes just density), of a continuous
random variable is a function that describes the relative likelihood for this random variable to
take on a given value.
The multivariate normal distribution, or multivariate Gaussian distribution, of a 𝑃 -dimensional
random vector x = [𝑥1 , 𝑥2 , . . . , 𝑥𝑃 ]𝑇 is
1 1
𝒩 (x|𝜇, Σ) = exp{− (x − 𝜇)𝑇 Σ−1 (x − 𝜇)}.
(2𝜋)𝑃/2 |Σ|1/2 2
import numpy as np
import scipy.stats
from scipy.stats import multivariate_normal
from mpl_toolkits.mplot3d import Axes3D


def multivariate_normal_pdf(X, mean, sigma):
"""Multivariate normal probability density function over X (n_samples x n_features)"""
P = X.shape[1]
det = np.linalg.det(sigma)
norm_const = 1.0 / (((2*np.pi) ** (P/2)) * np.sqrt(det))
X_mu = X - mu
inv = np.linalg.inv(sigma)
d2 = np.sum(np.dot(X_mu, inv) * X_mu, axis=1)
return norm_const * np.exp(-0.5 * d2)
# mean and covariance

mu = np.array([0, 0])
sigma = np.array([[1, -.5],
[-.5, 1]])
# x, y grid
x, y = np.mgrid[-3:3:.1, -3:3:.1]
X = np.stack((x.ravel(), y.ravel())).T
norm = multivariate_normal_pdf(X, mean, sigma).reshape(x.shape)
# Do it with scipy
norm_scpy = multivariate_normal(mu, sigma).pdf(np.stack((x, y), axis=2))
assert np.allclose(norm, norm_scpy)
# Plot
fig = plt.figure(figsize=(10, 7))
ax = fig.gca(projection='3d')
surf = ax.plot_surface(x, y, norm, rstride=3,
cstride=3, cmap=plt.cm.coolwarm,
linewidth=1, antialiased=False
)
ax.set_zlim(0, 0.2)
ax.zaxis.set_major_locator(plt.LinearLocator(10))
ax.zaxis.set_major_formatter(plt.FormatStrFormatter('%.02f'))
ax.set_xlabel('X')
ax.set_ylabel('Y')
ax.set_zlabel('p(x)')
plt.title('Bivariate Normal/Gaussian distribution')

fig.colorbar(surf, shrink=0.5, aspect=7, cmap=plt.cm.coolwarm)
plt.show()

4.3.8 Exercises
Dot product and Euclidean norm
Given a = [2, 1]𝑇 and b = [1, 1]𝑇

1. Write a function euclidean(x) that computes the Euclidean norm of vector, x.
2. Compute the Euclidean norm of a.
3. Compute the Euclidean distance of ‖a − b‖2 .
4. Compute the projection of b in the direction of vector a: 𝑏𝑎 .
5. Simulate a dataset X of 𝑁 = 100 samples of 2-dimensional vectors.
6. Project all samples in the direction of the vector a.

Covariance matrix and Mahalanobis norm
1. Sample a dataset X of 𝑁 = 100 samples of 2-dimensional [︂ vectors

]︂ from the bivariate
1 0.8
normal distribution 𝒩 (𝜇, Σ) where 𝜇 = [1, 1]𝑇 and Σ = .
0.8, 1
2. Compute the mean vector x̄ and center X. Compare the estimated mean x̄ to the true
mean, 𝜇.
3. Compute the empirical covariance matrix S. Compare the estimated covariance matrix S
to the true covariance matrix, Σ.
4. Compute S−1 (Sinv) the inverse of the covariance matrix by using scipy.linalg.inv(S).
5. Write a function mahalanobis(x, xbar, Sinv) that computes the Mahalanobis distance
of a vector x to the mean, x̄.
6. Compute the Mahalanobis and Euclidean distances of each sample x𝑖 to the mean x̄. Store
the results in a 100 × 2 dataframe.
4.4 Time Series in python
Two libraries:
• Pandas: https://pandas.pydata.org/pandas-docs/stable/timeseries.html
• scipy http://www.statsmodels.org/devel/tsa.html
4.4.1 Stationarity
A TS is said to be stationary if its statistical properties such as mean, variance remain constant
over time.
• constant mean
• constant variance
• an autocovariance that does not depend on time.
what is making a TS non-stationary. There are 2 major reasons behind non-stationaruty of a
TS:
1. Trend – varying mean over time. For eg, in this case we saw that on average, the number
of passengers was growing over time.
2. Seasonality – variations at specific time-frames. eg people might have a tendency to buy
cars in a particular month because of pay increment or festivals.
4.4. Time Series in python 141

4.4.2 Pandas Time Series Data Structure
A Series is similar to a list or an array in Python. It represents a series of values (numeric

or otherwise) such as a column of data. It provides additional functionality, methods, and
operators, which make it a more powerful version of a list.
import pandas as pd
import numpy as np
# Create a Series from a list

ser = pd.Series([1, 3])
print(ser)
# String as index
prices = {'apple': 4.99,
'banana': 1.99,
'orange': 3.99}
ser = pd.Series(prices)
print(ser)
x = pd.Series(np.arange(1,3), index=[x for x in 'ab'])

print(x)
print(x['b'])
0 1
1 3
dtype: int64
apple 4.99
banana 1.99
orange 3.99
dtype: float64
a 1
b 2
dtype: int64
2
4.4.3 Time Series Analysis of Google Trends
source: https://www.datacamp.com/community/tutorials/time-series-analysis-tutorial
Get Google Trends data of keywords such as ‘diet’ and ‘gym’ and see how they vary over time
while learning about trends and seasonality in time series data.
In the Facebook Live code along session on the 4th of January, we checked out Google trends
data of keywords ‘diet’, ‘gym’ and ‘finance’ to see how they vary over time. We asked ourselves
if there could be more searches for these terms in January when we’re all trying to turn over a
new leaf?
In this tutorial, you’ll go through the code that we put together during the session step by step.
You’re not going to do much mathematics but you are going to do the following:
• Read data
• Recode data
• Exploratory Data Analysis

4.4.4 Read data
import numpy as np
import pandas as pd
# Plot appears on its own windows

%matplotlib inline
# Tools / Preferences / Ipython Console / Graphics / Graphics Backend / Backend:␣
˓→“automatic”
# Interactive Matplotlib Jupyter Notebook

# %matplotlib inline
try:
url = "https://raw.githubusercontent.com/datacamp/datacamp_facebook_live_ny_
˓→resolution/master/datasets/multiTimeline.csv"
df = pd.read_csv(url, skiprows=2)
except:
df = pd.read_csv("../datasets/multiTimeline.csv", skiprows=2)
print(df.head())
# Rename columns
df.columns = ['month', 'diet', 'gym', 'finance']
# Describe
print(df.describe())
Month diet: (Worldwide) gym: (Worldwide) finance: (Worldwide)

0 2004-01 100 31 48
1 2004-02 75 26 49
2 2004-03 67 24 47
3 2004-04 70 22 48
4 2004-05 72 22 43
diet gym finance
count 168.000000 168.000000 168.000000
mean 49.642857 34.690476 47.148810
std 8.033080 8.134316 4.972547
min 34.000000 22.000000 38.000000
25% 44.000000 28.000000 44.000000
50% 48.500000 32.500000 46.000000
75% 53.000000 41.000000 50.000000
max 100.000000 58.000000 73.000000
4.4.5 Recode data
Next, you’ll turn the ‘month’ column into a DateTime data type and make it the index of the
DataFrame.
Note that you do this because you saw in the result of the .info() method that the ‘Month’
column was actually an of data type object. Now, that generic data type encapsulates everything
from strings to integers, etc. That’s not exactly what you want when you want to be looking
at time series data. That’s why you’ll use .to_datetime() to convert the ‘month’ column in your
DataFrame to a DateTime.

Be careful! Make sure to include the inplace argument when you’re setting the index of the
DataFrame df so that you actually alter the original index and set it to the ‘month’ column.
df.month = pd.to_datetime(df.month)
df.set_index('month', inplace=True)
print(df.head())
diet gym finance

month
2004-01-01 100 31 48
2004-02-01 75 26 49
2004-03-01 67 24 47
2004-04-01 70 22 48
2004-05-01 72 22 43
4.4.6 Exploratory Data Analysis
You can use a built-in pandas visualization method .plot() to plot your data as 3 line plots on a
single figure (one for each column, namely, ‘diet’, ‘gym’, and ‘finance’).
df.plot()
plt.xlabel('Year');
# change figure parameters

# df.plot(figsize=(20,10), linewidth=5, fontsize=20)
# Plot single column

# df[['diet']].plot(figsize=(20,10), linewidth=5, fontsize=20)
# plt.xlabel('Year', fontsize=20);
Text(0.5, 0, 'Year')

Note that this data is relative. As you can read on Google trends:
Numbers represent search interest relative to the highest point on the chart for the given region
and time. A value of 100 is the peak popularity for the term. A value of 50 means that the term
is half as popular. Likewise a score of 0 means the term was less than 1% as popular as the
peak.
4.4.7 Resampling, Smoothing, Windowing, Rolling average: Trends
Rolling average, for each time point, take the average of the points on either side of it. Note
that the number of points is specified by a window size.
Remove Seasonality with pandas Series.
See: http://pandas.pydata.org/pandas-docs/stable/timeseries.html A: ‘year end frequency’
year frequency
diet = df['diet']
diet_resamp_yr = diet.resample('A').mean()
diet_roll_yr = diet.rolling(12).mean()
ax = diet.plot(alpha=0.5, style='-') # store axis (ax) for latter plots

diet_resamp_yr.plot(style=':', label='Resample at year frequency', ax=ax)
diet_roll_yr.plot(style='--', label='Rolling average (smooth), window size=12', ax=ax)
ax.legend()
<matplotlib.legend.Legend at 0x7f2a045af0d0>
Rolling average (smoothing) with Numpy
x = np.asarray(df[['diet']])
win = 12


win_half = int(win / 2)
# print([((idx-win_half), (idx+win_half)) for idx in np.arange(win_half, len(x))])
diet_smooth = np.array([x[(idx-win_half):(idx+win_half)].mean() for idx in np.arange(win_

˓→half, len(x))])
plt.plot(diet_smooth)
[<matplotlib.lines.Line2D at 0x7f2a0021ee90>]
Trends Plot Diet and Gym

Build a new DataFrame which is the concatenation diet and gym smoothed data
gym = df['gym']
df_avg = pd.concat([diet.rolling(12).mean(), gym.rolling(12).mean()], axis=1)

df_avg.plot()
plt.xlabel('Year')

Detrending
df_dtrend = df[["diet", "gym"]] - df_avg

df_dtrend.plot()
plt.xlabel('Year')

4.4.8 First-order differencing: Seasonal Patterns
# diff = original - shiftted data

# (exclude first term for some implementation details)
assert np.all((diet.diff() == diet - diet.shift())[1:])
df.diff().plot()
plt.xlabel('Year')
4.4.9 Periodicity and Correlation
df.plot()
plt.xlabel('Year');
print(df.corr())
diet gym finance

diet 1.000000 -0.100764 -0.034639
gym -0.100764 1.000000 -0.284279
finance -0.034639 -0.284279 1.000000

Plot correlation matrix
sns.heatmap(df.corr(), cmap="coolwarm")
<AxesSubplot:>
‘diet’ and ‘gym’ are negatively correlated! Remember that you have a seasonal and a trend
component. From the correlation coefficient, ‘diet’ and ‘gym’ are negatively correlated:
• trends components are negatively correlated.
• seasonal components would positively correlated and their
The actual correlation coefficient is actually capturing both of those.

Seasonal correlation: correlation of the first-order differences of these time series
df.diff().plot()
plt.xlabel('Year');
print(df.diff().corr())
diet gym finance

diet 1.000000 0.758707 0.373828
gym 0.758707 1.000000 0.301111
finance 0.373828 0.301111 1.000000
Plot correlation matrix
sns.heatmap(df.diff().corr(), cmap="coolwarm")
<AxesSubplot:>

Decomposing time serie in trend, seasonality and residuals
from statsmodels.tsa.seasonal import seasonal_decompose
x = gym
x = x.astype(float) # force float

decomposition = seasonal_decompose(x)
trend = decomposition.trend
seasonal = decomposition.seasonal
residual = decomposition.resid
plt.subplot(411)
plt.plot(x, label='Original')
plt.legend(loc='best')
plt.subplot(412)
plt.plot(trend, label='Trend')
plt.subplot(413)
plt.plot(seasonal,label='Seasonality')
plt.subplot(414)
plt.plot(residual, label='Residuals')
plt.tight_layout()

4.4.10 Autocorrelation
A time series is periodic if it repeats itself at equally spaced intervals, say, every 12 months.
Autocorrelation Function (ACF): It is a measure of the correlation between the TS with a lagged
version of itself. For instance at lag 5, ACF would compare series at time instant t1. . . t2 with
series at instant t1-5. . . t2-5 (t1-5 and t2 being end points).
Plot
# from pandas.plotting import autocorrelation_plot

from pandas.plotting import autocorrelation_plot
x = df["diet"].astype(float)
autocorrelation_plot(x)
<AxesSubplot:xlabel='Lag', ylabel='Autocorrelation'>

Compute Autocorrelation Function (ACF)
from statsmodels.tsa.stattools import acf
x_diff = x.diff().dropna() # first item is NA

lag_acf = acf(x_diff, nlags=36)
plt.plot(lag_acf)
plt.title('Autocorrelation Function')
/home/ed203246/anaconda3/lib/python3.7/site-packages/statsmodels/tsa/stattools.py:666:␣
˓→FutureWarning: fft=True will become the default after the release of the 0.12 release␣
˓→of statsmodels. To suppress this warning, explicitly set fft=False.
FutureWarning,
Text(0.5, 1.0, 'Autocorrelation Function')

ACF peaks every 12 months: Time series is correlated with itself shifted by 12 months.
4.4.11 Time Series Forecasting with Python using Autoregressive Moving Average
(ARMA) models
Source:
• https://www.packtpub.com/mapt/book/big_data_and_business_intelligence/
9781783553358/7/ch07lvl1sec77/arma-models
• http://en.wikipedia.org/wiki/Autoregressive%E2%80%93moving-average_model
• ARIMA: https://www.analyticsvidhya.com/blog/2016/02/
time-series-forecasting-codes-python/
ARMA models are often used to forecast a time series. These models combine autoregressive
and moving average models. In moving average models, we assume that a variable is the sum
of the mean of the time series and a linear combination of noise components.
The autoregressive and moving average models can have different orders. In general, we can
define an ARMA model with p autoregressive terms and q moving average terms as follows:
𝑝
∑︁ 𝑞
∑︁
𝑥𝑡 = 𝑎𝑖 𝑥𝑡−𝑖 + 𝑏𝑖 𝜀𝑡−𝑖 + 𝜀𝑡
𝑖 𝑖

Choosing p and q
Plot the partial autocorrelation functions for an estimate of p, and likewise using the autocorre-
lation functions for an estimate of q.
Partial Autocorrelation Function (PACF): This measures the correlation between the TS with a
lagged version of itself but after eliminating the variations already explained by the intervening
comparisons. Eg at lag 5, it will check the correlation but remove the effects already explained
by lags 1 to 4.
from statsmodels.tsa.stattools import acf, pacf
x = df["gym"].astype(float)
x_diff = x.diff().dropna() # first item is NA

# ACF and PACF plots:
lag_acf = acf(x_diff, nlags=20)

lag_pacf = pacf(x_diff, nlags=20, method='ols')
#Plot ACF:
plt.subplot(121)
plt.plot(lag_acf)
plt.axhline(y=0,linestyle='--',color='gray')
plt.axhline(y=-1.96/np.sqrt(len(x_diff)),linestyle='--',color='gray')
plt.axhline(y=1.96/np.sqrt(len(x_diff)),linestyle='--',color='gray')
plt.title('Autocorrelation Function (q=1)')
#Plot PACF:
plt.subplot(122)
plt.plot(lag_pacf)
plt.axhline(y=0,linestyle='--',color='gray')
plt.axhline(y=-1.96/np.sqrt(len(x_diff)),linestyle='--',color='gray')
plt.axhline(y=1.96/np.sqrt(len(x_diff)),linestyle='--',color='gray')
plt.title('Partial Autocorrelation Function (p=1)')
plt.tight_layout()
/home/ed203246/anaconda3/lib/python3.7/site-packages/statsmodels/tsa/stattools.py:666:␣
˓→FutureWarning: fft=True will become the default after the release of the 0.12 release␣
˓→of statsmodels. To suppress this warning, explicitly set fft=False.
FutureWarning,

In this plot, the two dotted lines on either sides of 0 are the confidence interevals. These can be
used to determine the p and q values as:
• p: The lag value where the PACF chart crosses the upper confidence interval for the first
time, in this case p=1.
• q: The lag value where the ACF chart crosses the upper confidence interval for the first
time, in this case q=1.
Fit ARMA model with statsmodels
1. Define the model by calling ARMA() and passing in the p and q parameters.
2. The model is prepared on the training data by calling the fit() function.
3. Predictions can be made by calling the predict() function and specifying the index of the
time or times to be predicted.
from statsmodels.tsa.arima_model import ARMA
model = ARMA(x, order=(1, 1)).fit() # fit model
plt.plot(x)
plt.plot(model.predict(), color='red')
plt.title('RSS: %.4f'% sum((model.fittedvalues-x)**2))
/home/ed203246/anaconda3/lib/python3.7/site-packages/statsmodels/tsa/arima_model.py:472:␣
˓→FutureWarning:
statsmodels.tsa.arima_model.ARMA and statsmodels.tsa.arima_model.ARIMA have

been deprecated in favor of statsmodels.tsa.arima.model.ARIMA (note the .
between arima and model) and


statsmodels.tsa.SARIMAX. These will be removed after the 0.12 release.
statsmodels.tsa.arima.model.ARIMA makes use of the statespace framework and

is both well tested and maintained.
To silence this warning and continue using ARMA and ARIMA until they are
removed, use:
import warnings
warnings.filterwarnings('ignore', 'statsmodels.tsa.arima_model.ARMA',
FutureWarning)
warnings.filterwarnings('ignore', 'statsmodels.tsa.arima_model.ARIMA',
FutureWarning)
warnings.warn(ARIMA_DEPRECATION_WARN, FutureWarning)
/home/ed203246/anaconda3/lib/python3.7/site-packages/statsmodels/tsa/base/tsa_model.
˓→py:527: ValueWarning: No frequency information was provided, so inferred frequency MS␣
˓→will be used.
% freq, ValueWarning)
ARMA Model Results

==============================================================================
Dep. Variable: gym No. Observations: 168
Model: ARMA(1, 1) Log Likelihood -436.852
Method: css-mle S.D. of innovations 3.229
Date: Mon, 12 Oct 2020 AIC 881.704
Time: 00:54:34 BIC 894.200
Sample: 01-01-2004 HQIC 886.776
- 12-01-2017
==============================================================================
coef std err z P>|z| [0.025 0.975]
------------------------------------------------------------------------------
const 36.4315 8.827 4.127 0.000 19.131 53.732
ar.L1.gym 0.9967 0.005 220.566 0.000 0.988 1.006
ma.L1.gym -0.7494 0.054 -13.931 0.000 -0.855 -0.644
Roots
=============================================================================
Real Imaginary Modulus Frequency
-----------------------------------------------------------------------------
AR.1 1.0033 +0.0000j 1.0033 0.0000
MA.1 1.3344 +0.0000j 1.3344 0.0000
-----------------------------------------------------------------------------
Text(0.5, 1.0, 'RSS: 1794.4653')


CHAPTER
FIVE
MACHINE LEARNING
5.1 Dimension reduction and feature extraction
5.1.1 Introduction
In machine learning and statistics, dimensionality reduction or dimension reduction is the pro-
cess of reducing the number of features under consideration, and can be divided into feature
selection (not addressed here) and feature extraction.
Feature extraction starts from an initial set of measured data and builds derived values (fea-
tures) intended to be informative and non-redundant, facilitating the subsequent learning and
generalization steps, and in some cases leading to better human interpretations. Feature extrac-
tion is related to dimensionality reduction.
The input matrix X, of dimension 𝑁 × 𝑃 , is
⎡ ⎤
𝑥11 . . . 𝑥1𝑃
⎢ ⎥
⎢ .. .. ⎥
⎢ ⎥
⎢ . X . ⎥
⎢ ⎥
⎣ ⎦
𝑥𝑁 1 . . . 𝑥 𝑁 𝑃
where the rows represent the samples and columns represent the variables.
The goal is to learn a transformation that extracts a few relevant features. This is generally
done by exploiting the covariance ΣXX between the input features.
5.1.2 Singular value decomposition and matrix factorization
Matrix factorization principles
Decompose the data matrix X𝑁 ×𝑃 into a product of a mixing matrix U𝑁 ×𝐾 and a dictionary
matrix V𝑃 ×𝐾 .
X = UV𝑇 ,
If we consider only a subset of components 𝐾 < 𝑟𝑎𝑛𝑘(X) < min(𝑃, 𝑁 − 1) , X is approximated

by a matrix X̂:
X ≈ X̂ = UV𝑇 ,
159
Each line of xi is a linear combination (mixing ui ) of dictionary items V.

𝑁 𝑃 -dimensional data points lie in a space whose dimension is less than 𝑁 − 1 (2 dots lie on a
line, 3 on a plane, etc.).
Fig. 1: Matrix factorization
Singular value decomposition (SVD) principles
Singular-value decomposition (SVD) factorises the data matrix X𝑁 ×𝑃 into a product:
X = UDV𝑇 ,
where
⎡ ⎤ ⎡ ⎤
𝑥11 𝑥1𝑃 𝑢11⎡ 𝑢1𝐾
⎤⎡ ⎤
⎢ ⎥ ⎢ ⎥ 𝑑1 0 𝑣11 𝑣1𝑃
⎢ ⎥ ⎢ ⎥
⎢
⎢ X ⎥=⎢
⎥ ⎢ U ⎥
⎥
⎣ D ⎦ ⎣ V𝑇 ⎦.
⎣ ⎦ ⎣ ⎦ 0 𝑑𝐾 𝑣𝐾1 𝑣𝐾𝑃
𝑥𝑁 1 𝑥𝑁 𝑃 𝑢𝑁 1 𝑢𝑁 𝐾
U: right-singular
• V = [v1 , · · · , v𝐾 ] is a 𝑃 × 𝐾 orthogonal matrix.
• It is a dictionary of patterns to be combined (according to the mixing coefficients) to
reconstruct the original samples.
• V perfoms the initial rotations (projection) along the 𝐾 = min(𝑁, 𝑃 ) principal compo-
nent directions, also called loadings.
• Each v𝑗 performs the linear combination of the variables that has maximum sample vari-
ance, subject to being uncorrelated with the previous v𝑗−1 .
D: singular values
• D is a 𝐾 × 𝐾 diagonal matrix made of the singular values of X with 𝑑1 ≥ 𝑑2 ≥ · · · ≥
𝑑𝐾 ≥ 0.
• D scale the projection along the coordinate axes by 𝑑1 , 𝑑2 , · · · , 𝑑𝐾 .
160 Chapter 5. Machine Learning

• Singular values are the square roots of the eigenvalues of X𝑇 X.

V: left-singular vectors
• U = [u1 , · · · , u𝐾 ] is an 𝑁 × 𝐾 orthogonal matrix.
• Each row vi provides the mixing coefficients of dictionary items to reconstruct the sample
xi
• It may be understood as the coordinates on the new orthogonal basis (obtained after the
initial rotation) called principal components in the PCA.
SVD for variables transformation
V transforms correlated variables (X) into a set of uncorrelated ones (UD) that better expose
the various relationships among the original data items.
X = UDV𝑇 , (5.1)
𝑇
XV = UDV V, (5.2)
XV = UDI, (5.3)
XV = UD (5.4)
At the same time, SVD is a method for identifying and ordering the dimensions along which
data points exhibit the most variation.
import numpy as np
import scipy
from sklearn.decomposition import PCA
%matplotlib inline
np.random.seed(42)
# dataset
n_samples = 100
experience = np.random.normal(size=n_samples)
salary = 1500 + experience + np.random.normal(size=n_samples, scale=.5)
X = np.column_stack([experience, salary])
# PCA using SVD

X -= X.mean(axis=0) # Centering is required
U, s, Vh = scipy.linalg.svd(X, full_matrices=False)
# U : Unitary matrix having left singular vectors as columns.
# Of shape (n_samples,n_samples) or (n_samples,n_comps), depending on
# full_matrices.
#
# s : The singular values, sorted in non-increasing order. Of shape (n_comps,),
# with n_comps = min(n_samples, n_features).
#
# Vh: Unitary matrix having right singular vectors as rows.
# Of shape (n_features, n_features) or (n_comps, n_features) depending
# on full_matrices.
5.1. Dimension reduction and feature extraction 161

plt.figure(figsize=(9, 3))
plt.subplot(131)
plt.scatter(U[:, 0], U[:, 1], s=50)
plt.axis('equal')
plt.title("U: Rotated and scaled data")
plt.subplot(132)
# Project data
PC = np.dot(X, Vh.T)
plt.scatter(PC[:, 0], PC[:, 1], s=50)
plt.axis('equal')
plt.title("XV: Rotated data")
plt.xlabel("PC1")
plt.ylabel("PC2")
plt.subplot(133)
plt.scatter(X[:, 0], X[:, 1], s=50)
for i in range(Vh.shape[0]):
plt.arrow(x=0, y=0, dx=Vh[i, 0], dy=Vh[i, 1], head_width=0.2,
head_length=0.2, linewidth=2, fc='r', ec='r')
plt.text(Vh[i, 0], Vh[i, 1],'v%i' % (i+1), color="r", fontsize=15,
horizontalalignment='right', verticalalignment='top')
plt.axis('equal')
plt.ylim(-4, 4)
plt.title("X: original data (v1, v2:PC dir.)")

plt.xlabel("experience")
plt.ylabel("salary")
plt.tight_layout()

5.1.3 Principal components analysis (PCA)
Sources:
• C. M. Bishop Pattern Recognition and Machine Learning, Springer, 2006
• Everything you did and didn’t know about PCA
• Principal Component Analysis in 3 Simple Steps
Principles
• Principal components analysis is the main method used for linear dimension reduction.
• The idea of principal component analysis is to find the 𝐾 principal components di-
rections (called the loadings) V𝐾×𝑃 that capture the variation in the data as much as
possible.
• It converts a set of 𝑁 𝑃 -dimensional observations N𝑁 ×𝑃 of possibly correlated variables
into a set of 𝑁 𝐾-dimensional samples C𝑁 ×𝐾 , where the 𝐾 < 𝑃 . The new variables are
linearly uncorrelated. The columns of C𝑁 ×𝐾 are called the principal components.
• The dimension reduction is obtained by using only 𝐾 < 𝑃 components that exploit corre-
lation (covariance) among the original variables.
• PCA is mathematically defined as an orthogonal linear transformation V𝐾×𝑃 that trans-
forms the data to a new coordinate system such that the greatest variance by some projec-
tion of the data comes to lie on the first coordinate (called the first principal component),
the second greatest variance on the second coordinate, and so on.
C𝑁 ×𝐾 = X𝑁 ×𝑃 V𝑃 ×𝐾
• PCA can be thought of as fitting a 𝑃 -dimensional ellipsoid to the data, where each axis of
the ellipsoid represents a principal component. If some axis of the ellipse is small, then the
variance along that axis is also small, and by omitting that axis and its corresponding prin-
cipal component from our representation of the dataset, we lose only a commensurately
small amount of information.
• Finding the 𝐾 largest axes of the ellipse will permit to project the data onto a space having
dimensionality 𝐾 < 𝑃 while maximizing the variance of the projected data.
Dataset preprocessing
Centering
Consider a data matrix, X , with column-wise zero empirical mean (the sample mean of each
column has been shifted to zero), ie. X is replaced by X − 1x̄𝑇 .

Standardizing
Optionally, standardize the columns, i.e., scale them by their standard-deviation. Without stan-
dardization, a variable with a high variance will capture most of the effect of the PCA. The
principal direction will be aligned with this variable. Standardization will, however, raise noise
variables to the save level as informative variables.
The covariance matrix of centered standardized data is the correlation matrix.
Eigendecomposition of the data covariance matrix
To begin with, consider the projection onto a one-dimensional space (𝐾 = 1). We can define
the direction of this space using a 𝑃 -dimensional vector v, which for convenience (and without
loss of generality) we shall choose to be a unit vector so that ‖v‖2 = 1 (note that we are only
interested in the direction defined by v, not in the magnitude of v itself). PCA consists of two
mains steps:
Projection in the directions that capture the greatest variance
Each 𝑃 -dimensional data point x𝑖 is then projected onto v, where the coordinate (in the co-
ordinate system of v) is a scalar value, namely x𝑇𝑖 v. I.e., we want to find the vector v that
maximizes these coordinates along v, which we will see corresponds to maximizing the vari-
ance of the projected data. This is equivalently expressed as
1 ∑︁ (︀ 𝑇 )︀2
v = arg max x𝑖 v .
‖v‖=1 𝑁
𝑖
We can write this in matrix form as

1 1
v = arg max ‖Xv‖2 = v𝑇 X𝑇 Xv = v𝑇 SXX v,
‖v‖=1 𝑁 𝑁
where SXX is a biased estiamte of the covariance matrix of the data, i.e.
1 𝑇
SXX = X X.
𝑁
We now maximize the projected variance v𝑇 SXX v with respect to v. Clearly, this has to be a
constrained maximization to prevent ‖v2 ‖ → ∞. The appropriate constraint comes from the
normalization condition ‖v‖2 ≡ ‖v‖22 = v𝑇 v = 1. To enforce this constraint, we introduce a
Lagrange multiplier that we shall denote by 𝜆, and then make an unconstrained maximization
of
v𝑇 SXX v − 𝜆(v𝑇 v − 1).
By setting the gradient with respect to v equal to zero, we see that this quantity has a stationary
point when
SXX v = 𝜆v.
We note that v is an eigenvector of SXX .

If we left-multiply the above equation by v𝑇 and make use of v𝑇 v = 1, we see that the variance
is given by
v𝑇 SXX v = 𝜆,

and so the variance will be at a maximum when v is equal to the eigenvector corresponding to
the largest eigenvalue, 𝜆. This eigenvector is known as the first principal component.
We can define additional principal components in an incremental fashion by choosing each new
direction to be that which maximizes the projected variance amongst all possible directions that
are orthogonal to those already considered. If we consider the general case of a 𝐾-dimensional
projection space, the optimal linear projection for which the variance of the projected data is
maximized is now defined by the 𝐾 eigenvectors, v1 , . . . , vK , of the data covariance matrix
SXX that corresponds to the 𝐾 largest eigenvalues, 𝜆1 ≥ 𝜆2 ≥ · · · ≥ 𝜆𝐾 .
Back to SVD
The sample covariance matrix of centered data X is given by

1
SXX = X𝑇 X.
𝑁 −1
We rewrite X𝑇 X using the SVD decomposition of X as
X𝑇 X = (UDV𝑇 )𝑇 (UDV𝑇 )
= VD𝑇 U𝑇 UDV𝑇
= VD2 V𝑇
V𝑇 X𝑇 XV = D2
1 1
V𝑇 X𝑇 XV = D2
𝑁 −1 𝑁 −1
1
V𝑇 SXX V = D2
𝑁 −1
.
Considering only the 𝑘 𝑡ℎ right-singular vectors v𝑘 associated to the singular value 𝑑𝑘
1
vk 𝑇 SXX vk = 𝑑2 ,
𝑁 −1 𝑘
It turns out that if you have done the singular value decomposition then you already have
the Eigenvalue decomposition for X𝑇 X. Where - The eigenvectors of SXX are equivalent to
the right singular vectors, V, of X. - The eigenvalues, 𝜆𝑘 , of SXX , i.e. the variances of the
components, are equal to 𝑁 1−1 times the squared singular values, 𝑑𝑘 .
Moreover computing PCA with SVD do not require to form the matrix X𝑇 X, so computing the
SVD is now the standard way to calculate a principal components analysis from a data matrix,
unless only a handful of components are required.
PCA outputs
The SVD or the eigendecomposition of the data covariance matrix provides three main quanti-
ties:
1. Principal component directions or loadings are the eigenvectors of X𝑇 X. The V𝐾×𝑃
or the right-singular vectors of an SVD of X are called principal component directions of
X. They are generally computed using the SVD of X.

2. Principal components is the 𝑁 × 𝐾 matrix C which is obtained by projecting X onto the

principal components directions, i.e.
C𝑁 ×𝐾 = X𝑁 ×𝑃 V𝑃 ×𝐾 .
Since X = UDV𝑇 and V is orthogonal (V𝑇 V = I):
C𝑁 ×𝐾 = UDV𝑇𝑁 ×𝑃 V𝑃 ×𝐾 (5.5)
C𝑁 ×𝐾 = UD𝑇𝑁 ×𝐾 I𝐾×𝐾 (5.6)
C𝑁 ×𝐾 = UD𝑇𝑁 ×𝐾 (5.7)
(5.8)
Thus c𝑗 = Xv𝑗 = u𝑗 𝑑𝑗 , for 𝑗 = 1, . . . 𝐾. Hence u𝑗 is simply the projection of the row vectors of
X, i.e., the input predictor vectors, on the direction v𝑗 , scaled by 𝑑𝑗 .
⎡ ⎤
𝑥1,1 𝑣1,1 + . . . + 𝑥1,𝑃 𝑣1,𝑃
⎢ 𝑥2,1 𝑣1,1 + . . . + 𝑥2,𝑃 𝑣1,𝑃 ⎥
c1 = ⎢ ..
⎢ ⎥
⎥
⎣ . ⎦
𝑥𝑁,1 𝑣1,1 + . . . + 𝑥𝑁,𝑃 𝑣1,𝑃
3. The variance of each component is given by the eigen values 𝜆𝑘 , 𝑘 = 1, . . . 𝐾. It can be

obtained from the singular values:
1
𝑣𝑎𝑟(c𝑘 ) = (Xv𝑘 )2 (5.9)
𝑁 −1
1
= (u𝑘 𝑑𝑘 )2 (5.10)
𝑁 −1
1
= 𝑑2 (5.11)
𝑁 −1 𝑘
Determining the number of PCs
We must choose 𝐾 * ∈ [1, . . . , 𝐾], the number of required components. This can be done by
calculating the explained variance ratio of the 𝐾 * first components and by choosing 𝐾 * such
that the cumulative explained variance ratio is greater than some given threshold (e.g., ≈
90%). This is expressed as
∑︀𝐾 *
𝑗 𝑣𝑎𝑟(c𝑘 )
cumulative explained variance(c𝑘 ) = ∑︀𝐾 .
𝑗 𝑣𝑎𝑟(c𝑘 )
Interpretation and visualization
PCs
Plot the samples projeted on first the principal components as e.g. PC1 against PC2.
PC directions
Exploring the loadings associated with a component provides the contribution of each original
variable in the component.

Remark: The loadings (PC directions) are the coefficients of multiple regression of PC on origi-
nal variables:
c = Xv (5.12)
X𝑇 c = X𝑇 Xv (5.13)
𝑇 −1 𝑇
(X X) X c=v (5.14)
Another way to evaluate the contribution of the original variables in each PC can be obtained
by computing the correlation between the PCs and the original variables, i.e. columns of X,
denoted x𝑗 , for 𝑗 = 1, . . . , 𝑃 . For the 𝑘 𝑡ℎ PC, compute and plot the correlations with all original
variables
𝑐𝑜𝑟(c𝑘 , x𝑗 ), 𝑗 = 1 . . . 𝐾, 𝑗 = 1 . . . 𝐾.
These quantities are sometimes called the correlation loadings.
import numpy as np
from sklearn.decomposition import PCA
np.random.seed(42)
# dataset
n_samples = 100
experience = np.random.normal(size=n_samples)
salary = 1500 + experience + np.random.normal(size=n_samples, scale=.5)
X = np.column_stack([experience, salary])
# PCA with scikit-learn

pca = PCA(n_components=2)
pca.fit(X)
print(pca.explained_variance_ratio_)
PC = pca.transform(X)
plt.subplot(121)
plt.scatter(X[:, 0], X[:, 1])
plt.xlabel("x1"); plt.ylabel("x2")
plt.subplot(122)
plt.scatter(PC[:, 0], PC[:, 1])
plt.xlabel("PC1 (var=%.2f)" % pca.explained_variance_ratio_[0])
plt.ylabel("PC2 (var=%.2f)" % pca.explained_variance_ratio_[1])
plt.axis('equal')
plt.tight_layout()
[0.93646607 0.06353393]

5.1.4 Multi-dimensional Scaling (MDS)
Resources:
• http://www.stat.pitt.edu/sungkyu/course/2221Fall13/lec8_mds_combined.pdf
• https://en.wikipedia.org/wiki/Multidimensional_scaling
• Hastie, Tibshirani and Friedman (2009). The Elements of Statistical Learning: Data Mining,
Inference, and Prediction. New York: Springer, Second Edition.
The purpose of MDS is to find a low-dimensional projection of the data in which the pairwise
distances between data points is preserved, as closely as possible (in a least-squares sense).
• Let D be the (𝑁 × 𝑁 ) pairwise distance matrix where 𝑑𝑖𝑗 is a distance between points 𝑖
and 𝑗.
• The MDS concept can be extended to a wide variety of data types specified in terms of a
similarity matrix.
Given the dissimilarity (distance) matrix D𝑁 ×𝑁 = [𝑑𝑖𝑗 ], MDS attempts to find 𝐾-dimensional
projections of the 𝑁 points x1 , . . . , x𝑁 ∈ R𝐾 , concatenated in an X𝑁 ×𝐾 matrix, so that 𝑑𝑖𝑗 ≈
‖x𝑖 − x𝑗 ‖ are as close as possible. This can be obtained by the minimization of a loss function
called the stress function
∑︁
stress(X) = (𝑑𝑖𝑗 − ‖x𝑖 − x𝑗 ‖)2 .
𝑖̸=𝑗
This loss function is known as least-squares or Kruskal-Shepard scaling.

A modification of least-squares scaling is the Sammon mapping
∑︁ (𝑑𝑖𝑗 − ‖x𝑖 − x𝑗 ‖)2
stressSammon (X) = .
𝑑𝑖𝑗
𝑖̸=𝑗

The Sammon mapping performs better at preserving small distances compared to the least-
squares scaling.
Classical multidimensional scaling
Also known as principal coordinates analysis, PCoA.

• The distance matrix, D, is transformed to a similarity matrix, B, often using centered
inner products.
• The loss function becomes
∑︁ (︀ )︀2
stressclassical (X) = 𝑏𝑖𝑗 − ⟨x𝑖 , x𝑗 ⟩ .
𝑖̸=𝑗
• The stress function in classical MDS is sometimes called strain.

• The solution for the classical MDS problems can be found from the eigenvectors of the
similarity matrix.
• If the distances in D are Euclidean and double centered inner products are used, the
results are equivalent to PCA.
Example
The eurodist datset provides the road distances (in kilometers) between 21 cities in Europe.
Given this matrix of pairwise (non-Euclidean) distances D = [𝑑𝑖𝑗 ], MDS can be used to recover
the coordinates of the cities in some Euclidean referential whose orientation is arbitrary.
import pandas as pd
import numpy as np
# Pairwise distance between European cities

try:
url = '../datasets/eurodist.csv'
except:
url = 'https://raw.github.com/neurospin/pystatsml/master/datasets/eurodist.csv'
print(df.iloc[:5, :5])
city = df["city"]
D = np.array(df.iloc[:, 1:]) # Distance matrix
# Arbitrary choice of K=2 components

from sklearn.manifold import MDS
mds = MDS(dissimilarity='precomputed', n_components=2, random_state=40, max_iter=3000,␣
˓→eps=1e-9)
X = mds.fit_transform(D)
city Athens Barcelona Brussels Calais

0 Athens 0 3313 2963 3175
1 Barcelona 3313 0 1318 1326


2 Brussels 2963 1318 0 204
3 Calais 3175 1326 204 0
4 Cherbourg 3339 1294 583 460
Recover coordinates of the cities in Euclidean referential whose orientation is arbitrary:
from sklearn import metrics

Deuclidean = metrics.pairwise.pairwise_distances(X, metric='euclidean')
print(np.round(Deuclidean[:5, :5]))
[[ 0. 3116. 2994. 3181. 3428.]

[3116. 0. 1317. 1289. 1128.]
[2994. 1317. 0. 198. 538.]
[3181. 1289. 198. 0. 358.]
[3428. 1128. 538. 358. 0.]]
Plot the results:
# Plot: apply some rotation and flip

theta = 80 * np.pi / 180.
rot = np.array([[np.cos(theta), -np.sin(theta)],
[np.sin(theta), np.cos(theta)]])
Xr = np.dot(X, rot)
# flip x
Xr[:, 0] *= -1
plt.scatter(Xr[:, 0], Xr[:, 1])
for i in range(len(city)):
plt.text(Xr[i, 0], Xr[i, 1], city[i])
plt.axis('equal')
(-1894.0919178069155,
2914.3554370871234,
-1712.9733697197494,
2145.437068788015)

Determining the number of components
We must choose 𝐾 * ∈ {1, . . . , 𝐾} the number of required components. Plotting the values of
the stress function, obtained using 𝑘 ≤ 𝑁 − 1 components. In general, start with 1, . . . 𝐾 ≤ 4.
Choose 𝐾 * where you can clearly distinguish an elbow in the stress curve.
Thus, in the plot below, we choose to retain information accounted for by the first two compo-
nents, since this is where the elbow is in the stress curve.
k_range = range(1, min(5, D.shape[0]-1))

stress = [MDS(dissimilarity='precomputed', n_components=k,
random_state=42, max_iter=300, eps=1e-9).fit(D).stress_ for k in k_range]
print(stress)
plt.plot(k_range, stress)
plt.xlabel("k")
plt.ylabel("stress")
[48644495.28571428, 3356497.365752386, 2858455.495887962, 2756310.637628011]
Text(0, 0.5, 'stress')

5.1.5 Nonlinear dimensionality reduction
Sources:
• Scikit-learn documentation
• Wikipedia
Nonlinear dimensionality reduction or manifold learning cover unsupervised methods that
attempt to identify low-dimensional manifolds within the original 𝑃 -dimensional space that
represent high data density. Then those methods provide a mapping from the high-dimensional
space to the low-dimensional embedding.
Isomap
Isomap is a nonlinear dimensionality reduction method that combines a procedure to compute

the distance matrix with MDS. The distances calculation is based on geodesic distances evalu-
ated on neighborhood graph:
1. Determine the neighbors of each point. All points in some fixed radius or K nearest neigh-
bors.
2. Construct a neighborhood graph. Each point is connected to other if it is a K nearest
neighbor. Edge length equal to Euclidean distance.
3. Compute shortest path between pairwise of points 𝑑𝑖𝑗 to build the distance matrix D.
4. Apply MDS on D.

from sklearn import manifold, datasets

X, color = datasets.make_s_curve(1000, random_state=42)
fig = plt.figure(figsize=(10, 5))

plt.suptitle("Isomap Manifold Learning", fontsize=14)
ax = fig.add_subplot(121, projection='3d')
ax.scatter(X[:, 0], X[:, 1], X[:, 2], c=color, cmap=plt.cm.Spectral)
ax.view_init(4, -72)
plt.title('2D "S shape" manifold in 3D')
Y = manifold.Isomap(n_neighbors=10, n_components=2).fit_transform(X)
ax = fig.add_subplot(122)
plt.scatter(Y[:, 0], Y[:, 1], c=color, cmap=plt.cm.Spectral)
plt.title("Isomap")
plt.xlabel("First component")
plt.ylabel("Second component")
plt.axis('tight')
(-5.263300792026621, 5.40342656540886, -1.22283642981746, 1.278952725143333)

5.1.6 Exercises
PCA
Write a basic PCA class
Write a class BasicPCA with two methods:

• fit(X) that estimates the data mean, principal components directions V and the explained
variance of each component.
• transform(X) that projects the data onto the principal components.
Check that your BasicPCA gave similar results, compared to the results from sklearn.
Apply your Basic PCA on the iris dataset
The data set is available at: https://raw.github.com/neurospin/pystatsml/master/datasets/iris.

csv
• Describe the data set. Should the dataset been standardized?
• Describe the structure of correlations among variables.
• Compute a PCA with the maximum number of components.
• Compute the cumulative explained variance ratio. Determine the number of components
𝐾 by your computed values.
• Print the 𝐾 principal components directions and correlations of the 𝐾 principal compo-
nents with the original variables. Interpret the contribution of the original variables into
the PC.
• Plot the samples projected into the 𝐾 first PCs.
• Color samples by their species.
MDS
Apply MDS from sklearn on the iris dataset available at:

https://raw.github.com/neurospin/pystatsml/master/datasets/iris.csv
• Center and scale the dataset.
• Compute Euclidean pairwise distances matrix.
• Select the number of components.
• Show that classical MDS on Euclidean pairwise distances matrix is equivalent to PCA.

5.2 Clustering
Wikipedia: Cluster analysis or clustering is the task of grouping a set of objects in such a way
that objects in the same group (called a cluster) are more similar (in some sense or another)
to each other than to those in other groups (clusters). Clustering is one of the main task of
exploratory data mining, and a common technique for statistical data analysis, used in many
fields, including machine learning, pattern recognition, image analysis, information retrieval,
and bioinformatics.
Sources: http://scikit-learn.org/stable/modules/clustering.html
5.2.1 K-means clustering
Source: C. M. Bishop Pattern Recognition and Machine Learning, Springer, 2006

Suppose we have a data set 𝑋 = {𝑥1 , · · · , 𝑥𝑁 } that consists of 𝑁 observations of a random
𝐷-dimensional Euclidean variable 𝑥. Our goal is to partition the data set into some number, 𝐾,
of clusters, where we shall suppose for the moment that the value of 𝐾 is given. Intuitively, we
might think of a cluster as comprising a group of data points whose inter-point distances are
small compared to the distances to points outside of the cluster. We can formalize this notion
by first introducing a set of 𝐷-dimensional vectors 𝜇𝑘 , where 𝑘 = 1, . . . , 𝐾, in which 𝜇𝑘 is a
prototype associated with the 𝑘 𝑡ℎ cluster. As we shall see shortly, we can think of the 𝜇𝑘 as
representing the centres of the clusters. Our goal is then to find an assignment of data points
to clusters, as well as a set of vectors {𝜇𝑘 }, such that the sum of the squares of the distances of
each data point to its closest prototype vector 𝜇𝑘 , is at a minimum.
It is convenient at this point to define some notation to describe the assignment of data points to
clusters. For each data point 𝑥𝑖 , we introduce a corresponding set of binary indicator variables
𝑟𝑖𝑘 ∈ {0, 1}, where 𝑘 = 1, . . . , 𝐾, that describes which of the 𝐾 clusters the data point 𝑥𝑖 is
assigned to, so that if data point 𝑥𝑖 is assigned to cluster 𝑘 then 𝑟𝑖𝑘 = 1, and 𝑟𝑖𝑗 = 0 for 𝑗 ̸= 𝑘.
This is known as the 1-of-𝐾 coding scheme. We can then define an objective function, denoted
inertia, as
𝑁 ∑︁
∑︁ 𝐾
𝐽(𝑟, 𝜇) = 𝑟𝑖𝑘 ‖𝑥𝑖 − 𝜇𝑘 ‖22
𝑖 𝑘
which represents the sum of the squares of the Euclidean distances of each data point to its
assigned vector 𝜇𝑘 . Our goal is to find values for the {𝑟𝑖𝑘 } and the {𝜇𝑘 } so as to minimize the
function 𝐽. We can do this through an iterative procedure in which each iteration involves two
successive steps corresponding to successive optimizations with respect to the 𝑟𝑖𝑘 and the 𝜇𝑘
. First we choose some initial values for the 𝜇𝑘 . Then in the first phase we minimize 𝐽 with
respect to the 𝑟𝑖𝑘 , keeping the 𝜇𝑘 fixed. In the second phase we minimize 𝐽 with respect to
the 𝜇𝑘 , keeping 𝑟𝑖𝑘 fixed. This two-stage optimization process is then repeated until conver-
gence. We shall see that these two stages of updating 𝑟𝑖𝑘 and 𝜇𝑘 correspond respectively to the
expectation (E) and maximization (M) steps of the expectation-maximisation (EM) algorithm,
and to emphasize this we shall use the terms E step and M step in the context of the 𝐾-means
algorithm.
Consider first the determination of the 𝑟𝑖𝑘 . Because 𝐽 in is a linear function of 𝑟𝑖𝑘 , this opti-
mization can be performed easily to give a closed form solution. The terms involving different
𝑖 are independent and so we can optimize for each 𝑖 separately by choosing 𝑟𝑖𝑘 to be 1 for
whichever value of 𝑘 gives the minimum value of ||𝑥𝑖 − 𝜇𝑘 ||2 . In other words, we simply assign
the 𝑖th data point to the closest cluster centre. More formally, this can be expressed as
5.2. Clustering 175

{︃
1, if 𝑘 = arg min𝑗 ||𝑥𝑖 − 𝜇𝑗 ||2 .
𝑟𝑖𝑘 = (5.15)
0, otherwise.
Now consider the optimization of the 𝜇𝑘 with the 𝑟𝑖𝑘 held fixed. The objective function 𝐽 is a
quadratic function of 𝜇𝑘 , and it can be minimized by setting its derivative with respect to 𝜇𝑘 to
zero giving
∑︁
2 𝑟𝑖𝑘 (𝑥𝑖 − 𝜇𝑘 ) = 0
𝑖
which we can easily solve for 𝜇𝑘 to give

∑︀
𝑟𝑖𝑘 𝑥𝑖
𝜇𝑘 = ∑︀𝑖 .
𝑖 𝑟𝑖𝑘
The denominator in this expression is equal to the number of points assigned to cluster 𝑘, and so
this result has a simple interpretation, namely set 𝜇𝑘 equal to the mean of all of the data points
𝑥𝑖 assigned to cluster 𝑘. For this reason, the procedure is known as the 𝐾-means algorithm.
The two phases of re-assigning data points to clusters and re-computing the cluster means are
repeated in turn until there is no further change in the assignments (or until some maximum
number of iterations is exceeded). Because each phase reduces the value of the objective func-
tion 𝐽, convergence of the algorithm is assured. However, it may converge to a local rather than
global minimum of 𝐽.
from sklearn import cluster, datasets

%matplotlib inline
iris = datasets.load_iris()
X = iris.data[:, :2] # use only 'sepal length and sepal width'
y_iris = iris.target
km2 = cluster.KMeans(n_clusters=2).fit(X)
plt.subplot(131)
plt.scatter(X[:, 0], X[:, 1], c=km2.labels_)
plt.title("K=2, J=%.2f" % km2.inertia_)
plt.subplot(132)
plt.scatter(X[:, 0], X[:, 1], c=km3.labels_)
plt.subplot(133)
plt.scatter(X[:, 0], X[:, 1], c=km4.labels_)#.astype(np.float))
Text(0.5, 1.0, 'K=4, J=27.99')

Exercises
1. Analyse clusters
• Analyse the plot above visually. What would a good value of 𝐾 be?
• If you instead consider the inertia, the value of 𝐽, what would a good value of 𝐾 be?
• Explain why there is such difference.
• For 𝐾 = 2 why did 𝐾-means clustering not find the two “natural” clusters? See
the assumptions of 𝐾-means: http://scikit-learn.org/stable/auto_examples/cluster/plot_
kmeans_assumptions.html#example-cluster-plot-kmeans-assumptions-py
2. Re-implement the 𝐾-means clustering algorithm (homework)
Write a function kmeans(X, K) that return an integer vector of the samples’ labels.
5.2.2 Gaussian mixture models
The Gaussian mixture model (GMM) is a simple linear superposition of Gaussian components
over the data, aimed at providing a rich class of density models. We turn to a formulation of
Gaussian mixtures in terms of discrete latent variables: the 𝐾 hidden classes to be discovered.
Differences compared to 𝐾-means:
• Whereas the 𝐾-means algorithm performs a hard assignment of data points to clusters, in
which each data point is associated uniquely with one cluster, the GMM algorithm makes
a soft assignment based on posterior probabilities.
• Whereas the classic 𝐾-means is only based on Euclidean distances, classic GMM use a
Mahalanobis distances that can deal with non-spherical distributions. It should be noted
that Mahalanobis could be plugged within an improved version of 𝐾-Means clustering.
The Mahalanobis distance is unitless and scale-invariant, and takes into account the cor-
relations of the data set.
5.2. Clustering 177

The Gaussian mixture distribution can be written as a linear superposition of 𝐾 Gaussians in

the form:
𝐾
∑︁
𝑝(𝑥) = 𝒩 (𝑥 | 𝜇𝑘 , Σ𝑘 )𝑝(𝑘),
𝑘=1
where:
• The 𝑝(𝑘) are ∑︀
the mixing coefficients also know as the class probability of class 𝑘, and they
sum to one: 𝐾 𝑘=1 𝑝(𝑘) = 1.
• 𝒩 (𝑥 | 𝜇𝑘 , Σ𝑘 ) = 𝑝(𝑥 | 𝑘) is the conditional distribution of 𝑥 given a particular class 𝑘. It is

the multivariate Gaussian distribution defined over a 𝑃 -dimensional vector 𝑥 of continu-
ous variables.
The goal is to maximize the log-likelihood of the GMM:
𝑁 𝑁
{︃ 𝐾 }︃ 𝑁
{︃ 𝐾 }︃
∏︁ ∏︁ ∑︁ ∑︁ ∑︁
ln 𝑝(𝑥𝑖 ) = ln 𝒩 (𝑥𝑖 | 𝜇𝑘 , Σ𝑘 )𝑝(𝑘) = ln 𝒩 (𝑥𝑖 | 𝜇𝑘 , Σ𝑘 )𝑝(𝑘) .
𝑖=1 𝑖=1 𝑘=1 𝑖=1 𝑘=1
To compute the classes parameters: 𝑝(𝑘), 𝜇𝑘 , Σ𝑘 we sum over all samples, by weighting each
sample 𝑖 by its responsibility or contribution
∑︀ to class 𝑘: 𝑝(𝑘 | 𝑥𝑖 ) such that for each point its
contribution to all classes sum to one 𝑘 𝑝(𝑘 | 𝑥𝑖 ) = 1. This contribution is the conditional
probability of class 𝑘 given 𝑥: 𝑝(𝑘 | 𝑥) (sometimes called the posterior). It can be computed
using Bayes’ rule:
𝑝(𝑥 | 𝑘)𝑝(𝑘)
𝑝(𝑘 | 𝑥) = (5.16)
𝑝(𝑥)
𝒩 (𝑥 | 𝜇𝑘 , Σ𝑘 )𝑝(𝑘)
= ∑︀𝐾 (5.17)
𝑘=1 𝒩 (𝑥 | 𝜇𝑘 , Σ𝑘 )𝑝(𝑘)
Since the class parameters, 𝑝(𝑘), 𝜇𝑘 and Σ𝑘 , depend on the responsibilities 𝑝(𝑘 | 𝑥) and the
responsibilities depend on class parameters, we need a two-step iterative algorithm: the
expectation-maximization (EM) algorithm. We discuss this algorithm next.
### The expectation-maximization (EM) algorithm for Gaussian mixtures
Given a Gaussian mixture model, the goal is to maximize the likelihood function with respect
to the parameters (comprised of the means and covariances of the components and the mixing
coefficients).
Initialize the means 𝜇𝑘 , covariances Σ𝑘 and mixing coefficients 𝑝(𝑘)
1. E step. For each sample 𝑖, evaluate the responsibilities for each class 𝑘 using the current
parameter values
𝒩 (𝑥𝑖 | 𝜇𝑘 , Σ𝑘 )𝑝(𝑘)
𝑝(𝑘 | 𝑥𝑖 ) = ∑︀𝐾
𝑘=1 𝒩 (𝑥𝑖 | 𝜇𝑘 , Σ𝑘 )𝑝(𝑘)
2. M step. For each class, re-estimate the parameters using the current responsibilities

𝑁
1 ∑︁
𝜇new
𝑘 = 𝑝(𝑘 | 𝑥𝑖 )𝑥𝑖 (5.18)
𝑁𝑘
𝑖=1
𝑁
1 ∑︁
Σnew
𝑘 = 𝑝(𝑘 | 𝑥𝑖 )(𝑥𝑖 − 𝜇new new 𝑇
𝑘 )(𝑥𝑖 − 𝜇𝑘 ) (5.19)
𝑁𝑘
𝑖=1
new 𝑁𝑘
𝑝 (𝑘) = (5.20)
𝑁
3. Evaluate the log-likelihood
𝑁
{︃ 𝐾 }︃
∑︁ ∑︁
ln 𝒩 (𝑥|𝜇𝑘 , Σ𝑘 )𝑝(𝑘) ,
𝑖=1 𝑘=1
and check for convergence of either the parameters or the log-likelihood. If the convergence
criterion is not satisfied return to step 1.
import numpy as np
from sklearn import datasets
import sklearn
from sklearn.mixture import GaussianMixture
X = iris.data[:, :2] # 'sepal length (cm)''sepal width (cm)'
gmm2 = GaussianMixture(n_components=2, covariance_type='full').fit(X)

plt.subplot(131)
plt.scatter(X[:, 0], X[:, 1], c=[colors[lab] for lab in gmm2.predict(X)])#, color=colors)
for i in range(gmm2.covariances_.shape[0]):
pystatsml.plot_utils.plot_cov_ellipse(cov=gmm2.covariances_[i, :], pos=gmm2.means_[i,␣
˓→:],
facecolor='none', linewidth=2, edgecolor=colors[i])
plt.scatter(gmm2.means_[i, 0], gmm2.means_[i, 1], edgecolor=colors[i],
marker="o", s=100, facecolor="w", linewidth=2)
plt.title("K=2")
plt.subplot(132)
plt.scatter(X[:, 0], X[:, 1], c=[colors[lab] for lab in gmm3.predict(X)])
˓→:],
5.2. Clustering 179


plt.title("K=3")
plt.subplot(133)
plt.scatter(X[:, 0], X[:, 1], c=[colors[lab] for lab in gmm4.predict(X)]) # .astype(np.
˓→float))
˓→:],

_ = plt.title("K=4")
5.2.3 Model selection
Bayesian information criterion
In statistics, the Bayesian information criterion (BIC) is a criterion for model selection among
a finite set of models; the model with the lowest BIC is preferred. It is based, in part, on the
likelihood function and it is closely related to the Akaike information criterion (AIC).
X = iris.data
bic = list()
#print(X)
ks = np.arange(1, 10)
for k in ks:
gmm = GaussianMixture(n_components=k, covariance_type='full')
gmm.fit(X)
bic.append(gmm.bic(X))
k_chosen = ks[np.argmin(bic)]
plt.plot(ks, bic)
plt.xlabel("k")


plt.ylabel("BIC")
print("Choose k=", k_chosen)
Choose k= 2
5.2.4 Hierarchical clustering
Hierarchical clustering is an approach to clustering that build hierarchies of clusters in two main
approaches:
• Agglomerative: A bottom-up strategy, where each observation starts in their own cluster,
and pairs of clusters are merged upwards in the hierarchy.
• Divisive: A top-down strategy, where all observations start out in the same cluster, and
then the clusters are split recursively downwards in the hierarchy.
In order to decide which clusters to merge or to split, a measure of dissimilarity between clusters
is introduced. More specific, this comprise a distance measure and a linkage criterion. The
distance measure is just what it sounds like, and the linkage criterion is essentially a function of
the distances between points, for instance the minimum distance between points in two clusters,
the maximum distance between points in two clusters, the average distance between points in
two clusters, etc. One particular linkage criterion, the Ward criterion, will be discussed next.
5.2. Clustering 181

Ward clustering
Ward clustering belongs to the family of agglomerative hierarchical clustering algorithms. This
means that they are based on a “bottoms up” approach: each sample starts in its own cluster,
and pairs of clusters are merged as one moves up the hierarchy.
In Ward clustering, the criterion for choosing the pair of clusters to merge at each step is the
minimum variance criterion. Ward’s minimum variance criterion minimizes the total within-
cluster variance by each merge. To implement this method, at each step: find the pair of
clusters that leads to minimum increase in total within-cluster variance after merging. This
increase is a weighted squared distance between cluster centers.
The main advantage of agglomerative hierarchical clustering over 𝐾-means clustering is that
you can benefit from known neighborhood information, for example, neighboring pixels in an
image.
from sklearn import cluster, datasets

X = iris.data[:, :2] # 'sepal length (cm)''sepal width (cm)'
ward2 = cluster.AgglomerativeClustering(n_clusters=2, linkage='ward').fit(X)

plt.subplot(131)
plt.scatter(X[:, 0], X[:, 1], c=ward2.labels_)
plt.title("K=2")
plt.subplot(132)
plt.scatter(X[:, 0], X[:, 1], c=ward3.labels_)
plt.title("K=3")
plt.subplot(133)
plt.scatter(X[:, 0], X[:, 1], c=ward4.labels_) # .astype(np.float))
plt.title("K=4")
Text(0.5, 1.0, 'K=4')

5.2.5 Exercises
Perform clustering of the iris dataset based on all variables using Gaussian mixture models. Use
PCA to visualize clusters.
5.3 Linear methods for regression
Fig. 2: Linear regression
5.3. Linear methods for regression 183

5.3.1 Ordinary least squares
Linear regression models the output, or target variable 𝑦 ∈ R as a linear combination of the 𝑃 -
dimensional input x ∈ R𝑃 . Let X be the 𝑁 × 𝑃 matrix with each row an input vector (with a 1
in the first position), and similarly let y be the 𝑁 -dimensional vector of outputs in the training
set, the linear model will predict the y given x using the parameter vector, or weight vector
w ∈ R𝑃 according to
y = Xw + 𝜀,
where 𝜀 ∈ R𝑁 are the residuals, or the errors of the prediction. The w is found by minimizing
an objective function, which is the loss function, 𝐿(w), i.e. the error measured on the data.
This error is the sum of squared errors (SSE) loss.
𝐿(w) = SSE(w) (5.21)

𝑁
∑︁
= (𝑦𝑖 − x𝑇𝑖 w)2 (5.22)
𝑖
= (y − X𝑇 w)𝑇 (y − X𝑇 w) (5.23)
𝑇
= ‖y − X w‖22 , (5.24)
Minimizing the SSE is the Ordinary Least Square OLS regression as objective function. which is
a simple ordinary least squares (OLS) minimization whose analytic solution is:
wOLS = (X𝑇 X)−1 X𝑇 y
The gradient of the loss:
𝐿(w, X, y) ∑︁
𝜕 =2 x𝑖 (x𝑖 · w − 𝑦𝑖 )
𝜕w
𝑖
5.3.2 Linear regression with scikit-learn
Scikit learn offer many models for supervised learning, and they all follow the same application
programming interface (API), namely:
model = Estimator()
model.fit(X, y)
predictions = model.predict(X)
%matplotlib inline
#import warnings
#warnings.filterwarnings(action='once')

import numpy as np
import pandas as pd
import sklearn.linear_model as lm


import sklearn.metrics as metrics
%matplotlib inline
# Fit Ordinary Least Squares: OLS

csv = pd.read_csv('https://raw.githubusercontent.com/neurospin/pystatsml/master/datasets/
˓→Advertising.csv', index_col=0)
X = csv[['TV', 'Radio']]
y = csv['Sales']
lr = lm.LinearRegression().fit(X, y)
y_pred = lr.predict(X)
print("R-squared =", metrics.r2_score(y, y_pred))
print("Coefficients =", lr.coef_)
# Plot
fig = plt.figure()
ax = fig.add_subplot(111, projection='3d')
ax.scatter(csv['TV'], csv['Radio'], csv['Sales'], c='r', marker='o')
xx1, xx2 = np.meshgrid(

np.linspace(csv['TV'].min(), csv['TV'].max(), num=10),
np.linspace(csv['Radio'].min(), csv['Radio'].max(), num=10))
XX = np.column_stack([xx1.ravel(), xx2.ravel()])
yy = lr.predict(XX)
ax.plot_surface(xx1, xx2, yy.reshape(xx1.shape), color='None')
ax.set_xlabel('TV')
ax.set_ylabel('Radio')
_ = ax.set_zlabel('Sales')
R-squared = 0.8971942610828956
Coefficients = [0.04575482 0.18799423]

5.3.3 Overfitting
In statistics and machine learning, overfitting occurs when a statistical model describes random
errors or noise instead of the underlying relationships. Overfitting generally occurs when a
model is excessively complex, such as having too many parameters relative to the number
of observations. A model that has been overfit will generally have poor predictive performance,
as it can exaggerate minor fluctuations in the data.
A learning algorithm is trained using some set of training samples. If the learning algorithm has
the capacity to overfit the training samples the performance on the training sample set will
improve while the performance on unseen test sample set will decline.
The overfitting phenomenon has three main explanations: - excessively complex models, - mul-
ticollinearity, and - high dimensionality.
Model complexity
Complex learners with too many parameters relative to the number of observations may overfit
the training dataset.
Multicollinearity
Predictors are highly correlated, meaning that one can be linearly predicted from the others.
In this situation the coefficient estimates of the multiple regression may change erratically in
response to small changes in the model or the data. Multicollinearity does not reduce the
predictive power or reliability of the model as a whole, at least not within the sample data
set; it only affects computations regarding individual predictors. That is, a multiple regression
model with correlated predictors can indicate how well the entire bundle of predictors predicts
the outcome variable, but it may not give valid results about any individual predictor, or about
which predictors are redundant with respect to others. In case of perfect multicollinearity the
predictor matrix is singular and therefore cannot be inverted. Under these circumstances, for a

general linear model y = Xw + 𝜀, the ordinary least-squares estimator, w𝑂𝐿𝑆 = (X𝑇 X)−1 X𝑇 y,
does not exist.
An example where correlated predictor may produce an unstable model follows:
import numpy as np
bv = np.array([10, 20, 30, 40, 50]) # business volume

tax = .2 * bv # Tax
bp = .1 * bv + np.array([-.1, .2, .1, -.2, .1]) # business potential
X = np.column_stack([bv, tax])
beta_star = np.array([.1, 0]) # true solution
'''
Since tax and bv are correlated, there is an infinite number of linear combinations
leading to the same prediction.
'''
# 10 times the bv then subtract it 9 times using the tax variable:

beta_medium = np.array([.1 * 10, -.1 * 9 * (1/.2)])
# 100 times the bv then subtract it 99 times using the tax variable:
beta_large = np.array([.1 * 100, -.1 * 99 * (1/.2)])
# Check that all model lead to the same result

assert np.all(np.dot(X, beta_star) == np.dot(X, beta_medium))
assert np.all(np.dot(X, beta_star) == np.dot(X, beta_large))
Multicollinearity between the predictors: business volumes and

tax produces unstable models with arbitrary large coefficients.

Dealing with multicollinearity:

• Regularisation by e.g. ℓ2 shrinkage: Introduce a bias in the solution by making (𝑋 𝑇 𝑋)−1
non-singular. See ℓ2 shrinkage.
• Feature selection: select a small number of features. See: Isabelle Guyon and André
Elisseeff An introduction to variable and feature selection The Journal of Machine Learning
Research, 2003.
• Feature selection: select a small number of features using ℓ1 shrinkage.
• Extract few independent (uncorrelated) features using e.g. principal components analysis
(PCA), partial least squares regression (PLS-R) or regression methods that cut the number
of predictors to a smaller set of uncorrelated components.
High dimensionality
High dimensions means a large number of input features. Linear predictor associate one pa-
rameter to each input feature, so a high-dimensional situation (𝑃 , number of features, is large)
with a relatively small number of samples 𝑁 (so-called large 𝑃 small 𝑁 situation) generally
lead to an overfit of the training data. Thus it is generally a bad idea to add many input features
into the learner. This phenomenon is called the curse of dimensionality.
One of the most important criteria to use when choosing a learning algorithm is based on the
relative size of 𝑃 and 𝑁 .
• Remenber that the “covariance” matrix X𝑇 X used in the linear model is a 𝑃 × 𝑃 matrix of
rank min(𝑁, 𝑃 ). Thus if 𝑃 > 𝑁 the equation system is overparameterized and admit an

infinity of solutions that might be specific to the learning dataset. See also ill-conditioned
or singular matrices.
• The sampling density of 𝑁 samples in an 𝑃 -dimensional space is proportional to 𝑁 1/𝑃 .
Thus a high-dimensional space becomes very sparse, leading to poor estimations of sam-
ples densities.
• Another consequence of the sparse sampling in high dimensions is that all sample points
are close to an edge of the sample. Consider 𝑁 data points uniformly distributed in a
𝑃 -dimensional unit ball centered at the origin. Suppose we consider a nearest-neighbor
estimate at the origin. The median distance from the origin to the closest data point is
given by the expression
)︂1/𝑃
1𝑁
(︂
𝑑(𝑃, 𝑁 ) = 1 − .
2
A more complicated expression exists for the mean distance to the closest point. For N = 500,
P = 10 , 𝑑(𝑃, 𝑁 ) ≈ 0.52, more than halfway to the boundary. Hence most data points are
closer to the boundary of the sample space than to any other data point. The reason that
this presents a problem is that prediction is much more difficult near the edges of the training
sample. One must extrapolate from neighboring sample points rather than interpolate between
them. (Source: T Hastie, R Tibshirani, J Friedman. The Elements of Statistical Learning: Data
Mining, Inference, and Prediction. Second Edition, 2009.)
• Structural risk minimization provides a theoretical background of this phenomenon. (See
VC dimension.)
• See also bias–variance trade-off.
import seaborn # nicer plots
def fit_on_increasing_size(model):
n_samples = 100
n_features_ = np.arange(10, 800, 20)
r2_train, r2_test, snr = [], [], []
for n_features in n_features_:
# Sample the dataset (* 2 nb of samples)
n_features_info = int(n_features/10)
np.random.seed(42) # Make reproducible
X = np.random.randn(n_samples * 2, n_features)
beta = np.zeros(n_features)
beta[:n_features_info] = 1
Xbeta = np.dot(X, beta)
eps = np.random.randn(n_samples * 2)
y = Xbeta + eps
# Split the dataset into train and test sample
Xtrain, Xtest = X[:n_samples, :], X[n_samples:, :]
ytrain, ytest = y[:n_samples], y[n_samples:]
# fit/predict
lr = model.fit(Xtrain, ytrain)
y_pred_train = lr.predict(Xtrain)
y_pred_test = lr.predict(Xtest)
snr.append(Xbeta.std() / eps.std())
r2_train.append(metrics.r2_score(ytrain, y_pred_train))
r2_test.append(metrics.r2_score(ytest, y_pred_test))
return n_features_, np.array(r2_train), np.array(r2_test), np.array(snr)


def plot_r2_snr(n_features_, r2_train, r2_test, xvline, snr, ax):
"""
Two scales plot. Left y-axis: train test r-squared. Right y-axis SNR.
"""
ax.plot(n_features_, r2_train, label="Train r-squared", linewidth=2)
ax.plot(n_features_, r2_test, label="Test r-squared", linewidth=2)
ax.axvline(x=xvline, linewidth=2, color='k', ls='--')
ax.axhline(y=0, linewidth=1, color='k', ls='--')
ax.set_ylim(-0.2, 1.1)
ax.set_xlabel("Number of input features")
ax.set_ylabel("r-squared")
ax.legend(loc='best')
ax.set_title("Prediction perf.")
ax_right = ax.twinx()
ax_right.plot(n_features_, snr, 'r-', label="SNR", linewidth=1)
ax_right.set_ylabel("SNR", color='r')
for tl in ax_right.get_yticklabels():
tl.set_color('r')
# Model = linear regression

mod = lm.LinearRegression()
# Fit models on dataset

n_features, r2_train, r2_test, snr = fit_on_increasing_size(model=mod)
argmax = n_features[np.argmax(r2_test)]
# plot
fig, axis = plt.subplots(1, 2, figsize=(9, 3))
# Left pane: all features

plot_r2_snr(n_features, r2_train, r2_test, argmax, snr, axis[0])
# Right pane: Zoom on 100 first features

plot_r2_snr(n_features[n_features <= 100],
r2_train[n_features <= 100], r2_test[n_features <= 100],
argmax,
snr[n_features <= 100],
axis[1])
plt.tight_layout()

Exercises
Study the code above and:

• Describe the datasets: 𝑁 : nb_samples, 𝑃 : nb_features.
• What is n_features_info?
• Give the equation of the generative model.
• What is modified by the loop?
• What is the SNR?
Comment the graph above, in terms of training and test performances:
• How does the train and test performance changes as a function of 𝑥?
• Is it the expected results when compared to the SNR?
• What can you conclude?
5.3.4 Ridge regression (ℓ2 -regularization)
Regarding linear models, overfitting generally leads to excessively complex solutions (coeffi-
cient vectors), accounting for noise or spurious correlations within predictors. Regularization
aims to alleviate this phenomenon by constraining (biasing or reducing) the capacity of the
learning algorithm in order to promote simple solutions. Regularization penalizes “large” solu-
tions forcing the coefficients to be small, i.e. to shrink them toward zeros.
The objective function 𝐽(w) to minimize with respect to w is composed of a loss function 𝐿(w)
for goodness-of-fit and a penalty term Ω(w) (regularization to avoid overfitting). This is a
trade-off where the respective contribution of the loss and the penalty terms is controlled by
the regularization parameter 𝜆.
Therefore the loss function 𝐿(w) (generally the SSE) is combined with a penalty function
Ω(w) leading to the general form:
𝐽(w) = 𝐿(w) + 𝜆Ω(w)
The respective contribution of the loss and the penalty is controlled by the regularization
parameter 𝜆.
Ridge regression impose a ℓ2 penalty on the coefficients, i.e. it penalizes with the Euclidean
norm of the coefficients while minimizing SSE. The objective function becomes:
𝑁
∑︁
Ridge(w) = (𝑦𝑖 − x𝑇𝑖 w)2 + 𝜆‖w‖22 (5.25)
𝑖
= ‖y − xw‖22 + 𝜆‖w‖22 . (5.26)
The w that minimises 𝐹𝑅𝑖𝑑𝑔𝑒 (w) can be found by the following derivation:

∇w Ridge(w) = 0 (5.27)
∇w (y − Xw) (y − Xw) + 𝜆w𝑇 w = 0
𝑇
(︀ )︀
(5.28)
∇w (y𝑇 y − 2w𝑇 X𝑇 y + w𝑇 X𝑇 Xw + 𝜆w𝑇 w) = 0
(︀ )︀
(5.29)
𝑇 𝑇
−2X y + 2X Xw + 2𝜆w = 0 (5.30)
−X𝑇 y + (X𝑇 X + 𝜆I)w = 0 (5.31)
𝑇 𝑇
(X X + 𝜆I)w = x y (5.32)
w = (X𝑇 X + 𝜆I)−1 x𝑇 y (5.33)
• The solution adds a positive constant to the diagonal of X𝑇 X before inversion. This makes
the problem nonsingular, even if X𝑇 X is not of full rank, and was the main motivation
behind ridge regression.
• Increasing 𝜆 shrinks the w coefficients toward 0.
• This approach penalizes the objective function by the Euclidian (:math:èll_2`) norm
of the coefficients such that solutions with large coefficients become unattractive.
The gradient of the loss:
𝐿(w, X, y) ∑︁
𝜕 = 2( x𝑖 (x𝑖 · w − 𝑦𝑖 ) + 𝜆w)
𝜕w
𝑖

import numpy as np
# lambda is alpha!
mod = lm.Ridge(alpha=10)

# plot


argmax,
axis[1])
plt.tight_layout()

Exercice
What benefit has been obtained by using ℓ2 regularization?
5.3.5 Lasso regression (ℓ1 -regularization)
Lasso regression penalizes the coefficients by the ℓ1 norm. This constraint will reduce (bias)
the capacity of the learning algorithm. To add such a penalty forces the coefficients to be small,
i.e. it shrinks them toward zero. The objective function to minimize becomes:
Lasso(w) = ‖y − Xw‖22 + 𝜆‖w‖1 . (5.34)
This penalty forces some coefficients to be exactly zero, providing a feature selection property.

import numpy as np
# lambda is alpha !
mod = lm.Lasso(alpha=.1)

# plot


argmax,
axis[1])
plt.tight_layout()

Sparsity of the ℓ1 norm
Occam’s razor
Occam’s razor (also written as Ockham’s razor, and lex parsimoniae in Latin, which means
law of parsimony) is a problem solving principle attributed to William of Ockham (1287-1347),
who was an English Franciscan friar and scholastic philosopher and theologian. The principle
can be interpreted as stating that among competing hypotheses, the one with the fewest
assumptions should be selected.
Principle of parsimony
The simplest of two competing theories is to be preferred. Definition of parsimony: Economy

of explanation in conformity with Occam’s razor.
Among possible models with similar loss, choose the simplest one:
• Choose the model with the smallest coefficient vector, i.e. smallest ℓ2 (‖w‖2 ) or ℓ1 (‖w‖1 )
norm of w, i.e. ℓ2 or ℓ1 penalty. See also bias-variance tradeoff.
• Choose the model that uses the smallest number of predictors. In other words, choose the
model that has many predictors with zero weights. Two approaches are available to obtain
this: (i) Perform a feature selection as a preprocessing prior to applying the learning
algorithm, or (ii) embed the feature selection procedure within the learning process.
Sparsity-induced penalty or embedded feature selection with the ℓ1 penalty
The penalty based on the ℓ1 norm promotes sparsity (scattered, or not dense): it forces many
coefficients to be exactly zero. This also makes the coefficient vector scattered.
The figure bellow illustrates the OLS loss under a constraint acting on the ℓ1 norm of the coef-
ficient vector. I.e., it illustrates the following optimization problem:
minimize ‖y − Xw‖22
w
subject to ‖w‖1 ≤ 1.

Fig. 3: Sparsity of L1 norm
Optimization issues
Section to be completed
• No more closed-form solution.
• Convex but not differentiable.
• Requires specific optimization algorithms, such as the fast iterative shrinkage-thresholding
algorithm (FISTA): Amir Beck and Marc Teboulle, A Fast Iterative Shrinkage-Thresholding
Algorithm for Linear Inverse Problems SIAM J. Imaging Sci., 2009.
The ridge penalty shrinks the coefficients toward zero. The figure illustrates: the OLS solution
on the left. The ℓ1 and ℓ2 penalties in the middle pane. The penalized OLS in the right pane.
The right pane shows how the penalties shrink the coefficients toward zero. The black points
are the minimum found in each case, and the white points represents the true solution used to
generate the data.

Fig. 4: ℓ1 and ℓ2 shrinkages
5.3.6 Elastic-net regression (ℓ2 -ℓ1 -regularization)
The Elastic-net estimator combines the ℓ1 and ℓ2 penalties, and results in the problem to
Enet(w) = ‖y − X𝑇 w‖22 + 𝛼 𝜌 ‖w‖1 + (1 − 𝜌) ‖w‖22 ,

(︀ )︀
(5.35)
where 𝛼 acts as a global penalty and 𝜌 as an ℓ1 /ℓ2 ratio.
Rational
• If there are groups of highly correlated variables, Lasso tends to arbitrarily select only
one from each group. These models are difficult to interpret because covariates that are
strongly associated with the outcome are not included in the predictive model. Conversely,
the elastic net encourages a grouping effect, where strongly correlated predictors tend to
be in or out of the model together.
• Studies on real world data and simulation studies show that the elastic net often outper-
forms the lasso, while enjoying a similar sparsity of representation.

import numpy as np
mod = lm.ElasticNet(alpha=.5, l1_ratio=.5)



# plot


argmax,
axis[1])
plt.tight_layout()
5.4 Linear classification
Fig. 5: Linear (logistic) classification
5.4. Linear classification 197

Given a training set of 𝑁 samples, 𝐷 = {(𝑥1 , 𝑦1 ), . . . , (𝑥𝑁 , 𝑦𝑁 )} , where 𝑥𝑖 is a multidimen-

sional input vector with dimension 𝑃 and class label (target or response).
Multiclass Classification problems can be seen as several binary classification problems 𝑦𝑖 ∈
{0, 1} where the classifier aims to discriminate the sample of the current class (label 1) versus
the samples of other classes (label 0).
Therfore, for each class the classifier seek for a vector of parameters 𝑤 that performs a linear
combination of the input variables, 𝑥𝑇 𝑤. This step performs a projection or a rotation of input
sample into a good discriminative one-dimensional sub-space, that best discriminate sample of
current class vs sample of other classes.
This score (a.k.a decision function) is tranformed, using the nonlinear activation funtion 𝑓 (.), to
a “posterior probabilities” of class 1: 𝑝(𝑦 = 1|𝑥) = 𝑓 (𝑥𝑇 𝑤), where, 𝑝(𝑦 = 1|𝑥) = 1 − 𝑝(𝑦 = 0|𝑥).
The decision surfaces (orthogonal hyperplan to 𝑤) correspond to 𝑓 (𝑥) = constant, so that
𝑥𝑇 𝑤 = constant and hence the decision surfaces are linear functions of 𝑥, even if the function
𝑓 (.) is nonlinear.
A thresholding of the activation (shifted by the bias or intercept) provides the predicted class
label.
The vector of parameters, that defines the discriminative axis, minimizes an objective function
𝐽(𝑤) that is a sum of of loss function 𝐿(𝑤) and some penalties on the weights vector Ω(𝑤).
∑︁
min 𝐽 = 𝐿(𝑦𝑖 , 𝑓 (𝑥𝑖 𝑇 𝑤)) + Ω(𝑤),
𝑤
𝑖
5.4.1 Fisher’s linear discriminant with equal class covariance
This geometric method does not make any probabilistic assumptions, instead it relies on dis-
tances. It looks for the linear projection of the data points onto a vector, 𝑤, that maximizes
the between/within variance ratio, denoted 𝐹 (𝑤). Under a few assumptions, it will provide the
same results as linear discriminant analysis (LDA), explained below.
Suppose two classes of observations, 𝐶0 and 𝐶1 , have means 𝜇0 and 𝜇1 and the same total
within-class scatter (“covariance”) matrix,
∑︁ ∑︁
𝑆𝑊 = (𝑥𝑖 − 𝜇0 )(𝑥𝑖 − 𝜇0 )𝑇 + (𝑥𝑗 − 𝜇1 )(𝑥𝑗 − 𝜇1 )𝑇 (5.36)
𝑖∈𝐶0 𝑗∈𝐶1
𝑇
= 𝑋𝑐 𝑋𝑐 , (5.37)
where 𝑋𝑐 is the (𝑁 × 𝑃 ) matrix of data centered on their respective means:

[︂ ]︂
𝑋0 − 𝜇0
𝑋𝑐 = ,
𝑋1 − 𝜇1
where 𝑋0 and 𝑋1 are the (𝑁0 × 𝑃 ) and (𝑁1 × 𝑃 ) matrices of samples of classes 𝐶0 and 𝐶1 .
Let 𝑆𝐵 being the scatter “between-class” matrix, given by
𝑆𝐵 = (𝜇1 − 𝜇0 )(𝜇1 − 𝜇0 )𝑇 .

The linear combination of features 𝑤𝑇 𝑥 have means 𝑤𝑇 𝜇𝑖 for 𝑖 = 0, 1, and variance

𝑤𝑇 𝑋𝑐𝑇 𝑋𝑐 𝑤. Fisher defined the separation between these two distributions to be the ratio
of the variance between the classes to the variance within the classes:
2
𝜎between
𝐹Fisher (𝑤) = 2 (5.38)
𝜎within
(𝑤𝑇 𝜇1 − 𝑤𝑇 𝜇0 )2
= (5.39)
𝑤𝑇 𝑋𝑐𝑇 𝑋𝑐 𝑤
(𝑤𝑇 (𝜇1 − 𝜇0 ))2
= (5.40)
𝑤𝑇 (𝜇1 − 𝜇0 )(𝜇1 − 𝜇0 )𝑇 𝑤
= (5.41)
𝑤 𝑇 𝑆𝐵 𝑤
= 𝑇 . (5.42)
𝑤 𝑆𝑊 𝑤
The Fisher most discriminant projection
In the two-class case, the maximum separation occurs by a projection on the (𝜇1 − 𝜇0 ) using
the Mahalanobis metric 𝑆𝑊 −1 , so that
𝑤 ∝ 𝑆𝑊 −1 (𝜇1 − 𝜇0 ).
Demonstration
Differentiating 𝐹Fisher (𝑤) with respect to 𝑤 gives
∇𝑤 𝐹Fisher (𝑤) = 0
(︂ 𝑇 )︂
𝑤 𝑆𝐵 𝑤
∇𝑤 =0
𝑤 𝑇 𝑆𝑊 𝑤
(𝑤𝑇 𝑆𝑊 𝑤)(2𝑆𝐵 𝑤) − (𝑤𝑇 𝑆𝐵 𝑤)(2𝑆𝑊 𝑤) = 0
(𝑤𝑇 𝑆𝑊 𝑤)(𝑆𝐵 𝑤) = (𝑤𝑇 𝑆𝐵 𝑤)(𝑆𝑊 𝑤)
𝑤 𝑇 𝑆𝐵 𝑤
𝑆𝐵 𝑤 = (𝑆𝑊 𝑤)
𝑤 𝑇 𝑆𝑊 𝑤
𝑆𝐵 𝑤 = 𝜆(𝑆𝑊 𝑤)
𝑆𝑊 −1 𝑆𝐵 𝑤 = 𝜆𝑤.
Since we do not care about the magnitude of 𝑤, only its direction, we replaced the scalar factor
(𝑤𝑇 𝑆𝐵 𝑤)/(𝑤𝑇 𝑆𝑊 𝑤) by 𝜆.
In the multiple-class case, the solutions 𝑤 are determined by the eigenvectors of 𝑆𝑊 −1 𝑆𝐵 that
correspond to the 𝐾 − 1 largest eigenvalues.
However, in the two-class case (in which 𝑆𝐵 = (𝜇1 − 𝜇0 )(𝜇1 − 𝜇0 )𝑇 ) it is easy to show that
𝑤 = 𝑆𝑊 −1 (𝜇1 − 𝜇0 ) is the unique eigenvector of 𝑆𝑊 −1 𝑆𝐵 :

𝑆𝑊 −1 (𝜇1 − 𝜇0 )(𝜇1 − 𝜇0 )𝑇 𝑤 = 𝜆𝑤
𝑆𝑊 −1 (𝜇1 − 𝜇0 )(𝜇1 − 𝜇0 )𝑇 𝑆𝑊 −1 (𝜇1 − 𝜇0 ) = 𝜆𝑆𝑊 −1 (𝜇1 − 𝜇0 ),
where here 𝜆 = (𝜇1 − 𝜇0 )𝑇 𝑆𝑊 −1 (𝜇1 − 𝜇0 ). Which leads to the result
𝑤 ∝ 𝑆𝑊 −1 (𝜇1 − 𝜇0 ).
The separating hyperplane
The separating hyperplane is a 𝑃 − 1-dimensional hyper surface, orthogonal to the projection

vector, 𝑤. There is no single best way to find the origin of the plane along 𝑤, or equivalently
the classification threshold that determines whether a point should be classified as belonging
to 𝐶0 or to 𝐶1 . However, if the projected points have roughly the same distribution, then the
threshold can be chosen as the hyperplane exactly between the projections of the two means,
i.e. as
1
𝑇 = 𝑤 · (𝜇1 − 𝜇0 ).
2

import warnings
warnings.filterwarnings('ignore')
%matplotlib inline
5.4.2 Linear discriminant analysis (LDA)
Linear discriminant analysis (LDA) is a probabilistic generalization of Fisher’s linear discrimi-

nant. It uses Bayes’ rule to fix the threshold based on prior probabilities of classes.
1. First compute the class-conditional distributions of 𝑥 given class 𝐶𝑘 : 𝑝(𝑥|𝐶𝑘 ) =
𝒩 (𝑥|𝜇𝑘 , 𝑆𝑊 ). Where 𝒩 (𝑥|𝜇𝑘 , 𝑆𝑊 ) is the multivariate Gaussian distribution defined over
a P-dimensional vector 𝑥 of continuous variables, which is given by
1 1
𝒩 (𝑥|𝜇𝑘 , 𝑆𝑊 ) = exp{− (𝑥 − 𝜇𝑘 )𝑇 𝑆𝑊 −1 (𝑥 − 𝜇𝑘 )}
(2𝜋)𝑃/2 |𝑆 𝑊 |1/2 2
2. Estimate the prior probabilities of class 𝑘, 𝑝(𝐶𝑘 ) = 𝑁𝑘 /𝑁 .
3. Compute posterior probabilities (ie. the probability of a each class given a sample)
combining conditional with priors using Bayes’ rule:
𝑝(𝐶𝑘 )𝑝(𝑥|𝐶𝑘 )
𝑝(𝐶𝑘 |𝑥) =
𝑝(𝑥)
Where 𝑝(𝑥) is the marginal distribution obtained by suming of classes: As usual, the denom-
inator in Bayes’ theorem can be found in terms of the quantities appearing in the numerator,
because
∑︁
𝑝(𝑥) = 𝑝(𝑥|𝐶𝑘 )𝑝(𝐶𝑘 )
𝑘

Fig. 6: The Fisher most discriminant projection

4. Classify 𝑥 using the Maximum-a-Posteriori probability: 𝐶𝑘 = arg max𝐶𝑘 𝑝(𝐶𝑘 |𝑥)

LDA is a generative model since the class-conditional distributions cal be used to generate
samples of each classes.
LDA is useful to deal with imbalanced group sizes (eg.: 𝑁1 ≫ 𝑁0 ) since priors probabilities can
be used to explicitly re-balance the classification by setting 𝑝(𝐶0 ) = 𝑝(𝐶1 ) = 1/2 or whatever
seems relevant.
LDA can be generalised to the multiclass case with 𝐾 > 2.
With 𝑁1 = 𝑁0 , LDA lead to the same solution than Fisher’s linear discriminant.
Exercise
How many parameters are required to estimate to perform a LDA ?
import numpy as np
from sklearn.discriminant_analysis import LinearDiscriminantAnalysis as LDA
# Dataset
mean0, mean1 = np.array([0, 0]), np.array([0, 2])
Cov = np.array([[1, .8],[.8, 1]])
np.random.seed(42)
X0 = np.random.multivariate_normal(mean0, Cov, n_samples)
X1 = np.random.multivariate_normal(mean1, Cov, n_samples)
X = np.vstack([X0, X1])
y = np.array([0] * X0.shape[0] + [1] * X1.shape[0])
# LDA with scikit-learn

lda = LDA()
proj = lda.fit(X, y).transform(X)
y_pred_lda = lda.predict(X)
errors = y_pred_lda != y
print("Nb errors=%i, error rate=%.2f" % (errors.sum(), errors.sum() / len(y_pred_lda)))
Nb errors=10, error rate=0.05
5.4.3 Logistic regression
Logistic regression is called a generalized linear models. ie.: it is a linear model with a link
function that maps the output of linear multiple regression to the posterior probability of class
1 𝑝(1|𝑥) using the logistic sigmoid function:
1
𝑝(1|𝑤, 𝑥𝑖 ) =
1 + exp(−𝑤 · 𝑥𝑖 )
def logistic(x): return 1 / (1 + np.exp(-x))

def logistic_loss(x): return np.log(1 + np.exp(-x))
x = np.linspace(-6, 6, 100)
plt.subplot(121)


plt.plot(x, logistic(x))
plt.grid(True)
plt.title('Logistic (sigmoid)')
x = np.linspace(-3, 3, 100)
plt.subplot(122)
plt.plot(x, logistic_loss(x), label='Logistic loss')
plt.plot(x, np.maximum(0, 1 - x), label='Hinge loss')
plt.legend()
plt.title('Losses')
plt.grid(True)
The Loss function for sample 𝑖 is the negative log of the probability:
{︃
− log(𝑝(1|𝑤, 𝑥𝑖 )) if 𝑦𝑖 = 1
𝐿(𝑤, 𝑥𝑖 , 𝑦𝑖 ) =
− log(1 − 𝑝(1|𝑤, 𝑥𝑖 ) if 𝑦𝑖 = 0
For the whole dataset 𝑋, 𝑦 = {𝑥𝑖 , 𝑦𝑖 } the loss function to minimize 𝐿(𝑤, 𝑋, 𝑦) is the negative
negative log likelihood (nll) that can be simplied using a 0/1 coding of the label in the case of
binary classification:
𝐿(𝑤, 𝑋, 𝑦) = − log ℒ(𝑤, 𝑋, 𝑦) (5.43)

𝑦𝑖 (1−𝑦𝑖 )
= − log Π𝑖 {𝑝(1|𝑤, 𝑥𝑖 ) * (1 − 𝑝(1|𝑤, 𝑥𝑖 ) } (5.44)
∑︁
= {𝑦𝑖 log 𝑝(1|𝑤, 𝑥𝑖 ) + (1 − 𝑦𝑖 ) log(1 − 𝑝(1|𝑤, 𝑥𝑖 ))} (5.45)
𝑖
∑︁
= {𝑦𝑖 𝑤 · 𝑥𝑖 ) − log(1 + exp(𝑤 · 𝑥𝑖 ))} (5.46)
𝑖
(5.47)

This is solved by numerical method using the gradient of the loss:
𝐿(𝑤, 𝑋, 𝑦) ∑︁
𝜕 = 𝑥𝑖 (𝑦𝑖 − 𝑝(1|𝑤, 𝑥𝑖 ))
𝜕𝑤
𝑖
Logistic regression is a discriminative model since it focuses only on the posterior probability
of each class 𝑝(𝐶𝑘 |𝑥). It only requires to estimate the 𝑃 weight of the 𝑤 vector. Thus it should
be favoured over LDA with many input features. In small dimension and balanced situations it
would provide similar predictions than LDA.
However imbalanced group sizes cannot be explicitly controlled. It can be managed using a
reweighting of the input samples.
from sklearn import linear_model

logreg = linear_model.LogisticRegression(C=1e8, solver='lbfgs')
# This class implements regularized logistic regression. C is the Inverse of␣
˓→regularization strength.
# Large value => no regularization.
logreg.fit(X, y)
y_pred_logreg = logreg.predict(X)
errors = y_pred_logreg != y
print("Nb errors=%i, error rate=%.2f" % (errors.sum(), errors.sum() / len(y_pred_logreg)))
print(logreg.coef_)

[[-5.1516729 5.57303883]]
Exercise
Explore the Logistic Regression parameters and proposes a solution in cases of highly im-
balanced training dataset 𝑁1 ≫ 𝑁0 when we know that in reality both classes have the same
probability 𝑝(𝐶1 ) = 𝑝(𝐶0 ).
5.4.4 Overfitting
VC dimension (for Vapnik–Chervonenkis dimension) is a measure of the capacity (complexity,

expressive power, richness, or flexibility) of a statistical classification algorithm, defined as the
cardinality of the largest set of points that the algorithm can shatter.
Theorem: Linear classifier in 𝑅𝑃 have VC dimension of 𝑃 + 1. Hence in dimension two (𝑃 = 2)
any random partition of 3 points can be learned.
Fig. 7: In 2D we can shatter any three non-collinear points

5.4.5 Ridge Fisher’s linear classification (L2-regularization)
When the matrix 𝑆𝑊 is not full rank or 𝑃 ≫ 𝑁 , the The Fisher most discriminant projection
estimate of the is not unique. This can be solved using a biased version of 𝑆𝑊 :
𝑆𝑊 𝑅𝑖𝑑𝑔𝑒 = 𝑆𝑊 + 𝜆𝐼
where 𝐼 is the 𝑃 × 𝑃 identity matrix. This leads to the regularized (ridge) estimator of the
Fisher’s linear discriminant analysis:
𝑤𝑅𝑖𝑑𝑔𝑒 ∝ (𝑆𝑊 + 𝜆𝐼)−1 (𝜇1 − 𝜇0 )
Fig. 8: The Ridge Fisher most discriminant projection
Increasing 𝜆 will:
• Shrinks the coefficients toward zero.
• The covariance will converge toward the diagonal matrix, reducing the contribution of
the pairwise covariances.
5.4.6 Ridge logistic regression (L2-regularization)
The objective function to be minimized is now the combination of the logistic loss (negative
log likelyhood) − log ℒ(𝑤) with a penalty of the L2 norm of the weights vector. In the two-class
case, using the 0/1 coding we obtain:
min Logistic ridge(𝑤) = − log ℒ(𝑤, 𝑋, 𝑦) + 𝜆 ‖𝑤‖2

𝑤
# Dataset
# Build a classification task using 3 informative features
X, y = datasets.make_classification(n_samples=100,
n_features=20,
n_informative=3,
n_redundant=0,
n_repeated=0,


n_classes=2,
random_state=0,
shuffle=False)
import numpy as np

lr = linear_model.LogisticRegression(C=1, solver='lbfgs')
lr.fit(X, y)
y_pred_lr = lr.predict(X)
# Retrieve proba from coef vector

print(lr.coef_.shape)
probas = 1 / (1 + np.exp(- (np.dot(X, lr.coef_.T) + lr.intercept_))).ravel()
print("Diff", np.max(np.abs(lr.predict_proba(X)[:, 1] - probas)))
#plt.plot(lr.predict_proba(X)[:, 1], probas, "ob")
errors = y_pred_lr != y
print("Nb errors=%i, error rate=%.2f" % (errors.sum(), errors.sum() / len(y)))
print(lr.coef_)
(1, 20)
Diff 0.0
[[-0.12899737 0.7579822 -0.01228473 -0.11412421 0.25491221 0.4329847
0.14564739 0.16763962 0.85071394 0.02116803 -0.1611039 -0.0146019
-0.03399884 0.43127728 -0.05831644 -0.0812323 0.15877844 0.29387389
0.54659524 0.03376169]]
5.4.7 Lasso logistic regression (L1-regularization)
The objective function to be minimized is now the combination of the logistic loss − log ℒ(𝑤)
with a penalty of the L1 norm of the weights vector. In the two-class case, using the 0/1 coding
we obtain:
min Logistic Lasso(𝑤) = − log ℒ(𝑤, 𝑋, 𝑦) + 𝜆 ‖𝑤‖1

𝑤

lrl1 = linear_model.LogisticRegression(penalty='l1', solver='saga')
lrl1.fit(X, y)
y_pred_lrl1 = lrl1.predict(X)
errors = y_pred_lrl1 != y


print("Nb errors=%i, error rate=%.2f" % (errors.sum(), errors.sum() / len(y_pred_lrl1)))
print(lrl1.coef_)

[[-0.12618366 0.71711208 0. -0.00530107 0.20418502 0.38240135
0.07003549 0.06296028 0.76841952 0. -0.10625209 0.
0. 0.34010421 0. 0. 0.08038212 0.19202935
0.47988515 0. ]]
5.4.8 Ridge linear Support Vector Machine (L2-regularization)
Support Vector Machine seek for separating hyperplane with maximum margin to enforce ro-
bustness against noise. Like logistic regression it is a discriminative method that only focuses
of predictions.
Here we present the non separable case of Maximum Margin Classifiers with ±1 coding (ie.:
𝑦𝑖 {−1, +1}). In the next figure the legend aply to samples of “dot” class.
Fig. 9: Linear lar margin classifiers
Linear SVM for classification (also called SVM-C or SVC) minimizes:

min Linear SVM(𝑤) = penalty(𝑤) + 𝐶 Hinge loss(𝑤)
∑︀𝑁
= ‖𝑤‖22 + 𝐶 𝑖 𝜉𝑖
with ∀𝑖 𝑦𝑖 (𝑤 · 𝑥𝑖 ) ≥ 1 − 𝜉𝑖
Here we introduced the slack variables: 𝜉𝑖 , with 𝜉𝑖 = 0 for points that are on or inside the
correct margin boundary and 𝜉𝑖 = |𝑦𝑖 − (𝑤 𝑐𝑑𝑜𝑡 · 𝑥𝑖 )| for other points. Thus:
1. If 𝑦𝑖 (𝑤 · 𝑥𝑖 ) ≥ 1 then the point lies outside the margin but on the correct side of the
decision boundary. In this case 𝜉𝑖 = 0. The constraint is thus not active for this point. It
does not contribute to the prediction.
2. If 1 > 𝑦𝑖 (𝑤 · 𝑥𝑖 ) ≥ 0 then the point lies inside the margin and on the correct side of the
decision boundary. In this case 0 < 𝜉𝑖 ≤ 1. The constraint is active for this point. It does
contribute to the prediction as a support vector.

3. If 0 < 𝑦𝑖 (𝑤 · 𝑥𝑖 )) then the point is on the wrong side of the decision boundary (missclassi-
fication). In this case 0 < 𝜉𝑖 > 1. The constraint is active for this point. It does contribute
to the prediction as a support vector.
This loss is called the hinge loss, defined as:
max(0, 1 − 𝑦𝑖 (𝑤 · 𝑥𝑖 ))
So linear SVM is closed to Ridge logistic regression, using the hinge loss instead of the logistic
loss. Both will provide very similar predictions.
from sklearn import svm
svmlin = svm.LinearSVC()
# Remark: by default LinearSVC uses squared_hinge as loss
svmlin.fit(X, y)
y_pred_svmlin = svmlin.predict(X)
errors = y_pred_svmlin != y
print("Nb errors=%i, error rate=%.2f" % (errors.sum(), errors.sum() / len(y_pred_svmlin)))
print(svmlin.coef_)

[[-0.05612121 0.31189716 0.00272117 -0.05148767 0.09939817 0.17726852
0.06519342 0.08921434 0.35339727 0.0060124 -0.06201436 -0.00741298
-0.02157145 0.18271762 -0.02162947 -0.04060884 0.07204494 0.13083561
0.23721824 0.00824139]]
5.4.9 Lasso linear Support Vector Machine (L1-regularization)
Linear SVM for classification (also called SVM-C or SVC) with l1-regularization
∑︀𝑁
min 𝐹Lasso linear SVM (𝑤) = ||𝑤||1 + 𝐶 𝑖 𝜉𝑖
with ∀𝑖 𝑦𝑖 (𝑤 · 𝑥𝑖 ) ≥ 1 − 𝜉𝑖
from sklearn import svm
svmlinl1 = svm.LinearSVC(penalty='l1', dual=False)

# Remark: by default LinearSVC uses squared_hinge as loss
svmlinl1.fit(X, y)
y_pred_svmlinl1 = svmlinl1.predict(X)
errors = y_pred_svmlinl1 != y
print("Nb errors=%i, error rate=%.2f" % (errors.sum(), errors.sum() / len(y_pred_
˓→svmlinl1)))
print(svmlinl1.coef_)

[[-0.05333926 0.29934673 0. -0.03541846 0.09261443 0.1676322
0.05808032 0.07587832 0.34065463 0. -0.05559105 -0.00194167
-0.01312533 0.16866547 -0.01450459 -0.02500777 0.06074154 0.11739131
0.22485568 0.00473276]]

## Exercise
Compare predictions of Logistic regression (LR) and their SVM counterparts, ie.: L2 LR vs L2
SVM and L1 LR vs L1 SVM
• Compute the correlation between pairs of weights vectors.
• Compare the predictions of two classifiers using their decision function:
– Give the equation of the decision function for a linear classifier, assuming that their
is no intercept.
– Compute the correlation decision function.
– Plot the pairwise decision function of the classifiers.
• Conclude on the differences between Linear SVM and logistic regression.
5.4.10 Elastic-net classification (L2-L1-regularization)
The objective function to be minimized is now the combination of the logistic loss log 𝐿(𝑤) or
the hinge loss with combination of L1 and L2 penalties. In the two-class case, using the 0/1
coding we obtain:
min Logistic enet(𝑤) = − log ℒ(𝑤, 𝑋, 𝑦) + 𝛼 𝜌 ‖𝑤‖1 + (1 − 𝜌) ‖𝑤‖22

(︀ )︀
(5.48)
min Hinge enet(𝑤) = Hinge loss(𝑤) + 𝛼 𝜌 ‖𝑤‖1 + (1 − 𝜌) ‖𝑤‖22
(︀ )︀
(5.49)

from sklearn import linear_model as lm
n_features=20,
n_informative=3,
n_redundant=0,
n_repeated=0,
n_classes=2,
random_state=0,
shuffle=False)
enetloglike = lm.SGDClassifier(loss="log", penalty="elasticnet",

alpha=0.0001, l1_ratio=0.15, class_weight='balanced')
enetloglike.fit(X, y)
enethinge = lm.SGDClassifier(loss="hinge", penalty="elasticnet",

alpha=0.0001, l1_ratio=0.15, class_weight='balanced')
enethinge.fit(X, y)
SGDClassifier(class_weight='balanced', penalty='elasticnet')

Exercise
Compare predictions of Elastic-net Logistic regression (LR) and Hinge-loss Elastic-net

• Compute the correlation between pairs of weights vectors.
• Compare the predictions of two classifiers using their decision function:
– Compute the correlation decision function.
– Plot the pairwise decision function of the classifiers.
• Conclude on the differences between the two losses.
5.4.11 Metrics of classification performance evaluation
Metrics for binary classification
source: https://en.wikipedia.org/wiki/Sensitivity_and_specificity
Imagine a study evaluating a new test that screens people for a disease. Each person taking the
test either has or does not have the disease. The test outcome can be positive (classifying the
person as having the disease) or negative (classifying the person as not having the disease). The
test results for each subject may or may not match the subject’s actual status. In that setting:
• True positive (TP): Sick people correctly identified as sick
• False positive (FP): Healthy people incorrectly identified as sick
• True negative (TN): Healthy people correctly identified as healthy
• False negative (FN): Sick people incorrectly identified as healthy
• Accuracy (ACC):
ACC = (TP + TN) / (TP + FP + FN + TN)
• Sensitivity (SEN) or recall of the positive class or true positive rate (TPR) or hit rate:
SEN = TP / P = TP / (TP+FN)
• Specificity (SPC) or recall of the negative class or true negative rate:
SPC = TN / N = TN / (TN+FP)
• Precision or positive predictive value (PPV):
PPV = TP / (TP + FP)
• Balanced accuracy (bACC):is a useful performance measure is the balanced accuracy
which avoids inflated performance estimates on imbalanced datasets (Brodersen, et
al. (2010). “The balanced accuracy and its posterior distribution”). It is defined as the
arithmetic mean of sensitivity and specificity, or the average accuracy obtained on either
class:
bACC = 1/2 * (SEN + SPC)
• F1 Score (or F-score) which is a weighted average of precision and recall are usefull to
deal with imballaced datasets

The four outcomes can be formulated in a 2×2 contingency table or confusion matrix https:
//en.wikipedia.org/wiki/Sensitivity_and_specificity
For more precision see: http://scikit-learn.org/stable/modules/model_evaluation.html

y_pred = [0, 1, 0, 0]
y_true = [0, 1, 0, 1]
metrics.accuracy_score(y_true, y_pred)
# The overall precision an recall

metrics.precision_score(y_true, y_pred)
metrics.recall_score(y_true, y_pred)
# Recalls on individual classes: SEN & SPC

recalls = metrics.recall_score(y_true, y_pred, average=None)
recalls[0] # is the recall of class 0: specificity
recalls[1] # is the recall of class 1: sensitivity
# Balanced accuracy
b_acc = recalls.mean()
# The overall precision an recall on each individual class

p, r, f, s = metrics.precision_recall_fscore_support(y_true, y_pred)
Significance of classification rate
P-value associated to classification rate. Compared the number of correct classifications (=ac-
curacy ×𝑁 ) to the null hypothesis of Binomial distribution of parameters 𝑝 (typically 50% of
chance level) and 𝑁 (Number of observations).
Is 65% of accuracy a significant prediction rate among 70 observations?
Since this is an exact, two-sided test of the null hypothesis, the p-value can be divided by 2
since we test that the accuracy is superior to the chance level.
import scipy.stats
acc, N = 0.65, 70
pval = scipy.stats.binom_test(x=int(acc * N), n=N, p=0.5) / 2
print(pval)
0.01123144774625465

Area Under Curve (AUC) of Receiver operating characteristic (ROC)
Some classifier may have found a good discriminative projection 𝑤. However if the threshold
to decide the final predicted class is poorly adjusted, the performances will highlight an high
specificity and a low sensitivity or the contrary.
In this case it is recommended to use the AUC of a ROC analysis which basically provide a mea-
sure of overlap of the two classes when points are projected on the discriminative axis. For more
detail on ROC and AUC see:https://en.wikipedia.org/wiki/Receiver_operating_characteristic.

score_pred = np.array([.1 ,.2, .3, .4, .5, .6, .7, .8])
y_true = np.array([0, 0, 0, 0, 1, 1, 1, 1])
thres = .9
y_pred = (score_pred > thres).astype(int)
print("Predictions:", y_pred)
metrics.accuracy_score(y_true, y_pred)
# The overall precision an recall on each individual class

p, r, f, s = metrics.precision_recall_fscore_support(y_true, y_pred)
print("Recalls:", r)
# 100% of specificity, 0% of sensitivity
# However AUC=1 indicating a perfect separation of the two classes

auc = metrics.roc_auc_score(y_true, score_pred)
print("AUC:", auc)
Predictions: [0 0 0 0 0 0 0 0]
Recalls: [1. 0.]
AUC: 1.0
5.4.12 Imbalanced classes
Learning with discriminative (logistic regression, SVM) methods is generally based on minimiz-
ing the misclassification of training samples, which may be unsuitable for imbalanced datasets
where the recognition might be biased in favor of the most numerous class. This problem
can be addressed with a generative approach, which typically requires more parameters to be
determined leading to reduced performances in high dimension.
Dealing with imbalanced class may be addressed by three main ways (see Japkowicz and
Stephen (2002) for a review), resampling, reweighting and one class learning.
In sampling strategies, either the minority class is oversampled or majority class is undersam-
pled or some combination of the two is deployed. Undersampling (Zhang and Mani, 2003) the
majority class would lead to a poor usage of the left-out samples. Sometime one cannot afford
such strategy since we are also facing a small sample size problem even for the majority class.
Informed oversampling, which goes beyond a trivial duplication of minority class samples, re-
quires the estimation of class conditional distributions in order to generate synthetic samples.
Here generative models are required. An alternative, proposed in (Chawla et al., 2002) generate
samples along the line segments joining any/all of the k minority class nearest neighbors. Such
procedure blindly generalizes the minority area without regard to the majority class, which may
be particularly problematic with high-dimensional and potentially skewed class distribution.

Reweighting, also called cost-sensitive learning, works at an algorithmic level by adjusting

the costs of the various classes to counter the class imbalance. Such reweighting can be im-
plemented within SVM (Chang and Lin, 2001) or logistic regression (Friedman et al., 2010)
classifiers. Most classifiers of Scikit learn offer such reweighting possibilities.
The class_weight parameter can be positioned into the "balanced" mode which uses the values
of 𝑦 to automatically adjust weights inversely proportional to class frequencies in the input data
as 𝑁/(2𝑁𝑘 ).
import numpy as np
# dataset
n_features=5,
n_informative=2,
n_redundant=0,
n_repeated=0,
n_classes=2,
random_state=1,
shuffle=False)
print(*["#samples of class %i = %i;" % (lev, np.sum(y == lev)) for lev in np.unique(y)])
print('# No Reweighting balanced dataset')

lr_inter = linear_model.LogisticRegression(C=1)
lr_inter.fit(X, y)
p, r, f, s = metrics.precision_recall_fscore_support(y, lr_inter.predict(X))
print("SPC: %.3f; SEN: %.3f" % tuple(r))
print('# => The predictions are balanced in sensitivity and specificity\n')
# Create imbalanced dataset, by subsampling sample of class 0: keep only 10% of

# class 0's samples and all class 1's samples.
n0 = int(np.rint(np.sum(y == 0) / 20))
subsample_idx = np.concatenate((np.where(y == 0)[0][:n0], np.where(y == 1)[0]))
Ximb = X[subsample_idx, :]
yimb = y[subsample_idx]
print(*["#samples of class %i = %i;" % (lev, np.sum(yimb == lev)) for lev in
np.unique(yimb)])
print('# No Reweighting on imbalanced dataset')

lr_inter = linear_model.LogisticRegression(C=1)
lr_inter.fit(Ximb, yimb)
p, r, f, s = metrics.precision_recall_fscore_support(yimb, lr_inter.predict(Ximb))
print('# => Sensitivity >> specificity\n')
print('# Reweighting on imbalanced dataset')

lr_inter_reweight = linear_model.LogisticRegression(C=1, class_weight="balanced")
lr_inter_reweight.fit(Ximb, yimb)
p, r, f, s = metrics.precision_recall_fscore_support(yimb,
lr_inter_reweight.predict(Ximb))
print('# => The predictions are balanced in sensitivity and specificity\n')

#samples of class 0 = 250; #samples of class 1 = 250;

# No Reweighting balanced dataset
SPC: 0.940; SEN: 0.928
# => The predictions are balanced in sensitivity and specificity
#samples of class 0 = 12; #samples of class 1 = 250;

# No Reweighting on imbalanced dataset
SPC: 0.750; SEN: 0.996
# => Sensitivity >> specificity
# Reweighting on imbalanced dataset

SPC: 1.000; SEN: 0.980
# => The predictions are balanced in sensitivity and specificity
5.4.13 Exercise
Fisher linear discriminant rule
Write a class FisherLinearDiscriminant that implements the Fisher’s linear discriminant anal-
ysis. This class must be compliant with the scikit-learn API by providing two methods: - fit(X,
y) which fits the model and returns the object itself; - predict(X) which returns a vector of the
predicted values. Apply the object on the dataset presented for the LDA.
5.5 Non linear learning algorithms
5.5.1 Support Vector Machines (SVM)
SVM are based kernel methods require only a user-specified kernel function 𝐾(𝑥𝑖 , 𝑥𝑗 ), i.e., a
similarity function over pairs of data points (𝑥𝑖 , 𝑥𝑗 ) into kernel (dual) space on which learning
algorithms operate linearly, i.e. every operation on points is a linear combination of 𝐾(𝑥𝑖 , 𝑥𝑗 ).
Outline of the SVM algorithm:
1. Map points 𝑥 into kernel space using a kernel function: 𝑥 → 𝐾(𝑥, .).
2. Learning algorithms operates linearly by dot product into high-kernel space 𝐾(., 𝑥𝑖 ) ·
𝐾(., 𝑥𝑗 ).
• Using the kernel trick (Mercer’s Theorem) replaces dot product in high dimensional
space by a simpler operation such that 𝐾(., 𝑥𝑖 ) · 𝐾(., 𝑥𝑗 ) = 𝐾(𝑥𝑖 , 𝑥𝑗 ). Thus we only
need to compute a similarity measure for each pairs of point and store in a 𝑁 × 𝑁
Gram matrix.
• Finally, The learning process consist of estimating the 𝛼𝑖 of the decision function that
maximises the hinge loss (of 𝑓 (𝑥)) plus some penalty when applied on all training
points.
(︃ 𝑁 )︃
∑︁
𝑓 (𝑥) = sign 𝛼𝑖 𝑦𝑖 𝐾(𝑥𝑖 , 𝑥) .
𝑖
3. Predict a new point 𝑥 using the decision function.

Fig. 10: Support Vector Machines.
Gaussian kernel (RBF, Radial Basis Function):
One of the most commonly used kernel is the Radial Basis Function (RBF) Kernel. For a pair of
points 𝑥𝑖 , 𝑥𝑗 the RBF kernel is defined as:
‖𝑥𝑖 − 𝑥𝑗 ‖2
(︂ )︂
𝐾(𝑥𝑖 , 𝑥𝑗 ) = exp − (5.50)
2𝜎 2
= exp −𝛾 ‖𝑥𝑖 − 𝑥𝑗 ‖2
(︀ )︀
(5.51)
Where 𝜎 (or 𝛾) defines the kernel width parameter. Basically, we consider a Gaussian function
centered on each training sample 𝑥𝑖 . it has a ready interpretation as a similarity measure as it
decreases with squared Euclidean distance between the two feature vectors.
Non linear SVM also exists for regression problems.
%matplotlib inline
import warnings
warnings.filterwarnings(action='once')
import numpy as np
from sklearn.svm import SVC
# dataset
X, y = datasets.make_classification(n_samples=10, n_features=2,n_redundant=0,
n_classes=2,
5.5. Non linear learning algorithms 215


random_state=1,
shuffle=False)
clf = SVC(kernel='rbf')#, gamma=1)
clf.fit(X, y)
print("#Errors: %i" % np.sum(y != clf.predict(X)))
clf.decision_function(X)
# Usefull internals:
# Array of support vectors
clf.support_vectors_
# indices of support vectors within original X

np.all(X[clf.support_,:] == clf.support_vectors_)
/home/ed203246/anaconda3/lib/python3.7/site-packages/ipykernel/ipkernel.py:287:␣
˓→DeprecationWarning: should_run_async will not call transform_cell automatically␣
˓→in the future. Please pass the result to transformed_cell argument and any␣
˓→exception that happen during thetransform in preprocessing_exc_tuple in IPython␣
˓→7.17 and above.
and should_run_async(code)
/home/ed203246/anaconda3/lib/python3.7/importlib/_bootstrap.py:219:␣
˓→RuntimeWarning: numpy.ufunc size changed, may indicate binary incompatibility.␣
˓→Expected 192 from C header, got 216 from PyObject
return f(*args, **kwds)


#Errors: 0
/home/ed203246/anaconda3/lib/python3.7/importlib/_bootstrap.py:219: RuntimeWarning: numpy.

˓→ufunc size changed, may indicate binary incompatibility. Expected 192 from C header,␣
˓→got 216 from PyObject
True

5.5.2 Decision tree
A tree can be “learned” by splitting the training dataset into subsets based on an features value
test.
Each internal node represents a “test” on an feature resulting on the split of the current sample.
At each step the algorithm selects the feature and a cutoff value that maximises a given metric.
Different metrics exist for regression tree (target is continuous) or classification tree (the target
is qualitative).
This process is repeated on each derived subset in a recursive manner called recursive partition-
ing. The recursion is completed when the subset at a node has all the same value of the target
variable, or when splitting no longer adds value to the predictions. This general principle is
implemented by many recursive partitioning tree algorithms.
Fig. 11: Classification tree.
Decision trees are simple to understand and interpret however they tend to overfit the data.
However decision trees tend to overfit the training set. Leo Breiman propose random forest to
deal with this issue.
A single decision tree is usually overfits the data it is learning from because it learn from only
one pathway of decisions. Predictions from a single decision tree usually don’t make accurate
predictions on new data.
5.5. Non linear learning algorithms 217

5.5.3 Random forest
A random forest is a meta estimator that fits a number of decision tree learners on various
sub-samples of the dataset and use averaging to improve the predictive accuracy and control
over-fitting.
Random forest models reduce the risk of overfitting by introducing randomness by:
• building multiple trees (n_estimators)
• drawing observations with replacement (i.e., a bootstrapped sample)
• splitting nodes on the best split among a random subset of the features selected at every
node
from sklearn.ensemble import RandomForestClassifier
forest = RandomForestClassifier(n_estimators = 100)

forest.fit(X, y)
print("#Errors: %i" % np.sum(y != forest.predict(X)))
#Errors: 0
/home/ed203246/anaconda3/lib/python3.7/site-packages/ipykernel/ipkernel.py:287:␣
˓→DeprecationWarning: should_run_async will not call transform_cell automatically␣
˓→in the future. Please pass the result to transformed_cell argument and any␣
˓→exception that happen during thetransform in preprocessing_exc_tuple in IPython␣
˓→7.17 and above.
and should_run_async(code)

5.5.4 Extra Trees (Low Variance)
Extra Trees is like Random Forest, in that it builds multiple trees and splits nodes using random
subsets of features, but with two key differences: it does not bootstrap observations (meaning
it samples without replacement), and nodes are split on random splits, not best splits. So, in
summary, ExtraTrees: builds multiple trees with bootstrap = False by default, which means it
samples without replacement nodes are split based on random splits among a random subset of
the features selected at every node In Extra Trees, randomness doesn’t come from bootstrapping
of data, but rather comes from the random splits of all observations. ExtraTrees is named for
(Extremely Randomized Trees).

5.6 Resampling Methods

import warnings
warnings.filterwarnings('ignore')
%matplotlib inline
5.6.1 Train, Validation and Test Sets
Machine learning algorithms overfit taining data. Predictive performances MUST be evaluated
on independant hold-out dataset.
Fig. 12: Train, Validation and Test Sets.
1. Training dataset: Dataset used to fit the model (set the model parameters like weights).
The training error can be easily calculated by applying the statistical learning method to
the observations used in its training. But because of overfitting, the training error rate
can dramatically underestimate the error that would be obtained on new samples.
2. Validation dataset: Dataset used to provide an unbiased evaluation of a model fit on the
training dataset while tuning model hyperparameters. The validation error is the aver-
age error that results from a learning method to predict the response on a new (validation)
samples that is, on samples that were not used in training the method.
3. Test Dataset: Dataset used to provide an unbiased evaluation of a final model fit on the
training dataset. It is only used once a model is completely trained(using the train and
validation sets).
What is the Difference Between Test and Validation Datasets? by Jason Brownlee
Thus the original dataset is generally split in a training, validation and a test data sets. Large
training+validation set (80%) small test set (20%) might provide a poor estimation of the pre-
dictive performances (same argument stands for train vs validation samples). On the contrary,
5.6. Resampling Methods 219

large test set and small training set might produce a poorly estimated learner. This is why, on
situation where we cannot afford such split, it recommended to use cross-validation scheme to
estimate the predictive power of a learning algorithm.
5.6.2 Cross-Validation (CV)
Cross-Validation scheme randomly divides the set of observations into 𝐾 groups, or folds, of
approximately equal size. The first fold is treated as a validation set, and the method 𝑓 () is
fitted on the remaining union of 𝐾 − 1 folds: (𝑓 (𝑋 −𝐾 , 𝑦 −𝐾 )).
The measure of performance (the score function 𝒮), either a error measure or an correct predic-
tion measure is an average of a loss error or correct prediction measure, noted ℒ, between a true
target value and the predicted target value. The score function is evaluated of the on the obser-
vations in the held-out fold. For each sample 𝑖 we consider the model estimated 𝑓 (𝑋 −𝑘(𝑖) , 𝑦 −𝑘(𝑖)
on the data set without the group 𝑘 that contains 𝑖 noted −𝑘(𝑖). This procedure is repeated 𝐾
times; each time, a different group of observations is treated as a test set. Then we compare the
predicted value (𝑓−𝑘(𝑖) (𝑥𝑖 ) = 𝑦ˆ𝑖 ) with true value 𝑦𝑖 using a Error or Loss function ℒ(𝑦, 𝑦ˆ).
For 10-fold we can either average over 10 values (Macro measure) or concatenate the 10 ex-
periments and compute the micro measures.
Two strategies micro vs macro estimates:
Micro measure: average(individual scores): compute a score 𝒮 for each sample and average
over all samples. It is simillar to average score(concatenation): an averaged score computed
over all concatenated samples.
𝑁
1 ∑︁ (︁ )︁
𝒮(𝑓 ) = ℒ 𝑦𝑖 , 𝑓 (𝑥−𝑘(𝑖) , 𝑦 −𝑘(𝑖) ) .
𝑁
𝑖
Macro measure mean(CV scores) (the most commonly used method): compute a score 𝒮 on
each each fold 𝑘 and average accross folds:
𝐾
1 ∑︁
𝒮(𝑓 ) = 𝒮𝑘 (𝑓 ).
𝐾
𝑘
𝐾
1 ∑︁ 1 ∑︁ (︁ )︁
𝒮(𝑓 ) = ℒ 𝑦𝑖 , 𝑓 (𝑥−𝑘(𝑖) , 𝑦 −𝑘(𝑖) ) .
𝐾 𝑁𝑘
𝑘 𝑖∈𝑘
These two measures (an average of average vs. a global average) are generaly similar. They
may differ slightly is folds are of different sizes.
This validation scheme is known as the K-Fold CV. Typical choices of 𝐾 are 5 or 10, [Kohavi
1995]. The extreme case where 𝐾 = 𝑁 is known as leave-one-out cross-validation, LOO-CV.

CV for regression
Usually the error function ℒ() is the r-squared score. However other function could be used.
%matplotlib inline
import warnings
warnings.filterwarnings(action='once')
import numpy as np
from sklearn.model_selection import KFold
X, y = datasets.make_regression(n_samples=100, n_features=100,
n_informative=10, random_state=42)
estimator = lm.Ridge(alpha=10)
cv = KFold(n_splits=5, random_state=42)
r2_train, r2_test = list(), list()
for train, test in cv.split(X):

estimator.fit(X[train, :], y[train])
r2_train.append(metrics.r2_score(y[train], estimator.predict(X[train, :])))
r2_test.append(metrics.r2_score(y[test], estimator.predict(X[test, :])))
print("Train r2:%.2f" % np.mean(r2_train))

print("Test r2:%.2f" % np.mean(r2_test))
Train r2:0.99
Test r2:0.73
Scikit-learn provides user-friendly function to perform CV:
from sklearn.model_selection import cross_val_score
scores = cross_val_score(estimator=estimator, X=X, y=y, cv=5)

print("Test r2:%.2f" % scores.mean())
# provide a cv
scores = cross_val_score(estimator=estimator, X=X, y=y, cv=cv)
Test r2:0.73
Test r2:0.73

CV for classification
With classification problems it is essential to sample folds where each set contains approxi-
mately the same percentage of samples of each target class as the complete set. This is called
stratification. In this case, we will use StratifiedKFold with is a variation of k-fold which
returns stratified folds.
Usually the error function 𝐿() are, at least, the sensitivity and the specificity. However other
function could be used.
import numpy as np
from sklearn.model_selection import StratifiedKFold
X, y = datasets.make_classification(n_samples=100, n_features=100,
estimator = lm.LogisticRegression(C=1, solver='lbfgs')
cv = StratifiedKFold(n_splits=5)
# Lists to store scores by folds (for macro measure only)

recalls_train, recalls_test, acc_test = list(), list(), list()
# Or vector of test predictions (for both macro and micro measures, not for training␣
˓→samples)
y_test_pred = np.zeros(len(y))
for train, test in cv.split(X, y):

estimator.fit(X[train, :], y[train])
recalls_train.append(metrics.recall_score(y[train], estimator.predict(X[train, :]),␣
˓→average=None))
recalls_test.append(metrics.recall_score(y[test], estimator.predict(X[test, :]),␣

˓→average=None))
acc_test.append(metrics.accuracy_score(y[test], estimator.predict(X[test, :])))
# Store test predictions (for micro measures)

y_test_pred[test] = estimator.predict(X[test, :])
print("== Macro measures ==")

# Use lists of scores
recalls_train = np.array(recalls_train)
recalls_test = np.array(recalls_test)
print("Train SPC:%.2f; SEN:%.2f" % tuple(recalls_train.mean(axis=0)))
print("Test SPC:%.2f; SEN:%.2f" % tuple(recalls_test.mean(axis=0)), )
print("Test ACC:%.2f, ballanced ACC:%.2f" %
(np.mean(acc_test), recalls_test.mean(axis=1).mean()), "Folds:", acc_test)
# Or use vector to test predictions

acc_test = [metrics.accuracy_score(y[test], y_test_pred[test]) for train, test in cv.
˓→split(X, y)]
print("Test ACC:%.2f" % np.mean(acc_test), "Folds:", acc_test)
print("== Micro measures ==")

print("Test SPC:%.2f; SEN:%.2f" % \


tuple(metrics.recall_score(y, y_test_pred, average=None)))
print("Test ACC:%.2f" % metrics.accuracy_score(y, y_test_pred))
== Macro measures ==
Train SPC:1.00; SEN:1.00
Test SPC:0.78; SEN:0.82
Test ACC:0.80, ballanced ACC:0.80 Folds: [0.9, 0.7, 0.95, 0.7, 0.75]
Test ACC:0.80 Folds: [0.9, 0.7, 0.95, 0.7, 0.75]
== Micro measures ==
Test SPC:0.78; SEN:0.82
Test ACC:0.80
Scikit-learn provides user-friendly function to perform CV:
scores = cross_val_score(estimator=estimator, X=X, y=y, cv=5)

scores.mean()
# provide CV and score

def balanced_acc(estimator, X, y, **kwargs):
'''
Balanced acuracy scorer
'''
return metrics.recall_score(y, estimator.predict(X), average=None).mean()
scores = cross_val_score(estimator=estimator, X=X, y=y, cv=5, scoring=balanced_acc)

print("Test ACC:%.2f" % scores.mean())
Test ACC:0.80
Note that with Scikit-learn user-friendly function we average the scores’ average obtained on
individual folds which may provide slightly different results that the overall average presented
earlier.
5.6.3 Parallel computation with joblib
Dataset
import numpy as np
from sklearn.model_selection import StratifiedKFold
X, y = datasets.make_classification(n_samples=20, n_features=5, n_informative=2, random_
˓→state=42)
cv = StratifiedKFold(n_splits=5)

Use cross_validate function
from sklearn.model_selection import cross_validate

cv_results = cross_validate(estimator, X, y, cv=cv, n_jobs=5)
print(np.mean(cv_results['test_score']), cv_results['test_score'])
0.8 [0.5 0.5 1. 1. 1. ]
### Sequential computation

If we want have full control of the operations performed within each fold (retrieve the models
parameters, etc.). We would like to parallelize the folowing sequetial code:

y_test_pred_seq = np.zeros(len(y)) # Store predictions in the original order
coefs_seq = list()
X_train, X_test, y_train, y_test = X[train, :], X[test, :], y[train], y[test]
estimator.fit(X_train, y_train)
y_test_pred_seq[test] = estimator.predict(X_test)
coefs_seq.append(estimator.coef_)
test_accs = [metrics.accuracy_score(y[test], y_test_pred_seq[test]) for train, test in cv.

˓→split(X, y)]
print(np.mean(test_accs), test_accs)
coefs_cv = np.array(coefs_seq)
print(coefs_cv)
print(coefs_cv.mean(axis=0))
print("Std Err of the coef")
print(coefs_cv.std(axis=0) / np.sqrt(coefs_cv.shape[0]))
0.8 [0.5, 0.5, 1.0, 1.0, 1.0]

[[[-0.87692513 0.6260013 1.18714373 -0.30685978 -0.38037393]]
[[-0.7464993 0.62138165 1.10144804 0.19800115 -0.40112109]]
[[-0.96020317 0.51135134 1.1210943 0.08039112 -0.2643663 ]]
[[-0.85755505 0.52010552 1.06637346 -0.10994258 -0.29152132]]
[[-0.89914467 0.51481483 1.08675378 -0.24767837 -0.27899525]]]

[[-0.86806546 0.55873093 1.11256266 -0.07721769 -0.32327558]]
Std Err of the coef
[[0.03125544 0.02376198 0.01850211 0.08566194 0.02510739]]

Parallel computation with joblib
from sklearn.externals.joblib import Parallel, delayed

from sklearn.base import is_classifier, clone
def _split_fit_predict(estimator, X, y, train, test):

estimator.fit(X_train, y_train)
return [estimator.predict(X_test), estimator.coef_]
parallel = Parallel(n_jobs=5)
cv_ret = parallel(
delayed(_split_fit_predict)(
clone(estimator), X, y, train, test)
for train, test in cv.split(X, y))
y_test_pred_cv, coefs_cv = zip(*cv_ret)
# Retrieve predictions in the original order

y_test_pred = np.zeros(len(y))
for i, (train, test) in enumerate(cv.split(X, y)):
y_test_pred[test] = y_test_pred_cv[i]
test_accs = [metrics.accuracy_score(y[test], y_test_pred[test]) for train, test in cv.

˓→split(X, y)]
print(np.mean(test_accs), test_accs)
0.8 [0.5, 0.5, 1.0, 1.0, 1.0]
Test same predictions and same coeficients
assert np.all(y_test_pred == y_test_pred_seq)

assert np.allclose(np.array(coefs_cv).squeeze(), np.array(coefs_seq).squeeze())
5.6.4 CV for model selection: setting the hyper parameters
It is important to note CV may be used for two separate goals:

1. Model assessment: having chosen a final model, estimating its prediction error (gener-
alization error) on new data.
2. Model selection: estimating the performance of different models in order to choose the
best one. One special case of model selection is the selection model’s hyper parameters.
Indeed remember that most of learning algorithm have a hyper parameters (typically the
regularization parameter) that has to be set.
Generally we must address the two problems simultaneously. The usual approach for both
problems is to randomly divide the dataset into three parts: a training set, a validation set, and
a test set.
• The training set (train) is used to fit the models;
• the validation set (val) is used to estimate prediction error for model selection or to
determine the hyper parameters over a grid of possible values.

• the test set (test) is used for assessment of the generalization error of the final chosen
model.
Grid search procedure
Model selection of the best hyper parameters over a grid of possible values
For each possible values of hyper parameters 𝛼𝑘 :
1. Fit the learner on training set: 𝑓 (𝑋𝑡𝑟𝑎𝑖𝑛 , 𝑦𝑡𝑟𝑎𝑖𝑛 , 𝛼𝑘 )
2. Evaluate the model on the validation set and keep the parameter(s) that minimises the
error measure
𝛼* = arg min 𝐿(𝑓 (𝑋𝑡𝑟𝑎𝑖𝑛 ), 𝑦𝑣𝑎𝑙 , 𝛼𝑘 )
3. Refit the learner on all training + validation data using the best hyper parameters: 𝑓 * ≡
𝑓 (𝑋𝑡𝑟𝑎𝑖𝑛∪𝑣𝑎𝑙 , 𝑦𝑡𝑟𝑎𝑖𝑛∪𝑣𝑎𝑙 , 𝛼* )
4. ** Model assessment ** of 𝑓 * on the test set: 𝐿(𝑓 * (𝑋𝑡𝑒𝑠𝑡 ), 𝑦𝑡𝑒𝑠𝑡 )
Nested CV for model selection and assessment
Most of time, we cannot afford such three-way split. Thus, again we will use CV, but in this case
we need two nested CVs.
One outer CV loop, for model assessment. This CV performs 𝐾 splits of the dataset into
training plus validation (𝑋−𝐾 , 𝑦−𝐾 ) set and a test set 𝑋𝐾 , 𝑦𝐾
One inner CV loop, for model selection. For each run of the outer loop, the inner loop loop
performs 𝐿 splits of dataset (𝑋−𝐾 , 𝑦−𝐾 ) into training set: (𝑋−𝐾,−𝐿 , 𝑦−𝐾,−𝐿 ) and a validation
set: (𝑋−𝐾,𝐿 , 𝑦−𝐾,𝐿 ).
Implementation with scikit-learn
Note that the inner CV loop combined with the learner form a new learner with an automatic
model (parameter) selection procedure. This new learner can be easily constructed using Scikit-
learn. The learned is wrapped inside a GridSearchCV class.
Then the new learned can be plugged into the classical outer CV loop.
import numpy as np
from sklearn.model_selection import GridSearchCV
from sklearn.model_selection import KFold
# Dataset
noise_sd = 10
X, y, coef = datasets.make_regression(n_samples=50, n_features=100, noise=noise_sd,
n_informative=2, random_state=42, coef=True)
# Use this to tune the noise parameter such that snr < 5
print("SNR:", np.std(np.dot(X, coef)) / noise_sd)

# param grid over alpha & l1_ratio

param_grid = {'alpha': 10. ** np.arange(-3, 3), 'l1_ratio':[.1, .5, .9]}
# Warp
model = GridSearchCV(lm.ElasticNet(max_iter=10000), param_grid, cv=5)
SNR: 2.6358469446381614
Regression models with built-in cross-validation
Sklearn will automatically select a grid of parameters, most of time use the defaults values.
n_jobs is the number of CPUs to use during the cross validation. If -1, use all the CPUs.
1) Biased usage: fit on all data, ommit outer CV loop
model.fit(X, y)
print("Train r2:%.2f" % metrics.r2_score(y, model.predict(X)))
print(model.best_params_)
Train r2:0.96
{'alpha': 1.0, 'l1_ratio': 0.9}
2) User made outer CV, useful to extract specific information
r2_train, r2_test = list(), list()
alphas = list()

model.fit(X_train, y_train)
r2_test.append(metrics.r2_score(y_test, model.predict(X_test)))
r2_train.append(metrics.r2_score(y_train, model.predict(X_train)))
alphas.append(model.best_params_)
print("Train r2:%.2f" % np.mean(r2_train))

print("Test r2:%.2f" % np.mean(r2_test))
print("Selected alphas:", alphas)
Train r2:1.00
Test r2:0.55
Selected alphas: [{'alpha': 0.001, 'l1_ratio': 0.9}, {'alpha': 0.001, 'l1_ratio': 0.9}, {
˓→'alpha': 0.001, 'l1_ratio': 0.9}, {'alpha': 0.01, 'l1_ratio': 0.9}, {'alpha': 0.001,
˓→'l1_ratio': 0.9}]
3) User-friendly sklearn for outer CV

scores = cross_val_score(estimator=model, X=X, y=y, cv=cv)

Test r2:0.55

# Dataset
X, y, coef = datasets.make_regression(n_samples=50, n_features=100, noise=10,
n_informative=2, random_state=42, coef=True)
print("== Ridge (L2 penalty) ==")

model = lm.RidgeCV(cv=3)
# Let sklearn select a list of alphas with default LOO-CV
scores = cross_val_score(estimator=model, X=X, y=y, cv=5)
print("== Lasso (L1 penalty) ==")

model = lm.LassoCV(n_jobs=-1, cv=3)
# Let sklearn select a list of alphas with default 3CV
print("== ElasticNet (L1 penalty) ==")

model = lm.ElasticNetCV(l1_ratio=[.1, .5, .9], n_jobs=-1, cv=3)
# Let sklearn select a list of alphas with default 3CV
== Ridge (L2 penalty) ==

Test r2:0.16
== Lasso (L1 penalty) ==
Test r2:0.74
== ElasticNet (L1 penalty) ==
Test r2:0.58
Classification models with built-in cross-validation

X, y = datasets.make_classification(n_samples=100, n_features=100,
# provide CV and score

def balanced_acc(estimator, X, y, **kwargs):
'''
Balanced accuracy scorer
'''
return metrics.recall_score(y, estimator.predict(X), average=None).mean()


print("== Logistic Ridge (L2 penalty) ==")
model = lm.LogisticRegressionCV(class_weight='balanced', scoring=balanced_acc, n_jobs=-1,␣
˓→cv=3)
# Let sklearn select a list of alphas with default LOO-CV
print("Test ACC:%.2f" % scores.mean())
== Logistic Ridge (L2 penalty) ==

Test ACC:0.77
5.6.5 Random Permutations
A permutation test is a type of non-parametric randomization test in which the null distribution
of a test statistic is estimated by randomly permuting the observations.
Permutation tests are highly attractive because they make no assumptions other than that the
observations are independent and identically distributed under the null hypothesis.
1. Compute a observed statistic 𝑡𝑜𝑏𝑠 on the data.
2. Use randomization to compute the distribution of 𝑡 under the null hypothesis: Perform 𝑁
random permutation of the data. For each sample of permuted data, 𝑖 the data compute
the statistic 𝑡𝑖 . This procedure provides the distribution of 𝑡 under the null hypothesis 𝐻0 :
𝑃 (𝑡|𝐻0 )
3. Compute the p-value = 𝑃 (𝑡 > 𝑡𝑜𝑏𝑠 |𝐻0 ) |{𝑡𝑖 > 𝑡𝑜𝑏𝑠 }|, where 𝑡𝑖 ’s include 𝑡𝑜𝑏𝑠 .
Example with a correlation
The statistic is the correlation.
import numpy as np
%matplotlib inline
#%matplotlib qt
np.random.seed(42)
x = np.random.normal(loc=10, scale=1, size=100)
y = x + np.random.normal(loc=-3, scale=3, size=100) # snr = 1/2
# Permutation: simulate the null hypothesis

nperm = 10000
perms = np.zeros(nperm + 1)
perms[0] = np.corrcoef(x, y)[0, 1]
for i in range(1, nperm):

perms[i] = np.corrcoef(np.random.permutation(x), y)[0, 1]
# Plot
# Re-weight to obtain distribution


weights = np.ones(perms.shape[0]) / perms.shape[0]
plt.hist([perms[perms >= perms[0]], perms], histtype='stepfilled',
bins=100, label=["t>t obs (p-value)", "t<t obs"],
weights=[weights[perms >= perms[0]], weights])
plt.xlabel("Statistic distribution under null hypothesis")

plt.axvline(x=perms[0], color='blue', linewidth=1, label="observed statistic")
_ = plt.legend(loc="upper left")
# One-tailed empirical p-value

pval_perm = np.sum(perms >= perms[0]) / perms.shape[0]
# Compare with Pearson's correlation test

_, pval_test = stats.pearsonr(x, y)
print("Permutation two tailed p-value=%.5f. Pearson test p-value=%.5f" % (2*pval_perm,␣

˓→pval_test))
Permutation two tailed p-value=0.06959. Pearson test p-value=0.07355
Exercise
Given the logistic regression presented above and its validation given a 5 folds CV.
1. Compute the p-value associated with the prediction accuracy using a permutation test.
2. Compute the p-value associated with the prediction accuracy using a parametric test.

5.6.6 Bootstrapping
Bootstrapping is a random sampling with replacement strategy which provides an non-

parametric method to assess the variability of performances scores such standard errors or
confidence intervals.
A great advantage of bootstrap is its simplicity. It is a straightforward way to derive estimates of
standard errors and confidence intervals for complex estimators of complex parameters of the
distribution, such as percentile points, proportions, odds ratio, and correlation coefficients.
1. Perform 𝐵 sampling, with replacement, of the dataset.
2. For each sample 𝑖 fit the model and compute the scores.
3. Assess standard errors and confidence intervals of scores using the scores obtained on the
𝐵 resampled dataset.
import numpy as np
import pandas as pd
# Regression dataset
n_features = 5
n_features_info = 2
n_samples = 100
X = np.random.randn(n_samples, n_features)
beta = np.zeros(n_features)
beta[:n_features_info] = 1
Xbeta = np.dot(X, beta)
eps = np.random.randn(n_samples)
y = Xbeta + eps
# Fit model on all data (!! risk of overfit)

model = lm.RidgeCV()
model.fit(X, y)
print("Coefficients on all data:")
print(model.coef_)
# Bootstrap loop
nboot = 100 # !! Should be at least 1000
scores_names = ["r2"]
scores_boot = np.zeros((nboot, len(scores_names)))
coefs_boot = np.zeros((nboot, X.shape[1]))
orig_all = np.arange(X.shape[0])
for boot_i in range(nboot):
boot_tr = np.random.choice(orig_all, size=len(orig_all), replace=True)
boot_te = np.setdiff1d(orig_all, boot_tr, assume_unique=False)
Xtr, ytr = X[boot_tr, :], y[boot_tr]
Xte, yte = X[boot_te, :], y[boot_te]
model.fit(Xtr, ytr)
y_pred = model.predict(Xte).ravel()
scores_boot[boot_i, :] = metrics.r2_score(yte, y_pred)
coefs_boot[boot_i, :] = model.coef_


# Compute Mean, SE, CI
scores_boot = pd.DataFrame(scores_boot, columns=scores_names)
scores_stat = scores_boot.describe(percentiles=[.975, .5, .025])
print("r-squared: Mean=%.2f, SE=%.2f, CI=(%.2f %.2f)" %\

tuple(scores_stat.loc[["mean", "std", "5%", "95%"], "r2"]))
coefs_boot = pd.DataFrame(coefs_boot)
coefs_stat = coefs_boot.describe(percentiles=[.975, .5, .025])
print("Coefficients distribution")
print(coefs_stat)
Coefficients on all data:

[ 1.0257263 1.11323 -0.0499828 -0.09263008 0.15267576]
r-squared: Mean=0.61, SE=0.10, CI=(nan nan)
Coefficients distribution
0 1 2 3 4
count 100.000000 100.000000 100.000000 100.000000 100.000000
mean 1.012534 1.132775 -0.056369 -0.100046 0.164236
std 0.094269 0.104934 0.111308 0.095098 0.095656
min 0.759189 0.836394 -0.290386 -0.318755 -0.092498
2.5% 0.814260 0.948158 -0.228483 -0.268790 -0.044067
50% 1.013097 1.125304 -0.057039 -0.099281 0.164194
97.5% 1.170183 1.320637 0.158680 0.085064 0.331809
max 1.237874 1.340585 0.291111 0.151059 0.450812
5.7 Ensemble learning: bagging, boosting and stacking
These methods are Ensemble learning techniques. These models are machine learning
paradigms where multiple models (often called “weak learners”) are trained to solve the same
problem and combined to get better results. The main hypothesis is that when weak models
are correctly combined we can obtain more accurate and/or robust models.
5.7.1 Single weak learner
In machine learning, no matter if we are facing a classification or a regression problem, the

choice of the model is extremely important to have any chance to obtain good results. This
choice can depend on many variables of the problem: quantity of data, dimensionality of the
space, distribution hypothesis. . .
A low bias and a low variance, although they most often vary in opposite directions, are the
two most fundamental features expected for a model. Indeed, to be able to “solve” a problem,
we want our model to have enough degrees of freedom to resolve the underlying complexity
of the data we are working with, but we also want it to have not too much degrees of freedom
to avoid high variance and be more robust. This is the well known bias-variance tradeoff.
Illustration of the bias-variance tradeoff.
In ensemble learning theory, we call weak learners (or base models) models that can be used
as building blocks for designing more complex models by combining several of them. Most
of the time, these basics models perform not so well by themselves either because they have

Fig. 13: towardsdatascience blog
a high bias (low degree of freedom models, for example) or because they have too much
variance to be robust (high degree of freedom models, for example). Then, the idea of ensem-
ble methods is to combining several of them together in order to create a strong learner (or
ensemble model) that achieves better performances.
Usually, ensemble models are used in order to :
• decrease the variance for bagging (Bootstrap Aggregating) technique
• reduce bias for the boosting technique
• improving the predictive force for stacking technique.
To understand these techniques, first, we will explore what is boostrapping and its different
hypothesis.
5.7.2 Bootstrapping
Bootstrapping is a statistical technique which consists in generating samples of size B (called

bootstrap samples) from an initial dataset of size N by randomly drawing with replacement
B observations.
Illustration of the bootstrapping process.
Under some assumptions, these samples have pretty good statistical properties: in first ap-
proximation, they can be seen as being drawn both directly from the true underlying (and often
unknown) data distribution and independently from each others. So, they can be considered
as representative and independent samples of the true data distribution. The hypothesis
that have to be verified to make this approximation valid are twofold: - First, the size N of the
initial dataset should be large enough to capture most of the complexity of the underlying
5.7. Ensemble learning: bagging, boosting and stacking 233

distribution so that sampling from the dataset is a good approximation of sampling from the
real distribution (representativity).
• Second, the size N of the dataset should be large enough compared to the size B of the
bootstrap samples so that samples are not too much correlated (independence).
Bootstrap samples are often used, for example, to evaluate variance or confidence intervals
of a statistical estimators. By definition, a statistical estimator is a function of some observa-
tions and, so, a random variable with variance coming from these observations. In order to
estimate the variance of such an estimator, we need to evaluate it on several independent sam-
ples drawn from the distribution of interest. In most of the cases, considering truly independent
samples would require too much data compared to the amount really available. We can then use
bootstrapping to generate several bootstrap samples that can be considered as being “almost-
representative” and “almost-independent” (almost i.i.d. samples). These bootstrap samples will
allow us to approximate the variance of the estimator, by evaluating its value for each of them.
Bootstrapping is often used to evaluate variance or confidence interval of some statistical esti-

mators.
5.7.3 Bagging
In parallel methods we fit the different considered learners independently from each others
and, so, it is possible to train them concurrently. The most famous such approach is “bagging”
(standing for “bootstrap aggregating”) that aims at producing an ensemble model that is more
robust than the individual models composing it.
When training a model, no matter if we are dealing with a classification or a regression problem,
we obtain a function that takes an input, returns an output and that is defined with respect to
the training dataset.
The idea of bagging is then simple: we want to fit several independent models and “average”
their predictions in order to obtain a model with a lower variance. However, we can’t, in
practice, fit fully independent models because it would require too much data. So, we rely on
the good “approximate properties” of bootstrap samples (representativity and independence)
to fit models that are almost independent.
First, we create multiple bootstrap samples so that each new bootstrap sample will act as
another (almost) independent dataset drawn from true distribution. Then, we can fit a weak
learner for each of these samples and finally aggregate them such that we kind of “aver-
age” their outputs and, so, obtain an ensemble model with less variance that its components.
Roughly speaking, as the bootstrap samples are approximatively independent and identically
distributed (i.i.d.), so are the learned base models. Then, “averaging” weak learners outputs
do not change the expected answer but reduce its variance.
So, assuming that we have L bootstrap samples (approximations of L independent datasets) of
size B denoted
Fig. 16: Medium Science Blog
Each {. . . .} is a bootstrap sample of B observation

we can fit L almost independent weak learners (one on each dataset)
and then aggregate them into some kind of averaging process in order to get an ensemble model
with a lower variance. For example, we can define our strong model such that

There are several possible ways to aggregate the multiple models fitted in parallel. - For a
regression problem, the outputs of individual models can literally be averaged to obtain the
output of the ensemble model. - For classification problem the class outputted by each model
can be seen as a vote and the class that receives the majority of the votes is returned by
the ensemble model (this is called hard-voting). Still for a classification problem, we can
also consider the probabilities of each classes returned by all the models, average these
probabilities and keep the class with the highest average probability (this is called soft-
voting). –> Averages or votes can either be simple or weighted if any relevant weights can be
used.
Finally, we can mention that one of the big advantages of bagging is that it can be parallelised.
As the different models are fitted independently from each others, intensive parallelisation tech-
niques can be used if required.
Bagging consists in fitting several base models on different bootstrap samples and build an
ensemble model that “average” the results of these weak learners.
Question : - Can you name an algorithms based on Bagging technique , Hint : leaf
###### Examples
Here, we are trying some example of stacking
• Bagged Decision Trees for Classification
import pandas
from sklearn import model_selection
from sklearn.ensemble import BaggingClassifier
from sklearn.tree import DecisionTreeClassifier
names = ['preg', 'plas', 'pres', 'skin', 'test', 'mass', 'pedi', 'age', 'class']
dataframe = pandas.read_csv("https://raw.githubusercontent.com/jbrownlee/Datasets/master/
˓→pima-indians-diabetes.data.csv",names=names)
array = dataframe.values
x = array[:,0:8]
y = array[:,8]


max_features = 3
kfold = model_selection.KFold(n_splits=10, random_state=2020)

rf = DecisionTreeClassifier(max_features=max_features)
num_trees = 100
model = BaggingClassifier(base_estimator=rf, n_estimators=num_trees, random_state=2020)

results = model_selection.cross_val_score(model, x, y, cv=kfold)
print("Accuracy: %0.2f (+/- %0.2f)" % (results.mean(), results.std()))
• Random Forest Classification
import pandas
from sklearn import model_selection
names = ['preg', 'plas', 'pres', 'skin', 'test', 'mass', 'pedi', 'age', 'class']
dataframe = pandas.read_csv("https://raw.githubusercontent.com/jbrownlee/Datasets/master/
˓→pima-indians-diabetes.data.csv",names=names)
array = dataframe.values
x = array[:,0:8]
y = array[:,8]

rf = DecisionTreeClassifier()
num_trees = 100
max_features = 3

model = RandomForestClassifier(n_estimators=num_trees, max_features=max_features)
results = model_selection.cross_val_score(model, x, y, cv=kfold)
print("Accuracy: %0.2f (+/- %0.2f)" % (results.mean(), results.std()))
Both of these algorithms will print, Accuracy: 0.77 (+/- 0.07). They are equivalent.
5.7.4 Boosting
In sequential methods the different combined weak models are no longer fitted indepen-
dently from each others. The idea is to fit models iteratively such that the training of model at
a given step depends on the models fitted at the previous steps. “Boosting” is the most famous
of these approaches and it produces an ensemble model that is in general less biased than the
weak learners that compose it.
Boosting methods work in the same spirit as bagging methods: we build a family of models
that are aggregated to obtain a strong learner that performs better.
However, unlike bagging that mainly aims at reducing variance, boosting is a technique
that consists in fitting sequentially multiple weak learners in a very adaptative way: each
model in the sequence is fitted giving more importance to observations in the dataset that
were badly handled by the previous models in the sequence. Intuitively, each new model
focus its efforts on the most difficult observations to fit up to now, so that we obtain, at the
end of the process, a strong learner with lower bias (even if we can notice that boosting can
also have the effect of reducing variance).

–> Boosting, like bagging, can be used for regression as well as for classification problems.
Being mainly focused at reducing bias, the base models that are often considered for boosting
are* *models with low variance but high bias. For example, if we want to usetreesas our base
models, we will choosemost of the time shallow decision trees with only a few depths.**
Another important reason that motivates the use of low variance but high bias models as weak
learners for boosting is that these models are in general less computationally expensive to fit
(few degrees of freedom when parametrised). Indeed, as computations to fit the different mod-
els can’t be done in parallel (unlike bagging), it could become too expensive to fit sequentially
several complex models.
Once the weak learners have been chosen, we still need to define how they will be sequentially
fitted and how they will be aggregated. We will discuss these questions in the two follow-
ing subsections, describing more especially two important boosting algorithms: adaboost and
gradient boosting.
In a nutshell, these two meta-algorithms differ on how they create and aggregate the weak
learners during the sequential process. Adaptive boosting updates the weights attached to
each of the training dataset observations whereas gradient boosting updates the value of
these observations. This main difference comes from the way both methods try to solve the
optimisation problem of finding the best model that can be written as a weighted sum of weak
learners.
Boosting consists in, iteratively, fitting a weak learner, aggregate it to the ensemble model and
“update” the training dataset to better take into account the strengths and weakness of the
current ensemble model when fitting the next base model.

1/ Adaptative boosting
In adaptative boosting (often called “adaboost”), we try to define our ensemble model as a
weighted sum of L weak learners
Finding the best ensemble model with this form is a difficult optimisation problem. Then,
instead of trying to solve it in one single shot (finding all the coefficients and weak learners
that give the best overall additive model), we make use of an iterative optimisation process
that is much more tractable, even if it can lead to a sub-optimal solution. More especially,
we add the weak learners one by one, looking at each iteration for the best possible pair
(coefficient, weak learner) to add to the current ensemble model. In other words, we define
recurrently the (s_l)’s such that
Fig. 22: towardsdatascience Blog
where c_l and w_l are chosen such that s_l is the model that fit the best the training data and,
so, that is the best possible improvement over s_(l-1). We can then denote
Fig. 23: towardsdatascience Blog
where E(.) is the fitting error of the given model and e(.,.) is the loss/error function. Thus,
instead of optimising “globally” over all the L models in the sum, we approximate the optimum
by optimising “locally” building and adding the weak learners to the strong model one by one.
More especially, when considering a binary classification, we can show that the adaboost algo-
rithm can be re-written into a process that proceeds as follow. First, it updates the observations
weights in the dataset and train a new weak learner with a special focus given to the obser-
vations misclassified by the current ensemble model. Second, it adds the weak learner to the
weighted sum according to an update coefficient that expresse the performances of this weak
model: the better a weak learner performs, the more it contributes to the strong learner.
So, assume that we are facing a binary classification problem, with N observations in our
dataset and we want to use adaboost algorithm with a given family of weak models. At the
very beginning of the algorithm (first model of the sequence), all the observations have
the same weights 1/N. Then, we repeat L times (for the L learners in the sequence) the
following steps:
fit the best possible weak model with the current observations weights
compute the value of the update coefficient that is some kind of scalar evaluation metric of the
weak learner that indicates how much this weak learner should be taken into account into the

ensemble model
update the strong learner by adding the new weak learner multiplied by its update coefficient
compute new observations weights that expresse which observations we would like to focus
on at the next iteration (weights of observations wrongly predicted by the aggregated model
increase and weights of the correctly predicted observations decrease)
Repeating these steps, we have then build sequentially our L models and aggregate them into
a simple linear combination weighted by coefficients expressing the performance of each
learner.
Notice that there exists variants of the initial adaboost algorithm such that LogitBoost (classifi-
cation) or L2Boost (regression) that mainly differ by their choice of loss function.
Adaboost updates weights of the observations at each iteration. Weights of well classified obser-
vations decrease relatively to weights of misclassified observations. Models that perform better
have higher weights in the final ensemble model.
2/ Gradient boosting
In gradient boosting, the ensemble model we try to build is also a weighted sum of weak
learners
Just as we mentioned for adaboost, finding the optimal model under this form is too difficult
and an iterative approach is required. The main difference with adaptative boosting is in
the definition of the sequential optimisation process. Indeed, gradient boosting casts the

problem into a gradient descent one: at each iteration we fit a weak learner to the opposite
of the gradient of the current fitting error with respect to the current ensemble model.
Let’s try to clarify this last point. First, theoretical gradient descent process over the ensemble
model can be written
where E(.) is the fitting error of the given model, c_l is a coefficient corresponding to the step
size and
This entity is the opposite of the gradient of the fitting error with respect to the ensemble
model at step l-1. This opposite of the gradient is a function that can, in practice, only be
evaluated for observations in the training dataset (for which we know inputs and outputs):
these evaluations are called pseudo-residuals attached to each observations. Moreover, even if
we know for the observations the values of these pseudo-residuals, we don’t want to add to
our ensemble model any kind of function: we only want to add a new instance of weak model.
So, the natural thing to do is to fit a weak learner to the pseudo-residuals computed for each
observation. Finally, the coefficient c_l is computed following a one dimensional optimisation
process (line-search to obtain the best step size c_l).
So, assume that we want to use gradient boosting technique with a given family of weak models.
At the very beginning of the algorithm (first model of the sequence), the pseudo-residuals are
set equal to the observation values. Then, we repeat L times (for the L models of the sequence)
the following steps:
fit the best possible weak model to pseudo-residuals (approximate the opposite of the gradient
with respect to the current strong learner)
compute the value of the optimal step size that defines by how much we update the ensemble
model in the direction of the new weak learner
update the ensemble model by adding the new weak learner multiplied by the step size (make
a step of gradient descent)
compute new pseudo-residuals that indicate, for each observation, in which direction we would
like to update next the ensemble model predictions
Repeating these steps, we have then build sequentially our L models and aggregate them fol-
lowing a gradient descent approach. Notice that, while adaptative boosting tries to solve at
each iteration exactly the “local” optimisation problem (find the best weak learner and its
coefficient to add to the strong model), gradient boosting uses instead a gradient descent
approach and can more easily be adapted to large number of loss functions. Thus, gradi-
ent boosting can be considered as a generalization of adaboost to arbitrary differentiable
loss functions.
Note There is an algorithm which gained huge popularity after a Kaggle’s competitions. It is
XGBoost (Extreme Gradient Boosting). This is a gradient boosting algorithm which has more

flexibility (varying number of terminal nodes and left weights) parameters to avoid sub-
learners correlations. Having these important qualities, XGBOOST is one of the most used
algorithm in data science. LIGHTGBM is a recent implementation of this algorithm. It was
published by Microsoft and it gives us the same scores (if parameters are equivalents) but it
runs quicker than a classic XGBOOST.
Gradient boosting updates values of the observations at each iteration. Weak learners are
trained to fit the pseudo-residuals that indicate in which direction to correct the current en-
semble model predictions to lower the error.
Examples
Here, we are trying an example of Boosting and compare it to a Bagging. Both of algorithms
take the same weak learners to build the macro-model
• Adaboost Classifier
from sklearn.ensemble import AdaBoostClassifier

from sklearn.datasets import load_breast_cancer
import pandas as pd
import numpy as np
from sklearn.model_selection import train_test_split
from sklearn.metrics import confusion_matrix
from sklearn.preprocessing import LabelEncoder
from sklearn.metrics import accuracy_score


from sklearn.metrics import f1_score
breast_cancer = load_breast_cancer()
x = pd.DataFrame(breast_cancer.data, columns=breast_cancer.feature_names)
y = pd.Categorical.from_codes(breast_cancer.target, breast_cancer.target_names)
# Transforming string Target to an int
encoder = LabelEncoder()
binary_encoded_y = pd.Series(encoder.fit_transform(y))
#Train Test Split

train_x, test_x, train_y, test_y = train_test_split(x, binary_encoded_y, random_state=1)
clf_boosting = AdaBoostClassifier(
DecisionTreeClassifier(max_depth=1),
n_estimators=200
)
clf_boosting.fit(train_x, train_y)
predictions = clf_boosting.predict(test_x)
print("For Boosting : F1 Score {}, Accuracy {}".format(round(f1_score(test_y,predictions),
˓→2),round(accuracy_score(test_y,predictions),2)))
• Random Forest as a bagging classifier

import pandas as pd
import numpy as np
#Train Test Split

train_x, test_x, train_y, test_y = train_test_split(x, binary_encoded_y, random_state=1)
clf_bagging = RandomForestClassifier(n_estimators=200, max_depth=1)
clf_bagging.fit(train_x, train_y)
predictions = clf_bagging.predict(test_x)
print("For Bagging : F1 Score {}, Accuracy {}".format(round(f1_score(test_y,predictions),
˓→2),round(accuracy_score(test_y,predictions),2)))
Comparaison
Metric Bagging Boosting

Accuracy 0.91 0.97
F1-Score 0.88 0.95

5.7.5 Overview of stacking
Stacking mainly differ from bagging and boosting on two points : - First stacking often con-
siders heterogeneous weak learners (different learning algorithms are combined) whereas
bagging and boosting consider mainly homogeneous weak learners. - Second, stacking learns
to combine the base models using a meta-model whereas bagging and boosting combine weak
learners following deterministic algorithms.
As we already mentioned, the idea of stacking is to learn several different weak learners and
combine them by training a meta-model to output predictions based on the multiple predic-
tions returned by these weak models. So, we need to define two things in order to build our
stacking model: the L learners we want to fit and the meta-model that combines them.
For example, for a classification problem, we can choose as weak learners a KNN classifier, a
logistic regression and a SVM, and decide to learn a neural network as meta-model. Then, the
neural network will take as inputs the outputs of our three weak learners and will learn to
return final predictions based on it.
So, assume that we want to fit a stacking ensemble composed of L weak learners. Then we have
to follow the steps thereafter:
• split the training data in two folds
• choose L weak learners and fit them to data of the first fold
• for each of the L weak learners, make predictions for observations in the second fold
• fit the meta-model on the second fold, using predictions made by the weak learners
as inputs
In the previous steps, we split the dataset in two folds because predictions on data that have
been used for the training of the weak learners are not relevant for the training of the meta-
model.

Stacking consists in training a meta-model to produce outputs based on the outputs returned
by some lower layer weak learners.
A possible extension of stacking is multi-level stacking. It consists in doing stacking with
multiple layers. As an example,
Multi-level stacking considers several layers of stacking: some meta-models are trained on out-
puts returned by lower layer meta-models and so on. Here we have represented a 3-layers
stacking model.
Examples
Here, we are trying an example of Stacking and compare it to a Bagging & a Boosting. We note
that, many other applications (datasets) would show more difference between these techniques.
import pandas as pd
import numpy as np
from sklearn.linear_model import LogisticRegression

#Train Test Split

train_x, test_x, train_y, test_y = train_test_split(x, binary_encoded_y, random_
˓→state=2020) (continues on next page)

boosting_clf_ada_boost= AdaBoostClassifier(
DecisionTreeClassifier(max_depth=1),
n_estimators=3
)
bagging_clf_rf = RandomForestClassifier(n_estimators=200, max_depth=1,random_state=2020)
clf_rf = RandomForestClassifier(n_estimators=200, max_depth=1,random_state=2020)

clf_ada_boost = AdaBoostClassifier(
DecisionTreeClassifier(max_depth=1,random_state=2020),
n_estimators=3
)
clf_logistic_reg = LogisticRegression(solver='liblinear',random_state=2020)
#Customizing and Exception message

class NumberOfClassifierException(Exception):
pass
#Creating a stacking class

class Stacking():
'''
This is a test class for stacking !
Please fill Free to change it to fit your needs
We suppose that at least the First N-1 Classifiers have
a predict_proba function.
'''
def __init__(self,classifiers):
if(len(classifiers) < 2):
raise numberOfClassifierException("You must fit your classifier with 2␣
˓→classifiers at least");
else:
self._classifiers = classifiers
def fit(self,data_x,data_y):
stacked_data_x = data_x.copy()
for classfier in self._classifiers[:-1]:
classfier.fit(data_x,data_y)
stacked_data_x = np.column_stack((stacked_data_x,classfier.predict_proba(data_
˓→ x)))
last_classifier = self._classifiers[-1]
last_classifier.fit(stacked_data_x,data_y)
def predict(self,data_x):
stacked_data_x = data_x.copy()
for classfier in self._classifiers[:-1]:
prob_predictions = classfier.predict_proba(data_x)


stacked_data_x = np.column_stack((stacked_data_x,prob_predictions))
last_classifier = self._classifiers[-1]
return last_classifier.predict(stacked_data_x)
bagging_clf_rf.fit(train_x, train_y)
boosting_clf_ada_boost.fit(train_x, train_y)
classifers_list = [clf_rf,clf_ada_boost,clf_logistic_reg]
clf_stacking = Stacking(classifers_list)
clf_stacking.fit(train_x,train_y)
predictions_bagging = bagging_clf_rf.predict(test_x)
predictions_boosting = boosting_clf_ada_boost.predict(test_x)
predictions_stacking = clf_stacking.predict(test_x)
print("For Bagging : F1 Score {}, Accuracy {}".format(round(f1_score(test_y,predictions_

˓→bagging),2),round(accuracy_score(test_y,predictions_bagging),2)))
print("For Boosting : F1 Score {}, Accuracy {}".format(round(f1_score(test_y,predictions_

˓→boosting),2),round(accuracy_score(test_y,predictions_boosting),2)))
print("For Stacking : F1 Score {}, Accuracy {}".format(round(f1_score(test_y,predictions_
˓→stacking),2),round(accuracy_score(test_y,predictions_stacking),2)))
Comparaison
Metric Bagging Boosting Stacking

Accuracy 0.90 0.94 0.98
F1-Score 0.88 0.93 0.98
5.8 Gradient descent
Gradient descent is an optimization algorithm used to minimize some function by iteratively

moving in the direction of steepest descent as defined by the negative of the gradient. In
machine learning, we use gradient descent to update the parameters of our model. Parame-
ters refer to coefficients in Linear Regression and weights in neural networks.
This section aims to provide you an explanation of gradient descent and intuitions towards the
behaviour of different algorithms for optimizing it. These explanations will help you put
them to use.
We are first going to introduce the gradient descent, solve it for a regression problem and look at
its different variants. Then, we will then briefly summarize challenges during training. Finally,
we will introduce the most common optimization algorithms by showing their motivation to
resolve these challenges and list some advices for facilitate the algorithm choice.
5.8. Gradient descent 247

5.8.1 Introduction
Consider the 3-dimensional graph below in the context of a cost function. Our goal is to move
from the mountain in the top right corner (high cost) to the dark blue sea in the bottom
left (low cost). The arrows represent the direction of steepest descent (negative gradient)
from any given point–the direction that decreases the cost function as quickly as possible
Fig. 31: adalta.it
Gradient descent intuition.

Starting at the top of the mountain, we take our first step downhill in the direction specified by
the negative gradient. Next we recalculate the negative gradient (passing in the coordinates
of our new point) and take another step in the direction it specifies. We continue this process
iteratively until we get to the bottom of our graph, or to a point where we can no longer
move downhill–a local minimum.
Learning rate
The size of these steps is called the learning rate. With a high learning rate we can cover
more ground each step, but we risk overshooting the lowest point since the slope of the hill is
constantly changing. With a very low learning rate, we can confidently move in the direction
of the negative gradient since we are recalculating it so frequently. A low learning rate is
more precise, but calculating the gradient is time-consuming, so it will take us a very long
time to get to the bottom.
impacts of learning rate choice.

Fig. 32: jeremyjordan
Cost function
A Loss Function (Error function) tells us “how good” our model is at making predictions for
a given set of parameters. The cost function has its own curve and its own gradients. The
slope of this curve tells us how to update our parameters to make the model more accurate.
5.8.2 Numerical solution for gradient descent
Let’s run gradient descent using a linear regression cost function.

There are two parameters in our cost function we can control: - $ .. raw:: latex
beta`_0$ : (the bias) - $:raw-latex:beta_1 $ : (weight or coefficient)
Since we need to consider the impact each one has on the final prediction, we
need to use partial derivatives. We calculate the partial derivatives of the cost
function with respect to each parameter and store the results in a gradient.
Given the cost function
𝑛 𝑛
1 𝜕𝑀 𝑆𝐸 1 ∑︁ 1 ∑︁
𝑓 (𝛽0 , 𝛽1 ) = = (𝑦𝑖 − (𝛽1 𝑥𝑖 + 𝛽0 ))2 = ((𝛽1 𝑥𝑖 + 𝛽0 ) − 𝑦𝑖 )2
2 𝜕𝛽 2𝑁 2𝑁
𝑖=1 𝑖=1
The gradient can be calculated as

[︃ ]︃ [︂
𝜕𝑓 1 ∑︀
]︂ [︂ −1 ∑︀ ]︂
′ 𝜕𝛽0 2𝑁 −2((𝛽1 𝑥𝑖 + 𝛽0 ) − 𝑦𝑖 ) 𝑁 ((𝛽1 𝑥𝑖 + 𝛽0 ) − 𝑦𝑖 )
𝑓 (𝛽0 , 𝛽1 ) = 𝜕𝑓 = 1 ∑︀ = −1 ∑︀
𝜕𝛽 2𝑁 −2𝑥𝑖 ((𝛽1 𝑥𝑖 + 𝛽0 ) − 𝑦𝑖 ) 𝑁 𝑥𝑖 ((𝛽1 𝑥𝑖 + 𝛽0 ) − 𝑦𝑖 )
1
To solve for the gradient, we iterate through our data points using our
:math:beta_1 and :math:beta_0 values and compute the
partial derivatives. This new gradient tells us the slope of our cost function at our cur-
rent position (current parameter values) and the direction we should move to update our
parameters. The size of our update is controlled by the learning rate.
Pseudocode of this algorithm

Function gradient_descent(X, Y, learning_rate, number_iterations):
m : 1
b : 1
m_deriv : 0
b_deriv : 0
data_length : length(X)
loop i : 1 --> number_iterations:
loop i : 1 -> data_length :
m_deriv : m_deriv -X[i] * ((m*X[i] + b) - Y[i])
b_deriv : b_deriv - ((m*X[i] + b) - Y[i])
m : m - (m_deriv / data_length) * learning_rate
b : b - (b_deriv / data_length) * learning_rate
return m, b
5.8.3 Gradient descent variants
There are three variants of gradient descent, which differ in how much data we use to
compute the gradient of the objective function. Depending on the amount of data, we make
a trade-off between the accuracy of the parameter update and the time it takes to perform
an update.
Batch gradient descent
Batch gradient descent, known also as Vanilla gradient descent, computes the gradient of the
cost function with respect to the parameters 𝜃 for the entire training dataset :
𝜃 = 𝜃 − 𝜂 · ∇𝜃 𝐽(𝜃)
As we need to calculate the gradients for the whole dataset to perform just one update, batch
gradient descent can be very slow and is intractable for datasets that don’t fit in memory.
Batch gradient descent also doesn’t allow us to update our model online.
Stochastic gradient descent
Stochastic gradient descent (SGD) in contrast performs a parameter update for each training
example 𝑥(𝑖) and label 𝑦 (𝑖)
• Choose an initial vector of parameters 𝑤 and learning rate 𝜂.
• Repeat until an approximate minimum is obtained:
– Randomly shuffle examples in the training set.
– For 𝑖 ∈ 1, . . . , 𝑛
(𝑖) (𝑖)
* 𝜃 = 𝜃 − 𝜂 · ∇𝜃 𝐽(𝜃; 𝑥 ; 𝑦 )
Batch gradient descent performs redundant computations for large datasets, as it recom-
putes gradients for similar examples before each parameter update. SGD does away with
this redundancy by performing one update at a time. It is therefore usually much faster
and can also be used to learn online. SGD performs frequent updates with a high variance
that cause the objective function to fluctuate heavily as in the image below.

Fig. 33: Wikipedia

SGD fluctuation.
While batch gradient descent converges to the minimum of the basin the parameters are
placed in, SGD’s fluctuation, on the one hand, enables it to jump to new and potentially
better local minima. On the other hand, this ultimately complicates convergence to the
exact minimum, as SGD will keep overshooting. However, it has been shown that when we
slowly decrease the learning rate, SGD shows the same convergence behaviour as batch
gradient descent, almost certainly converging to a local or the global minimum for non-
convex and convex optimization respectively.
Mini-batch gradient descent
Mini-batch gradient descent finally takes the best of both worlds and performs an update for
every mini-batch of n training examples:
𝜃 = 𝜃 − 𝜂 · ∇𝜃 𝐽(𝜃; 𝑥(𝑖:𝑖+𝑛) ; 𝑦 (𝑖:𝑖+𝑛) )
This way, it :
• reduces the variance of the parameter updates, which can lead to more stable con-
vergence.
• can make use of highly optimized matrix optimizations common to state-of-the-art deep
learning libraries that make computing the gradient very efficient. Common mini-batch
sizes range between 50 and 256, but can vary for different applications.
Mini-batch gradient descent is typically the algorithm of choice when training a neural
network.
5.8.4 Gradient Descent challenges
Vanilla mini-batch gradient descent, however, does not guarantee good convergence, but offers
a few challenges that need to be addressed:
• Choosing a proper learning rate can be difficult. A learning rate that is too small leads
to painfully slow convergence, while a learning rate that is too large can hinder conver-
gence and cause the loss function to fluctuate around the minimum or even to diverge.
• Learning rate schedules try to adjust the learning rate during training by e.g. an-
nealing, i.e. reducing the learning rate according to a pre-defined schedule or when the
change in objective between epochs falls below a threshold. These schedules and thresh-
olds, however, have to be defined in advance and are thus unable to adapt to a dataset’s
characteristics.
• Additionally, the same learning rate applies to all parameter updates. If our data is sparse
and our features have very different frequencies, we might not want to update all of
them to the same extent, but perform a larger update for rarely occurring features.
• Another key challenge of minimizing highly non-convex error functions common for
neural networks is avoiding getting trapped in their numerous suboptimal local min-
ima. These saddle points (local minimas) are usually surrounded by a plateau of the
same error, which makes it notoriously hard for SGD to escape, as the gradient is close
to zero in all dimensions.

5.8.5 Gradient descent optimization algorithms
In the following, we will outline some algorithms that are widely used by the deep learning
community to deal with the aforementioned challenges.
Momentum
SGD has trouble navigating ravines (areas where the surface curves much more steeply in
one dimension than in another), which are common around local optima. In these scenarios,
SGD oscillates across the slopes of the ravine while only making hesitant progress along
the bottom towards the local optimum as in the image below.
Fig. 34: Wikipedia
SGD and momentum.

Source
Fig. 35: No momentum: oscillations toward local largest gradient
No momentum: moving toward local largest gradient create oscillations.

With momentum: accumulate velocity to avoid oscillations.
Momentum is a method that helps accelerate SGD in the relevant direction and dampens
oscillations as can be seen in image above. It does this by adding a fraction :math:`gamma`
of the update vector of the past time step to the current update vector
𝑣𝑡 = 𝜌𝑣𝑡−1 + ∇𝜃 𝐽(𝜃)
(5.52)
𝜃 = 𝜃 − 𝑣𝑡
vx = 0
while True:
dx = gradient(J, x)

Fig. 36: With momentum: accumulate velocity to avoid oscillations

vx = rho * vx + dx
x -= learning_rate * vx
Note: The momentum term :math:`rho` is usually set to 0.9 or a similar value.
Essentially, when using momentum, we push a ball down a hill. The ball accumulates mo-
mentum as it rolls downhill, becoming faster and faster on the way (until it reaches its
terminal velocity if there is air resistance, i.e. :math:`rho` <1 ).
The same thing happens to our parameter updates: The momentum term increases for dimen-
sions whose gradients point in the same directions and reduces updates for dimensions
whose gradients change directions. As a result, we gain faster convergence and reduced
oscillation.
AdaGrad: adaptive learning rates
• Added element-wise scaling of the gradient based on the historical sum of squares in each
dimension.
• “Per-parameter learning rates” or “adaptive learning rates”
grad_squared = 0
while True:
dx = gradient(J, x)
grad_squared += dx * dx
x -= learning_rate * dx / (np.sqrt(grad_squared) + 1e-7)
• Progress along “steep” directions is damped.

• Progress along “flat” directions is accelerated.
• Problem: step size over long time => Decays to zero.

RMSProp: “Leaky AdaGrad”
grad_squared = 0
while True:
dx = gradient(J, x)
grad_squared += decay_rate * grad_squared + (1 - decay_rate) * dx * dx
x -= learning_rate * dx / (np.sqrt(grad_squared) + 1e-7)
• decay_rate = 1: gradient descent

• decay_rate = 0: AdaGrad
Nesterov accelerated gradient
However, a ball that rolls down a hill, blindly following the slope, is highly unsatisfactory. We’d
like to have a smarter ball, a ball that has a notion of where it is going so that it knows to
slow down before the hill slopes up again. Nesterov accelerated gradient (NAG) is a way to
give our momentum term this kind of prescience. We know that we will use our momentum
term 𝛾𝑣𝑡−1 to move the parameters 𝜃.
Computing 𝜃 − 𝛾𝑣𝑡−1 thus gives us an approximation of the next position of the parameters
(the gradient is missing for the full update), a rough idea where our parameters are going to
be. We can now effectively look ahead by calculating the gradient not w.r.t. to our current
parameters 𝜃 but w.r.t. the approximate future position of our parameters:
𝑣𝑡 = 𝛾𝑣𝑡−1 + 𝜂∇𝜃 𝐽(𝜃 − 𝛾𝑣𝑡−1 )

(5.53)
𝜃 = 𝜃 − 𝑣𝑡
Again, we set the momentum term 𝛾 to a value of around 0.9. While Momentum first com-
putes the current gradient and then takes a big jump in the direction of the updated
accumulated gradient , NAG first makes a big jump in the direction of the previous ac-
cumulated gradient, measures the gradient and then makes a correction, which results
in the complete NAG update. This anticipatory update prevents us from going too fast and
results in increased responsiveness, which has significantly increased the performance of
RNNs on a number of tasks
Adam
Adaptive Moment Estimation (Adam) is a method that computes adaptive learning rates for
each parameter. In addition to storing an exponentially decaying average of past squared
gradients :math:`v_t`, Adam also keeps an exponentially decaying average of past gradi-
ents :math:`m_t`, similar to momentum. Whereas momentum can be seen as a ball running
down a slope, Adam behaves like a heavy ball with friction, which thus prefers flat minima
in the error surface. We compute the decaying averages of past and past squared gradients 𝑚𝑡
and 𝑣𝑡 respectively as follows:
𝑚𝑡 = 𝛽1 𝑚𝑡−1 + (1 − 𝛽1 )∇𝜃 𝐽(𝜃)

(5.54)
𝑣𝑡 = 𝛽2 𝑣𝑡−1 + (1 − 𝛽2 )∇𝜃 𝐽(𝜃)2

𝑚𝑡 and 𝑣𝑡 are estimates of the first moment (the mean) and the second moment (the uncentered
variance) of the gradients respectively, hence the name of the method. Adam (almost)
first_moment = 0
second_moment = 0
while True:
dx = gradient(J, x)
# Momentum:
first_moment = beta1 * first_moment + (1 - beta1) * dx
# AdaGrad/RMSProp
second_moment = beta2 * second_moment + (1 - beta2) * dx * dx
x -= learning_rate * first_moment / (np.sqrt(second_moment) + 1e-7)
As 𝑚𝑡 and 𝑣𝑡 are initialized as vectors of 0’s, the authors of Adam observe that they are biased
towards zero, especially during the initial time steps, and especially when the decay rates are
small (i.e. 𝛽1 and 𝛽2 are close to 1). They counteract these biases by computing bias-corrected
first and second moment estimates:
𝑚𝑡
𝑚
ˆ𝑡 = (5.55)
1 − 𝛽1𝑡
𝑣𝑡
𝑣ˆ𝑡 = (5.56)
1 − 𝛽2𝑡
They then use these to update the parameters (Adam update rule):
𝜂
𝜃𝑡+1 = 𝜃𝑡 − √ 𝑚
ˆ𝑡
𝑣ˆ𝑡 + 𝜖
• 𝑚
ˆ 𝑡 Accumulate gradient: velocity.
• 𝑣ˆ𝑡 Element-wise scaling of the gradient based on the historical sum of squares in each
dimension.
• Choose Adam as default optimizer
• Default values of 0.9 for 𝛽1 , 0.999 for 𝛽2 , and 10−7 for 𝜖.
• learning rate in a range between 1𝑒 − 3 and 5𝑒 − 4

CHAPTER
SIX
DEEP LEARNING
6.1 Backpropagation
6.1.1 Course outline:
1. Backpropagation and chaine rule

2. Lab: with numpy and pytorch
%matplotlib inline
6.1.2 Backpropagation and chaine rule
We will set up a two layer network source pytorch tuto :
Y = max(XW(1) , 0)W(2)
A fully-connected ReLU network with one hidden layer and no biases, trained to predict y from
x using Euclidean error.
Chaine rule
Forward pass with local partial derivatives of ouput given inputs:
𝑥 → 𝑧 (1) = 𝑥𝑇 𝑤(1) → ℎ(1) = max(𝑧 (1) , 0) → 𝑧 (2) = ℎ(1)𝑇 𝑤(2) → 𝐿(𝑧 (2) , 𝑦) = (𝑧 (2) − 𝑦)2
𝑤(1) ↗ 𝑤(2) ↗
𝜕𝑧 (1) 𝜕ℎ(1) 1 if 𝑧 (1) >0 𝜕𝑧
(2) 𝜕𝐿
=𝑥 = {else 0 = ℎ(1) = 2(𝑧 (2) − 𝑦)
𝜕𝑤(1) 𝜕𝑧 (1) 𝜕𝑤(2) 𝜕𝑧 (2)
𝜕𝑧 (1) 𝜕𝑧 (2)
= 𝑤(1) = 𝑤(2)
𝜕𝑥 𝜕ℎ(1)
Backward: compute gradient of the loss given each parameters vectors applying chaine rule
from the loss downstream to the parameters:
For 𝑤(2) :
257
𝜕𝐿 𝜕𝐿 𝜕𝑧 (2)
= (6.1)
𝜕𝑤(2) 𝜕𝑧 (2) 𝜕𝑤(2)
=2(𝑧 (2) − 𝑦)ℎ(1) (6.2)
For 𝑤(1) :
𝜕𝐿 𝜕𝐿 𝜕𝑧 (2) 𝜕ℎ(1) 𝜕𝑧 (1)

= (6.3)
𝜕𝑤(1) 𝜕𝑧 (2) 𝜕ℎ(1) 𝜕𝑧 (1) 𝜕𝑤(1)
if 𝑧 (1) >0
=2(𝑧 (2) − 𝑦)𝑤(2) {1else 0 𝑥 (6.4)
Recap: Vector derivatives
Given a function 𝑧 = 𝑥 with 𝑧 the output, 𝑥 the input and 𝑤 the coeficients.
• Scalar to Scalar: 𝑥 ∈ R, 𝑧 ∈ R, 𝑤 ∈ R
Regular derivative:
𝜕𝑧
=𝑥∈R
𝜕𝑤
If 𝑤 changes by a small amount, how much will 𝑧 change?
• Vector to Scalar: 𝑥 ∈ R𝑁 , 𝑧 ∈ R, 𝑤 ∈ R𝑁
𝜕𝑧
Derivative is Gradient of partial derivative: 𝜕𝑤 ∈ R𝑁
⎡ 𝜕𝑧 ⎤
𝜕𝑤1
⎢ .. ⎥
⎢ . ⎥
𝜕𝑧 ⎢ 𝜕𝑧 ⎥
= ∇𝑤 𝑧 = ⎢
⎢ 𝜕𝑤𝑖
⎥ (6.5)
𝜕𝑤 ⎢ ..
⎥
⎥
⎣ . ⎦
𝜕𝑧
𝜕𝑤𝑁
For each element 𝑤𝑖 of 𝑤, if it changes by a small amount then how much will y change?
• Vector to Vector: 𝑤 ∈ R𝑁 , 𝑧 ∈ R𝑀
Derivative is Jacobian of partial derivative:
TO COMPLETE
𝜕𝑧
𝜕𝑤 ∈ R𝑁 ×𝑀
258 Chapter 6. Deep Learning

Backpropagation summary
Backpropagation algorithm in a graph: 1. Forward pass, for each node compute local partial
derivatives of ouput given inputs 2. Backward pass: apply chain rule from the end to each
parameters - Update parameter with gradient descent using the current upstream gradient and
the current local gradient - Compute upstream gradient for the backward nodes
Think locally and remember that at each node: - For the loss the gradient is the error - At each
step, the upstream gradient is obtained by multiplying the upstream gradient (an error) with
the current parameters (vector of matrix). - At each step, the current local gradient equal the
input, therfore the current update is the current upstream gradient time the input.
import numpy as np
import sklearn.model_selection
6.1.3 Lab: with numpy and pytorch
Load iris data set
Goal: Predict Y = [petal_length, petal_width] = f(X = [sepal_length, sepal_width])

• Plot data with seaborn
• Remove setosa samples
• Recode ‘versicolor’:1, ‘virginica’:2
• Scale X and Y
• Split data in train/test 50%/50%
iris = sns.load_dataset("iris")
#g = sns.pairplot(iris, hue="species")
df = iris[iris.species != "setosa"]
g = sns.pairplot(df, hue="species")
df['species_n'] = iris.species.map({'versicolor':1, 'virginica':2})
# Y = 'petal_length', 'petal_width'; X = 'sepal_length', 'sepal_width')

X_iris = np.asarray(df.loc[:, ['sepal_length', 'sepal_width']], dtype=np.float32)
Y_iris = np.asarray(df.loc[:, ['petal_length', 'petal_width']], dtype=np.float32)
label_iris = np.asarray(df.species_n, dtype=int)
# Scale
from sklearn.preprocessing import StandardScaler
scalerx, scalery = StandardScaler(), StandardScaler()
X_iris = scalerx.fit_transform(X_iris)
Y_iris = StandardScaler().fit_transform(Y_iris)
# Split train test

X_iris_tr, X_iris_val, Y_iris_tr, Y_iris_val, label_iris_tr, label_iris_val = \
sklearn.model_selection.train_test_split(X_iris, Y_iris, label_iris, train_size=0.5,␣
˓→stratify=label_iris)
6.1. Backpropagation 259

/home/edouard/anaconda3/lib/python3.7/site-packages/ipykernel_launcher.py:5:␣
˓→SettingWithCopyWarning:
A value is trying to be set on a copy of a slice from a DataFrame.
Try using .loc[row_indexer,col_indexer] = value instead
See the caveats in the documentation: http://pandas.pydata.org/pandas-docs/stable/

˓→indexing.html#indexing-view-versus-copy
"""
Backpropagation with numpy
This implementation uses numpy to manually compute the forward pass, loss, and backward
pass.
# X=X_iris_tr; Y=Y_iris_tr; X_val=X_iris_val; Y_val=Y_iris_val
def two_layer_regression_numpy_train(X, Y, X_val, Y_val, lr, nite):

# N is batch size; D_in is input dimension;
# H is hidden dimension; D_out is output dimension.
# N, D_in, H, D_out = 64, 1000, 100, 10
N, D_in, H, D_out = X.shape[0], X.shape[1], 100, Y.shape[1]

W1 = np.random.randn(D_in, H)
W2 = np.random.randn(H, D_out)
losses_tr, losses_val = list(), list()
learning_rate = lr
for t in range(nite):
# Forward pass: compute predicted y
z1 = X.dot(W1)
h1 = np.maximum(z1, 0)
Y_pred = h1.dot(W2)
# Compute and print loss

loss = np.square(Y_pred - Y).sum()
# Backprop to compute gradients of w1 and w2 with respect to loss

grad_y_pred = 2.0 * (Y_pred - Y)
grad_w2 = h1.T.dot(grad_y_pred)
grad_h1 = grad_y_pred.dot(W2.T)
grad_z1 = grad_h1.copy()
grad_z1[z1 < 0] = 0
grad_w1 = X.T.dot(grad_z1)
# Update weights
W1 -= learning_rate * grad_w1
# Forward pass for validation set: compute predicted y

z1 = X_val.dot(W1)
h1 = np.maximum(z1, 0)
y_pred_val = h1.dot(W2)
loss_val = np.square(y_pred_val - Y_val).sum()
losses_tr.append(loss)
losses_val.append(loss_val)
if t % 10 == 0:
print(t, loss, loss_val)
return W1, W2, losses_tr, losses_val
W1, W2, losses_tr, losses_val = two_layer_regression_numpy_train(X=X_iris_tr, Y=Y_iris_tr,

˓→ X_val=X_iris_val, Y_val=Y_iris_val,
lr=1e-4, nite=50)
plt.plot(np.arange(len(losses_tr)), losses_tr, "-b", np.arange(len(losses_val)), losses_
˓→val, "-r")
0 15126.224825529907 2910.260853330454
10 71.5381374591153 104.97056197642135
20 50.756938353833334 80.02800827986354
30 46.546510744624236 72.85211241738614
40 44.41413064447564 69.31127324764276

[<matplotlib.lines.Line2D at 0x7f960cf5e9b0>,
<matplotlib.lines.Line2D at 0x7f960cf5eb00>]
Backpropagation with PyTorch Tensors
source
Numpy is a great framework, but it cannot utilize GPUs to accelerate its numerical compu-
tations. For modern deep neural networks, GPUs often provide speedups of 50x or greater,
so unfortunately numpy won’t be enough for modern deep learning. Here we introduce the
most fundamental PyTorch concept: the Tensor. A PyTorch Tensor is conceptually identical to a
numpy array: a Tensor is an n-dimensional array, and PyTorch provides many functions for op-
erating on these Tensors. Behind the scenes, Tensors can keep track of a computational graph
and gradients, but they’re also useful as a generic tool for scientific computing. Also unlike
numpy, PyTorch Tensors can utilize GPUs to accelerate their numeric computations. To run a
PyTorch Tensor on GPU, you simply need to cast it to a new datatype. Here we use PyTorch
Tensors to fit a two-layer network to random data. Like the numpy example above we need to
manually implement the forward and backward passes through the network:
import torch
def two_layer_regression_tensor_train(X, Y, X_val, Y_val, lr, nite):
dtype = torch.float
device = torch.device("cpu")
# device = torch.device("cuda:0") # Uncomment this to run on GPU


# Create random input and output data

X = torch.from_numpy(X)
Y = torch.from_numpy(Y)
X_val = torch.from_numpy(X_val)
Y_val = torch.from_numpy(Y_val)
# Randomly initialize weights

W1 = torch.randn(D_in, H, device=device, dtype=dtype)
W2 = torch.randn(H, D_out, device=device, dtype=dtype)
learning_rate = lr
# Forward pass: compute predicted y
z1 = X.mm(W1)
h1 = z1.clamp(min=0)
y_pred = h1.mm(W2)
# Compute and print loss

loss = (y_pred - Y).pow(2).sum().item()
# Backprop to compute gradients of w1 and w2 with respect to loss

grad_y_pred = 2.0 * (y_pred - Y)
grad_w2 = h1.t().mm(grad_y_pred)
grad_h1 = grad_y_pred.mm(W2.t())
grad_z1 = grad_h1.clone()
grad_z1[z1 < 0] = 0
grad_w1 = X.t().mm(grad_z1)
# Update weights using gradient descent

# Forward pass for validation set: compute predicted y

z1 = X_val.mm(W1)
h1 = z1.clamp(min=0)
y_pred_val = h1.mm(W2)
loss_val = (y_pred_val - Y_val).pow(2).sum().item()
losses_tr.append(loss)
losses_val.append(loss_val)
if t % 10 == 0:
print(t, loss, loss_val)
W1, W2, losses_tr, losses_val = two_layer_regression_tensor_train(X=X_iris_tr, Y=Y_iris_

˓→tr, X_val=X_iris_val, Y_val=Y_iris_val,
lr=1e-4, nite=50)

˓→val, "-r")

0 8086.1591796875 5429.57275390625
10 225.77589416503906 331.83734130859375
20 86.46501159667969 117.72447204589844
30 52.375606536865234 73.84156036376953
40 43.16458511352539 64.0667495727539
[<matplotlib.lines.Line2D at 0x7f960033c470>,
<matplotlib.lines.Line2D at 0x7f960033c5c0>]
Backpropagation with PyTorch: Tensors and autograd
source
A fully-connected ReLU network with one hidden layer and no biases, trained to predict y
from x by minimizing squared Euclidean distance. This implementation computes the forward
pass using operations on PyTorch Tensors, and uses PyTorch autograd to compute gradients.
A PyTorch Tensor represents a node in a computational graph. If x is a Tensor that has x.
requires_grad=True then x.grad is another Tensor holding the gradient of x with respect to
some scalar value.
import torch

# del X, Y, X_val, Y_val
def two_layer_regression_autograd_train(X, Y, X_val, Y_val, lr, nite):
dtype = torch.float
device = torch.device("cpu")
# device = torch.device("cuda:0") # Uncomment this to run on GPU



# Setting requires_grad=False indicates that we do not need to compute gradients

# with respect to these Tensors during the backward pass.
# Create random Tensors for weights.

# Setting requires_grad=True indicates that we want to compute gradients with
# respect to these Tensors during the backward pass.
W1 = torch.randn(D_in, H, device=device, dtype=dtype, requires_grad=True)
W2 = torch.randn(H, D_out, device=device, dtype=dtype, requires_grad=True)
learning_rate = lr
# Forward pass: compute predicted y using operations on Tensors; these
# are exactly the same operations we used to compute the forward pass using
# Tensors, but we do not need to keep references to intermediate values since
# we are not implementing the backward pass by hand.
y_pred = X.mm(W1).clamp(min=0).mm(W2)
# Compute and print loss using operations on Tensors.

# Now loss is a Tensor of shape (1,)
# loss.item() gets the scalar value held in the loss.
loss = (y_pred - Y).pow(2).sum()
# Use autograd to compute the backward pass. This call will compute the
# gradient of loss with respect to all Tensors with requires_grad=True.
# After this call w1.grad and w2.grad will be Tensors holding the gradient
# of the loss with respect to w1 and w2 respectively.
loss.backward()
# Manually update weights using gradient descent. Wrap in torch.no_grad()

# because weights have requires_grad=True, but we don't need to track this
# in autograd.
# An alternative way is to operate on weight.data and weight.grad.data.
# Recall that tensor.data gives a tensor that shares the storage with
# tensor, but doesn't track history.
# You can also use torch.optim.SGD to achieve this.
with torch.no_grad():
W1 -= learning_rate * W1.grad
W2 -= learning_rate * W2.grad
# Manually zero the gradients after updating weights

W1.grad.zero_()
W2.grad.zero_()
y_pred = X_val.mm(W1).clamp(min=0).mm(W2)
# Compute and print loss using operations on Tensors.

# Now loss is a Tensor of shape (1,)
# loss.item() gets the scalar value held in the loss.


loss_val = (y_pred - Y).pow(2).sum()
if t % 10 == 0:
print(t, loss.item(), loss_val.item())
losses_tr.append(loss.item())
losses_val.append(loss_val.item())
W1, W2, losses_tr, losses_val = two_layer_regression_autograd_train(X=X_iris_tr, Y=Y_iris_

lr=1e-4, nite=50)
˓→val, "-r")
0 8307.1806640625 2357.994873046875
10 111.97289276123047 250.04209899902344
20 65.83244323730469 201.63694763183594
30 53.70908737182617 183.17051696777344
40 48.719329833984375 173.3616943359375
[<matplotlib.lines.Line2D at 0x7f95ff2ad978>,
<matplotlib.lines.Line2D at 0x7f95ff2adac8>]

Backpropagation with PyTorch: nn
source
This implementation uses the nn package from PyTorch to build the network. PyTorch autograd
makes it easy to define computational graphs and take gradients, but raw autograd can be a bit
too low-level for defining complex neural networks; this is where the nn package can help. The
nn package defines a set of Modules, which you can think of as a neural network layer that has
produces output from input and may have some trainable weights.
import torch

# del X, Y, X_val, Y_val
def two_layer_regression_nn_train(X, Y, X_val, Y_val, lr, nite):

# Use the nn package to define our model as a sequence of layers. nn.Sequential

# is a Module which contains other Modules, and applies them in sequence to
# produce its output. Each Linear Module computes output from input using a
# linear function, and holds internal Tensors for its weight and bias.
model = torch.nn.Sequential(
torch.nn.Linear(D_in, H),
torch.nn.ReLU(),
torch.nn.Linear(H, D_out),
)
# The nn package also contains definitions of popular loss functions; in this

# case we will use Mean Squared Error (MSE) as our loss function.
loss_fn = torch.nn.MSELoss(reduction='sum')
learning_rate = lr
# Forward pass: compute predicted y by passing x to the model. Module objects
# override the __call__ operator so you can call them like functions. When
# doing so you pass a Tensor of input data to the Module and it produces
# a Tensor of output data.
y_pred = model(X)
# Compute and print loss. We pass Tensors containing the predicted and true
# values of y, and the loss function returns a Tensor containing the
# loss.
loss = loss_fn(y_pred, Y)
# Zero the gradients before running the backward pass.

model.zero_grad()

# Backward pass: compute gradient of the loss with respect to all the learnable
# parameters of the model. Internally, the parameters of each Module are stored
# in Tensors with requires_grad=True, so this call will compute gradients for
# all learnable parameters in the model.
loss.backward()
# Update the weights using gradient descent. Each parameter is a Tensor, so

# we can access its gradients like we did before.
for param in model.parameters():
param -= learning_rate * param.grad
y_pred = model(X_val)
loss_val = (y_pred - Y_val).pow(2).sum()
if t % 10 == 0:
return model, losses_tr, losses_val
model, losses_tr, losses_val = two_layer_regression_nn_train(X=X_iris_tr, Y=Y_iris_tr, X_

˓→val=X_iris_val, Y_val=Y_iris_val,
lr=1e-4, nite=50)

˓→val, "-r")
0 82.32025146484375 91.3389892578125
10 50.322200775146484 63.563087463378906
20 40.825225830078125 57.13555145263672
30 37.53572082519531 55.74506378173828
40 36.191200256347656 55.499732971191406
[<matplotlib.lines.Line2D at 0x7f95ff296668>,
<matplotlib.lines.Line2D at 0x7f95ff2967b8>]

Backpropagation with PyTorch optim
This implementation uses the nn package from PyTorch to build the network. Rather than man-
ually updating the weights of the model as we have been doing, we use the optim package to
define an Optimizer that will update the weights for us. The optim package defines many op-
timization algorithms that are commonly used for deep learning, including SGD+momentum,
RMSProp, Adam, etc.
import torch
def two_layer_regression_nn_optim_train(X, Y, X_val, Y_val, lr, nite):

# Use the nn package to define our model and loss function.

model = torch.nn.Sequential(
torch.nn.Linear(D_in, H),
torch.nn.ReLU(),
torch.nn.Linear(H, D_out),
)
loss_fn = torch.nn.MSELoss(reduction='sum')
# Use the optim package to define an Optimizer that will update the weights of


# the model for us. Here we will use Adam; the optim package contains many other
# optimization algoriths. The first argument to the Adam constructor tells the
# optimizer which Tensors it should update.
learning_rate = lr
optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
# Forward pass: compute predicted y by passing x to the model.
y_pred = model(X)
# Compute and print loss.

loss = loss_fn(y_pred, Y)
# Before the backward pass, use the optimizer object to zero all of the
# gradients for the variables it will update (which are the learnable
# weights of the model). This is because by default, gradients are
# accumulated in buffers( i.e, not overwritten) whenever .backward()
# is called. Checkout docs of torch.autograd.backward for more details.
optimizer.zero_grad()
# Backward pass: compute gradient of the loss with respect to model

# parameters
loss.backward()
# Calling the step function on an Optimizer makes an update to its

# parameters
optimizer.step()
y_pred = model(X_val)
loss_val = loss_fn(y_pred, Y_val)
if t % 10 == 0:
return model, losses_tr, losses_val
model, losses_tr, losses_val = two_layer_regression_nn_optim_train(X=X_iris_tr, Y=Y_iris_

lr=1e-3, nite=50)
˓→val, "-r")
0 92.271240234375 83.96189880371094
10 64.25907135009766 59.872535705566406
20 47.6252555847168 50.228126525878906
30 40.33802032470703 50.60377502441406
40 38.19448471069336 54.03163528442383
[<matplotlib.lines.Line2D at 0x7f95ff200080>,
<matplotlib.lines.Line2D at 0x7f95ff2001d0>]

6.2 Multilayer Perceptron (MLP)
6.2.1 Course outline:
1. Recall of linear classifier

2. MLP with scikit-learn
3. MLP with pytorch
4. Test several MLP architectures
5. Limits of MLP
Sources:
Deep learning
• cs231n.stanford.edu
Pytorch
• WWW tutorials
• github tutorials
• github examples
MNIST and pytorch:
• MNIST nextjournal.com/gkoehler/pytorch-mnist
• MNIST github/pytorch/examples
• MNIST kaggle
6.2. Multilayer Perceptron (MLP) 271

%matplotlib inline
import os
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torch.optim import lr_scheduler
import torchvision
from torchvision import transforms
from torchvision import datasets
from torchvision import models
#
from pathlib import Path
# Device configuration
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
print(device)
cuda:0
Hyperparameters
6.2.2 Dataset: MNIST Handwritten Digit Recognition

WD = os.path.join(Path.home(), "data", "pystatml", "dl_mnist_pytorch")
os.chdir(WD)
print("Working dir is:", os.getcwd())
os.makedirs("data", exist_ok=True)
os.makedirs("models", exist_ok=True)
def load_mnist(batch_size_train, batch_size_test):
train_loader = torch.utils.data.DataLoader(
datasets.MNIST('data', train=True, download=True,
transform=transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.1307,), (0.3081,)) # Mean and Std of␣
˓→the MNIST dataset
])),
batch_size=batch_size_train, shuffle=True)
val_loader = torch.utils.data.DataLoader(
datasets.MNIST('data', train=False, transform=transforms.Compose([
transforms.Normalize((0.1307,), (0.3081,)) # Mean and Std of the MNIST dataset
])),
batch_size=batch_size_test, shuffle=True)
return train_loader, val_loader

train_loader, val_loader = load_mnist(64, 10000)
dataloaders = dict(train=train_loader, val=val_loader)
# Info about the dataset

D_in = np.prod(dataloaders["train"].dataset.data.shape[1:])
D_out = len(dataloaders["train"].dataset.targets.unique())
print("Datasets shapes:", {x: dataloaders[x].dataset.data.shape for x in ['train', 'val']}
˓→)
print("N input features:", D_in, "Output classes:", D_out)
Working dir is: /volatile/duchesnay/data/pystatml/dl_mnist_pytorch

Datasets shapes: {'train': torch.Size([60000, 28, 28]), 'val': torch.Size([10000, 28,␣
˓→28])}
N input features: 784 Output classes: 10
Now let’s take a look at some mini-batches examples.
batch_idx, (example_data, example_targets) = next(enumerate(train_loader))

print("Train batch:", example_data.shape, example_targets.shape)
batch_idx, (example_data, example_targets) = next(enumerate(val_loader))
print("Val batch:", example_data.shape, example_targets.shape)
Train batch: torch.Size([64, 1, 28, 28]) torch.Size([64])

Val batch: torch.Size([10000, 1, 28, 28]) torch.Size([10000])
So one test data batch is a tensor of shape: . This means we have 1000 examples of 28x28 pixels
in grayscale (i.e. no rgb channels, hence the one). We can plot some of them using matplotlib.
def show_data_label_prediction(data, y_true, y_pred=None, shape=(2, 3)):

y_pred = [None] * len(y_true) if y_pred is None else y_pred
fig = plt.figure()
for i in range(np.prod(shape)):
plt.subplot(*shape, i+1)
plt.tight_layout()
plt.imshow(data[i][0], cmap='gray', interpolation='none')
plt.title("True: {} Pred: {}".format(y_true[i], y_pred[i]))
plt.xticks([])
plt.yticks([])
show_data_label_prediction(data=example_data, y_true=example_targets, y_pred=None,␣

˓→shape=(2, 3))

6.2.3 Recall of linear classifier
Binary logistic regression
1 neuron as output layer
𝑓 (𝑥) = 𝜎(𝑥𝑇 𝑤)
Softmax Classifier (Multinomial Logistic Regression)
• Input 𝑥: a vector of dimension (0) (layer 0).

• Ouput 𝑓 (𝑥) a vector of (1) (layer 1) possible labels
The model as (1) neurons as output layer
𝑓 (𝑥) = softmax(𝑥𝑇 𝑊 + 𝑏)
Where 𝑊 is a (0) × (1) of coefficients and 𝑏 is a (1)-dimentional vector of bias.

MNIST classfification using multinomial logistic
source: Logistic regression MNIST
Here we fit a multinomial logistic regression with L2 penalty on a subset of the MNIST digits
classification task.
source: scikit-learn.org
X_train = train_loader.dataset.data.numpy()
#print(X_train.shape)
X_train = X_train.reshape((X_train.shape[0], -1))


y_train = train_loader.dataset.targets.numpy()
X_test = val_loader.dataset.data.numpy()
X_test = X_test.reshape((X_test.shape[0], -1))
y_test = val_loader.dataset.targets.numpy()
print(X_train.shape, y_train.shape)
(60000, 784) (60000,)

import numpy as np
#from sklearn.datasets import fetch_openml

from sklearn.linear_model import LogisticRegression
#from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler
from sklearn.utils import check_random_state
scaler = StandardScaler()
X_train = scaler.fit_transform(X_train)
X_test = scaler.transform(X_test)
# Turn up tolerance for faster convergence

clf = LogisticRegression(C=50., multi_class='multinomial', solver='sag', tol=0.1)
clf.fit(X_train, y_train)
#sparsity = np.mean(clf.coef_ == 0) * 100
score = clf.score(X_test, y_test)
print("Test score with penalty: %.4f" % score)
Test score with penalty: 0.9035
coef = clf.coef_.copy()
scale = np.abs(coef).max()
for i in range(10):
l1_plot = plt.subplot(2, 5, i + 1)
l1_plot.imshow(coef[i].reshape(28, 28), interpolation='nearest',
cmap=plt.cm.RdBu, vmin=-scale, vmax=scale)
l1_plot.set_xticks(())
l1_plot.set_yticks(())
l1_plot.set_xlabel('Class %i' % i)
plt.suptitle('Classification vector for...')
plt.show()

6.2.4 Model: Two Layer MLP
MLP with Scikit-learn
from sklearn.neural_network import MLPClassifier
mlp = MLPClassifier(hidden_layer_sizes=(100, ), max_iter=5, alpha=1e-4,

solver='sgd', verbose=10, tol=1e-4, random_state=1,
learning_rate_init=0.01, batch_size=64)
mlp.fit(X_train, y_train)
print("Training set score: %f" % mlp.score(X_train, y_train))
print("Test set score: %f" % mlp.score(X_test, y_test))
print("Coef shape=", len(mlp.coefs_))
fig, axes = plt.subplots(4, 4)

# use global min / max to ensure all weights are shown on the same scale
vmin, vmax = mlp.coefs_[0].min(), mlp.coefs_[0].max()
for coef, ax in zip(mlp.coefs_[0].T, axes.ravel()):
ax.matshow(coef.reshape(28, 28), cmap=plt.cm.gray, vmin=.5 * vmin,
vmax=.5 * vmax)
ax.set_xticks(())
ax.set_yticks(())
plt.show()
Iteration 1, loss = 0.28611761


/home/ed203246/anaconda3/lib/python3.7/site-packages/sklearn/neural_network/multilayer_
˓→perceptron.py:562: ConvergenceWarning: Stochastic Optimizer: Maximum iterations (5)␣
˓→reached and the optimization hasn't converged yet.
% self.max_iter, ConvergenceWarning)
Training set score: 0.989067

Test set score: 0.971900
Coef shape= 2
MLP with pytorch
class TwoLayerMLP(nn.Module):
def __init__(self, d_in, d_hidden, d_out):

super(TwoLayerMLP, self).__init__()
self.d_in = d_in
self.linear1 = nn.Linear(d_in, d_hidden)

self.linear2 = nn.Linear(d_hidden, d_out)
def forward(self, X):

X = X.view(-1, self.d_in)
X = self.linear1(X)
return F.log_softmax(self.linear2(X), dim=1)

Train the Model
• First we want to make sure our network is in training mode.

• Iterate over epochs
• Alternate train and validation dataset
• Iterate over all training/val data once per epoch. Loading the individual batches is han-
dled by the DataLoader.
• Set the gradients to zero using optimizer.zero_grad() since PyTorch by default accumu-
lates gradients.
• Forward pass:
– model(inputs): Produce the output of our network.
– torch.max(outputs, 1): softmax predictions.
– criterion(outputs, labels): loss between the output and the ground truth label.
• In training mode, backward pass backward(): collect a new set of gradients which we
propagate back into each of the network’s parameters using optimizer.step().
• We’ll also keep track of the progress with some printouts. In order to create a nice training
curve later on we also create two lists for saving training and testing losses. On the x-axis
we want to display the number of training examples the network has seen during training.
• Save model state: Neural network modules as well as optimizers have the ability
to save and load their internal state using .state_dict(). With this we can con-
tinue training from previously saved state dicts if needed - we’d just need to call .
load_state_dict(state_dict).
# %load train_val_model.py
import numpy as np
import torch
import time
import copy
def train_val_model(model, criterion, optimizer, dataloaders, num_epochs=25,

scheduler=None, log_interval=None):
since = time.time()
best_model_wts = copy.deepcopy(model.state_dict())
best_acc = 0.0
# Store losses and accuracies accross epochs

losses, accuracies = dict(train=[], val=[]), dict(train=[], val=[])
for epoch in range(num_epochs):

if log_interval is not None and epoch % log_interval == 0:
print('Epoch {}/{}'.format(epoch, num_epochs - 1))
print('-' * 10)


# Each epoch has a training and validation phase
for phase in ['train', 'val']:
if phase == 'train':
model.train() # Set model to training mode
else:
model.eval() # Set model to evaluate mode
running_loss = 0.0
running_corrects = 0
# Iterate over data.

nsamples = 0
for inputs, labels in dataloaders[phase]:
inputs = inputs.to(device)
labels = labels.to(device)
nsamples += inputs.shape[0]
# zero the parameter gradients

# forward
# track history if only in train
with torch.set_grad_enabled(phase == 'train'):
outputs = model(inputs)
_, preds = torch.max(outputs, 1)
loss = criterion(outputs, labels)
# backward + optimize only if in training phase

loss.backward()
optimizer.step()
# statistics
running_loss += loss.item() * inputs.size(0)
running_corrects += torch.sum(preds == labels.data)
if scheduler is not None and phase == 'train':

scheduler.step()
#nsamples = dataloaders[phase].dataset.data.shape[0]
epoch_loss = running_loss / nsamples
epoch_acc = running_corrects.double() / nsamples
losses[phase].append(epoch_loss)
accuracies[phase].append(epoch_acc)
print('{} Loss: {:.4f} Acc: {:.2f}%'.format(
phase, epoch_loss, 100 * epoch_acc))
# deep copy the model

if phase == 'val' and epoch_acc > best_acc:
best_acc = epoch_acc
print()


time_elapsed = time.time() - since
print('Training complete in {:.0f}m {:.0f}s'.format(
time_elapsed // 60, time_elapsed % 60))
print('Best val Acc: {:.2f}%'.format(100 * best_acc))
# load best model weights

model.load_state_dict(best_model_wts)
return model, losses, accuracies
Run one epoch and save the model
model = TwoLayerMLP(D_in, 50, D_out).to(device)

print(next(model.parameters()).is_cuda)
optimizer = optim.SGD(model.parameters(), lr=0.01, momentum=0.5)
criterion = nn.NLLLoss()
# Explore the model

for parameter in model.parameters():
print(parameter.shape)
print("Total number of parameters =", np.sum([np.prod(parameter.shape) for parameter in␣

˓→model.parameters()]))
model, losses, accuracies = train_val_model(model, criterion, optimizer, dataloaders,

num_epochs=1, log_interval=1)
print(next(model.parameters()).is_cuda)
torch.save(model.state_dict(), 'models/mod-%s.pth' % model.__class__.__name__)
True
torch.Size([50, 784])
torch.Size([50])
torch.Size([10])
Total number of parameters = 39760
Epoch 0/0
----------
train Loss: 0.4472 Acc: 87.65%
val Loss: 0.3115 Acc: 91.25%
Training complete in 0m 10s

Best val Acc: 91.25%
True
Use the model to make new predictions. Consider the device, ie, load data on device
example_data.to(device) from prediction, then move back to cpu example_data.cpu().
batch_idx, (example_data, example_targets) = next(enumerate(val_loader))

example_data = example_data.to(device)
output = model(example_data).cpu()
example_data = example_data.cpu()

# print(output.is_cuda)
# Softmax predictions
preds = output.argmax(dim=1)
print("Output shape=", output.shape, "label shape=", preds.shape)

print("Accuracy = {:.2f}%".format((example_targets == preds).sum().item() * 100. /␣
˓→len(example_targets)))
show_data_label_prediction(data=example_data, y_true=example_targets, y_pred=preds,␣

˓→shape=(3, 4))
Output shape= torch.Size([10000, 10]) label shape= torch.Size([10000])

Accuracy = 91.25%
Plot missclassified samples
errors = example_targets != preds

#print(errors, np.where(errors))
print("Nb errors = {}, (Error rate = {:.2f}%)".format(errors.sum(), 100 * errors.sum().
˓→item() / len(errors)))
err_idx = np.where(errors)[0]
show_data_label_prediction(data=example_data[err_idx], y_true=example_targets[err_idx],
y_pred=preds[err_idx], shape=(3, 4))
Nb errors = 875, (Error rate = 8.75%)

Continue training from checkpoints: reload the model and run 10 more epochs
model = TwoLayerMLP(D_in, 50, D_out)

model.load_state_dict(torch.load('models/mod-%s.pth' % model.__class__.__name__))
model.to(device)


_ = plt.plot(losses['train'], '-b', losses['val'], '--r')
Epoch 0/9
----------
val Loss: 0.2877 Acc: 91.92%
Epoch 2/9
----------
val Loss: 0.2797 Acc: 92.05%
Epoch 4/9
----------
val Loss: 0.2797 Acc: 92.11%
Epoch 6/9
----------


val Loss: 0.2717 Acc: 92.25%
Epoch 8/9
----------
val Loss: 0.2684 Acc: 92.44%

6.2.5 Test several MLP architectures
• Define a MultiLayerMLP([D_in, 512, 256, 128, 64, D_out]) class that take the size of
the layers as parameters of the constructor.
• Add some non-linearity with relu acivation function
class MLP(nn.Module):
def __init__(self, d_layer):

super(MLP, self).__init__()
self.d_layer = d_layer
layer_list = [nn.Linear(d_layer[l], d_layer[l+1]) for l in range(len(d_layer) -␣
˓→1)]
self.linears = nn.ModuleList(layer_list)
def forward(self, X):

X = X.view(-1, self.d_layer[0])
# relu(Wl x) for all hidden layer
for layer in self.linears[:-1]:
X = F.relu(layer(X))
# softmax(Wl x) for output layer
return F.log_softmax(self.linears[-1](X), dim=1)

model = MLP([D_in, 512, 256, 128, 64, D_out]).to(device)


Epoch 0/9
----------
val Loss: 0.3407 Acc: 89.73%
Epoch 2/9
----------
val Loss: 0.1510 Acc: 95.47%
Epoch 4/9
----------
val Loss: 0.1070 Acc: 96.76%
Epoch 6/9
----------
val Loss: 0.0967 Acc: 96.98%
Epoch 8/9
----------
val Loss: 0.0822 Acc: 97.55%


6.2.6 Reduce the size of training dataset
Reduce the size of the training dataset by considering only 10 minibatche for size16.
train_size = 10 * 16
# Stratified sub-sampling
targets = train_loader.dataset.targets.numpy()
nclasses = len(set(targets))
indices = np.concatenate([np.random.choice(np.where(targets == lab)[0], int(train_size /␣

˓→nclasses),replace=False)
for lab in set(targets)])
np.random.shuffle(indices)
train_loader = torch.utils.data.DataLoader(train_loader.dataset, batch_size=16,

sampler=torch.utils.data.SubsetRandomSampler(indices))
# Check train subsampling

train_labels = np.concatenate([labels.numpy() for inputs, labels in train_loader])
print("Train size=", len(train_labels), " Train label count=", {lab:np.sum(train_labels␣
˓→== lab) for lab in set(train_labels)})
print("Batch sizes=", [inputs.size(0) for inputs, labels in train_loader])
# Put together train and val


print("Datasets shape", {x: dataloaders[x].dataset.data.shape for x in ['train', 'val']})
print("N input features", D_in, "N output", D_out)

Train size= 160 Train label count= {0: 16, 1: 16, 2: 16, 3: 16, 4: 16, 5: 16, 6: 16, 7:␣
˓→16, 8: 16, 9: 16}
Batch sizes= [16, 16, 16, 16, 16, 16, 16, 16, 16, 16]
Datasets shape {'train': torch.Size([60000, 28, 28]), 'val': torch.Size([10000, 28, 28])}
N input features 784 N output 10


Epoch 0/99
----------
val Loss: 2.3058 Acc: 10.34%
Epoch 20/99
----------
val Loss: 2.1397 Acc: 51.34%
Epoch 40/99
----------
val Loss: 0.8372 Acc: 73.63%
Epoch 60/99
----------
train Loss: 0.0539 Acc: 100.00%
val Loss: 0.8384 Acc: 75.46%
Epoch 80/99
----------
train Loss: 0.0142 Acc: 100.00%
val Loss: 0.9417 Acc: 75.55%


Use an opimizer with an adaptative learning rate: Adam

optimizer = torch.optim.Adam(model.parameters(), lr=0.001)

Epoch 0/99
----------
val Loss: 2.0853 Acc: 45.51%
Epoch 20/99
----------
train Loss: 0.0010 Acc: 100.00%
val Loss: 1.0113 Acc: 78.08%
Epoch 40/99
----------
train Loss: 0.0002 Acc: 100.00%
val Loss: 1.1456 Acc: 78.12%
Epoch 60/99
----------
train Loss: 0.0001 Acc: 100.00%
val Loss: 1.2630 Acc: 77.98%
Epoch 80/99
----------
train Loss: 0.0000 Acc: 100.00%
val Loss: 1.3446 Acc: 77.87%


6.2.7 Run MLP on CIFAR-10 dataset
The CIFAR-10 dataset consists of 60000 32x32 colour images in 10 classes, with 6000 images
per class. There are 50000 training images and 10000 test images.
The dataset is divided into five training batches and one test batch, each with 10000 images.
The test batch contains exactly 1000 randomly-selected images from each class. The training
batches contain the remaining images in random order, but some training batches may contain
more images from one class than another. Between them, the training batches contain exactly
5000 images from each class.
Here are the classes in the dataset, as well as 10 random images from each: - airplane
- automobile
- bird
- cat
- deer
- dog
- frog
- horse
- ship
- truck
Load CIFAR-10 dataset


WD = os.path.join(Path.home(), "data", "pystatml", "dl_cifar10_pytorch")
os.chdir(WD)
import numpy as np
import torch
import torchvision
import torchvision.transforms as transforms
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# Hyper-parameters
num_epochs = 5
learning_rate = 0.001
# Image preprocessing modules

transform = transforms.Compose([
transforms.Pad(4),
transforms.RandomHorizontalFlip(),
transforms.RandomCrop(32),
transforms.ToTensor()])
# CIFAR-10 dataset
train_dataset = torchvision.datasets.CIFAR10(root='data/',
train=True,
transform=transform,
download=True)
val_dataset = torchvision.datasets.CIFAR10(root='data/',
train=False,
transform=transforms.ToTensor())
# Data loader
train_loader = torch.utils.data.DataLoader(dataset=train_dataset,
batch_size=100,
shuffle=True)
val_loader = torch.utils.data.DataLoader(dataset=val_dataset,
batch_size=100,
shuffle=False)


D_out = len(set(dataloaders["train"].dataset.targets))
print("Datasets shape:", {x: dataloaders[x].dataset.data.shape for x in ['train', 'val']})
print("N input features:", D_in, "N output:", D_out)

Working dir is: /volatile/duchesnay/data/pystatml/dl_cifar10_pytorch

Files already downloaded and verified
Datasets shape: {'train': (50000, 32, 32, 3), 'val': (10000, 32, 32, 3)}
N input features: 3072 N output: 10


Epoch 0/49
----------
val Loss: 1.8761 Acc: 30.84%
Epoch 10/49
----------
val Loss: 1.5853 Acc: 43.07%
Epoch 20/49
----------
val Loss: 1.4210 Acc: 48.88%
Epoch 30/49
----------
val Loss: 1.3890 Acc: 50.16%
Epoch 40/49
----------
val Loss: 1.3625 Acc: 51.41%


6.3 Convolutional neural network
6.3.1 Outline
2. Architecures
3. Train and test functions
4. CNN models
5. MNIST
6. CIFAR-10
Sources:
Deep learning - cs231n.stanford.edu
CNN - Stanford cs231n
Pytorch - WWW tutorials - github tutorials - github examples
MNIST and pytorch: - MNIST nextjournal.com/gkoehler/pytorch-mnist - MNIST
github/pytorch/examples - MNIST kaggle
6.3. Convolutional neural network 291

6.3.2 Architectures
Sources:
• cv-tricks.com
• [zhenye-na.github.io(]https://zhenye-na.github.io/2018/12/01/cnn-deep-leearning-ai-
week2.html)
LeNet
The first Convolutional Networks were developed by Yann LeCun in 1990’s.
Fig. 1: LeNet
AlexNet
(2012, Alex Krizhevsky, Ilya Sutskever and Geoff Hinton)
Fig. 2: AlexNet
• Deeper, bigger,
• Featured Convolutional Layers stacked on top of each other (previously it was common to
only have a single CONV layer always immediately followed by a POOL layer).
• ReLu(Rectified Linear Unit) for the non-linear part, instead of a Tanh or Sigmoid.

Fig. 3: AlexNet architecture
The advantage of the ReLu over sigmoid is that it trains much faster than the latter because the
derivative of sigmoid becomes very small in the saturating region and therefore the updates to
the weights almost vanish. This is called vanishing gradient problem.
• Dropout: reduces the over-fitting by using a Dropout layer after every FC layer. Dropout
layer has a probability,(p), associated with it and is applied at every neuron of the response
map separately. It randomly switches off the activation with the probability p.
Fig. 4: Dropout
Why does DropOut work?

The idea behind the dropout is similar to the model ensembles. Due to the dropout layer, dif-
ferent sets of neurons which are switched off, represent a different architecture and all these
different architectures are trained in parallel with weight given to each subset and the sum-
mation of weights being one. For n neurons attached to DropOut, the number of subset ar-
chitectures formed is 2^n. So it amounts to prediction being averaged over these ensembles
of models. This provides a structured model regularization which helps in avoiding the over-

fitting. Another view of DropOut being helpful is that since neurons are randomly chosen, they
tend to avoid developing co-adaptations among themselves thereby enabling them to develop
meaningful features, independent of others.
• Data augmentation is carried out to reduce over-fitting. This Data augmentation includes
mirroring and cropping the images to increase the variation in the training data-set.
GoogLeNet. (Szegedy et al. from Google 2014) was a Convolutional Network . Its main contri-
bution was the development of an
• Inception Module that dramatically reduced the number of parameters in the network
(4M, compared to AlexNet with 60M).
Fig. 5: Inception Module
• There are also several followup versions to the GoogLeNet, most recently Inception-v4.
VGGNet. (Karen Simonyan and Andrew Zisserman 2014)
• 16 CONV/FC layers and, appealingly, features an extremely homogeneous architecture.
• Only performs 3x3 convolutions and 2x2 pooling from the beginning to the end. Replace
large kernel-sized filters(11 and 5 in the first and second convolutional layer, respectively)
with multiple 3X3 kernel-sized filters one after another.
With a given receptive field(the effective area size of input image on which output depends),
multiple stacked smaller size kernel is better than the one with a larger size kernel because
multiple non-linear layers increases the depth of the network which enables it to learn more
complex features, and that too at a lower cost. For example, three 3X3 filters on top of each
other with stride 1 ha a receptive size of 7, but the number of parameters involved is 3*(9^2)
in comparison to 49^2 parameters of kernels with a size of 7.
• Lot more memory and parameters (140M)
ResNet. (Kaiming He et al. 2015)
Resnet block variants (Source):
• Skip connections

Fig. 6: VGGNet
Fig. 7: VGGNet architecture

Fig. 8: ResNet block
Fig. 9: ResNet 18

Fig. 10: ResNet 18 architecture
• Batch normalization.
• State of the art CNN models and are the default choice (as of May 10, 2016). In particular,
also see more
• Recent developments that tweak the original architecture from Kaiming He et al. Identity
Mappings in Deep Residual Networks (published March 2016).
Models in pytorch
6.3.3 Architecures general guidelines
• ConvNets stack CONV,POOL,FC layers

• Trend towards smaller filters and deeper architectures: stack 3x3, instead of 5x5
• Trend towards getting rid of POOL/FC layers (just CONV)
• Historically architectures looked like [(CONV-RELU) x N POOL?] x M (FC-RELU) x K,
SOFTMAX where N is usually up to ~5, M is large, 0 <= K <= 2.
• but recent advances such as ResNet/GoogLeNet have challenged this paradigm

6.3.4 Train function
%matplotlib inline
import os
import numpy as np
import torch
import torchvision
#
import numpy as np
import torch
import time
import copy

since = time.time()
best_acc = 0.0


print('-' * 10)

else:
running_loss = 0.0

nsamples = 0



# forward

loss.backward()
optimizer.step()
# statistics

scheduler.step()
print('{} Loss: {:.4f} Acc: {:.2f}%'.format(
phase, epoch_loss, 100 * epoch_acc))

print()

print('Best val Acc: {:.2f}%'.format(100 * best_acc))


6.3.5 CNN models
LeNet-5
Here we implement LeNet-5 with relu activation. Sources: (1), (2).

class LeNet5(nn.Module):
"""
layers: (nb channels in input layer,
nb channels in 1rst conv,
nb channels in 2nd conv,
nb neurons for 1rst FC: TO BE TUNED,
nb neurons for 2nd FC,
nb neurons for 3rd FC,
nb neurons output FC TO BE TUNED)
"""
def __init__(self, layers = (1, 6, 16, 1024, 120, 84, 10), debug=False):
super(LeNet5, self).__init__()
self.layers = layers
self.debug = debug
self.conv1 = nn.Conv2d(layers[0], layers[1], 5, padding=2)
self.conv2 = nn.Conv2d(layers[1], layers[2], 5)
self.fc1 = nn.Linear(layers[3], layers[4])
def forward(self, x):

x = F.max_pool2d(F.relu(self.conv1(x)), 2) # same shape / 2
x = F.max_pool2d(F.relu(self.conv2(x)), 2) # -4 / 2
if self.debug:
print("### DEBUG: Shape of last convnet=", x.shape[1:], ". FC size=", np.
˓→prod(x.shape[1:]))
x = x.view(-1, self.layers[3])
x = F.relu(self.fc1(x))
x = self.fc3(x)
return F.log_softmax(x, dim=1)
VGGNet like: conv-relu blocks
# Defining the network (LeNet-5)

class MiniVGGNet(torch.nn.Module):
def __init__(self, layers=(1, 16, 32, 1024, 120, 84, 10), debug=False):
super(MiniVGGNet, self).__init__()
self.layers = layers
self.debug = debug
# Conv block 1


self.conv11 = nn.Conv2d(in_channels=layers[0], out_channels=layers[1], kernel_
˓→size=3,
stride=1, padding=0, bias=True)

˓→size=3,
# Conv block 2
˓→size=3,
˓→size=3,
# Fully connected layer


x = F.relu(self.conv11(x))
x = F.max_pool2d(x, 2)
x = F.max_pool2d(x, 2)
if self.debug:
print("### DEBUG: Shape of last convnet=", x.shape[1:], ". FC size=", np.
˓→prod(x.shape[1:]))
x = x.view(-1, self.layers[3])
x = self.fc3(x)
return F.log_softmax(x, dim=1)
ResNet-like Model:
Stack multiple resnet blocks
# ---------------------------------------------------------------------------- #
# An implementation of https://arxiv.org/pdf/1512.03385.pdf #
# See section 4.2 for the model architecture on CIFAR-10 #
# Some part of the code was referenced from below #
# https://github.com/pytorch/vision/blob/master/torchvision/models/resnet.py #
# ---------------------------------------------------------------------------- #
# 3x3 convolution
def conv3x3(in_channels, out_channels, stride=1):
return nn.Conv2d(in_channels, out_channels, kernel_size=3,


stride=stride, padding=1, bias=False)
# Residual block
class ResidualBlock(nn.Module):
def __init__(self, in_channels, out_channels, stride=1, downsample=None):
super(ResidualBlock, self).__init__()
self.conv1 = conv3x3(in_channels, out_channels, stride)
self.bn1 = nn.BatchNorm2d(out_channels)
self.relu = nn.ReLU(inplace=True)
self.conv2 = conv3x3(out_channels, out_channels)
self.bn2 = nn.BatchNorm2d(out_channels)
self.downsample = downsample

residual = x
out = self.conv1(x)
out = self.bn1(out)
out = self.relu(out)
out = self.conv2(out)
out = self.bn2(out)
if self.downsample:
residual = self.downsample(x)
out += residual
return out
# ResNet
class ResNet(nn.Module):
def __init__(self, block, layers, num_classes=10):
super(ResNet, self).__init__()
self.in_channels = 16
self.conv = conv3x3(3, 16)
self.bn = nn.BatchNorm2d(16)
self.relu = nn.ReLU(inplace=True)
self.layer1 = self.make_layer(block, 16, layers[0])
self.layer2 = self.make_layer(block, 32, layers[1], 2)
self.layer3 = self.make_layer(block, 64, layers[2], 2)
self.avg_pool = nn.AvgPool2d(8)
self.fc = nn.Linear(64, num_classes)
def make_layer(self, block, out_channels, blocks, stride=1):

downsample = None
if (stride != 1) or (self.in_channels != out_channels):
downsample = nn.Sequential(
conv3x3(self.in_channels, out_channels, stride=stride),
nn.BatchNorm2d(out_channels))
layers = []
layers.append(block(self.in_channels, out_channels, stride, downsample))
self.in_channels = out_channels
for i in range(1, blocks):
layers.append(block(out_channels, out_channels))
return nn.Sequential(*layers)

out = self.conv(x)
out = self.bn(out)


out = self.layer1(out)
out = self.avg_pool(out)
out = out.view(out.size(0), -1)
out = self.fc(out)
return F.log_softmax(out, dim=1)
#return out
ResNet9
• DAWNBench on cifar10
• ResNet9: train to 94% CIFAR10 accuracy in 100 seconds
6.3.6 MNIST digit classification

from torchvision import datasets, transforms
import os
WD = os.path.join(Path.home(), "data", "pystatml", "dl_mnist_pytorch")

os.chdir(WD)
def load_mnist(batch_size_train, batch_size_test):
train_loader = torch.utils.data.DataLoader(
datasets.MNIST('data', train=True, download=True,
transform=transforms.Compose([
transforms.Normalize((0.1307,), (0.3081,))
])),
batch_size=batch_size_train, shuffle=True)
test_loader = torch.utils.data.DataLoader(
datasets.MNIST('data', train=False, transform=transforms.Compose([
transforms.Normalize((0.1307,), (0.3081,))
])),
batch_size=batch_size_test, shuffle=True)
return train_loader, test_loader

data_shape = dataloaders["train"].dataset.data.shape[1:]
D_in = np.prod(data_shape)


D_out = len(dataloaders["train"].dataset.targets)
Working dir is: /volatile/duchesnay/data/pystatml/dl_mnist_pytorch

LeNet
Dry run in debug mode to get the shape of the last convnet layer.
model = LeNet5((1, 6, 16, 1, 120, 84, 10), debug=True)
batch_idx, (data_example, target_example) = next(enumerate(train_loader))
print(model)
_ = model(data_example)
LeNet5(
(conv1): Conv2d(1, 6, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2))
(conv2): Conv2d(6, 16, kernel_size=(5, 5), stride=(1, 1))
(fc1): Linear(in_features=1, out_features=120, bias=True)
)
### DEBUG: Shape of last convnet= torch.Size([16, 5, 5]) . FC size= 400
Set First FC layer to 400

model = LeNet5((1, 6, 16, 400, 120, 84, 10)).to(device)
# Explore the model



torch.Size([6, 1, 5, 5])
torch.Size([6])
torch.Size([16, 6, 5, 5])
torch.Size([16])
torch.Size([120])
torch.Size([84])


torch.Size([10])
Epoch 0/4
----------
val Loss: 0.1568 Acc: 94.88%
Epoch 2/4
----------
val Loss: 0.0561 Acc: 98.29%
Epoch 4/4
----------
val Loss: 0.0547 Acc: 98.23%

MiniVGGNet
model = MiniVGGNet(layers=(1, 16, 32, 1, 120, 84, 10), debug=True)
print(model)
MiniVGGNet(


)
Set First FC layer to 800
model = MiniVGGNet((1, 16, 32, 800, 120, 84, 10)).to(device)

# Explore the model



torch.Size([16, 1, 3, 3])
torch.Size([16])
torch.Size([16, 16, 3, 3])
torch.Size([16])
torch.Size([32, 16, 3, 3])
torch.Size([32])
torch.Size([32, 32, 3, 3])
torch.Size([32])
torch.Size([120])
torch.Size([84])
torch.Size([10])
Epoch 0/4
----------
val Loss: 0.2489 Acc: 91.97%
Epoch 2/4
----------
val Loss: 0.0653 Acc: 97.81%
Epoch 4/4
----------
val Loss: 0.0443 Acc: 98.60%


Reduce the size of training dataset
Reduce the size of the training dataset by considering only 10 minibatche for size16.
train_size = 10 * 16
# Stratified sub-sampling
targets = train_loader.dataset.targets.numpy()
nclasses = len(set(targets))
indices = np.concatenate([np.random.choice(np.where(targets == lab)[0], int(train_size /␣

˓→nclasses),replace=False)
for lab in set(targets)])
np.random.shuffle(indices)
train_loader = torch.utils.data.DataLoader(train_loader.dataset, batch_size=16,

sampler=torch.utils.data.SubsetRandomSampler(indices))
# Check train subsampling

train_labels = np.concatenate([labels.numpy() for inputs, labels in train_loader])
print("Train size=", len(train_labels), " Train label count=", {lab:np.sum(train_labels␣
˓→== lab) for lab in set(train_labels)})
print("Batch sizes=", [inputs.size(0) for inputs, labels in train_loader])




Train size= 160 Train label count= {0: 16, 1: 16, 2: 16, 3: 16, 4: 16, 5: 16, 6: 16, 7:␣
˓→16, 8: 16, 9: 16}
Batch sizes= [16, 16, 16, 16, 16, 16, 16, 16, 16, 16]
LeNet5
model = LeNet5((1, 6, 16, 400, 120, 84, D_out)).to(device)


Epoch 0/99
----------
val Loss: 2.2938 Acc: 10.90%
Epoch 20/99
----------
val Loss: 0.6946 Acc: 78.96%
Epoch 40/99
----------
train Loss: 0.0278 Acc: 100.00%
val Loss: 0.5514 Acc: 86.55%
Epoch 60/99
----------
train Loss: 0.0068 Acc: 100.00%
val Loss: 0.6252 Acc: 86.77%
Epoch 80/99
----------
train Loss: 0.0033 Acc: 100.00%
val Loss: 0.6711 Acc: 86.83%


MiniVGGNet
model = MiniVGGNet((1, 16, 32, 800, 120, 84, 10)).to(device)


Epoch 0/99
----------
val Loss: 2.3061 Acc: 9.82%
Epoch 20/99
----------
val Loss: 2.2860 Acc: 9.82%
Epoch 40/99
----------
val Loss: 0.7725 Acc: 72.44%
Epoch 60/99
----------
val Loss: 0.7235 Acc: 80.34%
Epoch 80/99
----------
train Loss: 0.0023 Acc: 100.00%
val Loss: 0.7130 Acc: 86.56%


6.3.7 CIFAR-10 dataset
Source Yunjey Choi

os.chdir(WD)
import numpy as np
import torch
import torchvision
# Hyper-parameters
num_epochs = 5
learning_rate = 0.001



transforms.Pad(4),
# CIFAR-10 dataset
train=True,
download=True)
val_dataset = torchvision.datasets.CIFAR10(root='data/',
train=False,
# Data loader
batch_size=100,
shuffle=True)
val_loader = torch.utils.data.DataLoader(dataset=val_dataset,
batch_size=100,
shuffle=False)


print("Datasets shape:", {x: dataloaders[x].dataset.data.shape for x in ['train', 'val']})
print("N input features:", D_in, "N output:", D_out)
Working dir is: /volatile/duchesnay/data/pystatml/dl_cifar10_pytorch

Datasets shape: {'train': (50000, 32, 32, 3), 'val': (10000, 32, 32, 3)}
N input features: 3072 N output: 10
LeNet
model = LeNet5((3, 6, 16, 1, 120, 84, D_out), debug=True)

batch_idx, (data_example, target_example) = next(enumerate(train_loader))
print(model)
LeNet5(
)


Set 576 neurons to the first FC layer

SGD with momentum lr=0.001, momentum=0.5
# Explore the model



torch.Size([6, 3, 5, 5])
torch.Size([6])
torch.Size([16, 6, 5, 5])
torch.Size([16])
torch.Size([120])
torch.Size([84])
torch.Size([10])
Epoch 0/24
----------
val Loss: 2.3043 Acc: 10.00%
Epoch 5/24
----------
val Loss: 2.3015 Acc: 10.36%
Epoch 10/24
----------
val Loss: 2.2979 Acc: 11.93%
Epoch 15/24
----------
val Loss: 2.2808 Acc: 10.26%
Epoch 20/24
----------


val Loss: 2.1761 Acc: 17.29%

Increase learning rate and momentum lr=0.01, momentum=0.9


Epoch 0/24
----------
val Loss: 1.9415 Acc: 29.59%
Epoch 5/24
----------
val Loss: 1.2294 Acc: 55.74%
Epoch 10/24
----------
val Loss: 1.1031 Acc: 60.68%
Epoch 15/24
----------


val Loss: 1.0167 Acc: 64.26%
Epoch 20/24
----------
val Loss: 0.9505 Acc: 66.62%

Adaptative learning rate: Adam


Epoch 0/24
----------
val Loss: 1.6223 Acc: 40.22%
Epoch 5/24
----------
val Loss: 1.2271 Acc: 55.97%
Epoch 10/24
----------


val Loss: 1.1393 Acc: 59.72%
Epoch 15/24
----------
val Loss: 1.0458 Acc: 63.05%
Epoch 20/24
----------
val Loss: 1.0047 Acc: 64.49%

MiniVGGNet
model = MiniVGGNet(layers=(3, 16, 32, 1, 120, 84, D_out), debug=True)

print(model)
MiniVGGNet(
)

Set 1152 neurons to the first FC layer

SGD with large momentum and learning rate
model = MiniVGGNet((3, 16, 32, 1152, 120, 84, D_out)).to(device)


Epoch 0/24
----------
val Loss: 2.3004 Acc: 10.31%
Epoch 5/24
----------
val Loss: 1.3726 Acc: 50.32%
Epoch 10/24
----------
val Loss: 1.0193 Acc: 64.00%
Epoch 15/24
----------
val Loss: 0.8297 Acc: 71.18%
Epoch 20/24
----------
val Loss: 0.7136 Acc: 75.42%


Adam
model = MiniVGGNet((3, 16, 32, 1152, 120, 84, D_out)).to(device)


Epoch 0/24
----------
val Loss: 1.5805 Acc: 41.62%
Epoch 5/24
----------
val Loss: 1.1027 Acc: 60.83%
Epoch 10/24
----------
val Loss: 0.8994 Acc: 68.29%
Epoch 15/24
----------
val Loss: 0.8465 Acc: 70.90%
Epoch 20/24
----------
val Loss: 0.7513 Acc: 74.03%


ResNet
model = ResNet(ResidualBlock, [2, 2, 2], num_classes=D_out).to(device) # 195738 parameters


Epoch 0/24
----------
val Loss: 1.7289 Acc: 38.18%
Epoch 5/24
----------
val Loss: 0.8672 Acc: 71.34%
Epoch 10/24
----------
val Loss: 0.5754 Acc: 80.47%
Epoch 15/24
----------


val Loss: 0.6208 Acc: 80.16%
Epoch 20/24
----------
val Loss: 0.4696 Acc: 84.20%
6.4 Transfer Learning Tutorial
Sources:
• cs231n @ Stanford
• Sasank Chilamkurthy
Quote cs231n @ Stanford:
In practice, very few people train an entire Convolutional Network from scratch (with random
initialization), because it is relatively rare to have a dataset of sufficient size. Instead, it is
common to pretrain a ConvNet on a very large dataset (e.g. ImageNet, which contains 1.2
million images with 1000 categories), and then use the ConvNet either as an initialization or a
fixed feature extractor for the task of interest.
These two major transfer learning scenarios look as follows:
• ConvNet as fixed feature extractor:
– Take a ConvNet pretrained on ImageNet,
6.4. Transfer Learning Tutorial 319

– Remove the last fully-connected layer (this layer’s outputs are the 1000 class scores
for a different task like ImageNet)
– Treat the rest of the ConvNet as a fixed feature extractor for the new dataset.
In practice:
– Freeze the weights for all of the network except that of the final fully connected layer.
This last fully connected layer is replaced with a new one with random weights and
only this layer is trained.
• Finetuning the convnet:
fine-tune the weights of the pretrained network by continuing the backpropagation. It is possi-
ble to fine-tune all the layers of the ConvNet
Instead of random initializaion, we initialize the network with a pretrained network, like the
one that is trained on imagenet 1000 dataset. Rest of the training looks as usual.
%matplotlib inline
import os
import numpy as np
import torch
import torchvision
#
6.4.1 Training function
Combine train and test/validation into a single function.

Now, let’s write a general function to train a model. Here, we will illustrate:
• Scheduling the learning rate
• Saving the best model
In the following, parameter scheduler is an LR scheduler object from torch.optim.
lr_scheduler.
import numpy as np
import torch
import time
import copy



since = time.time()
best_acc = 0.0


print('-' * 10)

else:
running_loss = 0.0

nsamples = 0

# forward

loss.backward()
optimizer.step()
# statistics

scheduler.step()


print('{} Loss: {:.4f} Acc: {:.4f}'.format(
phase, epoch_loss, epoch_acc))

print()

print('Best val Acc: {:4f}'.format(best_acc))

6.4.2 CIFAR-10 dataset
Source Yunjey Choi

os.chdir(WD)

transforms.Pad(4),
# CIFAR-10 dataset
train=True,
download=True)
test_dataset = torchvision.datasets.CIFAR10(root='data/',
train=False,
# Data loader


batch_size=100,
shuffle=True)
val_loader = torch.utils.data.DataLoader(dataset=test_dataset,
batch_size=100,
shuffle=False)


Working dir is: /home/edouard/data/pystatml/dl_cifar10_pytorch

Datasets shape {'train': (50000, 32, 32, 3), 'val': (10000, 32, 32, 3)}
Finetuning the convnet
• Load a pretrained model and reset final fully connected layer.

• SGD optimizer.
model_ft = models.resnet18(pretrained=True)
num_ftrs = model_ft.fc.in_features
# Here the size of each output sample is set to 10.
model_ft.fc = nn.Linear(num_ftrs, D_out)
model_ft = model_ft.to(device)
criterion = nn.CrossEntropyLoss()
# Observe that all parameters are being optimized

optimizer_ft = optim.SGD(model_ft.parameters(), lr=0.001, momentum=0.9)
# Decay LR by a factor of 0.1 every 7 epochs

exp_lr_scheduler = lr_scheduler.StepLR(optimizer_ft, step_size=7, gamma=0.1)
model, losses, accuracies = train_val_model(model_ft, criterion, optimizer_ft,

dataloaders, scheduler=exp_lr_scheduler, num_epochs=25, log_interval=5)
epochs = np.arange(len(losses['train']))
_ = plt.plot(epochs, losses['train'], '-b', epochs, losses['val'], '--r')
Epoch 0/24
----------
train Loss: 1.2478 Acc: 0.5603
val Loss: 0.9084 Acc: 0.6866

Epoch 5/24
----------
val Loss: 0.5918 Acc: 0.7951
Epoch 10/24
----------
val Loss: 0.5257 Acc: 0.8178
Epoch 15/24
----------
val Loss: 0.5205 Acc: 0.8201
Epoch 20/24
----------
val Loss: 0.5183 Acc: 0.8212

Best val Acc: 0.822800
Adam optimizer
model_ft = models.resnet18(pretrained=True)
num_ftrs = model_ft.fc.in_features
# Here the size of each output sample is set to 10.
model_ft.fc = nn.Linear(num_ftrs, D_out)
model_ft = model_ft.to(device)

# Observe that all parameters are being optimized

optimizer_ft = torch.optim.Adam(model_ft.parameters(), lr=0.001)

exp_lr_scheduler = lr_scheduler.StepLR(optimizer_ft, step_size=7, gamma=0.1)
model, losses, accuracies = train_val_model(model_ft, criterion, optimizer_ft,

Epoch 0/24
----------
val Loss: 0.8981 Acc: 0.6881
Epoch 5/24
----------
val Loss: 0.9076 Acc: 0.7147
Epoch 10/24
----------
val Loss: 0.4148 Acc: 0.8613
Epoch 15/24
----------
val Loss: 0.4019 Acc: 0.8646
Epoch 20/24
----------
val Loss: 0.4025 Acc: 0.8675


ResNet as a feature extractor
Freeze all the network except the final layer: requires_grad == False to freeze the parameters
so that the gradients are not computed in backward().
model_conv = torchvision.models.resnet18(pretrained=True)
for param in model_conv.parameters():
param.requires_grad = False
# Parameters of newly constructed modules have requires_grad=True by default

num_ftrs = model_conv.fc.in_features
model_conv.fc = nn.Linear(num_ftrs, D_out)
model_conv = model_conv.to(device)
# Observe that only parameters of final layer are being optimized as

# opposed to before.
optimizer_conv = optim.SGD(model_conv.fc.parameters(), lr=0.001, momentum=0.9)

exp_lr_scheduler = lr_scheduler.StepLR(optimizer_conv, step_size=7, gamma=0.1)
model, losses, accuracies = train_val_model(model_conv, criterion, optimizer_conv,
Epoch 0/24
----------
val Loss: 1.7982 Acc: 0.3798


Epoch 5/24
----------
val Loss: 1.7067 Acc: 0.4097
Epoch 10/24
----------
val Loss: 1.6737 Acc: 0.4269
Epoch 15/24
----------
val Loss: 1.6724 Acc: 0.4218
Epoch 20/24
----------
val Loss: 1.6755 Acc: 0.4250

Adam optimizer
model_conv = torchvision.models.resnet18(pretrained=True)
for param in model_conv.parameters():
param.requires_grad = False
# Parameters of newly constructed modules have requires_grad=True by default

num_ftrs = model_conv.fc.in_features
model_conv.fc = nn.Linear(num_ftrs, D_out)
model_conv = model_conv.to(device)

# Observe that only parameters of final layer are being optimized as

# opposed to before.
optimizer_conv = optim.Adam(model_conv.fc.parameters(), lr=0.001)

exp_lr_scheduler = lr_scheduler.StepLR(optimizer_conv, step_size=7, gamma=0.1)
model, losses, accuracies = train_val_model(model_conv, criterion, optimizer_conv,

exp_lr_scheduler, dataloaders, num_epochs=25, log_interval=5)
---------------------------------------------------------------------------
TypeError Traceback (most recent call last)
<ipython-input-16-dde92868b554> in <module>
19
20 model, losses, accuracies = train_val_model(model_conv, criterion, optimizer_conv,
---> 21 exp_lr_scheduler, dataloaders, num_epochs=25, log_interval=5)
22
23 epochs = np.arange(len(losses['train']))
TypeError: train_val_model() got multiple values for argument 'num_epochs'

CHAPTER
SEVEN
INDICES AND TABLES
• genindex
• modindex
• search
329

Statistics Machine Learning Python Draft

Uploaded by

Copyright:

Available Formats

Statistics Machine Learning Python Draft

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Statistics Machine Learning Python Draft

Uploaded by

Copyright:

Available Formats

What are the main topics covered in the document?

What are the main topics covered in the document?

What are the main Python libraries discussed for data science?

What are the main Python libraries discussed for data science?

Statistics and Machine Learning in

Edouard Duchesnay, Tommy Löfstedt, Feki Younes

Oct 29, 2020

5 Machine Learning 159

6 Deep Learning 257

7 Indices and tables 329

1.1 Python ecosystem for data-science

1.1.1 Python language

3. Add anaconda path in your PATH variable in your .bashrc file:

Managing with ``conda``

conda update conda

conda install pyqt

Install seaborn for graphics

conda install seaborn

List installed packages

Search available packages

conda search pyqt

conda create --name test

3. Add anaconda path in your PATH variable in your .bashrc file:

4. Install required packages

conda install -y scipy

Quite with CTL-D

Quite with CTL-D

pip install -U --user seaborn

pip install -U --user nibabel

spyder: IDE (integrated development environment):

text editor help/variable explorer

Shortcuts: - F9 run line/selection

1.1. Python ecosystem for data-science 3

Scipy: general scientific libraries with advanced solver:

Pandas: Manipulation of structured data (tables). input/output excel files, etc.

li- Arrays data, Structured Solvers: Solvers: Stats: Stats: Machine

1.2 Introduction to Machine Learning

1.2.1 Machine learning within data science

Machine learning covers two main types of data analysis:

1.2. Introduction to Machine Learning 5

1.2.2 IT/computing science tools

• High Performance Computing (HPC)

1.2.3 Statistics and applied mathematics

1.3 Data analysis methodology

1. Formalize customer’s needs into a learning problem:

1.3. Data analysis methodology 7

Note: Click here to download the full example code

Source Kevin Markham https://github.com/justmarkham/python-reference

2.1 Import libraries

# 'generic import' of math module

# import multiple functions at once

# import all functions in a module (generally discouraged)

# show all functions in math module

2.2 Basic operations

(continues on next page)

(continued from previous page)

# boolean operations (these return True)

2.3 Data types

# determine the type of an object