Python for R developers and data scientists

Installation Vectors Data Frames Analysis Visualization I/O Conclusion

Artur Matos

http://www.lambdatree.com

June 8, 2016

Outline

1 Getting Up and Running

2 Vectors

3 Data Frames

4 Analysis

5 Visualization

7 Conclusion

Section 1

Getting Up and Running

Which Python?

Python runtimesSeveral available: CPython, PyPy, Jython. . .CPython is the official runtime written in C.PyPy is a JIT-based runtime that runs significantly faster than CPython.For scientific computing, CPython is the only choice.

Python 2 vs Python 3Python 3 is not backwards compatibleAnswer today is Python 3 (might have answered differently last year)Unless you have other teams using Python 2. . .But all major packages support Python 3 already

Installation

Several options, ranging from simple to complex.We will use Anaconda here, which will get you up and running quickly.On Linux and Mac you can also install Python with your package manager.Use virtualenv to isolate Python environments (not covered here).

Installing Anaconda

https://www.continuum.io/downloads

Installing Anaconda (2)

Jupyter

Python syntax in 2 minutes

Convert input into uppercase;Cuts anything longer than 10 characters;Adds extra spaces if shorter than 10 characters;Add single quotes.

>>> quote_pad_string("This is rather long")’THIS IS RA’>>> quote_pad_string("Short")’SHORT ’

Python syntax in 2 minutes (2)

Function:

# WARNING: This is purely to cover some basic python syntax# there are better ways to do this in Pythondef quote_pad_string(a_string):

maximum_length = 10num_missing_characters = maximum_length - len(a_string)

if num_missing_characters < 0:num_missing_characters = 0

if num_missing_characters:for i in range(num_missing_characters):

a_string = a_string + " "else:

a_string = a_string[:maximum_length]

return "’" + a_string.upper() + "’"

def defines the body of a function. Python is dynamically typed:

Python uses indentation for code blocks instead of curly braces:

‘=’ for assignment:

‘if’ statement:

‘for’ statement:

len(a_string) is a function call, a_string.upper() is a method invocation:

Section 2

Vectors

Scalars - R

In R there are no real scalar types. They are just vectors of length 1:

> a <- 5 # Equivalent to a <- c(5)> a[1] 5> length(a)[1] 1

Scalars - Python

In Python scalars and vectors are not the same thing:

>>> a = 5 # Scalar5

>>> b = np.array([5]) # Array with one elementarray([5])>>> len(b) # Equivalent to ’length’ in R1

This won’t work:

>>> len(a)TypeError: object of type ’int’ has no len()

Vectors, matrices and arrays - R

In R, there’s ‘c’ for 1d vectors, ‘matrix’ for 2 dimensions, and ‘array’ for higher-orderdimensions:

> c(1,2,3,4)[1] 1 2 3 4> matrix(1:4, nrow=2,ncol=2)

[,1] [,2][1,] 1 3[2,] 2 4> array(1:3, c(2,4,6))...

Strangely enough, a 1d array is not the same as a vector:

> a <- as.array(1:3)[1] 1 2 3> is.vector(a)[1] FALSE

Vectors, matrices and arrays - Python

Python has no builtin vector or matrix type. You will need numpy:

>>> import numpy as np # Equivalent to ’library(numpy)’ in R.>>> np.array([1, 2, 3, 4]) # 1d vectorarray([1, 2, 3, 4])

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

[3, 4]])

np.array works with any dimension and it’s a single type (ndarray).(There’s also a matrix type specifically for two dimensions but it should be avoided.Always use ndarray.)

Generating regular sequences

R> 5:10 # Shortend for seq[1] 5 6 7 8 9 10> seq(0, 1, length.out = 11)[1] 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0

Python>>> np.arange(3.0)array([ 0., 1., 2.])>>> np.arange(3, 7)array([3, 4, 5, 6])>>> np.arange(3, 7, 2)array([3, 5])

Vector operations - Python

For the most part, vector operations in Python work just like in R:

>>> a = np.arange(5.0)array([ 0., 1., 2., 3., 4.])>>> 1.0 + a # Addingarray([ 1., 2., 3., 4., 5.])>>> a * a # Multiplying element wisearray([ 0., 1., 4., 9., 16.])>>> a ** 3 # to the power of 3array([ 0., 1., 8., 27., 64.])

Vector operations - Python (2)

For matrix multiplication use the ‘@’ operator:

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

[0, 1]])>> b = np.array([[4, 1], [2, 2]])array([[4, 1],

[2, 2]])>> a @ barray([[4, 1],

[2, 2]])

(In Python 2 use np.dot(a,b).)

Vector operations - Python (3)

Numpy also has the usual mathematical operations that work on vectors:

>> a = np.arange(5.0)array([ 0., 1., 2., 3., 4.])>> np.sin(a)array([ 0., 0.84147098, 0.90929743, 0.14112001, -0.7568025 ])

Full reference here:http://docs.scipy.org/doc/numpy/reference/routines.math.html

Recycling - R

When doing vector operations, R automatically extends the smallest element to be aslarge as the other:

> c(1,2) + c(1,2,3,4) # Equivalent to c(1,2,1,2) + c(1,2,3,4)[1] 2 4 4 6

You can do this even if the lengths aren’t multiples of one another, albeit with awarning:

> c(1,2) + c(1,2,3,4,5)[1] 2 4 4 6 6Warning message:In c(1, 2) + c(1,2,3,4,5) :

longer object length is not a multiple of shorter object length

Recycling - Python

This won’t work in Python however:

>> np.arange(2.0) + np.arange(4.0)----------------------------------------ValueError: operands could not be broadcast together with shapes(2,) (4,)

Numpy has much more strict recycling (aka broadcasting) rules.

Broadcasting Rules - Python

2x3 and 2x1: [0 1 23 4 5

[0 1 23 4 5

[0 0 01 1 1

]2x3 and 1x3: [

0 1 23 4 5

]+[0 1 2

[0 1 23 4 5

[0 1 20 1 2

]Adding a single element array or a scalar always works:[

0 1 23 4 5

[0 1 23 4 5

[0 0 00 0 0

]This won’t work (the dimensions need to match exactly or be 1):[

0 1 2 3]+[0 1

2x3 and 2x1: [0 1 23 4 5

[0 1 23 4 5

[0 0 01 1 1

2x3 and 1x3: [0 1 23 4 5

]+[0 1 2

[0 1 23 4 5

[0 1 20 1 2

0 1 23 4 5

[0 1 23 4 5

[0 0 00 0 0

0 1 2 3]+[0 1

2x3 and 2x1: [0 1 23 4 5

[0 1 23 4 5

[0 0 01 1 1

]2x3 and 1x3: [

0 1 23 4 5

]+[0 1 2

[0 1 23 4 5

[0 1 20 1 2

0 1 23 4 5

[0 1 23 4 5

[0 0 00 0 0

0 1 2 3]+[0 1

2x3 and 2x1: [0 1 23 4 5

[0 1 23 4 5

[0 0 01 1 1

]2x3 and 1x3: [

0 1 23 4 5

]+[0 1 2

[0 1 23 4 5

[0 1 20 1 2

Adding a single element array or a scalar always works:[0 1 23 4 5

[0 1 23 4 5

[0 0 00 0 0

0 1 2 3]+[0 1

2x3 and 2x1: [0 1 23 4 5

[0 1 23 4 5

[0 0 01 1 1

]2x3 and 1x3: [

0 1 23 4 5

]+[0 1 2

[0 1 23 4 5

[0 1 20 1 2

0 1 23 4 5

[0 1 23 4 5

[0 0 00 0 0

0 1 2 3]+[0 1

2x3 and 2x1: [0 1 23 4 5

[0 1 23 4 5

[0 0 01 1 1

]2x3 and 1x3: [

0 1 23 4 5

]+[0 1 2

[0 1 23 4 5

[0 1 20 1 2

0 1 23 4 5

]+[0]=

[0 1 23 4 5

[0 0 00 0 0

This won’t work (the dimensions need to match exactly or be 1):[0 1 2 3

]+[0 1

2x3 and 2x1: [0 1 23 4 5

[0 1 23 4 5

[0 0 01 1 1

]2x3 and 1x3: [

0 1 23 4 5

]+[0 1 2

[0 1 23 4 5

[0 1 20 1 2

0 1 23 4 5

]+[0]=

[0 1 23 4 5

[0 0 00 0 0

0 1 2 3]+[0 1

Indexing - i:j:k syntax

a = np.arange(10)

0 1 2 3 4 5 6 7 8 9]

a = np.arange(10)

0 1 2 3 4 5 6 7 8 9]

Indexing in Python starts from 0 (not 1):

>>> a[0]0.0

Indexing on a single value returns a scalar (not an array!)

a = np.arange(10)

0 1 2 3 4 5 6 7 8 9]

Use ‘i:j’ to index from position i to j-1:

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

a = np.arange(10)

0 1 2 3 4 5 6 7 8 9]

An optional ‘k’ element defines the step:

>>> a[1:7:2]array([1, 3, 5])

a = np.arange(10)

0 1 2 3 4 5 6 7 8 9]

i and j can be negative, which means they will start counting from the last:

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

a = np.arange(10)

0 1 2 3 4 5 6 7 8 9]

i and j can be negative, which means they will start counting from the last:

>>> a[-3:-1]array([7, 8])

a = np.arange(10)

0 1 2 3 4 5 6 7 8 9]

While a negative k will go in the opposite direction:

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

a = np.arange(10)

0 1 2 3 4 5 6 7 8 9]

Not all need to be included:

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

Indexing - multiple dimensions

Use ‘,’ for additional dimensions:

[1 2 34 5 6

]>>> x[0:2, 0:1]array([[1],

Indexing using conditions

Operators like ‘>’ or ‘<=’ operate element-wise and return a logical vector:

>>> a > 4array([False, False, False, False, False, True, True,True, True, True], dtype=bool)

These can be combined into more complex expressions:

>> (a > 2) && (b ** 2 <= a)...

And used as indexing too:

>> a[a > 4]array([5, 6, 7, 8, 9])

Assignment

Any index can be used together with ‘=’ for assignment:

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

Conditions work as well:

>>> a[a > 4] = 99array([99, 1, 2, 3, 4, 99, 99, 99, 99, 99])

Section 3

Data Frames

Pandas

Data frames in Python aren’t builtin. You will need pandas:

import pandas as pd

Loading the iris dataset:

>>> iris = pd.read\_csv("""https://raw.githubusercontent.com/pydata/pandas/master/pandas/tests/data/iris.csv""")>>> iris.head()

SepalLength SepalWidth PetalLength PetalWidth Name0 5.1 3.5 1.4 0.2 Iris-setosa1 4.9 3.0 1.4 0.2 Iris-setosa2 4.7 3.2 1.3 0.2 Iris-setosa3 4.6 3.1 1.5 0.2 Iris-setosa4 5.0 3.6 1.4 0.2 Iris-setosa

Selection

SepalLength SepalWidth PetalLength PetalWidth Name

0 5.1 3.5 1.4 0.2 Iris-setosa1 4.9 3.0 1.4 0.2 Iris-setosa2 4.7 3.2 1.3 0.2 Iris-setosa3 4.6 3.1 1.5 0.2 Iris-setosa4 5.0 3.6 1.4 0.2 Iris-setosa. . . . . . . . . . . . . . . . . .

Selection

‘.<columnName>’ returns only that column:

>>> iris.SepalLength0 5.11 4.92 4.7...

Selection

This also works:

>>> iris["SepalLength"]...

Selection

You can select multiple columns by passing a list:

>>> iris[["SepalWidth", "SepalLength"]]SepalWidth SepalLength

0 3.5 5.11 3.0 4.9...

Selection

Or if you pass a slice you can select rows:

>>> iris[1:3]SepalLength SepalWidth PetalLength PetalWidth Name

1 4.9 3.0 1.4 0.2 Iris-setosa2 4.7 3.2 1.3 0.2 Iris-setosa

Selection

With loc you can slice both rows and columns:

>>> iris.loc[1:3, ["SepalLength", "SepalWidth"]]SepalLength SepalWidth

1 4.9 3.02 4.7 3.23 4.6 3.1

loc is inclusive at the end.

Selection

As well as only one single row:

>>> iris.loc[3]SepalLength 4.6SepalWidth 3.1PetalLength 1.5PetalWidth 0.2Name Iris-setosaName: 3, dtype: object

Selection

iloc works with integer indices, similar to numpy arrays:

>>> iris.iloc[0:2, 0:3]SepalLength SepalWidth PetalLength

0 5.1 3.5 1.41 4.9 3.0 1.4

Selection

‘:’ will include all the rows (or all the columns):

>>> iris.iloc[0:2, :]SepalLength SepalWidth PetalLength PetalWidth Name

0 5.1 3.5 1.4 0.2 Iris-setosa1 4.9 3.0 1.4 0.2 Iris-setosa

Selection

Or you can use conditions like numpy:

>>> iris[iris.SepalLength < 5]SepalLength SepalWidth PetalLength PetalWidth Name

1 4.9 3.0 1.4 0.2 Iris-setosa2 4.7 3.2 1.3 0.2 Iris-setosa3 4.6 3.1 1.5 0.2 Iris-setosa...

Selection

Picking a single value returns a scalar:

>>> iris.iloc[0,0]5.0999999999999996

Normally it’s better to use at or iat (faster).

Assignment

Assignment works as expected:

>>> iris.loc[iris.SepalLength > 7.6, "Name"] = "Iris-orlando"

Beware that this doesn’t work:

>> iris[iris.SepalLength > 7.6].Name = "Iris-orlando"SettingWithCopyWarning: A value is trying to be set on a copyof a slice from a DataFrame.

Operations

Operators work as expected:

>>> iris["SepalD"] = iris["SepalLength"] * iris["SepalWidth"]

There’s also apply:

>>> iris[["SepalLength", "SepalWidth"]].apply(np.sqrt)SepalLength SepalWidth

0 2.258318 1.8708291 2.213594 1.732051...

Use axis=1 to apply function to each row.

SQL-like operations - group by

>>> iris.groupby("Name").mean()SepalLength SepalWidth PetalLength PetalWidth

NameIris-setosa 5.006 3.418 1.464 0.244Iris-versicolor 5.936 2.770 4.260 1.326Iris-virginica 6.588 2.974 5.552 2.026

SQL-like operations (2) - join

>> leftkey lval

0 foo 11 foo 2

>> rightkey rval

0 foo 41 foo 5

>> pd.merge(left, right, on=’key’)key lval rval

0 foo 1 41 foo 1 5...

Time Series

Pandas data frames can also work as time series, replacing R’s ts, xts or zoo.Downloading some stock data from Google finance:

>>> import pandas.io.data as web>>> import datetime

>>> aapl = web.DataReader("AAPL", ’google’,datetime.datetime(2013, 1, 1),datetime.datetime(2014, 1, 1))

Open High Low Close VolumeDate2013-01-02 79.12 79.29 77.38 78.43 1401248662013-01-03 78.27 78.52 77.29 77.44 88240950...

Time series are just regular pandas data frames but with time stamps as indices.

Time Series (2)

Use loc to select based on dates:

>>> aapl.loc[’20130131’:’20130217’]Open High Low Close Volume

Date2013-01-31 65.28 65.61 65.00 65.07 798332152013-02-01 65.59 65.64 64.05 64.80 134867089...

Use iloc as before for selecting based on numerical indices:

>>> aapl.iloc[1:3]Open High Low Close Volume

Date2013-01-03 78.27 78.52 77.29 77.44 882409502013-01-04 76.71 76.95 75.12 75.29 148581860

Section 4

Analysis

Statistical tests

Use scipy.stats for common statistical tests:

>>> from scipy import stats

>>> iris_virginica = iris[iris.Name == ’Iris-virginica’].SepalLength.values>>> iris_setosa = iris[iris.Name == ’Iris-setosa’].SepalLength.values>>> t_test = stats.ttest_ind(iris_virginica, iris_setosa)>>> t_test.pvalue6.8925460606740589e-28

Use scikits.bootstrap for bootstrapped confidence intervals.

Ordinary Least Squares

Use statsmodels:

import numpy as npimport statsmodels.api as smimport statsmodels.formula.api as smf

The formula API is very similar to R:

>>> results = smf.ols("PetalWidth ~ Name + PetalLength", data=iris).fit()

It automatically includes an intercept (just like R).Use smf.glm for generalized linear models.

Formula API

Very similar to R:

You can include arbitrary transformations, e.g. “np.log(PetalWidth)”.To remove the intercept add a “- 1” or “0 +”Use “C(a)” to coerce a number to a factorUse “a:b” for modelling interactions between a and b.“a*b” means “a + b + a:b”Strings are automatically coerced to factors (more on this later)

Decision trees

Use scikit-learn:

>>> from sklearn import tree>>> clf = tree.DecisionTreeClassifier()>>> clf = clf.fit(sk_iris.data, sk_iris.target)

After being fitted, the model can then be used to predict the class of samples:

>>> clf.predict([[5.1, 3.5, 1.4, 0.2]])array([0])

Support Vector Machines

scikit-learn has a very regular API. Here’s the same example using an SVM:

>>> from sklearn import svm

>>> clf_svm = svm.SVC()>>> clf_svm = clf_svm.fit(sk_iris.data, sk_iris.target)

>>> clf_svm.predict([[5.1, 3.5, 1.4, 0.2]])array([0])

K-Means clustering

Clustering follows the same pattern:

>>> from sklearn import cluster

>>> k_means = cluster.KMeans(n_clusters=3)>>> k_means.fit(sk_iris.data)KMeans(copy_x=True, init=’k-means++’, ...

labels_ contains the assigned categories, following the same order as the data:

>>> k_means.labels_array([1, 1, 1, 1, 1...

predict works the same as for the other models, and returns the predicted category.

Principal Component Analysis

from sklearn import decomposition

pca = decomposition.PCA(n_components=3)pca = pca.fit(sk_iris.data)

explained_variance_ratio_ and components_ will include the explained varianceand the PCA components respectively:

>>> pca.explained_variance_ratio_array([ 0.92461621, 0.05301557, 0.01718514])

>>>pca.components_array([[ 0.36158968, -0.08226889, 0.85657211, 0.35884393],

[-0.65653988, -0.72971237, 0.1757674 , 0.07470647],[ 0.58099728, -0.59641809, -0.07252408, -0.54906091]])

Cross validation

scikit-learn also includes extensive support for cross-validation. Here’s a simple splitinto training and out-of-sample:

>>> from sklearn import cross_validation

>>> X_train, X_test, y_train, y_test = cross_validation.train_test_split(... sk_iris.data, sk_iris.target, test_size=0.4, random_state=0)

>>> X_train.shape, y_train.shape((90, 4), (90,))>>> X_test.shape, y_test.shape((60, 4), (60,))

It also supports K-fold, stratified K-fold, shuffling, etc. . .

There’s no builtin NA in Python. You normally use NaN for NAs. numpy has a bunchof builtin functions to ignore NaNs:

>>> a = np.array([1.0, 3.0, np.NaN, 5.0])>>> a.sum()nan>>> np.nansum(a)9.0

Pandas usually ignores NaNs when computing sums, means, etc.. but propagatesthem accordingly.scikit-learn assumes there’s no missing data so be sure to pre-process them, e.g.remove them or set them to 0. Look at sklearn.preprocessing.Imputerstatsmodels also use NaNs for missing data, but only has basic support forhandling them (it can only ignore them or raise an error). See the missingattribute in the model class.

Factors

Similarly to NAs, Python has no builtin factor data type.Different packages handle them differently:

numpy has no support for factors. Use integers.Pandas has categoricals, which work fairly similar to factorsstatsmodels convert strings to their own internal factor type, very similar to R.There’s also the ‘C’ operator.scikit-learn doesn’t support factors internally, but has some tools to convert stringsinto dummy variables, e.g. DictVectorizer

Notorious Omissions

Bayesian modellingTime series analysisEconometricsSignal processing, i.e. filter designNatural language processing. . .

Section 5

Visualization

Section 6

Pandas

read_csv Reads data from CSV files

>>> pd.read_csv(’foo.csv’)Unnamed: 0 A B C D

0 2000-01-01 0.266457 -0.399641 -0.219582 1.1868601 2000-01-02 -1.170732 -0.345873 1.653061 -0.282953...

Conversely there is to_csv to write CSV files:

In [136]: df.to_csv(’foo.csv’)

Other options

For data frames:

HDF5: read_hdf5, to_hdf5Excel: read_excel, to_excelSQL: read_sql, to_sqlStata: read_stata, to_stataSAS: read_sas, to_sasREST APIs: read_json or alternatively use requests

For numpy arrays:

You can use load and save for saving into .npy formatNormally I prefer to use HDF5 with the h5py library

h5py - datasets

Creating a data set:

>>> import h5py>>> import numpy as np>>>>>> f = h5py.File("mytestfile.hdf5", "w")>>> dset = f.create_dataset("mydataset", (100,), dtype=’i’)

Datasets work similarly to numpy arrays:

>>> dset[...] = np.arange(100)>>> dset[0]0>>> dset[10]10>>> dset[0:100:10]array([ 0, 10, 20, 30, 40, 50, 60, 70, 80, 90])

Other options

pickle - Python standard serialization format. See also shelve.tinydb - local document-oriented database (good for NLP tasks)sqlalchemy - Heavy-duty SQL to relational mapper.

Section 7

Conclusion

Things I haven’t covered:

Python data structures: dicts, listsPython - R interoperability: RPyParallel computing: IPython.parallel, pysparkOptimizing python code: Cython, numba, numexpr

Hope you’ve enjoyed. Feel free to get in touch: amatos@lambdatree.com

Python for R developers and data scientists

Software

Top 10 Freelance Python Developers for Hire in December 2015

Python cheat sheet April 2021 - WebsiteSetup...Python Cheat....

INTERMEDIATE PYTHON FOR DEVELOPERS · 2020. 9. 23. ·...

INTERMEDIATE PYTHON FOR DEVELOPERS · Intermediate Python.....

An Introduction to Python for Scientists - Hands-On …...

Speeding up Python with Rust - Amazon S3 · Speeding up...

Introduction to Python for Plone developers

Python Programming and Visualization for Scientists

Python programming guide for Earth...

MongoDB quickstart for Java, PHP, and Python developers

software carpentry - Python€¦ · Modern programming...

PYTHON FOR SCIENTISTS & ENGINEERS€¦ · Python for...

INTERMEDIATE PYTHON FOR DEVELOPERS...powerful language...

Introduction to numerical programming: a practical guide for...

python-for-scientists Documentation

JavaScript for Python Developers · TypeScript. Overview...