// Databricks notebook source exported at Sun, 19 Jun 2016 02:24:27 UTC

Scalable Data Science

prepared by Raazesh Sainudiin and Sivanand Sivaram

supported by and

The html source url of this databricks HOMEWORK notebook and its recorded Uji :

The html source url of the context/parent databricks notebook (for this databricks HOMEWORK notebook) and its recorded Uji :

This is from

• https://github.com/scalanlp/breeze/wiki/Linear-Algebra-Cheat-Sheet

Core Concepts

Compared to other numerical computing environments, Breeze matrices default to column major ordering, like Matlab, but indexing is 0-based, like Numpy. Breeze has as its core concepts matrices and column vectors. Row vectors are normally stored as matrices with a single row. This allows for greater type safety with the downside that conversion of row vectors to column vectors is performed using a transpose-slice (a.t(::,0)) instead of a simple transpose (a.t).

[[UFunc|Universal Functions]]s are very important in Breeze. Once you get a feel for the syntax (i.e. what's in this section), it might be worthwhile to read the first half of the UFunc wiki page. (You can skip the last half that involves implementing your own UFuncs...until you're ready to contribute to Breeze!)

Quick Reference

The following table assumes that Numpy is used with from numpy import * and Breeze with:

import breeze.linalg._
import breeze.numerics._

Creation

DenseMatrix.zeros[Double](2,3)

%py
import numpy as np

%py
np.zeros((2,3))

%r
mat.or.vec(2,3)

Reading and writing Matrices

Currently, Breeze supports IO for Matrices in two ways: Java serialization and csv. The latter comes from two functions: breeze.linalg.csvread and breeze.linalg.csvwrite. csvread takes a File, and optionally parameters for how the CSV file is delimited (e.g. if it is actually a tsv file, you can set tabs as the field delimiter.) and returns a DenseMatrix. Similarly, csvwrite takes a File and a DenseMatrix, and writes the contents of a matrix to a file.

Indexing and Slicing

val matrix = DenseMatrix.rand(2, 3)

val two_one = matrix(1, 0) // Remember the index starts from zero

Other Manipulation

Operations

Sum

Boolean Operators

Linear Algebra Functions

Rounding and Signs

Constants

Complex numbers

If you make use of complex numbers, you will want to include a breeze.math._ import. This declares a i variable, and provides implicit conversions from Scala’s basic types to complex types.

Numeric functions

Breeze contains a fairly comprehensive set of special functions under the breeze.numerics._ import. These functions can be applied to single elements, vectors or matrices of Doubles. This includes versions of the special functions from scala.math that can be applied to vectors and matrices. Any function acting on a basic numeric type can “vectorized”, to a [[UFunc|Universal Functions]] function, which can act elementwise on vectors and matrices:

val v = DenseVector(1.0,2.0,3.0)
exp(v) // == DenseVector(2.7182818284590455, 7.38905609893065, 20.085536923187668)

UFuncs can also be used in-place on Vectors and Matrices:

val v = DenseVector(1.0,2.0,3.0)
exp.inPlace(v) // == DenseVector(2.7182818284590455, 7.38905609893065, 20.085536923187668)

See [[Universal Functions]] for more information.

Here is a (non-exhaustive) list of UFuncs in Breeze:

Trigonometry

• sin, sinh, asin, asinh
• cos, cosh, acos, acosh
• tan, tanh, atan, atanh
• atan2
• sinc(x) == sin(x)/x
• sincpi(x) == sinc(x * Pi)

Logarithm, Roots, and Exponentials

• log, exp log10
• log1p, expm1
• sqrt, sbrt
• pow

Gamma Function and its cousins

The gamma function is the extension of the factorial function to the reals. Numpy needs from scipy.special import * for this and subsequent sections.

Error Function

The error function...

Other functions

Map and Reduce

For most simple mapping tasks, one can simply use vectorized, or universal functions. Given a vector v, we can simply take the log of each element of a vector with log(v). Sometimes, however, we want to apply a somewhat idiosyncratic function to each element of a vector. For this, we can use the map function:

val v = DenseVector(1.0,2.0,3.0)
v.map( xi => foobar(xi) )

Breeze provides a number of built in reduction functions such as sum, mean. You can implement a custom reduction using the higher order function reduce. For instance, we can sum the first 9 integers as follows:

val v = linspace(0,9,10)
val s = v.reduce( _ + _ )

Broadcasting

Sometimes we want to apply an operation to every row or column of a matrix, as a unit. For instance, you might want to compute the mean of each row, or add a vector to every column. Adapting a matrix so that operations can be applied columnwise or rowwise is called broadcasting. Languages like R and numpy automatically and implicitly do broadcasting, meaning they won’t stop you if you accidentally add a matrix and a vector. In Breeze, you have to signal your intent using the broadcasting operator *. The * is meant to evoke “foreach” visually. Here are some examples:

    val dm = DenseMatrix((1.0,2.0,3.0),
                         (4.0,5.0,6.0))

    val res = dm(::, *) + DenseVector(3.0, 4.0)
    assert(res === DenseMatrix((4.0, 5.0, 6.0), (8.0, 9.0, 10.0)))

    res(::, *) := DenseVector(3.0, 4.0)
    assert(res === DenseMatrix((3.0, 3.0, 3.0), (4.0, 4.0, 4.0)))

    val m = DenseMatrix((1.0, 3.0), (4.0, 4.0))
    // unbroadcasted sums all elements
    assert(sum(m) === 12.0)
    assert(mean(m) === 3.0)

    assert(sum(m(*, ::)) === DenseVector(4.0, 8.0))
    assert(sum(m(::, *)) === DenseMatrix((5.0, 7.0)))

    assert(mean(m(*, ::)) === DenseVector(2.0, 4.0))
    assert(mean(m(::, *)) === DenseMatrix((2.5, 3.5)))

The UFunc trait is similar to numpy’s ufunc. See [[Universal Functions]] for more information on Breeze UFuncs.

Casting and type safety

Compared to Numpy and Matlab, Breeze requires you to be more explicit about the types of your variables. When you create a new vector for example, you must specify a type (such as in DenseVector.zeros[Double](n)) in cases where a type can not be inferred automatically. Automatic inference will occur when you create a vector by passing its initial values in (DenseVector). A common mistake is using integers for initialisation (e.g. DenseVector), which would give a matrix of integers instead of doubles. Both Numpy and Matlab would default to doubles instead.

Breeze will not convert integers to doubles for you in most expressions. Simple operations like a :+ 3 when a is a DenseVector[Double] will not compile. Breeze provides a convert function, which can be used to explicitly cast. You can also use v.mapValues(_.toDouble).

Casting

Performance

Breeze uses netlib-java for its core linear algebra routines. This includes all the cubic time operations, matrix-matrix and matrix-vector multiplication. Special efforts are taken to ensure that arrays are not copied.

Netlib-java will attempt to load system optimised BLAS/LAPACK if they are installed, falling back to the reference natives, falling back to pure Java. Set your logger settings to ALL for the com.github.fommil.netlib package to check the status, and to com.github.fommil.jniloader for a more detailed breakdown. Read the netlib-java project page for more details.

Currently vectors and matrices over types other than Double, Float and Int are boxed, so they will typically be a lot slower. If you find yourself needing other AnyVal types like Long or Short, please ask on the list about possibly adding support for them.

Scalable Data Science

prepared by Raazesh Sainudiin and Sivanand Sivaram

supported by and